diff options
| author | Vidhya Sudhan Loganathan <vidhyasudhan.loganathan@arm.com> | 2018-07-04 09:34:00 +0100 |
|---|---|---|
| committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:54:10 +0000 |
| commit | 7485d5a62685cb745ab50e970adb722cb71557ac (patch) | |
| tree | ba01b99ca466c93edc9a3f8c1e34394ff84be060 /src | |
| parent | 014333d73883c3872e458cedda5ccef586a7ccd4 (diff) | |
| download | ComputeLibrary-7485d5a62685cb745ab50e970adb722cb71557ac.tar.gz | |
COMPMID-970 : Remove QS8 / QS16 support
Removed fixed point related code.
Change-Id: I487acf138dace3b0450e0d72ca7071eaec254566
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/137678
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Diffstat (limited to 'src')
133 files changed, 351 insertions, 4778 deletions
diff --git a/src/core/CL/CLHelpers.cpp b/src/core/CL/CLHelpers.cpp index df06aff647..07f8bd7bcd 100644 --- a/src/core/CL/CLHelpers.cpp +++ b/src/core/CL/CLHelpers.cpp @@ -38,8 +38,6 @@ std::string get_cl_type_from_data_type(const DataType &dt) { case DataType::U8: return "uchar"; - case DataType::QS8: - return "qs8"; case DataType::S8: return "char"; case DataType::QASYMM8: @@ -48,8 +46,6 @@ std::string get_cl_type_from_data_type(const DataType &dt) return "ushort"; case DataType::S16: return "short"; - case DataType::QS16: - return "qs16"; case DataType::U32: return "uint"; case DataType::S32: @@ -75,13 +71,11 @@ std::string get_data_size_from_data_type(const DataType &dt) switch(dt) { case DataType::U8: - case DataType::QS8: case DataType::S8: case DataType::QASYMM8: return "8"; case DataType::U16: case DataType::S16: - case DataType::QS16: case DataType::F16: return "16"; case DataType::U32: @@ -101,10 +95,6 @@ std::string get_underlying_cl_type_from_data_type(const DataType &dt) { switch(dt) { - case DataType::QS8: - return "char"; - case DataType::QS16: - return "short"; case DataType::QS32: return "int"; default: diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp index db4b344935..42cf21350d 100644 --- a/src/core/CL/CLKernelLibrary.cpp +++ b/src/core/CL/CLKernelLibrary.cpp @@ -231,22 +231,16 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map = { "gemm_interleave4x4", "gemm.cl" }, { "gemm_ma_f16", "gemm.cl" }, { "gemm_ma_f32", "gemm.cl" }, - { "gemm_ma_qs8", "gemm.cl" }, - { "gemm_ma_qs16", "gemm.cl" }, { "gemm_mv", "gemv.cl" }, { "gemm_mv_quantized", "gemv.cl" }, { "gemm_mm_interleaved_transposed_f16", "gemm.cl" }, { "gemm_mm_interleaved_transposed_f16_bifrost", "gemm.cl" }, { "gemm_mm_interleaved_transposed_f32", "gemm.cl" }, { "gemm_mm_interleaved_transposed_f32_bifrost", "gemm.cl" }, - { "gemm_mm_interleaved_transposed_qs8", "gemm.cl" }, - { "gemm_mm_interleaved_transposed_qs16", "gemm.cl" }, { "gemm_mm_floating_point", "gemm.cl" }, { "gemm_mm_floating_point_f16_bifrost", "gemm.cl" }, { "gemm_mm_floating_point_f32_bifrost", "gemm.cl" }, { "gemm_mm_floating_point_f32_bifrost_1000", "gemm.cl" }, - { "gemm_mm_qs8", "gemm.cl" }, - { "gemm_mm_qs16", "gemm.cl" }, { "gemm_lc_vm_f32", "gemm.cl" }, { "gemm_transpose1xW", "gemm.cl" }, { "gemmlowp_matrix_a_reduction", "gemmlowp.cl" }, @@ -557,10 +551,6 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map = #include "./cl_kernels/fill_border.clembed" }, { - "fixed_point.h", -#include "./cl_kernels/fixed_point.hembed" - }, - { "floor.cl", #include "./cl_kernels/floor.clembed" }, diff --git a/src/core/CL/cl_kernels/activation_layer.cl b/src/core/CL/cl_kernels/activation_layer.cl index a8ea7387d6..373406a6da 100644 --- a/src/core/CL/cl_kernels/activation_layer.cl +++ b/src/core/CL/cl_kernels/activation_layer.cl @@ -25,23 +25,6 @@ #define TYPE VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) -#if defined(FIXED_POINT_POSITION) -#include "fixed_point.h" - -#define CONST_ONE (1 << FIXED_POINT_POSITION) -#define ABS_OP(a) ABS_SAT_OP_EXPAND((a), DATA_TYPE, VEC_SIZE) -#define ADD_OP(a, b) ADD_SAT_OP_EXPAND((a), (b), DATA_TYPE, VEC_SIZE) -#define SUB_OP(a, b) SUB_SAT_OP_EXPAND((a), (b), DATA_TYPE, VEC_SIZE) -#define MUL_OP(a, b) MUL_SAT_OP_EXPAND((a), (b), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION) -#define MLA_OP(a, b, c) MLA_SAT_OP_EXPAND((a), (b), (c), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION) -#define DIV_OP(a, b) DIV_SAT_OP_VEC_EXPAND((a), (b), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION) -#define EXP_OP(a) EXP_OP_EXPAND((a), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION) -#define LOG_OP(a) LOG_OP_EXPAND((a), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION) -#define SQRT_OP(a) DIV_OP(CONST_ONE, INVSQRT_OP_EXPAND((a), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)) -#define TANH_OP(a) TANH_OP_EXPAND((a), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION) - -#else /* FIXED_POINT_POSITION */ - #define CONST_ONE 1.f #define ABS_OP(a) fabs((a)) #define ADD_OP(a, b) ((a) + (b)) @@ -54,8 +37,6 @@ #define SQRT_OP(a) sqrt((a)) #define TANH_OP(a) tanh((a)) -#endif /* FIXED_POINT_POSITION */ - // Logistic Activation inline TYPE logistic_op(TYPE x) { @@ -125,9 +106,8 @@ inline TYPE linear_op(TYPE x) * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 * @note Activation function should be given as a preprocessor argument using -DACT=name. e.g. -DACT=TANH * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively. - * @note In case of fixed point calculations the fixed point position is passed using -DFIXED_POINT_POSITION=position. e.g. -DFIXED_POINT_POSITION=3. * - * @param[in] input_ptr Pointer to the source image. Supported data types: QS8/QS16/F16/F32 + * @param[in] input_ptr Pointer to the source image. Supported data types: F16/F32 * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) diff --git a/src/core/CL/cl_kernels/arithmetic_op.cl b/src/core/CL/cl_kernels/arithmetic_op.cl index 8bd28230b7..9efb71b199 100644 --- a/src/core/CL/cl_kernels/arithmetic_op.cl +++ b/src/core/CL/cl_kernels/arithmetic_op.cl @@ -23,10 +23,6 @@ */ #include "helpers.h" -#if defined(FIXED_POINT_POSITION) -#include "fixed_point.h" -#endif /* FIXED_POINT_POSITION */ - #ifdef SATURATE #define ADD(x, y) add_sat((x), (y)) #define SUB(x, y) sub_sat((x), (y)) @@ -43,7 +39,7 @@ * e.g. -DDATA_TYPE_IN1=uchar -DDATA_TYPE_IN2=uchar -DDATA_TYPE_OUT=short * @attention To perform saturating operation -DSATURATE has to be passed to the compiler otherwise wrapping policy will be used. * - * @param[in] in1_ptr Pointer to the source tensor. Supported data types: U8/QS8/QS16/S16/F16/F32 + * @param[in] in1_ptr Pointer to the source tensor. Supported data types: U8/S16/F16/F32 * @param[in] in1_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] in1_step_x in1_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] in1_stride_y Stride of the source tensor in Y dimension (in bytes) @@ -51,7 +47,7 @@ * @param[in] in1_stride_z Stride of the source tensor in Z dimension (in bytes) * @param[in] in1_step_z in1_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] in1_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[in] in2_ptr Pointer to the source tensor. Supported data types: U8/QS8 (only if @p in1_ptr is QS8), QS16 (only if @p in1_ptr is QS16), S16/F16/F32 + * @param[in] in2_ptr Pointer to the source tensor. Supported data types: U8/S16/F16/F32 * @param[in] in2_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] in2_step_x in2_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] in2_stride_y Stride of the source tensor in Y dimension (in bytes) @@ -59,7 +55,7 @@ * @param[in] in2_stride_z Stride of the source tensor in Z dimension (in bytes) * @param[in] in2_step_z in2_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] in2_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] out_ptr Pointer to the destination tensor. Supported data types: U8 (only if both inputs are U8), QS8 (only if both inputs are QS8), QS16 (only if both inputs are QS16), S16/F16/F32 + * @param[out] out_ptr Pointer to the destination tensor. Supported data types: U8 (only if both inputs are U8), S16/F16/F32 * @param[in] out_stride_x Stride of the destination tensor in X dimension (in bytes) * @param[in] out_step_x out_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] out_stride_y Stride of the destination tensor in Y dimension (in bytes) diff --git a/src/core/CL/cl_kernels/batchnormalization_layer.cl b/src/core/CL/cl_kernels/batchnormalization_layer.cl index 9c980da62a..5352af3c5a 100644 --- a/src/core/CL/cl_kernels/batchnormalization_layer.cl +++ b/src/core/CL/cl_kernels/batchnormalization_layer.cl @@ -25,25 +25,12 @@ #if defined(VEC_SIZE) && defined(DATA_TYPE) -#if defined(FIXED_POINT_POSITION) -#include "fixed_point.h" - -#define ADD_OP(a, b) ADD_SAT_OP_EXPAND((a), (b), DATA_TYPE, VEC_SIZE) -#define SUB_OP(a, b) SUB_SAT_OP_EXPAND((a), (b), DATA_TYPE, VEC_SIZE) -#define MUL_OP(a, b) MUL_SAT_OP_EXPAND((a), (b), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION) -#define INVSQRT_OP(a) INVSQRT_OP_EXPAND((a), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION) -#define SQCVT_SAT(a) SQCVT_SAT_OP_EXPAND((a), DATA_TYPE, FIXED_POINT_POSITION) - -#else /* FIXED_POINT_POSITION */ - #define ADD_OP(a, b) ((a) + (b)) #define SUB_OP(a, b) ((a) - (b)) #define MUL_OP(a, b) ((a) * (b)) #define INVSQRT_OP(a) rsqrt((a)) #define SQCVT_SAT(a) (a) -#endif /* FIXED_POINT_POSITION */ - #if defined(FUSED_ACTIVATION) #include "activation_layer.cl" #define ACTIVATION_FUNC(x) ACTIVATION_OP(FUSED_ACTIVATION, x) @@ -53,7 +40,7 @@ /** Apply batch normalization. * - * @param[in] input_ptr Pointer to the first source tensor. Supported data types: QS8/QS16/F16/F32 + * @param[in] input_ptr Pointer to the first source tensor. Supported data types: F16/F32 * @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes) * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes) @@ -163,7 +150,7 @@ __kernel void batchnormalization_layer_nchw(TENSOR3D_DECLARATION(input), /** Apply batch normalization on tensors with NHWC format. * - * @param[in] input_ptr Pointer to the first source tensor. Supported data types: QS8/QS16/F16/F32 + * @param[in] input_ptr Pointer to the first source tensor. Supported data types: F16/F32 * @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes) * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes) diff --git a/src/core/CL/cl_kernels/channel_shuffle.cl b/src/core/CL/cl_kernels/channel_shuffle.cl index 26cee9ccdd..23962e1c2e 100644 --- a/src/core/CL/cl_kernels/channel_shuffle.cl +++ b/src/core/CL/cl_kernels/channel_shuffle.cl @@ -38,7 +38,7 @@ * @note The number of channels in each group should be given as a preprocessor argument using -DK=num. e.g. -DK=1 * K is equal to num_channels / num_groups. * - * @param[in] src_ptr Pointer to the source matrix. Supported data types: U8/S8/QS8/QASYMM8/U16/S16/QS16/F16/U32/S32/F32 + * @param[in] src_ptr Pointer to the source matrix. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32 * @param[in] src_stride_x Stride of the first source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the first source tensor in Y dimension (in bytes) diff --git a/src/core/CL/cl_kernels/col2im.cl b/src/core/CL/cl_kernels/col2im.cl index 6e491f33cf..98bf8d1ed4 100644 --- a/src/core/CL/cl_kernels/col2im.cl +++ b/src/core/CL/cl_kernels/col2im.cl @@ -23,12 +23,7 @@ */ #include "helpers.h" -#if defined(FIXED_POINT_POSITION) -#include "fixed_point.h" -#endif // FIXED_POINT_POSITION - #if defined(DATA_TYPE) && defined(WIDTH_OUTPUT) && defined(ELEMENT_SIZE) && defined(WIDTH_INPUT) -#if !defined(FIXED_POINT_POSITION) #if ELEMENT_SIZE == 1 #define COND_DATA_TYPE char @@ -100,41 +95,4 @@ __kernel void col2im( *((__global DATA_TYPE *)(output_ptr + idx + x_clamped.s6 * dst_stride_z)) = data.s6; *((__global DATA_TYPE *)(output_ptr + idx + x_clamped.s7 * dst_stride_z)) = data.s7; } -#else // !defined(FIXED_POINT_POSITION) -/** This kernel performs a reshaping of the output of the convolution layer. - * - * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=qs8 - * @note The width of the output tensor must be passed at compile time using -DWIDTH_OUTPUT: e.g. -DWIDTH_OUTPUT=320 - * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QS16 - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes) - */ -__kernel void col2im( - IMAGE_DECLARATION(src), - TENSOR3D_DECLARATION(dst), - uint dst_stride_w) -{ - Image src = CONVERT_TO_IMAGE_STRUCT(src); - Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(dst); - - // Compute output offset - int idx = get_global_id(0) * dst.stride_z + (get_global_id(1) / WIDTH_OUTPUT) * dst_stride_y + (get_global_id(1) % WIDTH_OUTPUT) * dst_stride_x + get_global_id(2) * dst_stride_w; - - // Store value - *((__global DATA_TYPE *)(dst.ptr + idx)) = *((__global DATA_TYPE *)(src.ptr)); -} -#endif // !defined(FIXED_POINT_POSITION) #endif // defined(DATA_TYPE) && defined(WIDTH_OUTPUT) && defined(ELEMENT_SIZE) && defined(WIDTH_INPUT)
\ No newline at end of file diff --git a/src/core/CL/cl_kernels/concatenate.cl b/src/core/CL/cl_kernels/concatenate.cl index f97ae13a9a..6ec8383c52 100644 --- a/src/core/CL/cl_kernels/concatenate.cl +++ b/src/core/CL/cl_kernels/concatenate.cl @@ -25,7 +25,7 @@ /** This kernel concatenates the input tensor into the output tensor along the first dimension * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8, QASYMM8, QS16, F16, F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8, F16, F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) @@ -60,7 +60,7 @@ __kernel void concatenate_width( /** This kernel concatenates the input tensor into the output tensor along the third dimension * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8, QS16, F16, F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16, F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) diff --git a/src/core/CL/cl_kernels/convert_fc_weights.cl b/src/core/CL/cl_kernels/convert_fc_weights.cl index 3c3e8b0dc4..5aadfb36f9 100644 --- a/src/core/CL/cl_kernels/convert_fc_weights.cl +++ b/src/core/CL/cl_kernels/convert_fc_weights.cl @@ -32,7 +32,7 @@ * @attention Data type can be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float * @attention Original input tensor width*height and depth should be given as a preprocessor argument using -DFACTOR_1=size and -DFACTOR_2=size for NCHW and vice versa for NHWC. e.g. -DFACTOR_1=256 and -DFACTOR_2=128 * - * @param[in] src_ptr Pointer to the source image. Supported data types: U8, S8, QS8, QASYMM8, U16, S16, QS16, U32, S32, QS32, F16, F32 + * @param[in] src_ptr Pointer to the source image. Supported data types: U8, S8, QASYMM8, U16, S16, U32, S32, QS32, F16, F32 * @param[in] src_stride_x Stride of the source image in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source image in Y dimension (in bytes) diff --git a/src/core/CL/cl_kernels/convolution_layer.cl b/src/core/CL/cl_kernels/convolution_layer.cl index 6a70b009c8..2b83e5adf1 100644 --- a/src/core/CL/cl_kernels/convolution_layer.cl +++ b/src/core/CL/cl_kernels/convolution_layer.cl @@ -23,10 +23,6 @@ */ #include "helpers.h" -#if defined(FIXED_POINT_POSITION) -#include "fixed_point.h" -#endif // FIXED_POINT_POSITION - #if defined(DATA_TYPE) /** This kernel reshapes the tensor's low three dimensions to single column * diff --git a/src/core/CL/cl_kernels/depth_convert.cl b/src/core/CL/cl_kernels/depth_convert.cl index a9b7284c83..01491ec1b7 100644 --- a/src/core/CL/cl_kernels/depth_convert.cl +++ b/src/core/CL/cl_kernels/depth_convert.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016, 2017 ARM Limited. + * Copyright (c) 2016-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -23,23 +23,6 @@ */ #include "helpers.h" -#if defined(FIXED_POINT_POSITION) - -#include "fixed_point.h" - -#ifdef SATURATE -#define CONVERT_DOWN(x, in_type, out_type, fixed_point_position) CONVERT_DOWN1_SAT(x, in_type, out_type, fixed_point_position) -#define CONVERT_DOWN1_SAT(x, in_type, out_type, fixed_point_position) convert_##out_type##_##in_type##_sat(x, fixed_point_position) -#else /* SATURATE */ -#define CONVERT_DOWN(x, in_type, out_type, fixed_point_position) CONVERT_DOWN1(x, in_type, out_type, fixed_point_position) -#define CONVERT_DOWN1(x, in_type, out_type, fixed_point_position) convert_##out_type##_##in_type(x, fixed_point_position) -#endif /* SATURATE */ - -#define CONVERT_UP(x, in_type, out_type, fixed_point_position) CONVERT_UP1(x, in_type, out_type, fixed_point_position) -#define CONVERT_UP1(x, in_type, out_type, fixed_point_position) convert_##out_type##_##in_type(x, fixed_point_position) - -#else /* FIXED_POINT_POSITION */ - #ifdef SATURATE #define CONVERT_DOWN(x, type) CONVERT_SAT(x, type) #else /* SATURATE */ @@ -48,22 +31,18 @@ #define CONVERT_UP(x, type) CONVERT(x, type) -#endif /* FIXED_POINT_POSITION */ - /** This function performs a down-scaling depth conversion. * * @attention The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN and -DDATA_TYPE_OUT: * e.g. -DDATA_TYPE_IN=uchar -DDATA_TYPE_OUT=short * - * @note In case of fixed-point operation -DFIXED_POINT_POSITION=fixed_point_position must be provided: e.g. -DFIXED_POINT_POSITION=3 - * * @param[in] in_ptr Pointer to the source image. Supported data types: U8, U16, S16, U32, S32, F16, F32 * @param[in] in_stride_x Stride of the source image in X dimension (in bytes) * @param[in] in_step_x in_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes) * @param[in] in_step_y in_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] in_offset_first_element_in_bytes The offset of the first element in the source image - * @param[out] out_ptr Pointer to the destination image. Supported data types: QS8, U8, QS16, U16, S16, U32, S32 + * @param[out] out_ptr Pointer to the destination image. Supported data types: U8, U16, S16, U32, S32 * @param[in] out_stride_x Stride of the destination image in X dimension (in bytes) * @param[in] out_step_x out_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] out_stride_y Stride of the destination image in Y dimension (in bytes) @@ -84,11 +63,7 @@ __kernel void convert_depth_down( VEC_DATA_TYPE(DATA_TYPE_IN, 16) in_data = vload16(0, (__global DATA_TYPE_IN *)in.ptr); -#if defined(FIXED_POINT_POSITION) - vstore16(CONVERT_DOWN(in_data, VEC_DATA_TYPE(DATA_TYPE_IN, 16), VEC_DATA_TYPE(DATA_TYPE_OUT, 16), FIXED_POINT_POSITION), 0, (__global DATA_TYPE_OUT *)out.ptr); -#else /* FIXED_POINT_POSITION */ vstore16(CONVERT_DOWN(in_data >> shift, VEC_DATA_TYPE(DATA_TYPE_OUT, 16)), 0, (__global DATA_TYPE_OUT *)out.ptr); -#endif /* FIXED_POINT_POSITION */ } /** This function performs a up-scaling depth conversion. @@ -96,9 +71,7 @@ __kernel void convert_depth_down( * @attention The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN and -DDATA_TYPE_OUT: * e.g. -DDATA_TYPE_IN=uchar -DDATA_TYPE_OUT=short * - * @note In case of fixed-point operation -DFIXED_POINT_POSITION=fixed_point_position must be provided: e.g. -DFIXED_POINT_POSITION=3 - * - * @param[in] in_ptr Pointer to the source image. Supported data types: U8, QS8, U16, S16, QS16, U32 or S32 + * @param[in] in_ptr Pointer to the source image. Supported data types: U8, U16, S16, U32 or S32 * @param[in] in_stride_x Stride of the source image in X dimension (in bytes) * @param[in] in_step_x in_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes) @@ -125,9 +98,5 @@ __kernel void convert_depth_up( VEC_DATA_TYPE(DATA_TYPE_IN, 16) in_data = vload16(0, (__global DATA_TYPE_IN *)in.ptr); -#if defined(FIXED_POINT_POSITION) - vstore16(CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_IN, 16), VEC_DATA_TYPE(DATA_TYPE_OUT, 16), FIXED_POINT_POSITION), 0, (__global DATA_TYPE_OUT *)out.ptr); -#else /* FIXED_POINT_POSITION */ vstore16(CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, 16)) << shift, 0, (__global DATA_TYPE_OUT *)out.ptr); -#endif /* FIXED_POINT_POSITION */ } diff --git a/src/core/CL/cl_kernels/depthwise_convolution.cl b/src/core/CL/cl_kernels/depthwise_convolution.cl index f3aa0d6dd8..9a8b57e4c4 100644 --- a/src/core/CL/cl_kernels/depthwise_convolution.cl +++ b/src/core/CL/cl_kernels/depthwise_convolution.cl @@ -527,7 +527,7 @@ __kernel void depthwise_weights_reshape( * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float * @note The convolution information must be passed at compile time using -DSTRIDE_X, -DSTRIDE_Y, -DPAD_LEFT, -DPAD_TOP, -DPAD_RIGHT, -DPAD_BOTTOM, -DKERNEL_WIDHT, -DKERNEL_HEIGHT, -DSRC_WIDTH, -DSRC_HEIGHT, -DDEPTH_MULTIPLIER * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QS16/F16/F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) @@ -587,7 +587,7 @@ __kernel void depthwise_im2col(TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(d * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float * @note The convolution information must be passed at compile time using -DCONV_WIDTH, -DCONV_HEIGHT, e.g -DCONV_WIDTH=32, -DCONV_HEIGHT=42 * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QS16/F16/F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor diff --git a/src/core/CL/cl_kernels/dequantization_layer.cl b/src/core/CL/cl_kernels/dequantization_layer.cl index 21e9c873ac..4908bb0b31 100644 --- a/src/core/CL/cl_kernels/dequantization_layer.cl +++ b/src/core/CL/cl_kernels/dequantization_layer.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -25,7 +25,7 @@ /** This performs the dequantization of 8-bit unsigned integers to floating point. * - * @param[in] input_ptr Pointer to the source image. Supported data types: QS8/QS16/F16/F32 + * @param[in] input_ptr Pointer to the source image. Supported data types: F16/F32 * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) diff --git a/src/core/CL/cl_kernels/direct_convolution1x1.cl b/src/core/CL/cl_kernels/direct_convolution1x1.cl index 817c261ba2..7a308c99e2 100644 --- a/src/core/CL/cl_kernels/direct_convolution1x1.cl +++ b/src/core/CL/cl_kernels/direct_convolution1x1.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016, 2017 ARM Limited. + * Copyright (c) 2016-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -23,24 +23,12 @@ */ #include "helpers.h" -#if defined(FIXED_POINT_POSITION) -#include "fixed_point.h" - -#define ADD_OP(a, b) ADD_SAT_OP_EXPAND((a), (b), DATA_TYPE_PROMOTED, 8) -#define MUL_OP(a, b) MUL_SAT_OP_EXPAND(CONVERT((a), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)), CONVERT((b), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)), DATA_TYPE_PROMOTED, 8, FIXED_POINT_POSITION) - -// There is no need to have a larger intermediate type for qs32 because all the arguments are already promoted -MULQ_SAT_IMPL(qs32x8, qs32x8) - -#else /* FIXED_POINT_POSITION */ #undef CONVERT_SAT #define ADD_OP(a, b) ((a) + (b)) #define MUL_OP(a, b) ((a) * (b)) #define CONVERT_SAT(a, b) ((a)) -#endif /* FIXED_POINT_POSITION */ - #if defined(DATA_TYPE) && defined(DATA_SIZE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) #if STRIDE_X == 3 diff --git a/src/core/CL/cl_kernels/direct_convolution3x3.cl b/src/core/CL/cl_kernels/direct_convolution3x3.cl index a7abc9ff1d..824306f2ba 100644 --- a/src/core/CL/cl_kernels/direct_convolution3x3.cl +++ b/src/core/CL/cl_kernels/direct_convolution3x3.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016, 2017 ARM Limited. + * Copyright (c) 2016-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -23,25 +23,12 @@ */ #include "helpers.h" -#if defined(FIXED_POINT_POSITION) -#include "fixed_point.h" - -#define ADD_OP(a, b) ADD_SAT_OP_EXPAND((a), (b), DATA_TYPE_PROMOTED, 8) -#define MUL_OP(a, b) MUL_SAT_OP_EXPAND(CONVERT((a), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)), CONVERT((b), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)), DATA_TYPE_PROMOTED, 8, FIXED_POINT_POSITION) - -// There is no need to have a larger intermediate type for qs32 because all the arguments are already promoted -MULQ_SAT_IMPL(qs32x8, qs32x8) - -#else /* FIXED_POINT_POSITION */ - #undef CONVERT_SAT #define ADD_OP(a, b) ((a) + (b)) #define MUL_OP(a, b) ((a) * (b)) #define CONVERT_SAT(a, b) ((a)) -#endif /* FIXED_POINT_POSITION */ - #if defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) #if STRIDE_X == 1 @@ -86,7 +73,7 @@ MULQ_SAT_IMPL(qs32x8, qs32x8) * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH * @note If biases are used then -DHAS_BIAS has to be passed at compile time * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QS16/F16/F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) diff --git a/src/core/CL/cl_kernels/fill_border.cl b/src/core/CL/cl_kernels/fill_border.cl index 33a9495d66..9d6a2b8b5a 100644 --- a/src/core/CL/cl_kernels/fill_border.cl +++ b/src/core/CL/cl_kernels/fill_border.cl @@ -23,10 +23,6 @@ */ #include "helpers.h" -#if defined(FIXED_POINT_POSITION) -#include "fixed_point.h" -#endif /* FIXED_POINT_POSITION */ - /** Fill N pixel of the padding edge of a single channel image by replicating the closest valid pixel. * * @attention The DATA_TYPE needs to be passed at the compile time. diff --git a/src/core/CL/cl_kernels/fixed_point.h b/src/core/CL/cl_kernels/fixed_point.h deleted file mode 100644 index 46fa645c2b..0000000000 --- a/src/core/CL/cl_kernels/fixed_point.h +++ /dev/null @@ -1,518 +0,0 @@ -/* - * Copyright (c) 2017-2018 ARM Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_FIXED_POINT_H -#define ARM_COMPUTE_FIXED_POINT_H - -#define TYPE_ALIAS(type, alias) \ - typedef type alias; \ - typedef type alias##x##1; \ - typedef type##2 alias##x##2; \ - typedef type##3 alias##x##3; \ - typedef type##4 alias##x##4; \ - typedef type##8 alias##x##8; \ - typedef type##16 alias##x##16; - -TYPE_ALIAS(char, qs8) -TYPE_ALIAS(short, qs16) -TYPE_ALIAS(int, qs32) - -#define qs8_MIN ((char)CHAR_MIN) -#define qs8_MAX ((char)CHAR_MAX) -#define qs16_MIN ((short)SHRT_MIN) -#define qs16_MAX ((short)SHRT_MAX) -#define qs32_MIN ((int)INT_MIN) -#define qs32_MAX ((int)INT_MAX) - -#define qu8_MIN ((uchar)0) -#define qu8_MAX ((uchar)UCHAR_MAX) -#define qu16_MIN ((ushort)0) -#define qu16_MAX ((ushort)USHRT_MAX) -#define qu32_MIN ((uint)0) -#define qu32_MAX ((uint)UINT_MAX) - -#define qs8_TYPE char -#define qs8x1_TYPE char -#define qs8x2_TYPE char2 -#define qs8x3_TYPE char3 -#define qs8x4_TYPE char4 -#define qs8x8_TYPE char8 -#define qs8x16_TYPE char16 - -#define qs16_TYPE short -#define qs16x1_TYPE short -#define qs16x2_TYPE short2 -#define qs16x3_TYPE short3 -#define qs16x4_TYPE short4 -#define qs16x8_TYPE short8 -#define qs16x16_TYPE short16 - -#define qs32_TYPE int -#define qs32x1_TYPE int -#define qs32x2_TYPE int2 -#define qs32x3_TYPE int3 -#define qs32x4_TYPE int4 -#define qs32x8_TYPE int8 -#define qs32x16_TYPE int16 - -/* All internal constants are represented in the maximum supported fixed point format (QS16), - * thus we define an additional shift parameter required to convert the constant - * from the maximum supported format to the require one. - */ -#define qs8_SHIFT 8 -#define qs16_SHIFT 0 - -#undef VEC_DATA_TYPE_STR -#undef VEC_DATA_TYPE -#undef CONVERT_STR -#undef CONVERT -#undef CONVERT_SAT_STR -#undef CONVERT_SAT - -#define VEC_DATA_TYPE_STR(type, size) type##x##size -#define VEC_DATA_TYPE(type, size) VEC_DATA_TYPE_STR(type, size) - -#define CONVERT_STR3(x, type, rtype) (convert_##rtype((x))) -#define CONVERT_STR2(x, type, rtype) CONVERT_STR3(x, type, rtype) -#define CONVERT_STR(x, type) CONVERT_STR2(x, type, type##_TYPE) -#define CONVERT(x, type) CONVERT_STR(x, type) - -#define CONVERT_SAT_STR3(x, type, rtype) (convert_##rtype##_sat((x))) -#define CONVERT_SAT_STR2(x, type, rtype) CONVERT_SAT_STR3(x, type, rtype) -#define CONVERT_SAT_STR(x, type) CONVERT_SAT_STR2(x, type, type##_TYPE) -#define CONVERT_SAT(x, type) CONVERT_SAT_STR(x, type) - -/** Computes saturating absolute value of fixed point vector. - * - * @param[in] type the actual data type. - * - * @return The result of the fixed point absolute value. - */ -#define ABSQ_SAT_IMPL(type) \ - inline type abs_##type##_sat(type VopA) \ - { \ - return CONVERT_SAT(abs(VopA), type); \ - } - -ABSQ_SAT_IMPL(qs8x16) -ABSQ_SAT_IMPL(qs16x8) - -#define ABS_SAT_OP_EXPAND_STR(a, type, size) abs_##type##x##size##_sat((a)) -#define ABS_SAT_OP_EXPAND(a, type, size) ABS_SAT_OP_EXPAND_STR(a, type, size) - -/** Computes max of fixed point types. - * - * @param[in] type the actual data type. - * - * @return The result of the fixed point maximum. - */ -#define MAXQ_IMPL(type) \ - inline type max_##type(type VopA, type VopB) \ - { \ - return max(VopA, VopB); \ - } - -MAXQ_IMPL(qs8x1) -MAXQ_IMPL(qs8x2) -MAXQ_IMPL(qs8x4) -MAXQ_IMPL(qs8x8) -MAXQ_IMPL(qs8x16) -MAXQ_IMPL(qs16x1) -MAXQ_IMPL(qs16x2) -MAXQ_IMPL(qs16x4) -MAXQ_IMPL(qs16x8) -MAXQ_IMPL(qs16x16) - -#define MAX_OP_EXPAND_STR(a, b, type, size) max_##type##x##size((a), (b)) -#define MAX_OP_EXPAND(a, b, type, size) MAX_OP_EXPAND_STR(a, b, type, size) - -/** Computes saturated addition of fixed point types. - * - * @param[in] type the actual data type. - * - * @return The result of the fixed point addition. The result is saturated in case of overflow - */ -#define ADDQ_SAT_IMPL(type) \ - inline type add_sat_##type(type VopA, type VopB) \ - { \ - return add_sat(VopA, VopB); \ - } - -ADDQ_SAT_IMPL(qs8x1) -ADDQ_SAT_IMPL(qs8x2) -ADDQ_SAT_IMPL(qs8x4) -ADDQ_SAT_IMPL(qs8x8) -ADDQ_SAT_IMPL(qs8x16) -ADDQ_SAT_IMPL(qs16x1) -ADDQ_SAT_IMPL(qs16x2) -ADDQ_SAT_IMPL(qs16x4) -ADDQ_SAT_IMPL(qs16x8) -ADDQ_SAT_IMPL(qs16x16) -ADDQ_SAT_IMPL(qs32x1) -ADDQ_SAT_IMPL(qs32x2) -ADDQ_SAT_IMPL(qs32x4) -ADDQ_SAT_IMPL(qs32x8) -ADDQ_SAT_IMPL(qs32x16) - -#define ADD_SAT_OP_EXPAND_STR(a, b, type, size) add_sat_##type##x##size((a), (b)) -#define ADD_SAT_OP_EXPAND(a, b, type, size) ADD_SAT_OP_EXPAND_STR(a, b, type, size) - -/** Computes saturated subtraction of fixed point types. - * - * @param[in] type the actual data type. - * - * @return The result of the fixed point subtraction. The result is saturated in case of overflow - */ -#define SUBQ_SAT_IMPL(type) \ - inline type sub_sat_##type(type VopA, type VopB) \ - { \ - return sub_sat(VopA, VopB); \ - } - -SUBQ_SAT_IMPL(qs8x1) -SUBQ_SAT_IMPL(qs8x2) -SUBQ_SAT_IMPL(qs8x4) -SUBQ_SAT_IMPL(qs8x8) -SUBQ_SAT_IMPL(qs8x16) -SUBQ_SAT_IMPL(qs16x1) -SUBQ_SAT_IMPL(qs16x2) -SUBQ_SAT_IMPL(qs16x4) -SUBQ_SAT_IMPL(qs16x8) -SUBQ_SAT_IMPL(qs16x16) - -#define SUB_SAT_OP_EXPAND_STR(a, b, type, size) sub_sat_##type##x##size((a), (b)) -#define SUB_SAT_OP_EXPAND(a, b, type, size) SUB_SAT_OP_EXPAND_STR(a, b, type, size) - -/* Multiply of two fixed point numbers - * - * @param[in] type the actual data type. - * @param[in] itype the intermediate data type. - * - * @return The result of the fixed point multiplication. - */ -#define MULQ_IMPL(type, itype) \ - inline type mul_##type(type VopA, type VopB, int fixed_point_position) \ - { \ - itype round_val = (itype)(1 << (fixed_point_position - 1)); \ - itype res = CONVERT((VopA), itype) * CONVERT((VopB), itype) + round_val; \ - return CONVERT((res >> (itype)fixed_point_position), type); \ - } - -MULQ_IMPL(qs8x8, qs16x8) -MULQ_IMPL(qs16x8, qs32x8) -MULQ_IMPL(qs8x16, qs16x16) -MULQ_IMPL(qs16x16, qs32x16) - -#define MUL_OP_EXPAND_STR(a, b, type, size, position) mul_##type##x##size((a), (b), (position)) -#define MUL_OP_EXPAND(a, b, type, size, position) MUL_OP_EXPAND_STR(a, b, type, size, position) - -/* Saturate multiply of two fixed point numbers - * - * @param[in] type the actual data type. - * @param[in] itype the intermediate data type. - * - * @return The result of the fixed point multiplication. The result is saturated in case of overflow - */ -#define MULQ_SAT_IMPL(type, itype) \ - inline type mul_sat_##type(type VopA, type VopB, int fixed_point_position) \ - { \ - itype round_val = (itype)(1 << (fixed_point_position - 1)); \ - itype res = mad_sat(CONVERT((VopA), itype), CONVERT((VopB), itype), round_val); \ - return CONVERT_SAT((res >> (itype)fixed_point_position), type); \ - } - -MULQ_SAT_IMPL(qs8x1, qs16x1) -MULQ_SAT_IMPL(qs8x2, qs16x2) -MULQ_SAT_IMPL(qs8x3, qs16x3) -MULQ_SAT_IMPL(qs8x4, qs16x4) -MULQ_SAT_IMPL(qs8x8, qs16x8) -MULQ_SAT_IMPL(qs8x16, qs16x16) -MULQ_SAT_IMPL(qs16x1, qs32x1) -MULQ_SAT_IMPL(qs16x2, qs32x2) -MULQ_SAT_IMPL(qs16x3, qs32x3) -MULQ_SAT_IMPL(qs16x4, qs32x4) -MULQ_SAT_IMPL(qs16x8, qs32x8) -MULQ_SAT_IMPL(qs16x16, qs32x16) - -#define MUL_SAT_OP_EXPAND_STR(a, b, type, size, position) mul_sat_##type##x##size((a), (b), (position)) -#define MUL_SAT_OP_EXPAND(a, b, type, size, position) MUL_SAT_OP_EXPAND_STR(a, b, type, size, position) - -/** Saturate multiply-accumulate - * - * @param[in] type the actual data type. - * @param[in] itype the intermediate data type. - * - * @return The result of the fixed point multiply-accumulate. The result is saturated in case of overflow - */ -#define MLAQ_SAT_IMPL(type, itype) \ - type mla_sat_##type(type VopA, type VopB, type VopC, int fixed_point_position) \ - { \ - itype res = mad_sat(CONVERT(VopB, itype), CONVERT(VopC, itype), (itype)(1 << (fixed_point_position - 1))); \ - return add_sat(VopA, CONVERT_SAT(res >> (itype)fixed_point_position, type)); \ - } - -MLAQ_SAT_IMPL(qs8x8, qs16x8) -MLAQ_SAT_IMPL(qs8x16, qs16x16) -MLAQ_SAT_IMPL(qs16x8, qs32x8) - -#define MLA_SAT_OP_EXPAND_STR(a, b, c, type, size, position) mla_sat_##type##x##size((a), (b), (c), (position)) -#define MLA_SAT_OP_EXPAND(a, b, c, type, size, position) MLA_SAT_OP_EXPAND_STR(a, b, c, type, size, position) - -/** Saturate multiply-accumulate long - * - * @param[in] type the actual data type. - * @param[in] itype the intermediate data type. - * - * @return The result of the fixed point multiply-accumulate long. The result is saturated in case of overflow - */ -#define MLALQ_SAT_IMPL(type, itype) \ - itype mlal_sat_##type(itype VopA, type VopB, type VopC, int fixed_point_position) \ - { \ - itype res = mad_sat(CONVERT(VopB, itype), CONVERT(VopC, itype), (itype)(1 << (fixed_point_position - 1))); \ - return add_sat(VopA, res >> (itype)fixed_point_position); \ - } - -MLALQ_SAT_IMPL(qs8x8, qs16x8) -MLALQ_SAT_IMPL(qs16x8, qs32x8) - -#define MLAL_SAT_OP_EXPAND_STR(a, b, c, type, size, position) mlal_sat_##type##x##size((a), (b), (c), (position)) -#define MLAL_SAT_OP_EXPAND(a, b, c, type, size, position) MLAL_SAT_OP_EXPAND_STR(a, b, c, type, size, position) - -/** Saturate division of two fixed point vectors - * - * @param[in] stype the actual scalar data type. - * @param[in] type the actual data type. - * @param[in] itype the intermediate data type. - * - * @return The result of the fixed point division. The result is saturated in case of overflow - */ -#define DIVQ_SAT_IMPL(stype, type, itype) \ - inline type div_sat_##type(type VopA, type VopB, int fixed_point_position) \ - { \ - itype conv_a = CONVERT((VopA), itype); \ - itype denominator = CONVERT((VopB), itype); \ - itype numerator = conv_a << (itype)(fixed_point_position); \ - itype res = select((itype)(numerator / denominator), select((itype)stype##_MAX, (itype)stype##_MIN, (itype)(conv_a < (itype)0)), (itype)(denominator == (itype)0)); \ - return CONVERT_SAT((res), type); \ - } - -DIVQ_SAT_IMPL(qs8, qs8x16, qs16x16) -DIVQ_SAT_IMPL(qs16, qs16x8, qs32x8) -DIVQ_SAT_IMPL(qs16, qs16x16, qs32x16) -DIVQ_SAT_IMPL(qs8, qs8, qs16) -DIVQ_SAT_IMPL(qs16, qs16, qs32) - -#define DIV_SAT_OP_EXPAND_STR(a, b, type, position) div_sat_##type((a), (b), (position)) -#define DIV_SAT_OP_EXPAND(a, b, type, position) DIV_SAT_OP_EXPAND_STR(a, b, type, position) - -#define DIV_SAT_OP_VEC_EXPAND_STR(a, b, type, size, position) div_sat_##type##x##size((a), (b), (position)) -#define DIV_SAT_OP_VEC_EXPAND(a, b, type, size, position) DIV_SAT_OP_VEC_EXPAND_STR(a, b, type, size, position) - -/** Saturate exponential of a fixed point vector - * - * @note Implemented approach uses taylor polynomial to approximate the exponential function. - * - * @param[in] stype the actual scalar data type. - * @param[in] type the actual data type. - * @param[in] size the number of the calculated elements. - * - * @return The result of the fixed point exponential. The result is saturated in case of overflow - */ -#define EXPQ_IMPL(stype, type, size) \ - inline type exp_sat_##type(type VopA, int fixed_point_position) \ - { \ - type const_one = (type)(1 << (fixed_point_position)); \ - type ln2 = (type)((((0x58B9 >> (14 - fixed_point_position))) + 1) >> 1); \ - type inv_ln2 = (type)((((0x38AA >> (14 - fixed_point_position)) + 1) >> 1)) | const_one; \ - type A = (type)(((0x7FBA >> (14 - fixed_point_position)) + 1) >> 1); \ - type B = (type)(((0x3FE9 >> (14 - fixed_point_position)) + 1) >> 1); \ - type C = (type)(((0x1693 >> (14 - fixed_point_position)) + 1) >> 1); \ - type D = (type)(((0x0592 >> (14 - fixed_point_position)) + 1) >> 1); \ - type m = MUL_SAT_OP_EXPAND(VopA, inv_ln2, stype, size, fixed_point_position); \ - type dec_m = m >> (type)fixed_point_position; \ - type alpha = MUL_SAT_OP_EXPAND(dec_m << (type)fixed_point_position, ln2, stype, size, fixed_point_position); \ - alpha = CONVERT(abs_diff(VopA, alpha), type); \ - type sum = add_sat(MUL_SAT_OP_EXPAND(alpha, D, stype, size, fixed_point_position), C); \ - sum = add_sat(MUL_SAT_OP_EXPAND(alpha, sum, stype, size, fixed_point_position), B); \ - sum = add_sat(MUL_SAT_OP_EXPAND(alpha, sum, stype, size, fixed_point_position), A); \ - sum = add_sat(MUL_SAT_OP_EXPAND(alpha, sum, stype, size, fixed_point_position), const_one); \ - return select((type)stype##_MAX, select(sum << dec_m, sum >> -dec_m, dec_m < (type)0), clz(sum) > dec_m); /* Saturate result if needed */ \ - } - -EXPQ_IMPL(qs8, qs8x2, 2) -EXPQ_IMPL(qs8, qs8x4, 4) -EXPQ_IMPL(qs8, qs8x8, 8) -EXPQ_IMPL(qs8, qs8x16, 16) -EXPQ_IMPL(qs16, qs16x2, 2) -EXPQ_IMPL(qs16, qs16x4, 4) -EXPQ_IMPL(qs16, qs16x8, 8) -EXPQ_IMPL(qs16, qs16x16, 16) - -#define EXP_OP_EXPAND_STR(a, type, size, position) exp_sat_##type##x##size((a), (position)) -#define EXP_OP_EXPAND(a, type, size, position) EXP_OP_EXPAND_STR(a, type, size, position) - -/** Saturate logarithm of a fixed point vector - * - * @note Implemented approach uses taylor polynomial to approximate the logarithm function. - * - * @param[in] stype the actual scalar data type. - * @param[in] type the actual data type. - * @param[in] size the number of the calculated elements. - * - * @return The result of the fixed point logarithm. The result is saturated in case of overflow - */ -#define LOGQ_IMPL(stype, type, size) \ - inline type log_sat_##type(type VopA, int fixed_point_position) \ - { \ - type const_one = (type)(1 << (fixed_point_position)); \ - type ln2 = (type)(0x58B9 >> (15 - fixed_point_position)); /* 1.4384189 */ \ - type A = (type)(0x5C0F >> (14 - fixed_point_position)); /* 1.4384189 */ \ - type B = -(type)(0x56AE >> (15 - fixed_point_position)); /* -0.6771900 */ \ - type C = (type)(0x2933 >> (15 - fixed_point_position)); /* 0.3218538 */ \ - type D = -(type)(0x0AA7 >> (15 - fixed_point_position)); /* -0.0832229 */ \ - type inter_a = select(VopA, DIV_SAT_OP_VEC_EXPAND(const_one, VopA, stype, size, fixed_point_position), VopA < const_one); \ - type shift_val = (type)(15 - stype##_SHIFT) - clz(inter_a >> (type)fixed_point_position); \ - inter_a = inter_a >> shift_val; \ - inter_a = sub_sat(inter_a, const_one); \ - type sum = add_sat(MUL_SAT_OP_EXPAND(inter_a, D, stype, size, fixed_point_position), C); \ - sum = add_sat(MUL_SAT_OP_EXPAND(inter_a, sum, stype, size, fixed_point_position), B); \ - sum = add_sat(MUL_SAT_OP_EXPAND(inter_a, sum, stype, size, fixed_point_position), A); \ - sum = MUL_SAT_OP_EXPAND(inter_a, sum, stype, size, fixed_point_position); \ - sum = MUL_SAT_OP_EXPAND(add_sat(sum, shift_val << (type)fixed_point_position), ln2, stype, size, fixed_point_position); \ - return select(select(sum, -sum, VopA < const_one), (type)0, VopA < (type)0); /* Saturate result if needed */ \ - } - -LOGQ_IMPL(qs8, qs8x16, 16) -LOGQ_IMPL(qs16, qs16x8, 8) -LOGQ_IMPL(qs16, qs16x16, 16) - -#define LOG_OP_EXPAND_STR(a, type, size, position) log_sat_##type##x##size((a), (position)) -#define LOG_OP_EXPAND(a, type, size, position) LOG_OP_EXPAND_STR(a, type, size, position) - -/** Saturate inverse square root of a fixed point vector - * - * @note Implemented approach uses Newton's method to approximate the inverse square root function. - * - * @param[in] stype the actual scalar data type. - * @param[in] type the actual data type. - * @param[in] size the number of the calculated elements. - * - * @return The result of the fixed point inverse square root. The result is saturated in case of overflow - */ -#define INVSQRTQ_IMPL(stype, type, size) \ - inline type invsqrt_sat_##type(type VopA, int fixed_point_position) \ - { \ - type const_three = (type)(3 << (fixed_point_position)); \ - type shift_value = (type)(16 - stype##_SHIFT) - (clz(VopA) + (type)fixed_point_position); \ - type temp = select((type)(VopA >> shift_value), select((type)stype##_MAX, (type)(VopA << (-shift_value)), (type)(clz(VopA) > (-shift_value))), (type)(shift_value < (type)0)); \ - type x = temp; \ - x = MUL_SAT_OP_EXPAND(x, sub_sat(const_three, MUL_SAT_OP_EXPAND(MUL_SAT_OP_EXPAND(x, x, stype, size, fixed_point_position), temp, stype, size, fixed_point_position)), stype, size, fixed_point_position) >> 1; \ - x = MUL_SAT_OP_EXPAND(x, sub_sat(const_three, MUL_SAT_OP_EXPAND(MUL_SAT_OP_EXPAND(x, x, stype, size, fixed_point_position), temp, stype, size, fixed_point_position)), stype, size, fixed_point_position) >> 1; \ - x = MUL_SAT_OP_EXPAND(x, sub_sat(const_three, MUL_SAT_OP_EXPAND(MUL_SAT_OP_EXPAND(x, x, stype, size, fixed_point_position), temp, stype, size, fixed_point_position)), stype, size, fixed_point_position) >> 1; \ - if(sizeof((stype)(1)) > 1) /* Perform more iterations if datatype is QS16 */ \ - { \ - x = MUL_SAT_OP_EXPAND(x, sub_sat(const_three, MUL_SAT_OP_EXPAND(MUL_SAT_OP_EXPAND(x, x, stype, size, fixed_point_position), temp, stype, size, fixed_point_position)), stype, size, fixed_point_position) >> 1; \ - x = MUL_SAT_OP_EXPAND(x, sub_sat(const_three, MUL_SAT_OP_EXPAND(MUL_SAT_OP_EXPAND(x, x, stype, size, fixed_point_position), temp, stype, size, fixed_point_position)), stype, size, fixed_point_position) >> 1; \ - } \ - type shift_value2 = select(shift_value >> 1, (-shift_value) >> 1, shift_value < (type)0); \ - return select((type)(x >> shift_value2), select((type)stype##_MAX, (type)(x << shift_value2), (type)(clz(x) > shift_value2)), (type)(shift_value < (type)0)); /* Saturate result if needed */ \ - } - -INVSQRTQ_IMPL(qs8, qs8x1, 1) -INVSQRTQ_IMPL(qs16, qs16x1, 1) -INVSQRTQ_IMPL(qs8, qs8x16, 16) -INVSQRTQ_IMPL(qs16, qs16x8, 8) - -#define INVSQRT_OP_EXPAND_STR(a, type, size, position) invsqrt_sat_##type##x##size((a), (position)) -#define INVSQRT_OP_EXPAND(a, type, size, position) INVSQRT_OP_EXPAND_STR(a, type, size, position) - -/** Saturate hyperbolic tangent of a fixed point vector - * - * tanh(x) = (e^2x - 1)/(e^2x + 1) - * - * @param[in] stype the actual scalar data type. - * @param[in] type the actual data type. - * @param[in] size the number of the calculated elements. - * - * @return The result of the fixed point hyperbolic tangent. The result is saturated in case of overflow - */ -#define TANHQ_IMPL(stype, type, size) \ - inline type tanh_sat_##type(type VopA, int fixed_point_position) \ - { \ - type const_one = (type)(1 << (fixed_point_position)); \ - type const_two = (type)(2 << (fixed_point_position)); \ - type exp2x = EXP_OP_EXPAND(MUL_SAT_OP_EXPAND(const_two, VopA, stype, size, fixed_point_position), stype, size, fixed_point_position); \ - type num = SUB_SAT_OP_EXPAND(exp2x, const_one, stype, size); \ - type den = ADD_SAT_OP_EXPAND(exp2x, const_one, stype, size); \ - return DIV_SAT_OP_VEC_EXPAND(num, den, stype, size, fixed_point_position); \ - } - -TANHQ_IMPL(qs8, qs8x16, 16) -TANHQ_IMPL(qs16, qs16x8, 8) - -#define TANH_OP_EXPAND_STR(a, type, size, position) tanh_sat_##type##x##size((a), (position)) -#define TANH_OP_EXPAND(a, type, size, position) TANH_OP_EXPAND_STR(a, type, size, position) - -#define floatx16 float16 -#define float16_TYPE float16 - -#define CONVERTQ_DOWN_IMPL(in_type, out_type) \ - inline out_type convert_##out_type##_##in_type(in_type a, int fixed_point_position) \ - { \ - return CONVERT(a * (1 << fixed_point_position) + select((in_type)-0.5f, (in_type)0.5f, isgreater(a, (in_type)0)), out_type); \ - } - -CONVERTQ_DOWN_IMPL(float16, qs8x16) -CONVERTQ_DOWN_IMPL(float16, qs16x16) - -#define CONVERTQ_DOWN_SAT_IMPL(in_type, out_type) \ - inline out_type convert_##out_type##_##in_type##_sat(in_type a, int fixed_point_position) \ - { \ - return CONVERT_SAT(a * (1 << fixed_point_position) + select((in_type)-0.5f, (in_type)0.5f, isgreater(a, (in_type)0)), out_type); \ - } - -CONVERTQ_DOWN_SAT_IMPL(float16, qs8x16) -CONVERTQ_DOWN_SAT_IMPL(float16, qs16x16) - -#define CONVERTQ_UP_IMPL(in_type, out_type) \ - inline out_type convert_##out_type##_##in_type(in_type a, int fixed_point_position) \ - { \ - return CONVERT(a, out_type) / (1 << fixed_point_position); \ - } - -CONVERTQ_UP_IMPL(qs8x16, float16) -CONVERTQ_UP_IMPL(qs16x16, float16) - -#define SQCVT_SAT_IMPL(type) \ - inline type sqcvt_##type##_sat(float a, int fixed_point_position) \ - { \ - return CONVERT_SAT((a * (1 << fixed_point_position) + ((a < 0) ? -0.5f : 0.5f)), type); \ - } - -SQCVT_SAT_IMPL(qs8) -SQCVT_SAT_IMPL(qs16) - -#define SQCVT_SAT_OP_EXPAND_STR(a, type, position) sqcvt_##type##_sat((a), (position)) -#define SQCVT_SAT_OP_EXPAND(a, type, position) SQCVT_SAT_OP_EXPAND_STR((a), type, position) - -#endif // ARM_COMPUTE_FIXED_POINT_H diff --git a/src/core/CL/cl_kernels/gemm.cl b/src/core/CL/cl_kernels/gemm.cl index e969e847d7..f75161ca0a 100644 --- a/src/core/CL/cl_kernels/gemm.cl +++ b/src/core/CL/cl_kernels/gemm.cl @@ -23,10 +23,6 @@ */ #include "helpers.h" -#ifdef FIXED_POINT_POSITION -#include "fixed_point.h" -#endif // FIXED_POINT_POSITION - #if defined(TRANSPOSE_W) && defined(MULT_TRANSPOSE1XW_WIDTH) #if ELEMENT_SIZE == 1 @@ -44,7 +40,7 @@ * @note The transposition width must be passed at compile time using -DTRANSPOSE_W (i.e. -DTRANSPOSE_W) * @note The multiplication factor for the transposition width (mult_transpose1xW_width) must be passed at compile time using -DMULT_TRANSPOSE1XW_WIDTH (i.e. -DMULT_TRANSPOSE1XW_WIDTH=2) * - * @param[in] src_ptr Pointer to the source matrix. Supported data types: U8/S8/QS8/QASYMM8/U16/S16/QS16/F16/U32/S32/F32 + * @param[in] src_ptr Pointer to the source matrix. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32 * @param[in] src_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source matrix in Y dimension (in bytes) @@ -93,7 +89,7 @@ __kernel void gemm_transpose1xW(TENSOR3D_DECLARATION(src), * @note The data type must be passed at compile time using -DDATA_TYPE (i.e. -DDATA_TYPE=float) * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (i.e. -DMULT_INTERLEAVE4X4_HEIGHT=2) * - * @param[in] src_ptr Pointer to the source matrix. Supported data types: U8/S8/QS8/QASYMM8/U16/S16/QS16/F16/U32/S32/F32 + * @param[in] src_ptr Pointer to the source matrix. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32 * @param[in] src_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source matrix in Y dimension (in bytes) @@ -1085,248 +1081,6 @@ __kernel void gemm_mm_interleaved_transposed_f16_bifrost(IMAGE_DECLARATION(src0) #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) -#if defined(FIXED_POINT_POSITION) -/** This OpenCL kernel computes the matrix multiplication between matrix A (src0) and matrix B (src1) in 8 bit fixed point precision - * Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_8bit and @ref gemm_transpose1x16 before running the matrix multiplication - * - * @note The number of columns of matrix B and the optional alpha's value need to be passed at compile time using -DCOLS_B and -DALPHA - * @note The multiplication factor for the transposition width (mult_transpose1xW_width) must be passed at compile time using -DMULT_TRANSPOSE1XW_WIDTH (i.e. -DMULT_TRANSPOSE1XW_WIDTH=2) - * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (i.e. -DMULT_INTERLEAVE4X4_HEIGHT=2) - * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (i.e. -DMATRIX_B_DEPTH=16) - * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (i.e. a = [K, M, 16, Batches], b = [N, K, 16]) - * @note:ALPHA must be passed in 8 bit fixed point format - * - * @param[in] src0_ptr Pointer to the source matrix. Supported data types: QS8 - * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) - * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) - * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix - * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr - * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) - * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) - * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix - * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr - * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix - * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) - * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - */ -__kernel void gemm_mm_interleaved_transposed_qs8(IMAGE_DECLARATION(src0), - IMAGE_DECLARATION(src1), - IMAGE_DECLARATION(dst), - uint src0_stride_z, - uint src1_stride_z, - uint dst_stride_z) -{ - int x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH; - int y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT; - int z = get_global_id(2); - - // Offset - const int offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4; - const int offset_row_b = (get_global_id(0) % MULT_TRANSPOSE1XW_WIDTH) * 16; - - // src_addr_a = address of matrix A - // src_addr_b = address of matrix B - int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes; - int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes; - -#if defined(MATRIX_B_DEPTH) - // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 - src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z; -#else // defined(MATRIX_B_DEPTH) - src1_addr_in_bytes += z * src1_stride_z; -#endif // defined(MATRIX_B_DEPTH) - - __global char *src_addr_a = (__global char *)(src0_ptr + src0_addr_in_bytes); - __global char *src_addr_b = (__global char *)(src1_ptr + src1_addr_in_bytes); - - // Compute end row address for matrix B - __global char *src_end_addr_b = src_addr_b + COLS_B; - - src_addr_a += offset_row_a; - src_addr_b += offset_row_b; - - // Reset accumulators - short8 c00 = 0.0f; - short8 c10 = 0.0f; - short8 c20 = 0.0f; - short8 c30 = 0.0f; - short8 c01 = 0.0f; - short8 c11 = 0.0f; - short8 c21 = 0.0f; - short8 c31 = 0.0f; - - // This for loop performs 1 accumulation for each iteration - for(; src_addr_b < src_end_addr_b; src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 16 * MULT_TRANSPOSE1XW_WIDTH) - { - // Load values from matrix A (interleaved) and matrix B (transposed) - char4 a0 = vload4(0, src_addr_a); - char16 b0 = vload16(0, src_addr_b); - - c00 = mlal_sat_qs8x8(c00, (char8)a0.s0, b0.s01234567, FIXED_POINT_POSITION); - c10 = mlal_sat_qs8x8(c10, (char8)a0.s1, b0.s01234567, FIXED_POINT_POSITION); - c20 = mlal_sat_qs8x8(c20, (char8)a0.s2, b0.s01234567, FIXED_POINT_POSITION); - c30 = mlal_sat_qs8x8(c30, (char8)a0.s3, b0.s01234567, FIXED_POINT_POSITION); - - c01 = mlal_sat_qs8x8(c01, (char8)a0.s0, b0.s89ABCDEF, FIXED_POINT_POSITION); - c11 = mlal_sat_qs8x8(c11, (char8)a0.s1, b0.s89ABCDEF, FIXED_POINT_POSITION); - c21 = mlal_sat_qs8x8(c21, (char8)a0.s2, b0.s89ABCDEF, FIXED_POINT_POSITION); - c31 = mlal_sat_qs8x8(c31, (char8)a0.s3, b0.s89ABCDEF, FIXED_POINT_POSITION); - } - - // Compute destination address - Image dst = CONVERT_TO_IMAGE_STRUCT(dst); - - // Multiply by the weight of matrix product - char16 c00_qs8 = convert_char16_sat((short16)(c00, c01)); - char16 c10_qs8 = convert_char16_sat((short16)(c10, c11)); - char16 c20_qs8 = convert_char16_sat((short16)(c20, c21)); - char16 c30_qs8 = convert_char16_sat((short16)(c30, c31)); - -#if defined(ALPHA) - c00_qs8 = mul_sat_qs8x16(c00_qs8, (char16)ALPHA, FIXED_POINT_POSITION); - c10_qs8 = mul_sat_qs8x16(c10_qs8, (char16)ALPHA, FIXED_POINT_POSITION); - c20_qs8 = mul_sat_qs8x16(c20_qs8, (char16)ALPHA, FIXED_POINT_POSITION); - c30_qs8 = mul_sat_qs8x16(c30_qs8, (char16)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - - // Compute dst address - __global uchar *dst_addr = offset(&dst, 0, 0); - - // Add offset for batched GEMM - dst_addr += z * dst_stride_z; - - // Store 16x4 block - vstore16(c00_qs8, 0, (__global char *)(dst_addr + 0 * dst_stride_y)); - vstore16(c10_qs8, 0, (__global char *)(dst_addr + 1 * dst_stride_y)); - vstore16(c20_qs8, 0, (__global char *)(dst_addr + 2 * dst_stride_y)); - vstore16(c30_qs8, 0, (__global char *)(dst_addr + 3 * dst_stride_y)); -} - -/** This OpenCL kernel computes the matrix multiplication between matrix A (src0) and matrix B (src1) in 16 bit fixed point precision - * Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_16bit and @ref gemm_transpose1x8 before running the matrix multiplication - * - * @note The number of columns of matrix B and the optional alpha's value need to be passed at compile time using -DCOLS_B and -DALPHA - * @note The multiplication factor for the transposition width (mult_transpose1xW_width) must be passed at compile time using -DMULT_TRANSPOSE1XW_WIDTH (i.e. -DMULT_TRANSPOSE1XW_WIDTH=2) - * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (i.e. -DMULT_INTERLEAVE4X4_HEIGHT=2) - * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (i.e. -DMATRIX_B_DEPTH=16) - * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (i.e. a = [K, M, 16, Batches], b = [N, K, 16]) - * @note:ALPHA must be passed in 16 bit fixed point format - * - * @param[in] src0_ptr Pointer to the source matrix. Supported data types: QS16 - * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) - * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) - * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix - * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr - * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) - * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) - * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix - * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr - * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix - * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) - * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - */ -__kernel void gemm_mm_interleaved_transposed_qs16(IMAGE_DECLARATION(src0), - IMAGE_DECLARATION(src1), - IMAGE_DECLARATION(dst), - uint src0_stride_z, - uint src1_stride_z, - uint dst_stride_z) -{ - int x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH; - int y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT; - int z = get_global_id(2); - - // Offset - const int offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4; - const int offset_row_b = (get_global_id(0) % MULT_TRANSPOSE1XW_WIDTH) * 8; - - // src_addr_a = address of matrix A - // src_addr_b = address of matrix B - int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes; - int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes; - -#if defined(MATRIX_B_DEPTH) - // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 - src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z; -#else // defined(MATRIX_B_DEPTH) - src1_addr_in_bytes += z * src1_stride_z; -#endif // defined(MATRIX_B_DEPTH) - - __global short *src_addr_a = (__global short *)(src0_ptr + src0_addr_in_bytes); - __global short *src_addr_b = (__global short *)(src1_ptr + src1_addr_in_bytes); - - // Compute end row address for matrix B - __global short *src_end_addr_b = src_addr_b + COLS_B; - - src_addr_a += offset_row_a; - src_addr_b += offset_row_b; - - // Reset accumulators - int8 c00 = 0.0f; - int8 c10 = 0.0f; - int8 c20 = 0.0f; - int8 c30 = 0.0f; - - // This for loop performs 1 accumulation for each iteration - for(; src_addr_b < src_end_addr_b; src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH) - { - /* Load values from matrix A (interleaved) and matrix B (transposed) */ - short4 a0 = vload4(0, src_addr_a); - short8 b0 = vload8(0, src_addr_b); - - c00 = mlal_sat_qs16x8(c00, (short8)a0.s0, b0, FIXED_POINT_POSITION); - c10 = mlal_sat_qs16x8(c10, (short8)a0.s1, b0, FIXED_POINT_POSITION); - c20 = mlal_sat_qs16x8(c20, (short8)a0.s2, b0, FIXED_POINT_POSITION); - c30 = mlal_sat_qs16x8(c30, (short8)a0.s3, b0, FIXED_POINT_POSITION); - } - - // Compute destination address - Image dst = CONVERT_TO_IMAGE_STRUCT(dst); - - // Multiply by the weight of matrix product - short8 c00_qs16 = convert_short8_sat(c00); - short8 c10_qs16 = convert_short8_sat(c10); - short8 c20_qs16 = convert_short8_sat(c20); - short8 c30_qs16 = convert_short8_sat(c30); - -#if defined(ALPHA) - c00_qs16 = mul_sat_qs16x8(c00_qs16, (short8)ALPHA, FIXED_POINT_POSITION); - c10_qs16 = mul_sat_qs16x8(c10_qs16, (short8)ALPHA, FIXED_POINT_POSITION); - c20_qs16 = mul_sat_qs16x8(c20_qs16, (short8)ALPHA, FIXED_POINT_POSITION); - c30_qs16 = mul_sat_qs16x8(c30_qs16, (short8)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - - // Compute dst address - __global uchar *dst_addr = offset(&dst, 0, 0); - - // Add offset for batched GEMM - dst_addr += z * dst_stride_z; - - // Store 8x4 block - vstore8(c00_qs16, 0, (__global short *)(dst_addr + 0 * dst_stride_y)); - vstore8(c10_qs16, 0, (__global short *)(dst_addr + 1 * dst_stride_y)); - vstore8(c20_qs16, 0, (__global short *)(dst_addr + 2 * dst_stride_y)); - vstore8(c30_qs16, 0, (__global short *)(dst_addr + 3 * dst_stride_y)); -} -#endif // defined(FIXED_POINT_POSITION) #endif // defined(COLS_B) && defined(MULT_TRANSPOSE1XW_WIDTH) && defined(MULT_INTERLEAVE4X4_HEIGHT) #if defined(COLS_A) && defined(NUM_ELEMS_PROCESSED_PER_THREAD_X) && (NUM_ELEMS_PROCESSED_PER_THREAD_Y) @@ -2543,365 +2297,6 @@ __kernel void gemm_mm_floating_point_f16_bifrost(IMAGE_DECLARATION(src0), } #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) -#if defined(FIXED_POINT_POSITION) -/** This OpenCL kernel computes the matrix by matrix multiplication between the matrix A (src0) and matrix B (src1) in case both matrices have not beed reshaped - * - * @note This OpenCL kernel works with fixed point data types QS8 - * @note The number of elements processed along the x and y directions must be passed at compile time using -DNUM_ELEMS_PROCESSED_PER_THREAD_X and -DNUM_ELEMS_PROCESSED_PER_THREAD_Y - * @note The number matrix A columns, the number of elements processed per thread along the Y direction and the alpha's value need to be passed at compile time using -DCOLS_A, -DNUM_ELEMS_PROCESSED_PER_THREAD_Y and -DALPHA - * @note The fixed point position need to be passed at compile time using -DFIXED_POINT_POSITION - * @note The optional alpha value must be passed in 8 bit fixed point format using -DALPHA - * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (i.e. -DMATRIX_B_DEPTH=16) - * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (i.e. a = [K, M, 16, Batches], b = [N, K, 16]) - * - * @param[in] src0_ptr Pointer to the source matrix. Supported data types: QS8/QS16 - * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) - * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) - * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix - * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr - * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) - * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) - * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix - * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr - * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) - * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) - * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix - * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) - * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - */ -__kernel void gemm_mm_qs8(IMAGE_DECLARATION(src0), - IMAGE_DECLARATION(src1), - IMAGE_DECLARATION(dst), - uint src0_stride_z, - uint src1_stride_z, - uint dst_stride_z) -{ - int idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X; - - // Compute starting address for matrix A and Matrix B - int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); - - // Update address for the matrix A - src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y; - - // Update address for the matrix B - src_addr.s1 += idx * sizeof(char); - - // Add offset for batched GEMM - src_addr.s0 += get_global_id(2) * src0_stride_z; - -#if defined(MATRIX_B_DEPTH) - // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 - src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z; -#else // defined(MATRIX_B_DEPTH) - src_addr.s1 += get_global_id(2) * src1_stride_z; -#endif // defined(MATRIX_B_DEPTH) - - int end_row_vec_a = src_addr.s0 + (COLS_A * sizeof(char)); - - short8 acc00 = 0; - short8 acc01 = 0; -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - short8 acc10 = 0; - short8 acc11 = 0; -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - short8 acc20 = 0; - short8 acc21 = 0; -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - short8 acc30 = 0; - short8 acc31 = 0; -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - - // This for loop performs 4 accumulations per iteration - for(; src_addr.s0 <= (end_row_vec_a - 2); src_addr += (int2)(2, 2 * src1_stride_y)) - { - char2 a0 = vload2(0, (__global char *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - char2 a1 = vload2(0, (__global char *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - char2 a2 = vload2(0, (__global char *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - char2 a3 = vload2(0, (__global char *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - char16 b0 = vload16(0, (__global char *)(src1_ptr + src_addr.s1 + 0 * src1_stride_y)); - char16 b1 = vload16(0, (__global char *)(src1_ptr + src_addr.s1 + 1 * src1_stride_y)); - - acc00 = mlal_sat_qs8x8(acc00, (char8)a0.s0, b0.s01234567, FIXED_POINT_POSITION); - acc00 = mlal_sat_qs8x8(acc00, (char8)a0.s1, b1.s01234567, FIXED_POINT_POSITION); - acc01 = mlal_sat_qs8x8(acc01, (char8)a0.s0, b0.s89ABCDEF, FIXED_POINT_POSITION); - acc01 = mlal_sat_qs8x8(acc01, (char8)a0.s1, b1.s89ABCDEF, FIXED_POINT_POSITION); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - acc10 = mlal_sat_qs8x8(acc10, (char8)a1.s0, b0.s01234567, FIXED_POINT_POSITION); - acc10 = mlal_sat_qs8x8(acc10, (char8)a1.s1, b1.s01234567, FIXED_POINT_POSITION); - acc11 = mlal_sat_qs8x8(acc11, (char8)a1.s0, b0.s89ABCDEF, FIXED_POINT_POSITION); - acc11 = mlal_sat_qs8x8(acc11, (char8)a1.s1, b1.s89ABCDEF, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - acc20 = mlal_sat_qs8x8(acc20, (char8)a2.s0, b0.s01234567, FIXED_POINT_POSITION); - acc20 = mlal_sat_qs8x8(acc20, (char8)a2.s1, b1.s01234567, FIXED_POINT_POSITION); - acc21 = mlal_sat_qs8x8(acc21, (char8)a2.s0, b0.s89ABCDEF, FIXED_POINT_POSITION); - acc21 = mlal_sat_qs8x8(acc21, (char8)a2.s1, b1.s89ABCDEF, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - acc30 = mlal_sat_qs8x8(acc30, (char8)a3.s0, b0.s01234567, FIXED_POINT_POSITION); - acc30 = mlal_sat_qs8x8(acc30, (char8)a3.s1, b1.s01234567, FIXED_POINT_POSITION); - acc31 = mlal_sat_qs8x8(acc31, (char8)a3.s0, b0.s89ABCDEF, FIXED_POINT_POSITION); - acc31 = mlal_sat_qs8x8(acc31, (char8)a3.s1, b1.s89ABCDEF, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - } - - // Left-over accumulations - for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(1, src1_stride_y)) - { - char a0 = *((__global char *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - char a1 = *((__global char *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - char a2 = *((__global char *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - char a3 = *((__global char *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - char16 b0 = vload16(0, (__global char *)(src1_ptr + src_addr.s1)); - - acc00 = mlal_sat_qs8x8(acc00, (char8)a0, b0.s01234567, FIXED_POINT_POSITION); - acc01 = mlal_sat_qs8x8(acc01, (char8)a0, b0.s89ABCDEF, FIXED_POINT_POSITION); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - acc10 = mlal_sat_qs8x8(acc10, (char8)a1, b0.s01234567, FIXED_POINT_POSITION); - acc11 = mlal_sat_qs8x8(acc11, (char8)a1, b0.s89ABCDEF, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - acc20 = mlal_sat_qs8x8(acc20, (char8)a2, b0.s01234567, FIXED_POINT_POSITION); - acc21 = mlal_sat_qs8x8(acc21, (char8)a2, b0.s89ABCDEF, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - acc30 = mlal_sat_qs8x8(acc30, (char8)a3, b0.s01234567, FIXED_POINT_POSITION); - acc31 = mlal_sat_qs8x8(acc31, (char8)a3, b0.s89ABCDEF, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - } - - // Compute destination address - Image dst = CONVERT_TO_IMAGE_STRUCT(dst); - - // Compute dst address - __global uchar *dst_addr = offset(&dst, 0, 0); - - // Add offset for batched GEMM - dst_addr += get_global_id(2) * dst_stride_z; - - // Multiply by the weight of matrix product and store the result - char16 acc_qs8; - acc_qs8 = convert_char16_sat((short16)(acc00, acc01)); -#if defined(ALPHA) - acc_qs8 = mul_sat_qs8x16(acc_qs8, (char16)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - vstore16(acc_qs8, 0, (__global char *)(dst_addr + 0 * dst_stride_y)); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - acc_qs8 = convert_char16_sat((short16)(acc10, acc11)); -#if defined(ALPHA) - acc_qs8 = mul_sat_qs8x16(acc_qs8, (char16)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - vstore16(acc_qs8, 0, (__global char *)(dst_addr + 1 * dst_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - acc_qs8 = convert_char16_sat((short16)(acc20, acc21)); -#if defined(ALPHA) - acc_qs8 = mul_sat_qs8x16(acc_qs8, (char16)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - vstore16(acc_qs8, 0, (__global char *)(dst_addr + 2 * dst_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - acc_qs8 = convert_char16_sat((short16)(acc30, acc31)); -#if defined(ALPHA) - acc_qs8 = mul_sat_qs8x16(acc_qs8, (char16)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - vstore16(acc_qs8, 0, (__global char *)(dst_addr + 3 * dst_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 -} - -/** This OpenCL kernel computes the matrix by matrix multiplication between the matrix A (src0) and matrix B (src1) in case both matrices have not beed reshaped - * - * @note This OpenCL kernel works with fixed point data types QS16 - * @note The number of elements processed along the x and y directions must be passed at compile time using -DNUM_ELEMS_PROCESSED_PER_THREAD_X and -DNUM_ELEMS_PROCESSED_PER_THREAD_Y - * @note The number of matrix A columns, the number of elements processed per thread along the Y direction and the alpha's value need to be passed at compile time using -DCOLS_A, -DNUM_ELEMS_PROCESSED_PER_THREAD_Y and -DALPHA - * @note The fixed point position need to be passed at compile time using -DFIXED_POINT_POSITION - * @note The optional alpha value must be passed in 16 bit fixed point format using -DALPHA - * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (i.e. -DMATRIX_B_DEPTH=16) - * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (i.e. a = [K, M, 16, Batches], b = [N, K, 16]) - * - * @param[in] src0_ptr Pointer to the source matrix. Supported data types: QS8/QS16 - * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) - * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) - * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix - * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr - * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) - * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) - * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix - * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr - * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) - * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) - * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix - * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) - * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - */ -__kernel void gemm_mm_qs16(IMAGE_DECLARATION(src0), - IMAGE_DECLARATION(src1), - IMAGE_DECLARATION(dst), - uint src0_stride_z, - uint src1_stride_z, - uint dst_stride_z) -{ - int idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X; - - // Compute starting address for matrix A and Matrix B - int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); - - // Update address for the matrix A - src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y; - - // Update address for the matrix B - src_addr.s1 += idx * sizeof(short); - - // Add offset for batched GEMM - src_addr.s0 += get_global_id(2) * src0_stride_z; - -#if defined(MATRIX_B_DEPTH) - // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 - src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z; -#else // defined(MATRIX_B_DEPTH) - src_addr.s1 += get_global_id(2) * src1_stride_z; -#endif // defined(MATRIX_B_DEPTH) - - int end_row_vec_a = src_addr.s0 + (COLS_A * sizeof(short)); - - int8 acc0 = 0; -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - int8 acc1 = 0; -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - int8 acc2 = 0; -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - int8 acc3 = 0; -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - - // This for loop performs 4 accumulations per iteration - for(; src_addr.s0 <= (end_row_vec_a - 2 * (int)sizeof(short)); src_addr += (int2)(2 * sizeof(short), 2 * src1_stride_y)) - { - short2 a0 = vload2(0, (__global short *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - short2 a1 = vload2(0, (__global short *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - short2 a2 = vload2(0, (__global short *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - short2 a3 = vload2(0, (__global short *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - short8 b0 = vload8(0, (__global short *)(src1_ptr + src_addr.s1 + 0 * src1_stride_y)); - short8 b1 = vload8(0, (__global short *)(src1_ptr + src_addr.s1 + 1 * src1_stride_y)); - - acc0 = mlal_sat_qs16x8(acc0, (short8)a0.s0, b0, FIXED_POINT_POSITION); - acc0 = mlal_sat_qs16x8(acc0, (short8)a0.s1, b1, FIXED_POINT_POSITION); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - acc1 = mlal_sat_qs16x8(acc1, (short8)a1.s0, b0, FIXED_POINT_POSITION); - acc1 = mlal_sat_qs16x8(acc1, (short8)a1.s1, b1, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - acc2 = mlal_sat_qs16x8(acc2, (short8)a2.s0, b0, FIXED_POINT_POSITION); - acc2 = mlal_sat_qs16x8(acc2, (short8)a2.s1, b1, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - acc3 = mlal_sat_qs16x8(acc3, (short8)a3.s0, b0, FIXED_POINT_POSITION); - acc3 = mlal_sat_qs16x8(acc3, (short8)a3.s1, b1, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - } - - // Left-over accumulations - for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(sizeof(short), src1_stride_y)) - { - short a0 = *((__global short *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - short a1 = *((__global short *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - short a2 = *((__global short *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - short a3 = *((__global short *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - short8 b0 = vload8(0, (__global short *)(src1_ptr + src_addr.s1)); - - acc0 = mlal_sat_qs16x8(acc0, (short8)a0, b0, FIXED_POINT_POSITION); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - acc1 = mlal_sat_qs16x8(acc1, (short8)a1, b0, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - acc2 = mlal_sat_qs16x8(acc2, (short8)a2, b0, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - acc3 = mlal_sat_qs16x8(acc3, (short8)a3, b0, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - } - - // Compute destination address - Image dst = CONVERT_TO_IMAGE_STRUCT(dst); - - // Compute dst address - __global uchar *dst_addr = offset(&dst, 0, 0); - - // Add offset for batched GEMM - dst_addr += get_global_id(2) * dst_stride_z; - - // Multiply by the weight of matrix product and store the result - short8 acc_qs16; - acc_qs16 = convert_short8_sat(acc0); -#if defined(ALPHA) - acc_qs16 = mul_sat_qs16x8(acc_qs16, (short8)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - vstore8(acc_qs16, 0, (__global short *)(dst_addr + 0 * dst_stride_y)); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - acc_qs16 = convert_short8_sat(acc1); -#if defined(ALPHA) - acc_qs16 = mul_sat_qs16x8(acc_qs16, (short8)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - vstore8(acc_qs16, 0, (__global short *)(dst_addr + 1 * dst_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - acc_qs16 = convert_short8_sat(acc2); -#if defined(ALPHA) - acc_qs16 = mul_sat_qs16x8(acc_qs16, (short8)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - vstore8(acc_qs16, 0, (__global short *)(dst_addr + 2 * dst_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - acc_qs16 = convert_short8_sat(acc3); -#if defined(ALPHA) - acc_qs16 = mul_sat_qs16x8(acc_qs16, (short8)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - vstore8(acc_qs16, 0, (__global short *)(dst_addr + 3 * dst_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 -} -#endif // defined(FIXED_POINT_POSITION) #endif // defined(COLS_A) && defined(NUM_ELEMS_PROCESSED_PER_THREAD_X) && (NUM_ELEMS_PROCESSED_PER_THREAD_Y) #if defined(BETA) @@ -2988,94 +2383,6 @@ __kernel void gemm_ma_f16(TENSOR3D_DECLARATION(src), vstore8(out, 0, (__global half *)dst.ptr); } #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) - -#if defined(FIXED_POINT_POSITION) -/** This OpenCL kernel performs the in-place matrix addition between 2 matrices in 8 bit fixed point taking into account that the second matrix might be weighted by a scalar value beta: - * - * @note The beta's value and the fixed point position need to be passed at compile time using -DBETA and -DFIXED_POINT_POSITION - * - * @note: BETA must be passed in 8 bit fixed point format - * - * @param[in] src_ptr Pointer to the source matrix. Supported data types: QS8 - * @param[in] src_stride_x Stride of the source matrix in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source matrix in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] src_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source matrix - * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) - * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) - * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix - */ -__kernel void gemm_ma_qs8(TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst)) -{ - // Compute source and destination addresses - Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); - Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); - - // Load values from A x B - char16 alpha_ab = vload16(0, (__global char *)dst.ptr); - - // Load values from Matrix C - char16 c = vload16(0, (__global char *)src.ptr); - - // Computes alpha * axb + beta * c - char16 out = mla_sat_qs8x16(alpha_ab, (char16)BETA, c, FIXED_POINT_POSITION); - - // Store final result in axb matrix - vstore16(out, 0, (__global char *)dst.ptr); -} - -/** This OpenCL kernel performs the in-place matrix addition between 2 matrices in 16 bit fixed point taking into account that the second matrix might be weighted by a scalar value beta: - * - * @note The beta's value and the fixed point position need to be passed at compile time using -DBETA and -DFIXED_POINT_POSITION - * - * @note: BETA must be passed in 16 bit fixed point format - * - * @param[in] src_ptr Pointer to the source matrix. Supported data types: QS16 - * @param[in] src_stride_x Stride of the source matrix in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source matrix in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] src_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source matrix - * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) - * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) - * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix - */ -__kernel void gemm_ma_qs16(TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst)) -{ - // Compute source and destination addresses - Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); - Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); - - // Load values from A x B - short8 alpha_ab = vload8(0, (__global short *)dst.ptr); - - // Load values from Matrix C - short8 c = vload8(0, (__global short *)src.ptr); - - // Computes alpha * axb + beta * c - short8 out = mla_sat_qs16x8(alpha_ab, (short8)BETA, c, FIXED_POINT_POSITION); - - // Store final result in axb matrix - vstore8(out, 0, (__global short *)dst.ptr); -} -#endif // defined(FIXED_POINT_POSITION) #endif // defined(BETA) #if defined(WIDTH_VECTOR_A) @@ -3151,7 +2458,7 @@ __kernel void gemm_lc_vm_f32(IMAGE_DECLARATION(src0), * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=short. * @note The vector size must be passed at compile time using -DVECTOR_SIZE e.g. -DVECTOR_SIZE=16. * - * @param[in, out] accum_ptr Pointer to the accumulate tensor. Supported data type: U8/S8/QS8/U16/S16/F16/U32/S32/F32 + * @param[in, out] accum_ptr Pointer to the accumulate tensor. Supported data type: U8/S8/U16/S16/F16/U32/S32/F32 * @param[in] accum_stride_x Stride of the accmulate tensor in X dimension (in bytes) * @param[in] accum_step_x accum_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] accum_stride_y Stride of the accumlulate tensor in Y dimension (in bytes) @@ -3175,11 +2482,7 @@ __kernel void gemm_accumulate_biases( accum_value = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)accum.ptr); VEC_DATA_TYPE(DATA_TYPE, VECTOR_SIZE) biases_value = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)biases.ptr); -#ifdef FIXED_POINT_POSITION - accum_value = ADD_SAT_OP_EXPAND(biases_value, accum_value, DATA_TYPE, VECTOR_SIZE); -#else // FIXED_POINT_POSITION - accum_value = biases_value + accum_value; -#endif // FIXED_POINT_POSITION + accum_value = biases_value + accum_value; // Store result in the accumulate buffer VSTORE(VECTOR_SIZE) (accum_value, 0, (__global DATA_TYPE *)accum.ptr); diff --git a/src/core/CL/cl_kernels/im2col.cl b/src/core/CL/cl_kernels/im2col.cl index 6f25ad4b7a..d034b30b68 100644 --- a/src/core/CL/cl_kernels/im2col.cl +++ b/src/core/CL/cl_kernels/im2col.cl @@ -23,12 +23,7 @@ */ #include "helpers.h" -#if defined(FIXED_POINT_POSITION) -#include "fixed_point.h" -#endif // FIXED_POINT_POSITION - #if defined(DATA_TYPE) && defined(ELEMENT_SIZE) -#if !defined(FIXED_POINT_POSITION) #if ELEMENT_SIZE == 1 #define COND_DATA_TYPE char @@ -50,7 +45,7 @@ * @note The stride along the Y direction must be passed at compile time using -DSTRIDE_Y: e.g. -DSTRIDE_Y=1 * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row. * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/F16/F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/F16/F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) @@ -139,7 +134,7 @@ __kernel void im2col1x1_stridex1_dchw( * @note The dilation_x and dilation_y must be passed at compile time using -DDILATION_X and -DDILATION_Y: e.g. -DDILATION_X=1, -DDILATION_Y=1 * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row. * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/F16/F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/F16/F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) @@ -232,7 +227,7 @@ __kernel void im2col_generic_nhwc( * @note The stride along the X and Y directions must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y: e.g. -DSTRIDE_X=1 and -DSTRIDE_Y=1 * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row. * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/F16/F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/F16/F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) @@ -338,7 +333,7 @@ __kernel void im2col3x3_nhwc( * @note The stride along the X and Y directions must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y: e.g. -DSTRIDE_X=1 and -DSTRIDE_Y=1 * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row. * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/F16/F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/F16/F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) @@ -425,7 +420,7 @@ __kernel void im2col3x3_dchw( * @note The stride along the X and Y directions must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y: e.g. -DSTRIDE_X=1 and -DSTRIDE_Y=1 * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row. * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/F16/F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/F16/F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) @@ -598,7 +593,7 @@ __kernel void im2col5x5_dchw( * @note The stride along the X and Y directions must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y: e.g. -DSTRIDE_X=1 and -DSTRIDE_Y=1 * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row. * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/F16/F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/F16/F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) @@ -788,7 +783,6 @@ __kernel void im2col11x11_padx0_pady0_dchw( #endif // HAS_BIAS } #endif // defined(CONVOLVED_WIDTH) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(KERNEL_DEPTH) -#endif // !defined(FIXED_POINT_POSITION) #if defined(CONVOLVED_WIDTH) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(KERNEL_DEPTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(VECTOR_SIZE) && defined(WIDTH_MOD_VECTOR_SIZE) /** This kernel reshapes the input tensor to a tensor used to perform convolution using GEMM when @@ -799,7 +793,7 @@ __kernel void im2col11x11_padx0_pady0_dchw( * @note The width modulo vector size must be passed at compile time using -DWIDTH_MOD_VECTOR_SIZE e.g. -DWIDTH_MOD_VECTOR_SIZE=3. * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row. * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QS16/F16/F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) @@ -863,11 +857,7 @@ __kernel void im2col_generic_padx0_pady0_dchw( #ifdef HAS_BIAS if(ch == (KERNEL_DEPTH - 1)) { -#ifdef FIXED_POINT_POSITION - *output_ptr = (DATA_TYPE)(1 << FIXED_POINT_POSITION); -#else // FIXED_POINT_POSITION *output_ptr = 1.0f; -#endif // FIXED_POINT_POSITION } #endif // HAS_BIAS } @@ -886,7 +876,7 @@ __kernel void im2col_generic_padx0_pady0_dchw( * @note The dilation_x and dilation_y must be passed at compile time using -DDILATION_X and -DDILATION_Y: e.g. -DDILATION_X=1, -DDILATION_Y=1 * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row. * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/F16/F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/F16/F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) @@ -950,11 +940,7 @@ __kernel void im2col_generic_dchw( #ifdef HAS_BIAS if(ch == (KERNEL_DEPTH - 1)) { -#ifdef FIXED_POINT_POSITION - *output_ptr = (DATA_TYPE)(1 << FIXED_POINT_POSITION); -#else // FIXED_POINT_POSITION *output_ptr = 1.0f; -#endif // FIXED_POINT_POSITION } #endif // HAS_BIAS } @@ -966,7 +952,7 @@ __kernel void im2col_generic_dchw( * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float * @note In case biases will be added in late stage, -DHAS_BIAS has to be passed to append the final matrix with 1 in each row. * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/F16/F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/F16/F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) @@ -999,11 +985,7 @@ __kernel void im2col_reduced_dchw( if(get_global_id(0) == (get_global_size(0) - 1) && get_global_id(1) == (get_global_size(1) - 1) && get_global_id(2) == (get_global_size(2) - 1)) { tmp_out_ptr += dst_stride_x; -#ifdef FIXED_POINT_POSITION - *((__global DATA_TYPE *)tmp_out_ptr) = (DATA_TYPE)(1 << FIXED_POINT_POSITION); -#else // FIXED_POINT_POSITION *((__global DATA_TYPE *)tmp_out_ptr) = (DATA_TYPE)1.0f; -#endif // FIXED_POINT_POSITION } #endif // HAS_BIAS } diff --git a/src/core/CL/cl_kernels/l2_normalize.cl b/src/core/CL/cl_kernels/l2_normalize.cl index 8d47631019..f58e98bace 100644 --- a/src/core/CL/cl_kernels/l2_normalize.cl +++ b/src/core/CL/cl_kernels/l2_normalize.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016, 2017 ARM Limited. + * Copyright (c) 2016-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -28,11 +28,11 @@ * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float * @note The data size must be passed at compile time using -DDATA_SIZE e.g. -DDATA_SIZE=32 * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/F16/F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[in] sum_ptr Pointer to the source tensor. Supported data types: QS8/F16/F32 + * @param[in] sum_ptr Pointer to the source tensor. Supported data types: F16/F32 * @param[in] sum_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] sum_step_x sum_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] sum_offset_first_element_in_bytes The offset of the first element in the source tensor diff --git a/src/core/CL/cl_kernels/normalization_layer.cl b/src/core/CL/cl_kernels/normalization_layer.cl index bc00252fbd..dbdad27865 100644 --- a/src/core/CL/cl_kernels/normalization_layer.cl +++ b/src/core/CL/cl_kernels/normalization_layer.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -23,22 +23,6 @@ */ #include "helpers.h" -#if defined(FIXED_POINT_POSITION) - -#include "fixed_point.h" -#define MUL_OP(x, y) MUL_SAT_OP_EXPAND((x), (y), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION) -#define ADD_OP(x, y) ADD_SAT_OP_EXPAND((x), (y), DATA_TYPE, VEC_SIZE) -#define DIV_OP(x, y) DIV_SAT_OP_VEC_EXPAND((x), (y), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION) -#define EXP_OP(x) EXP_OP_EXPAND((x), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION) -#define LOG_OP(x) LOG_OP_EXPAND((x), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION) -#define POW_OP(x, y) EXP_OP(MUL_OP(LOG_OP((x)), (y))) -#define SQCVT_SAT(a) SQCVT_SAT_OP_EXPAND((a), DATA_TYPE, FIXED_POINT_POSITION) - -#define LOAD_OP(offset, ptr) vload16(offset, ptr) -#define STORE_OP(data, offset, ptr) vstore16(data, offset, ptr) - -#else // FIXED_POINT_POSITION - #define MUL_OP(x, y) ((x) * (y)) #define ADD_OP(x, y) ((x) + (y)) #define DIV_OP(x, y) ((x) / (y)) @@ -48,18 +32,15 @@ #define LOAD_OP(offset, ptr) vload4(offset, ptr) #define STORE_OP(data, offset, ptr) vstore4(data, offset, ptr) -#endif // FIXED_POINT_POSITION - /** Apply cross-map normalization. * * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size, e.g. -DVEC_SIZE=16 * @note The radius should be given as a preprocessor argument using -DRADIUS=size. e.g. -DRADIUS=5 * @note The number of slices should be given as a preprocessor argument using -DNUM_SLICES=size. e.g. -DNUM_SLICES=192 - * @note In case of fixed-point operation -DFIXED_POINT_POSITION=fixed_point_position must be provided: e.g. -DFIXED_POINT_POSITION=3 * @note Scaling coefficient (= alpha/norm_size), beta and kappa need to be passed at compile time using -DCOEFF, -DALPHA and -DKAPPA * - * @param[in] input_ptr Pointer to the first source tensor. Supported data types: QS8/QS16/F16/F32 + * @param[in] input_ptr Pointer to the first source tensor. Supported data types: F16/F32 * @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes) * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes) @@ -116,10 +97,9 @@ __kernel void normalization_layer_cross_map(TENSOR3D_DECLARATION(input), * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size, e.g. -DVEC_SIZE=16 * @note The radius should be given as a preprocessor argument using -DRADIUS=size. e.g. -DRADIUS=5 - * @note In case of fixed-point operation -DFIXED_POINT_POSITION=fixed_point_position must be provided: e.g. -DFIXED_POINT_POSITION=3 * @note Scaling coefficient (= alpha/norm_size), beta and kappa need to be passed at compile time using -DCOEFF, -DALPHA and -DKAPPA * - * @param[in] input_ptr Pointer to the first source tensor. Supported data types: QS8/F16/F32 + * @param[in] input_ptr Pointer to the first source tensor. Supported data types: F16/F32 * @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes) * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes) diff --git a/src/core/CL/cl_kernels/permute.cl b/src/core/CL/cl_kernels/permute.cl index 6f978c9b70..03fc15e4e8 100644 --- a/src/core/CL/cl_kernels/permute.cl +++ b/src/core/CL/cl_kernels/permute.cl @@ -29,7 +29,7 @@ * @attention Data type can be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float * @attention Input tensor depth should be given as a preprocessor argument using -DDEPTH_IN=size. e.g. -DDEPTH_IN=16 * - * @param[in] input_ptr Pointer to the source image. Supported data types: U8/S8/QS8/QASYMM8/U16/S16/QS16/F16/U32/S32/F32 + * @param[in] input_ptr Pointer to the source image. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32 * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) @@ -63,7 +63,7 @@ __kernel void permute_201( * @attention Data type can be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float * @attention Input tensor depth should be given as a preprocessor argument using -DDEPTH_IN=size. e.g. -DDEPTH_IN=16 * - * @param[in] input_ptr Pointer to the source image. Supported data types: U8/S8/QS8/QASYMM8/U16/S16/QS16/F16/U32/S32/F32 + * @param[in] input_ptr Pointer to the source image. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32 * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) @@ -97,7 +97,7 @@ __kernel void permute_120( * @attention Data type can be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float * @attention Input tensor depth should be given as a preprocessor argument using -DDEPTH_IN=size. e.g. -DDEPTH_IN=16 * - * @param[in] input_ptr Pointer to the source image. Supported data types: U8/S8/QS8/QASYMM8/U16/S16/QS16/F16/U32/S32/F32 + * @param[in] input_ptr Pointer to the source image. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32 * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) diff --git a/src/core/CL/cl_kernels/pixelwise_mul_int.cl b/src/core/CL/cl_kernels/pixelwise_mul_int.cl index b5734a39ed..c99a08a583 100644 --- a/src/core/CL/cl_kernels/pixelwise_mul_int.cl +++ b/src/core/CL/cl_kernels/pixelwise_mul_int.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016, 2017 ARM Limited. + * Copyright (c) 2016-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -23,18 +23,6 @@ */ #include "helpers.h" -#if defined(FIXED_POINT_POSITION) - -#include "fixed_point.h" - -#if defined(SATURATE) -#define MUL_OP(x, y, scale, type, size) MUL_SAT_OP_EXPAND((x), (y), type, size, FIXED_POINT_POSITION) -#else // SATURATE -#define MUL_OP(x, y, scale, type, size) MUL_OP_EXPAND((x), (y), type, size, FIXED_POINT_POSITION) -#endif // SATURATE - -#else // FIXED_POINT_POSITION - #if defined(SATURATE) #define CONVERT_OP_INT_STR(x, type, size) (convert_##type##size##_sat(x)) #else // SATURATE @@ -44,17 +32,14 @@ #define MUL_OP(x, y, scale, type, size) CONVERT_OP_INT((x) * (y) >> scale, type, size) -#endif // FIXED_POINT_POSITION - /** Performs a pixelwise multiplication with integer scale of integer inputs. * * @attention The inputs and output data types need to be passed at compile time using -DDATA_TYPE_IN1, -DDATA_TYPE_IN2 and -DDATA_TYPE_OUT: * e.g. -DDATA_TYPE_IN1=uchar -DDATA_TYPE_IN2=ushort -DDATA_TYPE_OUT=short * @attention The data_type of the intermediate result of the multiplication should passed as well using -DDATA_TYPE_RES. * e.g. If one of inputs is S16 -DDATA_TYPE_RES=int should be passed else -DDATA_TYPE_RES=short. - * @note In case of fixed-point operation -DFIXED_POINT_POSITION=fixed_point_position must be provided: e.g. -DFIXED_POINT_POSITION=3 * - * @param[in] in1_ptr Pointer to the source image. Supported data types: U8/QS8/QS16/S16 + * @param[in] in1_ptr Pointer to the source image. Supported data types: U8/S16 * @param[in] in1_stride_x Stride of the source image in X dimension (in bytes) * @param[in] in1_step_x in1_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] in1_stride_y Stride of the source image in Y dimension (in bytes) @@ -78,7 +63,7 @@ * @param[in] out_stride_z Stride of the destination image in Y dimension (in bytes) * @param[in] out_step_z out_stride_z * number of elements along Y processed per workitem(in bytes) * @param[in] out_offset_first_element_in_bytes The offset of the first element in the destination image - * @param[in] scale Integer scaling factor. Supported data types: S32 (ignored for QS8 and QS16 as the assumption is scale = 1). + * @param[in] scale Integer scaling factor. Supported data types: S32. */ __kernel void pixelwise_mul_int( TENSOR3D_DECLARATION(in1), diff --git a/src/core/CL/cl_kernels/pooling_layer.cl b/src/core/CL/cl_kernels/pooling_layer.cl index 2c7ddfdf23..c38a78ce3e 100644 --- a/src/core/CL/cl_kernels/pooling_layer.cl +++ b/src/core/CL/cl_kernels/pooling_layer.cl @@ -23,28 +23,6 @@ */ #include "helpers.h" -#ifdef FIXED_POINT_POSITION - -#include "fixed_point.h" - -#if defined(POOL_AVG) -#define POOL_OP(x, y) add_sat(x, y) -#else /* POOL_AVG */ -#define POOL_OP(x, y) (max((x), (y))) -#endif /* POOL_AVG */ - -#define DIV_OP1(x, y) DIV_SAT_OP_EXPAND((x), (y), DATA_TYPE, FIXED_POINT_POSITION) -#define DIV_OP(x, y) DIV_OP1(x, y << FIXED_POINT_POSITION) -#define SQRT_OP(x) DIV_OP1((1 << FIXED_POINT_POSITION), (INVSQRT_OP_EXPAND((x), DATA_TYPE, 1, FIXED_POINT_POSITION))) - -#if defined(POOL_L2) -#define POW2_OP(x, vec_size) MUL_SAT_OP_EXPAND((x), (x), DATA_TYPE, vec_size, FIXED_POINT_POSITION) -#else /* defined(POOL_L2) */ -#define POW2_OP(x, vec_size) (x) -#endif /* defined(POOL_L2) */ - -#else /* FIXED_POINT_POSITION */ - #if defined(POOL_AVG) || defined(POOL_L2) #define POOL_OP(x, y) ((x) + (y)) #else /* defined(POOL_AVG) || defined(POOL_L2) */ @@ -60,8 +38,6 @@ #define DIV_OP(x, y) (x * (1.f / y)) #define SQRT_OP(x) sqrt((x)) -#endif /* FIXED_POINT_POSITION */ - #define DIV_OP_NHWC(x, y) (x * (VEC_DATA_TYPE(DATA_TYPE, 8))(1.f / y)) #if STRIDE_X == 1 @@ -201,14 +177,14 @@ DATA_TYPE calculate_avg_scale(const int pool_size_x, const int pool_size_y, cons /** Performs a pooling function of pool size equal to 2. * - * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are QS8/QS16/F16/F32; + * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32; * @note In case of average pooling the following information must be passed at compile time: * -DPOOL_AVG or -DPOOL_L2 must be provided otherwise max pooling will be performed. * -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad) * -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions * -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension * - * @param[in] input_ptr Pointer to the source image. Supported data types: QS8/QS16/F16/F32 + * @param[in] input_ptr Pointer to the source image. Supported data types: F16/F32 * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) @@ -265,14 +241,14 @@ __kernel void pooling_layer_2( /** Performs a pooling function of pool size equal to 3 * - * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are QS8/QS16/F16/F32; + * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32; * @note In case of average pooling the following information must be passed at compile time: * -DPOOL_AVG or -DPOOL_L2 must be provided otherwise max pooling will be performed. * -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad) * -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions * -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension * - * @param[in] input_ptr Pointer to the source image. Supported data types: QS8/QS16/F16/F32 + * @param[in] input_ptr Pointer to the source image. Supported data types: F16/F32 * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) @@ -331,7 +307,7 @@ __kernel void pooling_layer_3( *(__global DATA_TYPE *)output.ptr = res; } -#if defined(POOLING3x3) && !defined(FIXED_POINT_POSITION) +#if defined(POOLING3x3) #define CONVERT_OP(data_type) convert_##data_type##4 #define CONVERT_VECTOR4(data_type) CONVERT_OP(data_type) @@ -353,7 +329,7 @@ calculate_avg_scale4(const int pool_size, const int upper_bound_w, const int upp /** Performs an optimized pooling function of pool size equal to 3 when the stride_x is less equal than 3 * - * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are QS8/QS16/F16/F32; + * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32; * @note In case of average pooling the following information must be passed at compile time: * -DPOOL_AVG or -DPOOL_L2 must be provided otherwise max pooling will be performed. * -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad) @@ -403,7 +379,7 @@ __kernel void pooling_layer_optimized_3( vstore4(res, 0, (__global DATA_TYPE *)output.ptr); } -#endif // defined(POOLING3x3) && !defined(FIXED_POINT_POSITION) +#endif // defined(POOLING3x3) #if defined(POOL_SIZE_X) && defined(POOL_SIZE_Y) @@ -411,23 +387,17 @@ __kernel void pooling_layer_optimized_3( #if defined(POOL_AVG) || defined(POOL_L2) #define INITIAL_VALUE 0 #else /* defined(POOL_AVG) || defined(POOL_L2) */ -#ifdef FIXED_POINT_POSITION -#define MIN_VAL_EXPAND(type) type##_MIN -#define MIN_VAL(type) MIN_VAL_EXPAND(type) -#define INITIAL_VALUE MIN_VAL(DATA_TYPE) -#else // FIXED_POINT_POSITION #if FP16 #define INITIAL_VALUE -HALF_MAX #else // FP16 #define INITIAL_VALUE -FLT_MAX #endif // FP16 -#endif // FIXED_POINT_POSITION #endif // POOL_AVG /** Performs a pooling function of pool size equal to N (NCHW) * - * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are QS8/QS16/F16/F32; + * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32; * @note -DFP16 must be passed at compile time if half float data type is used * @note Pool sizes must be passed using -DPOOL_SIZE_X and -DPOOL_SIZE_Y e.g. -DPOOL_SIZE_X=13; * @note In case of average pooling the following information must be passed at compile time: @@ -436,7 +406,7 @@ __kernel void pooling_layer_optimized_3( * -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions * -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension * - * @param[in] input_ptr Pointer to the source image. Supported data types: QS8/QS16/F16/F32 + * @param[in] input_ptr Pointer to the source image. Supported data types: F16/F32 * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) diff --git a/src/core/CL/cl_kernels/reshape_layer.cl b/src/core/CL/cl_kernels/reshape_layer.cl index 23eccbf817..11393d246d 100644 --- a/src/core/CL/cl_kernels/reshape_layer.cl +++ b/src/core/CL/cl_kernels/reshape_layer.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -27,7 +27,7 @@ * * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short * - * @param[in] input_ptr Pointer to the first source tensor. Supported data types: U8/S8/QS8/U16/S16/QS16/U32/S32/F16/F32 + * @param[in] input_ptr Pointer to the first source tensor. Supported data types: U8/S8/U16/S16/U32/S32/F16/F32 * @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes) * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes) diff --git a/src/core/CL/cl_kernels/softmax_layer.cl b/src/core/CL/cl_kernels/softmax_layer.cl index aa1fa01c53..e549b44245 100644 --- a/src/core/CL/cl_kernels/softmax_layer.cl +++ b/src/core/CL/cl_kernels/softmax_layer.cl @@ -23,23 +23,6 @@ */ #include "helpers.h" -#ifdef FIXED_POINT_POSITION - -#include "fixed_point.h" -#define MAX_OP(x, y, type, size) MAX_OP_EXPAND(x, y, type, size) -#define ADD_OP(x, y, type, size) ADD_SAT_OP_EXPAND((x), (y), type, size) -#define SUB_OP(x, y, type, size) SUB_SAT_OP_EXPAND((x), (y), type, size) -#define MUL_OP(x, y, type, size) MUL_SAT_OP_EXPAND((x), (y), type, size, FIXED_POINT_POSITION) -#define DIV_OP(x, y, type, size) DIV_SAT_OP_VEC_EXPAND((x), (y), type, size, FIXED_POINT_POSITION) -#define EXP_OP(x, type, size) EXP_OP_EXPAND((x), type, size, FIXED_POINT_POSITION) - -#define MIN_VAL_EXPAND(type) type##_MIN -#define MIN_VAL(type) MIN_VAL_EXPAND(type) -#define MINVAL MIN_VAL(DATA_TYPE) -#define SELECT_DATA_TYPE EXPAND(DATA_TYPE) - -#else /* FIXED_POINT_POSITION */ - #define MAX_OP(x, y, type, size) max((x), (y)) #define ADD_OP(x, y, type, size) ((x) + (y)) #define SUB_OP(x, y, type, size) ((x) - (y)) @@ -55,8 +38,6 @@ #define SELECT_DATA_TYPE int #endif /* USE_F16 */ -#endif /* FIXED_POINT_POSITION */ - /* Number of workitems in dimension 0. */ #if !defined(GRID_SIZE) #define GRID_SIZE 1 @@ -91,9 +72,8 @@ __constant uint4 idx4 = (uint4)(0, 1, 2, 3); /** Divides all the values of the input tensor by the sum calculated from softmax_layer_shift_exp_sum kernel. * * @note Datatype must be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short - * @note Fixed point position must be given as a preprocessor argument using -DFIXED_POINT_POSITION=pos. e.g. DFIXED_POINT_POSITION=4 * - * @param[in] src_ptr Pointer to the source tensor slice. Supported data types: QS8/QS16/F16/F32 + * @param[in] src_ptr Pointer to the source tensor slice. Supported data types: F16/F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) @@ -138,11 +118,10 @@ __kernel void softmax_layer_norm( * then gets the exponent of each element as sums all elements across each row. * * @note Datatype must be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short - * @note Fixed point position must be given as a preprocessor argument using -DFIXED_POINT_POSITION=pos. e.g. DFIXED_POINT_POSITION=4 * @note In case the input is not a multiple of VECTOR_SIZE (2,4,8,16) -DNON_MULTIPLE_OF_VECTOR_SIZE must be passed. * @note Beta can be optionally passed at compile time using -DBETA (by default, it is 1.0). * - * @param[in] src_ptr Pointer to the source tensor slice. Supported data types: QS8/QS16/F16/F32 + * @param[in] src_ptr Pointer to the source tensor slice. Supported data types: F16/F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) @@ -288,11 +267,10 @@ __kernel void softmax_layer_max_shift_exp_sum_serial( * then gets the exponent of each element as sums all elements across each row. * * @note Datatype must be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short - * @note Fixed point position must be given as a preprocessor argument using -DFIXED_POINT_POSITION=pos. e.g. DFIXED_POINT_POSITION=4 * @note In case the input is not a multiple of VECTOR_SIZE (2,4,8,16) -DNON_MULTIPLE_OF_VECTOR_SIZE must be passed. * @note Beta can be optionally passed at compile time using -DBETA (by default, it is 1.0). * - * @param[in] src_ptr Pointer to the source tensor slice. Supported data types: QS8/QS16/F16/F32 + * @param[in] src_ptr Pointer to the source tensor slice. Supported data types: F16/F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) diff --git a/src/core/CL/cl_kernels/softmax_layer_quantized.cl b/src/core/CL/cl_kernels/softmax_layer_quantized.cl index c055381fc5..95d6d4bcc5 100644 --- a/src/core/CL/cl_kernels/softmax_layer_quantized.cl +++ b/src/core/CL/cl_kernels/softmax_layer_quantized.cl @@ -230,10 +230,9 @@ __kernel void softmax_layer_max_shift_exp_sum_quantized_serial( * then gets the exponent of each element as sums all elements across each row. * * @note Datatype must be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short - * @note Fixed point position must be given as a preprocessor argument using -DFIXED_POINT_POSITION=pos. e.g. DFIXED_POINT_POSITION=4 * @note In case the input is not a multiple of VECTOR_SIZE (2,4,8,16) -DNON_MULTIPLE_OF_VECTOR_SIZE must be passed. * - * @param[in] src_ptr Pointer to the source tensor slice. Supported data types: QS8/QS16/F16/F32 + * @param[in] src_ptr Pointer to the source tensor slice. Supported data types: F16/F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) @@ -519,7 +518,6 @@ __kernel void softmax_layer_max_shift_exp_sum_quantized_parallel( /** Divides all the values of the input tensor by the sum calculated from softmax_layer_shift_exp_sum kernel. * - * @note Fixed point position must be given as a preprocessor argument using -DFIXED_POINT_POSITION=pos. e.g. DFIXED_POINT_POSITION=4 * @note Quantized beta can be optionally passed at compile time using -DINPUT_BETA_MULTIPLIER and -DINPUT_BETA_LEFT_SHIFT (if undefined, assume beta equals 1.0) * @note -DDIFF_MIN must be passed at compile time. It is threshold difference between maximum value of input data and current processed value, it defines whether the value will be taken into account or not. * diff --git a/src/core/CL/kernels/CLActivationLayerKernel.cpp b/src/core/CL/kernels/CLActivationLayerKernel.cpp index 3d8824aa2a..1ae1032cba 100644 --- a/src/core/CL/kernels/CLActivationLayerKernel.cpp +++ b/src/core/CL/kernels/CLActivationLayerKernel.cpp @@ -27,7 +27,6 @@ #include "arm_compute/core/CL/CLKernelLibrary.h" #include "arm_compute/core/CL/CLValidate.h" #include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/FixedPoint.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/IAccessWindow.h" #include "arm_compute/core/TensorInfo.h" @@ -47,7 +46,7 @@ namespace Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ActivationLayerInfo &act_info) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::QASYMM8, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG((input->data_type() == DataType::QASYMM8) && (act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) && (act_info.activation() != ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) && (act_info.activation() != ActivationLayerInfo::ActivationFunction::RELU), @@ -58,7 +57,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; @@ -118,7 +116,6 @@ void CLActivationLayerKernel::configure(ICLTensor *input, ICLTensor *output, Act const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size(); const DataType dt = input->info()->data_type(); - const int fixed_point_position = input->info()->fixed_point_position(); float a_const = act_info.a(); float b_const = act_info.b(); int a_const_int = 0; @@ -127,16 +124,8 @@ void CLActivationLayerKernel::configure(ICLTensor *input, ICLTensor *output, Act // Create quantized version of constants a, b if needed if(is_data_type_quantized(dt)) { - if(is_data_type_fixed_point(dt)) - { - a_const_int = static_cast<int>(lround(a_const * (1 << fixed_point_position))); - b_const_int = static_cast<int>(lround(b_const * (1 << fixed_point_position))); - } - else - { - a_const_int = input->info()->quantization_info().quantize(a_const, RoundingPolicy::TO_NEAREST_UP); - b_const_int = input->info()->quantization_info().quantize(b_const, RoundingPolicy::TO_NEAREST_UP); - } + a_const_int = input->info()->quantization_info().quantize(a_const, RoundingPolicy::TO_NEAREST_UP); + b_const_int = input->info()->quantization_info().quantize(b_const, RoundingPolicy::TO_NEAREST_UP); } // Set build options @@ -177,10 +166,6 @@ void CLActivationLayerKernel::configure(ICLTensor *input, ICLTensor *output, Act } build_opts.emplace((_run_in_place) ? "-DIN_PLACE" : ""); - if(is_data_type_fixed_point(dt)) - { - build_opts.emplace(("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(fixed_point_position))); - } // Create kernel std::string kernel_name = is_data_type_quantized_asymmetric(dt) ? std::string("activation_layer_qa8") : std::string("activation_layer"); diff --git a/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp b/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp index 011807ad88..78651f8679 100644 --- a/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp +++ b/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp @@ -37,9 +37,9 @@ Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, { ARM_COMPUTE_UNUSED(policy); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&input1); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::QASYMM8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&input2); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QASYMM8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::F32); const bool is_qasymm = is_data_type_quantized_asymmetric(input1.data_type()) || is_data_type_quantized_asymmetric(input2.data_type()); if(is_qasymm) @@ -50,18 +50,16 @@ Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const TensorShape out_shape = TensorShape::broadcast_shape(input1.tensor_shape(), input2.tensor_shape()); ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible"); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(&input1, &input2); // Validate in case of configured output if(output.total_size() > 0) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&output, 1, DataType::U8, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&output, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG((output.data_type() == DataType::U8) && ((input1.data_type() != DataType::U8) || (input2.data_type() != DataType::U8)), "Output can only be U8 if both inputs are U8"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output.tensor_shape(), 0), "Wrong shape for output"); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(&input1, &output); if(is_qasymm) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input1, &output); @@ -142,11 +140,7 @@ void CLArithmeticAdditionKernel::configure(const ICLTensor *input1, const ICLTen build_opts.emplace("-DDATA_TYPE_IN1=" + get_cl_type_from_data_type(input1->info()->data_type())); build_opts.emplace("-DDATA_TYPE_IN2=" + get_cl_type_from_data_type(input2->info()->data_type())); build_opts.emplace("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type())); - if(is_data_type_fixed_point(input1->info()->data_type())) - { - build_opts.emplace("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input1->info()->fixed_point_position())); - } - else if(is_data_type_quantized_asymmetric(input1->info()->data_type())) + if(is_data_type_quantized_asymmetric(input1->info()->data_type())) { build_opts.emplace("-DOFFSET=" + support::cpp11::to_string(input1->info()->quantization_info().offset)); kernel_name += "_quantized"; diff --git a/src/core/CL/kernels/CLArithmeticSubtractionKernel.cpp b/src/core/CL/kernels/CLArithmeticSubtractionKernel.cpp index db91bc0084..aeee6022a7 100644 --- a/src/core/CL/kernels/CLArithmeticSubtractionKernel.cpp +++ b/src/core/CL/kernels/CLArithmeticSubtractionKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016, 2017 ARM Limited. + * Copyright (c) 2016-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -44,21 +44,19 @@ Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2, { ARM_COMPUTE_UNUSED(policy); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input1); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input2); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, input2); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input1, input2); // Validate in case of configured output if((output != nullptr) && (output->total_size() != 0)) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() == DataType::U8 && (input1->data_type() != DataType::U8 || input2->data_type() != DataType::U8), "Output can only be U8 if both inputs are U8"); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input1, output); } return Status{}; @@ -122,10 +120,6 @@ void CLArithmeticSubtractionKernel::configure(const ICLTensor *input1, const ICL build_opts.emplace("-DDATA_TYPE_IN1=" + get_cl_type_from_data_type(input1->info()->data_type())); build_opts.emplace("-DDATA_TYPE_IN2=" + get_cl_type_from_data_type(input2->info()->data_type())); build_opts.emplace("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type())); - if(is_data_type_fixed_point(input1->info()->data_type())) - { - build_opts.emplace("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input1->info()->fixed_point_position())); - } // Create kernel _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("arithmetic_sub", build_opts)); diff --git a/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp b/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp index 391baef96a..5999c66056 100644 --- a/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp +++ b/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp @@ -27,7 +27,6 @@ #include "arm_compute/core/CL/CLKernelLibrary.h" #include "arm_compute/core/CL/CLValidate.h" #include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/FixedPoint.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Utils.h" @@ -46,22 +45,19 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, { ARM_COMPUTE_UNUSED(epsilon); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(mean, var); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, mean, var); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, mean, var); ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL)) != mean->dimension(0)); if(beta != nullptr) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(mean, beta); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, beta); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, beta); } if(gamma != nullptr) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(mean, gamma); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, gamma); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, gamma); } if(act_info.enabled()) @@ -78,7 +74,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; @@ -168,7 +163,6 @@ void CLBatchNormalizationLayerKernel::configure(ICLTensor *input, ICLTensor *out build_opts.add_option_if(act_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(act_info.a())); build_opts.add_option_if(act_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(act_info.b())); build_opts.add_option_if(_run_in_place, "-DIN_PLACE"); - build_opts.add_option_if(is_data_type_fixed_point(input->info()->data_type()), "-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position())); build_opts.add_option_if(beta == nullptr, "-DUSE_DEFAULT_BETA"); build_opts.add_option_if(gamma == nullptr, "-DUSE_DEFAULT_GAMMA"); diff --git a/src/core/CL/kernels/CLChannelShuffleLayerKernel.cpp b/src/core/CL/kernels/CLChannelShuffleLayerKernel.cpp index 1de987264c..5f0f0aebf8 100644 --- a/src/core/CL/kernels/CLChannelShuffleLayerKernel.cpp +++ b/src/core/CL/kernels/CLChannelShuffleLayerKernel.cpp @@ -39,8 +39,8 @@ namespace Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, unsigned int num_groups) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::QASYMM8, - DataType::U16, DataType::S16, DataType::QS16, + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8, + DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG(num_groups < 2, "Channel shuffling with less than 2 groups would be inefficient"); diff --git a/src/core/CL/kernels/CLCol2ImKernel.cpp b/src/core/CL/kernels/CLCol2ImKernel.cpp index 64e6a0b7d8..6274c9082a 100644 --- a/src/core/CL/kernels/CLCol2ImKernel.cpp +++ b/src/core/CL/kernels/CLCol2ImKernel.cpp @@ -44,14 +44,13 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, s { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); // Checks performed when output is configured if(output->total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), compute_col2im_shape(*input, convolved_dims)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output); } @@ -64,7 +63,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen // Output auto inizialitation if not yet initialized auto_init_if_empty(*output, input->clone()->set_tensor_shape(compute_col2im_shape(*input, convolved_dims))); - const unsigned int num_elems_read_per_iteration = is_data_type_fixed_point(input->data_type()) ? 1 : 8; + const unsigned int num_elems_read_per_iteration = 8; // Configure window Window win = calculate_max_window(*input, Steps(num_elems_read_per_iteration)); @@ -106,7 +105,6 @@ void CLCol2ImKernel::configure(const ICLTensor *input, ICLTensor *output, std::p build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(input->info()->element_size())); build_opts.add_option("-DWIDTH_INPUT=" + support::cpp11::to_string(input->info()->dimension(0))); build_opts.add_option("-DWIDTH_OUTPUT=" + support::cpp11::to_string(_convolved_dims.first)); - build_opts.add_option_if(is_data_type_fixed_point(data_type), "-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position())); _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("col2im", build_opts.options())); diff --git a/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.cpp b/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.cpp index c3cd494662..a39d1f4a0b 100644 --- a/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.cpp +++ b/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.cpp @@ -75,7 +75,7 @@ Status CLConvertFullyConnectedWeightsKernel::validate(const ITensorInfo *input, DataLayout data_layout) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::QASYMM8, DataType::U16, DataType::S16, DataType::QS16, DataType::U32, DataType::S32, + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8, DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::QS32, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); diff --git a/src/core/CL/kernels/CLDepthConcatenateLayerKernel.cpp b/src/core/CL/kernels/CLDepthConcatenateLayerKernel.cpp index 204f9aed6f..72dc21197d 100644 --- a/src/core/CL/kernels/CLDepthConcatenateLayerKernel.cpp +++ b/src/core/CL/kernels/CLDepthConcatenateLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -62,9 +62,8 @@ void CLDepthConcatenateLayerKernel::configure(const ICLTensor *input, unsigned i }; ARM_COMPUTE_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, output); ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) + depth_offset > output->info()->dimension(2)); ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) > output->info()->dimension(0)); ARM_COMPUTE_ERROR_ON(input->info()->dimension(1) > output->info()->dimension(1)); diff --git a/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp b/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp index 83908a1469..2f5b2466b1 100644 --- a/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp +++ b/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016, 2017 ARM Limited. + * Copyright (c) 2016-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -40,21 +40,15 @@ using namespace arm_compute; void CLDepthConvertLayerKernel::configure(const ICLTensor *input, ICLTensor *output, ConvertPolicy policy, uint32_t shift) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::U8, DataType::S16, DataType::QS16, - DataType::U16, DataType::U32, DataType::S32, DataType::F32); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QS8, DataType::U8, DataType::S16, DataType::QS16, + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S16, + DataType::U16, DataType::U32, DataType::S32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16, DataType::U16, DataType::U32, DataType::S32, DataType::F32); ARM_COMPUTE_ERROR_ON(input == output); ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == output->info()->data_type(), "Input and output data types must be different"); ARM_COMPUTE_ERROR_ON(shift >= 8); // Check if convertion is supported - ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::QS8 && output->info()->data_type() != DataType::F32, - "Only data types supported [in] QS8 -> [out] F32"); - ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::QS16 && (output->info()->data_type() != DataType::F32), - "Only data types supported [in] QS16 -> [out] F32"); - ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::F32 && ((output->info()->data_type() != DataType::QS8) && output->info()->data_type() != DataType::QS16), - "Only data types supported [in] F32 -> [out] QS8, QS16"); ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::U8 && (output->info()->data_type() != DataType::U16 && output->info()->data_type() != DataType::S16 && output->info()->data_type() != DataType::U32 && output->info()->data_type() != DataType::S32), "Only data types supported [in] U8 -> [out] U16, S16, U32, S32"); @@ -99,10 +93,6 @@ void CLDepthConvertLayerKernel::configure(const ICLTensor *input, ICLTensor *out } build_opts.emplace("-DDATA_TYPE_IN=" + get_cl_type_from_data_type(input->info()->data_type())); build_opts.emplace("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type())); - if(is_data_type_fixed_point(input->info()->data_type()) || is_data_type_fixed_point(output->info()->data_type())) - { - build_opts.emplace("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position())); - } // Create kernel _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts)); diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp index 1de08aa1a2..9d9c280182 100644 --- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp +++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp @@ -146,7 +146,6 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const ICLTensor *input, output_shape, 1, input->info()->data_type(), - input->info()->fixed_point_position(), input->info()->quantization_info()); ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, act_info)); diff --git a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp index bef13f9b1c..cab943629a 100644 --- a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp +++ b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp @@ -53,7 +53,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input->data_type()) && has_bias); ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(idx_c) * depth_multiplier) != output->dimension(2)); ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != (kernel_dims.width * kernel_dims.height + ((has_bias) ? 1 : 0))); diff --git a/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp b/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp index c97ecaf8e0..e124ee42f3 100644 --- a/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp +++ b/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp @@ -61,7 +61,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, s TensorShape output_shape = compute_output_shape(input->tensor_shape(), conv_w, conv_h, output->data_layout()); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; diff --git a/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp b/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp index fd3b75484a..c28be3fccf 100644 --- a/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp +++ b/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp @@ -46,7 +46,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input->data_type()) && (biases != nullptr)); ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_c) != output->dimension(1)); ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != (input->dimension(idx_w) * input->dimension(idx_h) + ((biases != nullptr) ? 1 : 0))); @@ -54,7 +53,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c if(biases != nullptr) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases); ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != input->dimension(idx_c)); ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); } diff --git a/src/core/CL/kernels/CLDequantizationLayerKernel.cpp b/src/core/CL/kernels/CLDequantizationLayerKernel.cpp index fa982d6cf2..fba721f50b 100644 --- a/src/core/CL/kernels/CLDequantizationLayerKernel.cpp +++ b/src/core/CL/kernels/CLDequantizationLayerKernel.cpp @@ -54,7 +54,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, ITensorInfo *min_max) { // Output tensor auto initialization if not yet initialized - auto_init_if_empty(*output, input->tensor_shape(), 1, DataType::F32, 0); + auto_init_if_empty(*output, input->tensor_shape(), 1, DataType::F32); constexpr unsigned int num_elems_processed_per_iteration = 4; diff --git a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp index d2794d7abd..dcb4ac1c5d 100644 --- a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp +++ b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp @@ -45,7 +45,7 @@ namespace Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(0) != weights->dimension(1), "Weights should have same width as length"); @@ -84,7 +84,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), misc::shape_calculator::compute_deep_convolution_shape(*input, *weights, conv_info)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; @@ -103,7 +102,6 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen auto_init_if_empty(*output, output_shape, 1, input->data_type(), - input->fixed_point_position(), input->quantization_info()); unsigned int conv_stride_x = std::get<0>(conv_info.stride()); @@ -265,7 +263,6 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL output_shape, 1, input->info()->data_type(), - input->info()->fixed_point_position(), input->info()->quantization_info()); // Perform validation step @@ -302,18 +299,14 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL } else { - bool is_quantized_fixed_point = is_data_type_fixed_point(data_type); - bool is_quantized_asymm = is_data_type_quantized_asymmetric(data_type); - DataType promoted_type = (is_quantized_fixed_point) ? get_promoted_data_type(data_type) : data_type; + bool is_quantized_asymm = is_data_type_quantized_asymmetric(data_type); build_options.add_option_if(is_quantized_asymm, std::string("-DKERNEL_SIZE=" + support::cpp11::to_string(kernel_size))); build_options.add_option(std::string("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type))); build_options.add_option(std::string("-DDATA_SIZE=" + get_data_size_from_data_type(data_type))); build_options.add_option(std::string("-DWEIGHTS_DEPTH=" + support::cpp11::to_string(_weights->info()->dimension(2)))); build_options.add_option(std::string("-DSTRIDE_X=" + support::cpp11::to_string(_conv_stride_x))); - build_options.add_option_if(is_quantized_fixed_point, - std::string("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position()))); - build_options.add_option(std::string("-DDATA_TYPE_PROMOTED=" + get_cl_type_from_data_type(promoted_type))); + build_options.add_option(std::string("-DDATA_TYPE_PROMOTED=" + get_cl_type_from_data_type(data_type))); // Create kernel _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(is_quantized_asymm ? "direct_convolution_1x1_3x3_5x5_quantized" : kernel_name.str(), diff --git a/src/core/CL/kernels/CLFillBorderKernel.cpp b/src/core/CL/kernels/CLFillBorderKernel.cpp index 66504e67b5..3b1edaf46c 100644 --- a/src/core/CL/kernels/CLFillBorderKernel.cpp +++ b/src/core/CL/kernels/CLFillBorderKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016, 2017 ARM Limited. + * Copyright (c) 2016-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -91,10 +91,6 @@ void CLFillBorderKernel::configure(ICLTensor *tensor, BorderSize border_size, Bo build_opts.emplace(("-DBORDER_SIZE_BOTTOM=" + support::cpp11::to_string(border_size.bottom))); build_opts.emplace(("-DBORDER_SIZE_LEFT=" + support::cpp11::to_string(border_size.left))); build_opts.emplace(("-DBORDER_SIZE_RIGHT=" + support::cpp11::to_string(border_size.right))); - if(is_data_type_fixed_point(tensor->info()->data_type())) - { - build_opts.emplace("-DFIXED_POINT_POSITION"); - } // Create kernel _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts)); @@ -125,14 +121,12 @@ void CLFillBorderKernel::configure(ICLTensor *tensor, BorderSize border_size, Bo case DataType::QASYMM8: set_constant_border<uint8_t>(idx, constant_border_value); break; - case DataType::QS8: case DataType::S8: set_constant_border<int8_t>(idx, constant_border_value); break; case DataType::U16: set_constant_border<uint16_t>(idx, constant_border_value); break; - case DataType::QS16: case DataType::S16: set_constant_border<int16_t>(idx, constant_border_value); break; diff --git a/src/core/CL/kernels/CLFloorKernel.cpp b/src/core/CL/kernels/CLFloorKernel.cpp index 11f8e33319..f6b0e829a0 100644 --- a/src/core/CL/kernels/CLFloorKernel.cpp +++ b/src/core/CL/kernels/CLFloorKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -45,7 +45,7 @@ void CLFloorKernel::configure(const ICLTensor *input, ICLTensor *output) ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); // Auto initialize output - auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type(), input->info()->fixed_point_position()); + auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type()); ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); diff --git a/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp b/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp index ba475f5819..12a40cd7dc 100644 --- a/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp +++ b/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp @@ -44,15 +44,14 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, i { ARM_COMPUTE_RETURN_ERROR_ON(mult_interleave4x4_height < 1); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::U8, DataType::S8, - DataType::QS16, DataType::U16, DataType::S16, DataType::U32, DataType::S32, + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::U8, DataType::S8, + DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32); if(output->total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), compute_interleaved_shape(*input, mult_interleave4x4_height)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; diff --git a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp index 3f705ac0a7..e040122663 100644 --- a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp +++ b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp @@ -172,7 +172,7 @@ void CLGEMMLowpMatrixMultiplyKernel::configure(const ICLTensor *input0, const IC tensor_shape.set(0, is_interleaved_transposed ? reshape_info.n() : input1->info()->dimension(0)); tensor_shape.set(1, is_interleaved_transposed ? reshape_info.m() : input0->info()->dimension(1)); - auto_init_if_empty(*output->info(), tensor_shape, 1, DataType::S32, 1, QuantizationInfo()); + auto_init_if_empty(*output->info(), tensor_shape, 1, DataType::S32, QuantizationInfo()); ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), is_interleaved_transposed, reshape_info)); diff --git a/src/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.cpp index 81e455fce8..04cf627818 100644 --- a/src/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.cpp +++ b/src/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.cpp @@ -41,9 +41,8 @@ namespace Status validate_arguments(const ITensorInfo *accum, const ITensorInfo *biases) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(accum); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(accum, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(accum, 1, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(biases, accum); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(biases, accum); ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() != 1); return Status{}; @@ -95,8 +94,6 @@ void CLGEMMMatrixAccumulateBiasesKernel::configure(ICLTensor *accum, const ICLTe CLBuildOptions build_opts; build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(accum->info()->data_type())); build_opts.add_option("-DVECTOR_SIZE=" + support::cpp11::to_string(vector_size)); - build_opts.add_option_if(is_data_type_fixed_point(accum->info()->data_type()), - "-DFIXED_POINT_POSITION=" + support::cpp11::to_string(accum->info()->fixed_point_position())); // Create kernel _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemm_accumulate_biases", build_opts.options())); diff --git a/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp index c50ee24a70..bcc3a01296 100644 --- a/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp +++ b/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp @@ -29,7 +29,6 @@ #include "arm_compute/core/CL/ICLTensor.h" #include "arm_compute/core/CL/OpenCL.h" #include "arm_compute/core/Error.h" -#include "arm_compute/core/FixedPoint.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/Types.h" #include "arm_compute/core/Window.h" @@ -64,7 +63,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, f ARM_COMPUTE_UNUSED(input, output, beta); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); @@ -88,19 +87,7 @@ void CLGEMMMatrixAdditionKernel::configure(const ICLTensor *input, ICLTensor *ou _output = output; std::ostringstream ma_arguments; - if(is_data_type_fixed_point(input->info()->data_type())) - { - ma_arguments << "-DBETA=" << (input->info()->data_type() == DataType::QS8 ? - sqcvt_qs8_f32(beta, input->info()->fixed_point_position()) : - sqcvt_qs16_f32(beta, input->info()->fixed_point_position())) - << " "; - ma_arguments << "-DFIXED_POINT_POSITION=" << input->info()->fixed_point_position(); - } - else - { - ma_arguments << "-DBETA=" << beta; - } - + ma_arguments << "-DBETA=" << beta; std::set<std::string> build_opts; build_opts.emplace(ma_arguments.str()); diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp index 2c2a92d070..814cbb631f 100644 --- a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp +++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp @@ -31,7 +31,6 @@ #include "arm_compute/core/CL/ICLTensor.h" #include "arm_compute/core/CL/OpenCL.h" #include "arm_compute/core/Error.h" -#include "arm_compute/core/FixedPoint.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Types.h" @@ -53,10 +52,8 @@ inline Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *i { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input0); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_data_type_fixed_point(input0->data_type()) && (reshape_info.depth_output_gemm3d() != 1), "GEMM3D only supports floating point data types"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the matrix A must be <= 4"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the matrix B must be <= 3"); @@ -95,7 +92,6 @@ inline Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *i const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, is_interleaved_transposed, reshape_info)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input0, output); } return Status{}; @@ -219,7 +215,6 @@ void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTen _slide_matrix_b = _input1->info()->num_dimensions() >= _input0->info()->num_dimensions(); const DataType data_type = input0->info()->data_type(); - const int fp_pos = input0->info()->fixed_point_position(); // Get target architecture GPUTarget gpu_target = get_target(); @@ -236,14 +231,11 @@ void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTen // Create build options CLBuildOptions build_opts; - build_opts.add_option_if(is_data_type_fixed_point(data_type), "-DFIXED_POINT_POSITION=" + support::cpp11::to_string(fp_pos)); // Only define ALPHA when alpha is not 1.0f. This avoids performing unnecessary multiplications. if(std::abs(1.0f - alpha) > 0.00001f) { - build_opts.add_option_if_else(is_data_type_fixed_point(data_type), - "-DALPHA=" + support::cpp11::to_string((data_type == DataType::QS8 ? sqcvt_qs8_f32(alpha, fp_pos) : sqcvt_qs16_f32(alpha, fp_pos))), - "-DALPHA=" + float_to_string_with_full_precision(alpha)); + build_opts.add_option("-DALPHA=" + float_to_string_with_full_precision(alpha)); } build_opts.add_option_if(_is_gemm3d, "-DREINTERPRET_OUTPUT_AS_3D"); build_opts.add_option_if(_is_gemm3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(1))); @@ -299,10 +291,6 @@ void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTen // via exhaustive autotuning over a range of representative layer configurations. _lws_hint = cl::NDRange(4); } - else if(is_data_type_fixed_point(data_type)) - { - kernel_name = "gemm_mm_" + lower_string(string_from_data_type(data_type)); - } else // (MIDGARD and F32) or (F16) { kernel_name = "gemm_mm_floating_point"; diff --git a/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp index d8ecd501b0..43a6cf25db 100644 --- a/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp +++ b/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp @@ -42,7 +42,6 @@ Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input0); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output); ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input0->data_type()) && (output->data_type() != DataType::S32)); ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(2) != input1->dimension(1)); diff --git a/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp b/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp index 7a8a1e529d..7e44fa7118 100644 --- a/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp +++ b/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp @@ -47,8 +47,8 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, i { ARM_COMPUTE_RETURN_ERROR_ON(mult_transpose1xW_width < 1); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::U8, DataType::S8, - DataType::QS16, DataType::U16, DataType::S16, DataType::U32, DataType::S32, + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::U8, DataType::S8, + DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32); if(output->total_size() != 0) @@ -56,7 +56,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, i ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), compute_transpose1xW_with_element_size_shape(*input, mult_transpose1xW_width)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; diff --git a/src/core/CL/kernels/CLIm2ColKernel.cpp b/src/core/CL/kernels/CLIm2ColKernel.cpp index 5d4e039e94..b54575ae30 100644 --- a/src/core/CL/kernels/CLIm2ColKernel.cpp +++ b/src/core/CL/kernels/CLIm2ColKernel.cpp @@ -48,7 +48,7 @@ namespace Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, bool has_bias, const Size2D &dilation) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::QASYMM8 && has_bias); ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1)); @@ -58,7 +58,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, b if(output->total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; @@ -136,7 +135,7 @@ CLIm2ColKernel::configure_window(const ICLTensor *input, ICLTensor *output, cons if(dilation == Size2D(1U, 1U)) { - if(squared_im2col && !is_data_type_fixed_point(data_type)) + if(squared_im2col) { // Check if we can run an optimized im2col switch(kernel_dims.width) @@ -304,7 +303,6 @@ void CLIm2ColKernel::configure(const ICLTensor *input, ICLTensor *output, const build_opts.add_option(("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type))); build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(input->info()->element_size())); build_opts.add_option_if(has_bias, "-DHAS_BIAS"); - build_opts.add_option_if(is_data_type_fixed_point(data_type), "-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position())); _num_elems_processed_per_iteration = 1; diff --git a/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp b/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp index 3d30350c59..39d9f958d3 100644 --- a/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp +++ b/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp @@ -26,7 +26,6 @@ #include "arm_compute/core/CL/CLHelpers.h" #include "arm_compute/core/CL/CLKernelLibrary.h" #include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/FixedPoint.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Utils.h" @@ -78,7 +77,7 @@ std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITe Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); // Output tensor auto initialization if not yet initialized - auto_init_if_empty(*output, input->tensor_shape(), 1, input->data_type(), input->fixed_point_position()); + auto_init_if_empty(*output, input->tensor_shape(), 1, input->data_type()); AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration); AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); diff --git a/src/core/CL/kernels/CLMinMaxLayerKernel.cpp b/src/core/CL/kernels/CLMinMaxLayerKernel.cpp index 60dd5e7de3..9493ddc878 100644 --- a/src/core/CL/kernels/CLMinMaxLayerKernel.cpp +++ b/src/core/CL/kernels/CLMinMaxLayerKernel.cpp @@ -62,7 +62,7 @@ std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITe TensorShape output_shape = compute_min_max_shape(input); // Output auto initialization if not yet initialized - auto_init_if_empty(*output, output_shape, 1, input->data_type(), input->fixed_point_position()); + auto_init_if_empty(*output, output_shape, 1, input->data_type()); const unsigned int num_elems_processed_per_iteration = 1; diff --git a/src/core/CL/kernels/CLNormalizationLayerKernel.cpp b/src/core/CL/kernels/CLNormalizationLayerKernel.cpp index 5456876ee8..df01eab240 100644 --- a/src/core/CL/kernels/CLNormalizationLayerKernel.cpp +++ b/src/core/CL/kernels/CLNormalizationLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -27,7 +27,6 @@ #include "arm_compute/core/CL/CLKernelLibrary.h" #include "arm_compute/core/CL/CLValidate.h" #include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/FixedPoint.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Utils.h" @@ -40,24 +39,16 @@ namespace Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, NormalizationLayerInfo norm_info) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); ARM_COMPUTE_RETURN_ERROR_ON_MSG(!(norm_info.norm_size() % 2), "Normalization size should be odd"); - if(is_data_type_fixed_point(input->data_type())) - { - ARM_COMPUTE_RETURN_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.beta(), input); - ARM_COMPUTE_RETURN_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.kappa(), input); - ARM_COMPUTE_RETURN_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.scale_coeff(), input); - } - // Checks performed when output is configured if(output->total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; @@ -74,7 +65,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen const unsigned int border_width = is_in_map ? std::min(norm_size / 2, 3U) : 0; const BorderSize border_size = BorderSize(0, border_width); - const unsigned int num_elems_processed_per_iteration = (is_data_type_fixed_point(input->data_type())) ? 16 : 4; + const unsigned int num_elems_processed_per_iteration = 4; const unsigned int num_elems_read_per_iteration = is_in_map ? (num_elems_processed_per_iteration + 2 * (norm_size / 2)) : num_elems_processed_per_iteration; Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); @@ -119,14 +110,12 @@ void CLNormalizationLayerKernel::configure(const ICLTensor *input, ICLTensor *ou const unsigned int border_width = _is_in_map ? std::min(norm_info.norm_size() / 2, 3U) : 0; _border_size = BorderSize(0, border_width); - const unsigned int num_elems_processed_per_iteration = (is_data_type_fixed_point(input->info()->data_type())) ? 16 : 4; + const unsigned int num_elems_processed_per_iteration = 4; const bool is_in_map_2D = (norm_info.type() == NormType::IN_MAP_2D); // Set build options CLBuildOptions build_opts; build_opts.add_option(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()))); - build_opts.add_option_if(is_data_type_fixed_point(input->info()->data_type()), - "-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position())); build_opts.add_option(("-DCOEFF=" + float_to_string_with_full_precision(norm_info.scale_coeff()))); build_opts.add_option(("-DBETA=" + float_to_string_with_full_precision(norm_info.beta()))); build_opts.add_option(("-DKAPPA=" + float_to_string_with_full_precision(norm_info.kappa()))); diff --git a/src/core/CL/kernels/CLPermuteKernel.cpp b/src/core/CL/kernels/CLPermuteKernel.cpp index 168ab81088..7c0c95be1c 100644 --- a/src/core/CL/kernels/CLPermuteKernel.cpp +++ b/src/core/CL/kernels/CLPermuteKernel.cpp @@ -52,8 +52,8 @@ TensorShape get_output_shape(const ITensorInfo *input, const PermutationVector & Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PermutationVector &perm) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::QASYMM8, - DataType::U16, DataType::S16, DataType::QS16, + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8, + DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG((perm != PermutationVector{ 2U, 0U, 1U }) @@ -68,7 +68,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; } diff --git a/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp b/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp index a9df36dfcc..4ea093fe04 100644 --- a/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp +++ b/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp @@ -51,36 +51,23 @@ Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2, ARM_COMPUTE_UNUSED(rounding_policy); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input1); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input2); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG(scale < 0, "Scale cannot be negative."); const TensorShape &out_shape = TensorShape::broadcast_shape(input1->tensor_shape(), input2->tensor_shape()); ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible"); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input1, input2); - - if(is_data_type_fixed_point(input1->data_type())) - { - // All data types must be all QS8 or all QS16 - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input1, input2); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(scale != 1, "Unsupported scaling factor for QS8/QS16. Scale must be 1."); - } // Validate in case of configured output if(output->total_size() > 0) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() == DataType::U8 && (input1->data_type() != DataType::U8 || input2->data_type() != DataType::U8), "Output can only be U8 if both inputs are U8"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output->tensor_shape(), 0), "Wrong shape for output"); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input1, output); - if(is_data_type_fixed_point(input1->data_type())) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input1, output); - } } return Status{}; @@ -174,14 +161,6 @@ void CLPixelWiseMultiplicationKernel::configure(const ICLTensor *input1, const I { compute_type = "int"; } - else if(input1->info()->data_type() == DataType::QS8) - { - compute_type = "qs8"; - } - else if(input1->info()->data_type() == DataType::QS16) - { - compute_type = "qs16"; - } else { compute_type = "ushort"; @@ -197,10 +176,6 @@ void CLPixelWiseMultiplicationKernel::configure(const ICLTensor *input1, const I std::set<std::string> build_opts; build_opts.emplace((overflow_policy == ConvertPolicy::WRAP || is_data_type_float(output->info()->data_type())) ? "-DWRAP" : "-DSATURATE"); build_opts.emplace((rounding_policy == RoundingPolicy::TO_ZERO) ? "-DROUND=_rtz" : "-DROUND=_rte"); - if(is_data_type_fixed_point(input1->info()->data_type())) - { - build_opts.emplace("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input1->info()->fixed_point_position())); - } build_opts.emplace("-DDATA_TYPE_IN1=" + get_cl_type_from_data_type(input1->info()->data_type())); build_opts.emplace("-DDATA_TYPE_IN2=" + get_cl_type_from_data_type(input2->info()->data_type())); build_opts.emplace("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type())); diff --git a/src/core/CL/kernels/CLPoolingLayerKernel.cpp b/src/core/CL/kernels/CLPoolingLayerKernel.cpp index 81c52ed53b..246ab68130 100644 --- a/src/core/CL/kernels/CLPoolingLayerKernel.cpp +++ b/src/core/CL/kernels/CLPoolingLayerKernel.cpp @@ -62,7 +62,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c switch(data_layout) { case DataLayout::NCHW: - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); break; case DataLayout::NHWC: ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); @@ -78,8 +78,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); - TensorInfo out_info(TensorInfo(compute_pool_shape(*input, pool_info), 1, output->data_type(), output->fixed_point_position())); + TensorInfo out_info(TensorInfo(compute_pool_shape(*input, pool_info), 1, output->data_type())); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &out_info); } @@ -214,8 +213,6 @@ void CLPoolingLayerKernel::configure(const ICLTensor *input, ICLTensor *output, CLBuildOptions build_opts; build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type)); build_opts.add_option("-DPOOL_" + string_from_pooling_type(pool_type)); - build_opts.add_option_if(is_data_type_fixed_point(data_type), - "-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position())); build_opts.add_option("-DSTRIDE_X=" + support::cpp11::to_string(pool_stride_x)); build_opts.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(pool_stride_y)); build_opts.add_option("-DPAD_X=" + support::cpp11::to_string(pool_pad_left)); @@ -240,7 +237,7 @@ void CLPoolingLayerKernel::configure(const ICLTensor *input, ICLTensor *output, { // Check if we have pool3x3 with stride_x less equal than 3. In these cases, run an optimized OpenCL kernel where // each thread computes 4 output elements - const bool is_pool3x3_stride_le3 = (pool_size_x == 3) && (pool_size_y == 3) && (pool_stride_x <= 3) && !is_data_type_fixed_point(data_type); + const bool is_pool3x3_stride_le3 = (pool_size_x == 3) && (pool_size_y == 3) && (pool_stride_x <= 3); std::string kernel_name = ((is_pool3x3_stride_le3) ? "pooling_layer_optimized_" : "pooling_layer_") + support::cpp11::to_string(pool_size_x); diff --git a/src/core/CL/kernels/CLQuantizationLayerKernel.cpp b/src/core/CL/kernels/CLQuantizationLayerKernel.cpp index 028e50821f..af751f4832 100644 --- a/src/core/CL/kernels/CLQuantizationLayerKernel.cpp +++ b/src/core/CL/kernels/CLQuantizationLayerKernel.cpp @@ -54,7 +54,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, ITensorInfo *min_max) { // Output tensor auto initialization if not yet initialized - auto_init_if_empty(*output, input->tensor_shape(), 1, DataType::U8, 0); + auto_init_if_empty(*output, input->tensor_shape(), 1, DataType::U8); constexpr unsigned int num_elems_processed_per_iteration = 4; diff --git a/src/core/CL/kernels/CLROIPoolingLayerKernel.cpp b/src/core/CL/kernels/CLROIPoolingLayerKernel.cpp index 51873ff66a..4048e927f5 100644 --- a/src/core/CL/kernels/CLROIPoolingLayerKernel.cpp +++ b/src/core/CL/kernels/CLROIPoolingLayerKernel.cpp @@ -56,7 +56,7 @@ void CLROIPoolingLayerKernel::configure(const ICLTensor *input, const ICLROIArra // Output auto inizialitation if not yet initialized TensorShape output_shape(pool_info.pooled_width(), pool_info.pooled_height(), input->info()->dimension(2), rois->num_values()); - auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position()); + auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type()); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); ARM_COMPUTE_ERROR_ON((output->info()->dimension(0) != pool_info.pooled_width()) || (output->info()->dimension(1) != pool_info.pooled_height())); diff --git a/src/core/CL/kernels/CLReductionOperationKernel.cpp b/src/core/CL/kernels/CLReductionOperationKernel.cpp index c44fced3e3..d64f0d89c5 100644 --- a/src/core/CL/kernels/CLReductionOperationKernel.cpp +++ b/src/core/CL/kernels/CLReductionOperationKernel.cpp @@ -27,7 +27,6 @@ #include "arm_compute/core/CL/CLHelpers.h" #include "arm_compute/core/CL/CLKernelLibrary.h" #include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/FixedPoint.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Utils.h" @@ -65,7 +64,7 @@ std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITe // Output tensor auto initialization if not yet initialized TensorShape output_shape{ input->tensor_shape() }; output_shape.set(axis, 1); - auto_init_if_empty(*output, output_shape, 1, input->data_type(), input->fixed_point_position()); + auto_init_if_empty(*output, output_shape, 1, input->data_type()); const unsigned int num_elems_processed_per_iteration = 16; @@ -118,10 +117,6 @@ void CLReductionOperationKernel::configure(const ICLTensor *input, ICLTensor *ou std::set<std::string> build_opts; build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()))); build_opts.emplace(("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration))); - if(is_data_type_fixed_point(input->info()->data_type())) - { - build_opts.emplace("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position())); - } switch(op) { diff --git a/src/core/CL/kernels/CLReshapeLayerKernel.cpp b/src/core/CL/kernels/CLReshapeLayerKernel.cpp index 15897c9dd7..ce9d7fff67 100644 --- a/src/core/CL/kernels/CLReshapeLayerKernel.cpp +++ b/src/core/CL/kernels/CLReshapeLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -47,12 +47,11 @@ CLReshapeLayerKernel::CLReshapeLayerKernel() void CLReshapeLayerKernel::configure(const ICLTensor *input, ICLTensor *output) { ARM_COMPUTE_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::QASYMM8, - DataType::U16, DataType::S16, DataType::QS16, + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8, + DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32); ARM_COMPUTE_ERROR_ON_NULLPTR(output); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); ARM_COMPUTE_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output); ARM_COMPUTE_ERROR_ON(input->info()->tensor_shape().total_size() != output->info()->tensor_shape().total_size()); diff --git a/src/core/CL/kernels/CLSoftmaxLayerKernel.cpp b/src/core/CL/kernels/CLSoftmaxLayerKernel.cpp index 6a18e5ffce..b9ebdc9583 100644 --- a/src/core/CL/kernels/CLSoftmaxLayerKernel.cpp +++ b/src/core/CL/kernels/CLSoftmaxLayerKernel.cpp @@ -82,11 +82,10 @@ CLBuildOptions prepare_quantized_softmax_build_options(float input_scale, float Status validate_arguments_1DMaxShiftExpSum(const ITensorInfo *input, const ITensorInfo *max, const ITensorInfo *output, const ITensorInfo *sum) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(max, sum, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, max); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, max); const bool is_quantized_asymmetric = is_data_type_quantized_asymmetric(input->data_type()); @@ -102,7 +101,6 @@ Status validate_arguments_1DMaxShiftExpSum(const ITensorInfo *input, const ITens ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); } ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, output); } // Checks performed when sum is configured @@ -117,7 +115,6 @@ Status validate_arguments_1DMaxShiftExpSum(const ITensorInfo *input, const ITens ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(max, sum); } ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(max, sum); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(max, sum); } return Status{}; @@ -126,10 +123,9 @@ Status validate_arguments_1DMaxShiftExpSum(const ITensorInfo *input, const ITens Status validate_arguments_1DNorm(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::S32, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(sum, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, sum); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, sum); // Note: output should always have a scale of 1/256 and offset 0 const QuantizationInfo allowed_quantization_info = QuantizationInfo(1.f / 256, 0); @@ -139,7 +135,6 @@ Status validate_arguments_1DNorm(const ITensorInfo *input, const ITensorInfo *su if(output->total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, output); if(!is_quantized_asymmetric) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); @@ -239,15 +234,11 @@ void CLLogits1DMaxShiftExpSumKernel::configure(const ICLTensor *input, ICLTensor const DataType dt = input->info()->data_type(); const size_t reduction_dim_size = input->info()->dimension(0); - auto beta_int = static_cast<int>(lround(beta * (1 << input->info()->fixed_point_position()))); // Set build options CLBuildOptions build_opts; build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(dt)); - build_opts.add_option_if(is_data_type_fixed_point(dt), - "-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position())); build_opts.add_option_if(dt == DataType::F16, "-DUSE_F16"); - build_opts.add_option_if(is_data_type_fixed_point(dt) && (beta != 1.0f), "-DBETA=" + support::cpp11::to_string(beta_int)); build_opts.add_option_if(is_data_type_float(dt) && (beta != 1.0f), "-DBETA=" + float_to_string_with_full_precision(beta)); build_opts.add_options_if(is_data_type_quantized_asymmetric(dt), prepare_quantized_softmax_build_options(input->info()->quantization_info().scale, beta).options()); @@ -364,8 +355,6 @@ void CLLogits1DNormKernel::configure(const ICLTensor *input, const ICLTensor *su // Set build options CLBuildOptions build_opts; build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())); - build_opts.add_option_if(is_data_type_fixed_point(input->info()->data_type()), - "-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position())); build_opts.add_options_if(is_quantized_asymmetric, prepare_quantized_softmax_build_options(input->info()->quantization_info().scale, beta).options()); diff --git a/src/core/CL/kernels/CLTransposeKernel.cpp b/src/core/CL/kernels/CLTransposeKernel.cpp index 8260606a7d..3d584345d7 100644 --- a/src/core/CL/kernels/CLTransposeKernel.cpp +++ b/src/core/CL/kernels/CLTransposeKernel.cpp @@ -57,8 +57,8 @@ TensorShape transposed_tensor_shape(const TensorShape &in) Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::QASYMM8, - DataType::U16, DataType::S16, DataType::QS16, + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8, + DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32); @@ -68,7 +68,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output) ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; diff --git a/src/core/CL/kernels/CLWeightsReshapeKernel.cpp b/src/core/CL/kernels/CLWeightsReshapeKernel.cpp index b012d58d59..5243c4099e 100644 --- a/src/core/CL/kernels/CLWeightsReshapeKernel.cpp +++ b/src/core/CL/kernels/CLWeightsReshapeKernel.cpp @@ -42,13 +42,12 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *biases, c { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); if(biases != nullptr) { ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input->data_type())); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases); ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 4) && (biases->num_dimensions() != 1)); ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 5) && (biases->num_dimensions() != 2)); ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 4) && (biases->dimension(0) != input->tensor_shape()[3])); @@ -60,7 +59,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *biases, c { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), compute_weights_reshaped_shape(*input, biases != nullptr)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output); } @@ -96,7 +94,6 @@ void CLWeightsReshapeKernel::configure(const ICLTensor *input, const ICLTensor * CLBuildOptions build_opts; build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type)); build_opts.add_option_if(biases != nullptr, "-DHAS_BIAS"); - build_opts.add_option_if(is_data_type_fixed_point(data_type), "-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position())); // Create kernel std::string kernel_name = std::string("reshape_to_columns_") + lower_string(string_from_data_layout(data_layout)); diff --git a/src/core/CL/kernels/CLWidthConcatenateLayerKernel.cpp b/src/core/CL/kernels/CLWidthConcatenateLayerKernel.cpp index 56d6ec8f16..587ba690c2 100644 --- a/src/core/CL/kernels/CLWidthConcatenateLayerKernel.cpp +++ b/src/core/CL/kernels/CLWidthConcatenateLayerKernel.cpp @@ -60,10 +60,9 @@ Status validate_arguments(const ITensorInfo *input, unsigned int width_offset, c { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::QASYMM8, DataType::U16, DataType::S16, DataType::QS16, DataType::F16, DataType::U32, + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8, DataType::U16, DataType::S16, DataType::F16, DataType::U32, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, output); ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) + width_offset > output->dimension(0)); for(size_t i = 1; i < Coordinates::num_max_dimensions; ++i) diff --git a/src/core/CPP/kernels/CPPPermuteKernel.cpp b/src/core/CPP/kernels/CPPPermuteKernel.cpp index 5c93f3e93a..17eaec2670 100644 --- a/src/core/CPP/kernels/CPPPermuteKernel.cpp +++ b/src/core/CPP/kernels/CPPPermuteKernel.cpp @@ -40,8 +40,8 @@ namespace { Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PermutationVector &perm) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::QASYMM8, - DataType::U16, DataType::S16, DataType::QS16, + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8, + DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG(perm.num_dimensions() > 4, "Only up to 4D permutation vectors are supported"); @@ -53,7 +53,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; diff --git a/src/core/GLES_COMPUTE/kernels/GCActivationLayerKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCActivationLayerKernel.cpp index 828782338c..874c3368a5 100644 --- a/src/core/GLES_COMPUTE/kernels/GCActivationLayerKernel.cpp +++ b/src/core/GLES_COMPUTE/kernels/GCActivationLayerKernel.cpp @@ -55,11 +55,10 @@ void GCActivationLayerKernel::configure(IGCTensor *input, IGCTensor *output, Act if(output != nullptr) { // Output auto inizialitation if not yet initialized - auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type(), input->info()->fixed_point_position()); + auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type()); ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); _output = output; } diff --git a/src/core/GLES_COMPUTE/kernels/GCBatchNormalizationLayerKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCBatchNormalizationLayerKernel.cpp index 9a592dfe00..c745f3ff3c 100644 --- a/src/core/GLES_COMPUTE/kernels/GCBatchNormalizationLayerKernel.cpp +++ b/src/core/GLES_COMPUTE/kernels/GCBatchNormalizationLayerKernel.cpp @@ -48,27 +48,23 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, mean, var); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, mean, var); ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(mean, var); if(output->total_size() != 0) { ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } if(beta != nullptr) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(mean, beta); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, beta); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, beta); } if(gamma != nullptr) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(mean, gamma); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, gamma); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, gamma); } if(act_info.enabled()) { @@ -86,7 +82,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen ITensorInfo *beta, ITensorInfo *gamma) { // Output tensor auto initialization if not yet initialized - auto_init_if_empty(*output, input->tensor_shape(), 1, input->data_type(), input->fixed_point_position()); + auto_init_if_empty(*output, input->tensor_shape(), 1, input->data_type()); unsigned int num_elems_processed_per_iteration = 1; if(input->data_type() == DataType::F16) diff --git a/src/core/GLES_COMPUTE/kernels/GCDepthwiseConvolutionLayer3x3Kernel.cpp b/src/core/GLES_COMPUTE/kernels/GCDepthwiseConvolutionLayer3x3Kernel.cpp index c2374096a2..a0d1876315 100644 --- a/src/core/GLES_COMPUTE/kernels/GCDepthwiseConvolutionLayer3x3Kernel.cpp +++ b/src/core/GLES_COMPUTE/kernels/GCDepthwiseConvolutionLayer3x3Kernel.cpp @@ -69,8 +69,7 @@ void GCDepthwiseConvolutionLayer3x3Kernel::configure(const IGCTensor *input, con auto_init_if_empty(*output->info(), output_shape, 1, - input->info()->data_type(), - input->info()->fixed_point_position()); + input->info()->data_type()); ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape); ARM_COMPUTE_ERROR_ON(output->info()->dimension(2) != weights->info()->dimension(2)); diff --git a/src/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.cpp index 67a1530431..ecff233382 100644 --- a/src/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.cpp +++ b/src/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.cpp @@ -80,12 +80,11 @@ void GCDirectConvolutionLayerKernel<kernel_size>::configure(const IGCTensor *inp output_shape.set(2, weights->info()->dimension(3)); // Output auto inizialitation if not yet initialized - auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position()); + auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type()); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); ARM_COMPUTE_ERROR_ON(!conv_info.padding_is_symmetric()); _conv_stride_x = std::get<0>(conv_info.stride()); diff --git a/src/core/GLES_COMPUTE/kernels/GCGEMMInterleave4x4Kernel.cpp b/src/core/GLES_COMPUTE/kernels/GCGEMMInterleave4x4Kernel.cpp index 171fbad702..efd5747985 100644 --- a/src/core/GLES_COMPUTE/kernels/GCGEMMInterleave4x4Kernel.cpp +++ b/src/core/GLES_COMPUTE/kernels/GCGEMMInterleave4x4Kernel.cpp @@ -51,7 +51,7 @@ void GCGEMMInterleave4x4Kernel::configure(const IGCTensor *input, IGCTensor *out output_shape.set(1, std::ceil(input->info()->dimension(1) / 4.0f)); // Output auto inizialitation if not yet initialized - auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position()); + auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type()); ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); diff --git a/src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp index d576c30f80..8ead05abfb 100644 --- a/src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp +++ b/src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp @@ -97,7 +97,6 @@ inline Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *i ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != static_cast<size_t>(n)); ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(1) != static_cast<size_t>(m)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input0, output); } } diff --git a/src/core/GLES_COMPUTE/kernels/GCGEMMTranspose1xWKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCGEMMTranspose1xWKernel.cpp index 5d9f9c2d3e..dfbd0216b4 100644 --- a/src/core/GLES_COMPUTE/kernels/GCGEMMTranspose1xWKernel.cpp +++ b/src/core/GLES_COMPUTE/kernels/GCGEMMTranspose1xWKernel.cpp @@ -49,7 +49,7 @@ void GCGEMMTranspose1xWKernel::configure(const IGCTensor *input, IGCTensor *outp output_shape.set(1, static_cast<size_t>(std::ceil((input->info()->dimension(0) / static_cast<float>(transpose_w))))); // Output tensor auto inizialitation if not yet initialized - auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position()); + auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type()); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape); diff --git a/src/core/GLES_COMPUTE/kernels/GCIm2ColKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCIm2ColKernel.cpp index 6c896168ed..21971903eb 100644 --- a/src/core/GLES_COMPUTE/kernels/GCIm2ColKernel.cpp +++ b/src/core/GLES_COMPUTE/kernels/GCIm2ColKernel.cpp @@ -53,7 +53,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output) if(output->total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; @@ -253,7 +252,7 @@ void GCIm2ColKernel::run_generic(const Window &window) if(_input->info()->data_type() == DataType::F16) { (dynamic_cast<TensorInfo *>(_input->info()))->init(_input->info()->tensor_shape(), _input->info()->num_channels(), _input->info()->data_type(), _input->info()->strides_in_bytes(), 0, - _input->info()->total_size(), _input->info()->fixed_point_position()); + _input->info()->total_size()); } _kernel.use(); diff --git a/src/core/GLES_COMPUTE/kernels/GCPoolingLayerKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCPoolingLayerKernel.cpp index 3a0944cd48..f225ebde6b 100644 --- a/src/core/GLES_COMPUTE/kernels/GCPoolingLayerKernel.cpp +++ b/src/core/GLES_COMPUTE/kernels/GCPoolingLayerKernel.cpp @@ -75,7 +75,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c if(output->total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); unsigned int pooled_w = 0; unsigned int pooled_h = 0; @@ -118,8 +117,7 @@ std::tuple<Status, Window, GCPoolingConfig> validate_and_configure_window(ITenso auto_init(input, output, pooled_w, pooled_h); - BorderSize border_size = BorderSize(pool_pad_y, pool_pad_x); - const DataType data_type = input->data_type(); + BorderSize border_size = BorderSize(pool_pad_y, pool_pad_x); const int input_width = input->dimension(0); const int input_height = input->dimension(1); @@ -131,7 +129,7 @@ std::tuple<Status, Window, GCPoolingConfig> validate_and_configure_window(ITenso { // Check if we have pool3x3 with stride_x less equal than 3. In these cases, run an optimized OpenGLES kernel where // each thread computes 4 output elements - const bool is_pool3x3_stride_le3 = (pool_size == 3) && (pool_stride_x <= 3) && !is_data_type_fixed_point(data_type); + const bool is_pool3x3_stride_le3 = (pool_size == 3) && (pool_stride_x <= 3); int num_elems_read_per_iteration = pool_size; @@ -261,8 +259,6 @@ void GCPoolingLayerKernel::configure(const IGCTensor *input, IGCTensor *output, _output = output; _pool_info = pool_info; - const DataType data_type = input->info()->data_type(); - // Set build options std::set<std::string> build_opts; build_opts.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(1)); @@ -293,7 +289,7 @@ void GCPoolingLayerKernel::configure(const IGCTensor *input, IGCTensor *output, { // Check if we have pool3x3 with stride_x less equal than 3. In these cases, run an optimized OpenGLES kernel where // each thread computes 4 output elements - const bool is_pool3x3_stride_le3 = (pool_size == 3) && (pool_stride_x <= 3) && !is_data_type_fixed_point(data_type); + const bool is_pool3x3_stride_le3 = (pool_size == 3) && (pool_stride_x <= 3); std::string kernel_name = "pooling_layer_" + support::cpp11::to_string(pool_size); if(is_pool3x3_stride_le3) diff --git a/src/core/GLES_COMPUTE/kernels/GCSoftmaxLayerKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCSoftmaxLayerKernel.cpp index 040a66358f..7ae2fc9fa5 100644 --- a/src/core/GLES_COMPUTE/kernels/GCSoftmaxLayerKernel.cpp +++ b/src/core/GLES_COMPUTE/kernels/GCSoftmaxLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -49,7 +49,7 @@ void GCLogits1DMaxKernel::configure(const IGCTensor *input, IGCTensor *output) output_shape.set(0, 1); // Output auto initialization if not yet initialized - auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position()); + auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type()); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape); @@ -110,8 +110,8 @@ void GCLogits1DShiftExpSumKernel::configure(const IGCTensor *input, const IGCTen ARM_COMPUTE_ERROR_ON_NULLPTR(max, sum, output); // Output auto initialization if not yet initialized - auto_init_if_empty(*sum->info(), max->info()->tensor_shape(), 1, input->info()->data_type(), input->info()->fixed_point_position()); - auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type(), input->info()->fixed_point_position()); + auto_init_if_empty(*sum->info(), max->info()->tensor_shape(), 1, input->info()->data_type()); + auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type()); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output, max, sum); ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); @@ -204,10 +204,9 @@ void GCLogits1DNormKernel::configure(const IGCTensor *input, const IGCTensor *su ARM_COMPUTE_ERROR_ON_NULLPTR(sum, output); // Output auto initialization if not yet initialized - auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type(), input->info()->fixed_point_position()); + auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type()); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, sum, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, sum, output); ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); _input = input; diff --git a/src/core/GLES_COMPUTE/kernels/GCTransposeKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCTransposeKernel.cpp index bda08e4238..7248891abe 100644 --- a/src/core/GLES_COMPUTE/kernels/GCTransposeKernel.cpp +++ b/src/core/GLES_COMPUTE/kernels/GCTransposeKernel.cpp @@ -49,7 +49,7 @@ void GCTransposeKernel::configure(const IGCTensor *input, IGCTensor *output) output_shape.set(1, h_out); // Output tensor auto inizialitation if not yet initialized - auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position()); + auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type()); ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); diff --git a/src/core/NEON/kernels/NEActivationLayerKernel.cpp b/src/core/NEON/kernels/NEActivationLayerKernel.cpp index ec125154a4..bdc93ed1b8 100644 --- a/src/core/NEON/kernels/NEActivationLayerKernel.cpp +++ b/src/core/NEON/kernels/NEActivationLayerKernel.cpp @@ -23,7 +23,6 @@ */ #include "arm_compute/core/NEON/kernels/NEActivationLayerKernel.h" -#include "arm_compute/core/FixedPoint.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/ITensor.h" #include "arm_compute/core/NEON/NEAsymm.h" @@ -46,14 +45,13 @@ namespace Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::QASYMM8, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::QASYMM8, DataType::F16, DataType::F32); // Checks performed when output is configured if((output != nullptr) && (output->total_size() != 0)) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; @@ -146,36 +144,6 @@ void NEActivationLayerKernel::configure(ITensor *input, ITensor *output, Activat }; #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC*/ - // Activation functions : QS8 - static std::map<ActivationFunction, ActivationFunctionExecutorPtr> act_map_qs8 = - { - { ActivationFunction::ABS, &NEActivationLayerKernel::activation<ActivationFunction::ABS, qint8_t> }, - { ActivationFunction::LINEAR, &NEActivationLayerKernel::activation<ActivationFunction::LINEAR, qint8_t> }, - { ActivationFunction::LOGISTIC, &NEActivationLayerKernel::activation<ActivationFunction::LOGISTIC, qint8_t> }, - { ActivationFunction::RELU, &NEActivationLayerKernel::activation<ActivationFunction::RELU, qint8_t> }, - { ActivationFunction::BOUNDED_RELU, &NEActivationLayerKernel::activation<ActivationFunction::BOUNDED_RELU, qint8_t> }, - { ActivationFunction::LU_BOUNDED_RELU, &NEActivationLayerKernel::activation<ActivationFunction::LU_BOUNDED_RELU, qint8_t> }, - { ActivationFunction::LEAKY_RELU, &NEActivationLayerKernel::activation<ActivationFunction::LEAKY_RELU, qint8_t> }, - { ActivationFunction::SOFT_RELU, &NEActivationLayerKernel::activation<ActivationFunction::SOFT_RELU, qint8_t> }, - { ActivationFunction::SQRT, &NEActivationLayerKernel::activation<ActivationFunction::SQRT, qint8_t> }, - { ActivationFunction::SQUARE, &NEActivationLayerKernel::activation<ActivationFunction::SQUARE, qint8_t> }, - { ActivationFunction::TANH, &NEActivationLayerKernel::activation<ActivationFunction::TANH, qint8_t> }, - }; - // Activation functions : QS16 - static std::map<ActivationFunction, ActivationFunctionExecutorPtr> act_map_qs16 = - { - { ActivationFunction::ABS, &NEActivationLayerKernel::activation<ActivationFunction::ABS, qint16_t> }, - { ActivationFunction::LINEAR, &NEActivationLayerKernel::activation<ActivationFunction::LINEAR, qint16_t> }, - { ActivationFunction::LOGISTIC, &NEActivationLayerKernel::activation<ActivationFunction::LOGISTIC, qint16_t> }, - { ActivationFunction::RELU, &NEActivationLayerKernel::activation<ActivationFunction::RELU, qint16_t> }, - { ActivationFunction::BOUNDED_RELU, &NEActivationLayerKernel::activation<ActivationFunction::BOUNDED_RELU, qint16_t> }, - { ActivationFunction::LU_BOUNDED_RELU, &NEActivationLayerKernel::activation<ActivationFunction::LU_BOUNDED_RELU, qint16_t> }, - { ActivationFunction::LEAKY_RELU, &NEActivationLayerKernel::activation<ActivationFunction::LEAKY_RELU, qint16_t> }, - { ActivationFunction::SOFT_RELU, &NEActivationLayerKernel::activation<ActivationFunction::SOFT_RELU, qint16_t> }, - { ActivationFunction::SQRT, &NEActivationLayerKernel::activation<ActivationFunction::SQRT, qint16_t> }, - { ActivationFunction::SQUARE, &NEActivationLayerKernel::activation<ActivationFunction::SQUARE, qint16_t> }, - { ActivationFunction::TANH, &NEActivationLayerKernel::activation<ActivationFunction::TANH, qint16_t> }, - }; // Activation functions : QASYMM8 static std::map<ActivationFunction, ActivationFunctionExecutorPtr> act_map_qasymm8 = { @@ -188,12 +156,6 @@ void NEActivationLayerKernel::configure(ITensor *input, ITensor *output, Activat case DataType::QASYMM8: _func = act_map_qasymm8[activation_info.activation()]; break; - case DataType::QS8: - _func = act_map_qs8[activation_info.activation()]; - break; - case DataType::QS16: - _func = act_map_qs16[activation_info.activation()]; - break; case DataType::F32: _func = act_map_f32[activation_info.activation()]; break; @@ -508,70 +470,6 @@ typename std::enable_if<std::is_same<T, float>::value, void>::type NEActivationL } template <ActivationLayerInfo::ActivationFunction F, typename T> -typename std::enable_if<std::is_same<T, int8_t>::value, void>::type NEActivationLayerKernel::activation(const Window &window) -{ - Iterator input(_input, window); - Iterator output(_output, window); - const int fixed_point_position = _input->info()->fixed_point_position(); - - static const qint8x16_t CONST_0 = vdupq_n_qs8(0); - const qint8x16_t CONST_1 = vdupq_n_qs8(sqcvt_qs8_f32(1.f, fixed_point_position)); - const qint8x16_t a = vdupq_n_qs8(sqcvt_qs8_f32(_act_info.a(), fixed_point_position)); - const qint8x16_t b = vdupq_n_qs8(sqcvt_qs8_f32(_act_info.b(), fixed_point_position)); - - execute_window_loop(window, [&](const Coordinates & id) - { - const auto input_ptr = reinterpret_cast<const int8_t *>(input.ptr()); - const auto output_ptr = reinterpret_cast<int8_t *>(output.ptr()); - - const qint8x16_t in = vld1q_qs8(input_ptr); - qint8x16_t tmp = {}; - - switch(F) - { - case ActivationFunction::ABS: - tmp = vqabsq_qs8(in); - break; - case ActivationFunction::LINEAR: - tmp = vqmlaq_qs8(b, a, in, fixed_point_position); - break; - case ActivationFunction::LOGISTIC: - tmp = vqrecipq_qs8(vqaddq_qs8(CONST_1, vqexpq_qs8(vnegq_s8(in), fixed_point_position)), fixed_point_position); - break; - case ActivationFunction::RELU: - tmp = vmaxq_qs8(CONST_0, in); - break; - case ActivationFunction::BOUNDED_RELU: - tmp = vminq_qs8(a, vmaxq_qs8(CONST_0, in)); - break; - case ActivationFunction::LU_BOUNDED_RELU: - tmp = vminq_qs8(a, vmaxq_qs8(b, in)); - break; - case ActivationFunction::LEAKY_RELU: - tmp = vbslq_s8(vcgtq_s8(in, CONST_0), in, vmulq_qs8(a, in, fixed_point_position)); - break; - case ActivationFunction::SOFT_RELU: - tmp = vlogq_qs8(vqaddq_qs8(CONST_1, vqexpq_qs8(in, fixed_point_position)), fixed_point_position); - break; - case ActivationFunction::SQRT: - tmp = vqrecipq_qs8(vqinvsqrtq_qs8(in, fixed_point_position), fixed_point_position); - break; - case ActivationFunction::SQUARE: - tmp = vqmulq_qs8(in, in, fixed_point_position); - break; - case ActivationFunction::TANH: - tmp = vqmulq_qs8(a, vqtanhq_qs8(vqmulq_qs8(b, in, fixed_point_position), fixed_point_position), fixed_point_position); - break; - default: - break; - } - - vst1q_qs8(output_ptr, tmp); - }, - input, output); -} - -template <ActivationLayerInfo::ActivationFunction F, typename T> typename std::enable_if<std::is_same<T, qasymm8_t>::value, void>::type NEActivationLayerKernel::activation(const Window &window) { Iterator input(_input, window); @@ -620,137 +518,6 @@ typename std::enable_if<std::is_same<T, qasymm8_t>::value, void>::type NEActivat input, output); } -template <ActivationLayerInfo::ActivationFunction F, typename T> -typename std::enable_if<std::is_same<T, qint16_t>::value, void>::type NEActivationLayerKernel::activation(const Window &window) -{ - Iterator input(_input, window); - Iterator output(_output, window); - const int fixed_point_position = _input->info()->fixed_point_position(); - - static const qint16x8_t CONST_0 = vdupq_n_qs16(0); - const qint16x8_t CONST_1 = vdupq_n_qs16(sqcvt_qs16_f32(1.f, fixed_point_position)); - const qint16x8_t a = vdupq_n_qs16(sqcvt_qs16_f32(_act_info.a(), fixed_point_position)); - const qint16x8_t b = vdupq_n_qs16(sqcvt_qs16_f32(_act_info.b(), fixed_point_position)); - - execute_window_loop(window, [&](const Coordinates & id) - { - const auto input_ptr = reinterpret_cast<const int16_t *>(input.ptr()); - const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr()); - - const qint16x8x2_t in = vld2q_s16(input_ptr); - qint16x8x2_t tmp = { {} }; - - switch(F) - { - case ActivationFunction::ABS: - tmp = - { - { - vqabsq_qs16(in.val[0]), - vqabsq_qs16(in.val[1]), - } - }; - break; - case ActivationFunction::LINEAR: - tmp = - { - { - vqmlaq_qs16(b, a, in.val[0], fixed_point_position), - vqmlaq_qs16(b, a, in.val[1], fixed_point_position), - } - }; - break; - case ActivationFunction::LOGISTIC: - tmp = - { - { - vqrecipq_qs16(vqaddq_qs16(CONST_1, vqexpq_qs16(vnegq_s16(in.val[0]), fixed_point_position)), fixed_point_position), - vqrecipq_qs16(vqaddq_qs16(CONST_1, vqexpq_qs16(vnegq_s16(in.val[1]), fixed_point_position)), fixed_point_position), - } - }; - break; - case ActivationFunction::RELU: - tmp = - { - { - vmaxq_qs16(CONST_0, in.val[0]), - vmaxq_qs16(CONST_0, in.val[1]), - } - }; - break; - case ActivationFunction::BOUNDED_RELU: - tmp = - { - { - vminq_qs16(a, vmaxq_qs16(CONST_0, in.val[0])), - vminq_qs16(a, vmaxq_qs16(CONST_0, in.val[1])), - } - }; - break; - case ActivationFunction::LU_BOUNDED_RELU: - tmp = - { - { - vminq_qs16(a, vmaxq_qs16(b, in.val[0])), - vminq_qs16(a, vmaxq_qs16(b, in.val[1])), - } - }; - break; - case ActivationFunction::LEAKY_RELU: - tmp = - { - { - vbslq_s16(vcgtq_s16(in.val[0], CONST_0), in.val[0], vmulq_qs16(a, in.val[0], fixed_point_position)), - vbslq_s16(vcgtq_s16(in.val[1], CONST_0), in.val[1], vmulq_qs16(a, in.val[1], fixed_point_position)), - } - }; - break; - case ActivationFunction::SOFT_RELU: - tmp = - { - { - vlogq_qs16(vqaddq_qs16(CONST_1, vqexpq_qs16(in.val[0], fixed_point_position)), fixed_point_position), - vlogq_qs16(vqaddq_qs16(CONST_1, vqexpq_qs16(in.val[1], fixed_point_position)), fixed_point_position), - } - }; - break; - case ActivationFunction::SQRT: - tmp = - { - { - vqrecipq_qs16(vqinvsqrtq_qs16(in.val[0], fixed_point_position), fixed_point_position), - vqrecipq_qs16(vqinvsqrtq_qs16(in.val[1], fixed_point_position), fixed_point_position), - } - }; - break; - case ActivationFunction::SQUARE: - tmp = - { - { - vqmulq_qs16(in.val[0], in.val[0], fixed_point_position), - vqmulq_qs16(in.val[1], in.val[1], fixed_point_position), - } - }; - break; - case ActivationFunction::TANH: - tmp = - { - { - vqmulq_qs16(a, vqtanhq_qs16(vqmulq_qs16(b, in.val[0], fixed_point_position), fixed_point_position), fixed_point_position), - vqmulq_qs16(a, vqtanhq_qs16(vqmulq_qs16(b, in.val[1], fixed_point_position), fixed_point_position), fixed_point_position), - } - }; - break; - default: - ARM_COMPUTE_ERROR("Function not implemented"); - break; - } - - vst2q_qs16(output_ptr, tmp); - }, - input, output); -} - Status NEActivationLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ActivationLayerInfo &act_info) { ARM_COMPUTE_UNUSED(act_info); diff --git a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp index a487090a98..f8e2b6d73e 100644 --- a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp +++ b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp @@ -48,38 +48,6 @@ namespace { constexpr unsigned int num_elems_processed_per_iteration = 16; -void add_wrap_QS8_QS8_QS8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) -{ - Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape())); - Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape())); - Iterator output(out, window); - - execute_window_loop(window, [&](const Coordinates & id) - { - const qint8x16_t a = vld1q_qs8(reinterpret_cast<const qint8_t *>(input1.ptr())); - const qint8x16_t b = vld1q_qs8(reinterpret_cast<const qint8_t *>(input2.ptr())); - - vst1q_qs8(reinterpret_cast<qint8_t *>(output.ptr()), vaddq_qs8(a, b)); - }, - input1, input2, output); -} - -void add_saturate_QS8_QS8_QS8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) -{ - Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape())); - Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape())); - Iterator output(out, window); - - execute_window_loop(window, [&](const Coordinates & id) - { - const qint8x16_t a = vld1q_qs8(reinterpret_cast<const qint8_t *>(input1.ptr())); - const qint8x16_t b = vld1q_qs8(reinterpret_cast<const qint8_t *>(input2.ptr())); - - vst1q_qs8(reinterpret_cast<qint8_t *>(output.ptr()), vqaddq_qs8(a, b)); - }, - input1, input2, output); -} - void add_wrap_U8_U8_U8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) { Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape())); @@ -362,28 +330,21 @@ Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, { ARM_COMPUTE_UNUSED(policy); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); const TensorShape out_shape = TensorShape::broadcast_shape(input1.tensor_shape(), input2.tensor_shape()); ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible"); - if(is_data_type_fixed_point(input1.data_type()) || is_data_type_fixed_point(input2.data_type())) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(&input1, &input2); - } - // Validate in case of configured output if(output.total_size() > 0) { ARM_COMPUTE_RETURN_ERROR_ON_MSG( - !(input1.data_type() == DataType::QS8 && input2.data_type() == DataType::QS8 && output.data_type() == DataType::QS8) - && !(input1.data_type() == DataType::U8 && input2.data_type() == DataType::U8 && output.data_type() == DataType::U8) + !(input1.data_type() == DataType::U8 && input2.data_type() == DataType::U8 && output.data_type() == DataType::U8) && !(input1.data_type() == DataType::U8 && input2.data_type() == DataType::U8 && output.data_type() == DataType::S16) && !(input1.data_type() == DataType::U8 && input2.data_type() == DataType::S16 && output.data_type() == DataType::S16) && !(input1.data_type() == DataType::S16 && input2.data_type() == DataType::U8 && output.data_type() == DataType::S16) - && !(input1.data_type() == DataType::QS16 && input2.data_type() == DataType::QS16 && output.data_type() == DataType::QS16) && !(input1.data_type() == DataType::S16 && input2.data_type() == DataType::S16 && output.data_type() == DataType::S16) && !(input1.data_type() == DataType::F32 && input2.data_type() == DataType::F32 && output.data_type() == DataType::F32) && !(input1.data_type() == DataType::F16 && input2.data_type() == DataType::F16 && output.data_type() == DataType::F16), @@ -391,11 +352,6 @@ Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output.tensor_shape(), 0), "Wrong shape for output"); - - if(is_data_type_fixed_point(input1.data_type()) || is_data_type_fixed_point(output.data_type())) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(&input1, &output); - } } return Status{}; @@ -460,8 +416,6 @@ void NEArithmeticAdditionKernel::configure(const ITensor *input1, const ITensor static std::map<std::string, AddFunction *> map_function = { - { "add_wrap_QS8_QS8_QS8", &add_wrap_QS8_QS8_QS8 }, - { "add_saturate_QS8_QS8_QS8", &add_saturate_QS8_QS8_QS8 }, { "add_wrap_U8_U8_U8", &add_wrap_U8_U8_U8 }, { "add_saturate_U8_U8_U8", &add_saturate_U8_U8_U8 }, { "add_wrap_S16_U8_S16", &add_wrap_S16_U8_S16 }, @@ -470,8 +424,6 @@ void NEArithmeticAdditionKernel::configure(const ITensor *input1, const ITensor { "add_saturate_U8_S16_S16", &add_saturate_U8_S16_S16 }, { "add_wrap_U8_U8_S16", &add_wrap_U8_U8_S16 }, { "add_saturate_U8_U8_S16", &add_saturate_U8_U8_S16 }, - { "add_wrap_QS16_QS16_QS16", &add_wrap_S16_S16_S16 }, - { "add_saturate_QS16_QS16_QS16", &add_saturate_S16_S16_S16 }, { "add_wrap_S16_S16_S16", &add_wrap_S16_S16_S16 }, { "add_saturate_S16_S16_S16", &add_saturate_S16_S16_S16 }, { "add_wrap_F32_F32_F32", &add_F32_F32_F32 }, diff --git a/src/core/NEON/kernels/NEArithmeticSubtractionKernel.cpp b/src/core/NEON/kernels/NEArithmeticSubtractionKernel.cpp index 3db80285c0..5a162e3b2c 100644 --- a/src/core/NEON/kernels/NEArithmeticSubtractionKernel.cpp +++ b/src/core/NEON/kernels/NEArithmeticSubtractionKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016, 2017 ARM Limited. + * Copyright (c) 2016-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -45,38 +45,6 @@ class Coordinates; namespace { -void sub_wrap_QS8_QS8_QS8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) -{ - Iterator input1(in1, window); - Iterator input2(in2, window); - Iterator output(out, window); - - execute_window_loop(window, [&](const Coordinates & id) - { - const qint8x16_t a = vld1q_qs8(reinterpret_cast<const qint8_t *>(input1.ptr())); - const qint8x16_t b = vld1q_qs8(reinterpret_cast<const qint8_t *>(input2.ptr())); - - vst1q_qs8(reinterpret_cast<qint8_t *>(output.ptr()), vsubq_qs8(a, b)); - }, - input1, input2, output); -} - -void sub_saturate_QS8_QS8_QS8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) -{ - Iterator input1(in1, window); - Iterator input2(in2, window); - Iterator output(out, window); - - execute_window_loop(window, [&](const Coordinates & id) - { - const qint8x16_t a = vld1q_qs8(reinterpret_cast<const qint8_t *>(input1.ptr())); - const qint8x16_t b = vld1q_qs8(reinterpret_cast<const qint8_t *>(input2.ptr())); - - vst1q_qs8(reinterpret_cast<qint8_t *>(output.ptr()), vqsubq_qs8(a, b)); - }, - input1, input2, output); -} - void sub_wrap_U8_U8_U8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) { Iterator input1(in1, window); @@ -353,23 +321,15 @@ inline Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *i { ARM_COMPUTE_UNUSED(policy); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, input2, output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::QS8, DataType::U8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::QS8, DataType::U8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QS8, DataType::U8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); - - if(is_data_type_fixed_point(input1->data_type()) || is_data_type_fixed_point(input2->data_type()) || is_data_type_fixed_point(output->data_type())) - { - // Check that all data types are the same and all fixed-point positions are the same - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input1, input2, output); - } + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG( - !(input1->data_type() == DataType::QS8 && input2->data_type() == DataType::QS8 && output->data_type() == DataType::QS8) - && !(input1->data_type() == DataType::U8 && input2->data_type() == DataType::U8 && output->data_type() == DataType::U8) + !(input1->data_type() == DataType::U8 && input2->data_type() == DataType::U8 && output->data_type() == DataType::U8) && !(input1->data_type() == DataType::U8 && input2->data_type() == DataType::U8 && output->data_type() == DataType::S16) && !(input1->data_type() == DataType::U8 && input2->data_type() == DataType::S16 && output->data_type() == DataType::S16) && !(input1->data_type() == DataType::S16 && input2->data_type() == DataType::U8 && output->data_type() == DataType::S16) - && !(input1->data_type() == DataType::QS16 && input2->data_type() == DataType::QS16 && output->data_type() == DataType::QS16) && !(input1->data_type() == DataType::S16 && input2->data_type() == DataType::S16 && output->data_type() == DataType::S16) && !(input1->data_type() == DataType::F32 && input2->data_type() == DataType::F32 && output->data_type() == DataType::F32) && !(input1->data_type() == DataType::F16 && input2->data_type() == DataType::F16 && output->data_type() == DataType::F16), @@ -432,8 +392,6 @@ void NEArithmeticSubtractionKernel::configure(const ITensor *input1, const ITens static std::map<std::string, NEArithmeticSubtractionKernel::SubFunction *> map_function = { - { "sub_wrap_QS8_QS8_QS8", &sub_wrap_QS8_QS8_QS8 }, - { "sub_saturate_QS8_QS8_QS8", &sub_saturate_QS8_QS8_QS8 }, { "sub_wrap_U8_U8_U8", &sub_wrap_U8_U8_U8 }, { "sub_wrap_U8_U8_S16", &sub_wrap_U8_U8_S16 }, { "sub_saturate_U8_U8_U8", &sub_saturate_U8_U8_U8 }, @@ -442,8 +400,6 @@ void NEArithmeticSubtractionKernel::configure(const ITensor *input1, const ITens { "sub_wrap_S16_U8_S16", &sub_wrap_S16_U8_S16 }, { "sub_saturate_U8_S16_S16", &sub_saturate_U8_S16_S16 }, { "sub_saturate_S16_U8_S16", &sub_saturate_S16_U8_S16 }, - { "sub_wrap_QS16_QS16_QS16", &sub_wrap_S16_S16_S16 }, - { "sub_saturate_QS16_QS16_QS16", &sub_saturate_S16_S16_S16 }, { "sub_wrap_S16_S16_S16", &sub_wrap_S16_S16_S16 }, { "sub_saturate_S16_S16_S16", &sub_saturate_S16_S16_S16 }, { "sub_wrap_F32_F32_F32", &sub_F32_F32_F32 }, diff --git a/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp b/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp index 6be50fdb0d..6aed41f3aa 100644 --- a/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp +++ b/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp @@ -43,7 +43,7 @@ validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const IT const ITensorInfo *beta, const ITensorInfo *gamma, float epsilon, ActivationLayerInfo act_info) { ARM_COMPUTE_UNUSED(epsilon); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); if(act_info.enabled()) @@ -60,22 +60,18 @@ validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const IT ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, mean, var); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, mean, var); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(mean, var); if(beta != nullptr) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, beta); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, beta); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(mean, beta); } if(gamma != nullptr) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, gamma); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, gamma); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(mean, gamma); } ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL)) != mean->dimension(0)); @@ -104,112 +100,6 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen } //namespace template <bool fused_activation> -void NEBatchNormalizationLayerKernel::batch_normalization_qs8(const Window &window) -{ - static_assert(!fused_activation, "Activation is not supported for QS8"); - - Iterator input(_input, window); - Iterator output(_output, window); - - // Hold information about the current feature map we are iterating. - // Only compute denominator and NEON vectors once per feature map. - int slice = -1; - - const int fixed_point_position = _input->info()->fixed_point_position(); - const auto input_mean = reinterpret_cast<const qint8_t *>(_mean->ptr_to_element(Coordinates(0, 0))); - const auto input_var = reinterpret_cast<const qint8_t *>(_var->ptr_to_element(Coordinates(0, 0))); - const auto input_gamma = (_gamma != nullptr) ? reinterpret_cast<const qint8_t *>(_gamma->ptr_to_element(Coordinates(0, 0))) : nullptr; - const auto input_beta = (_beta != nullptr) ? reinterpret_cast<const qint8_t *>(_beta->ptr_to_element(Coordinates(0, 0))) : nullptr; - - qint8x16_t mean_vec = vdupq_n_qs8(0); - qint8x16_t var_vec = vdupq_n_qs8(0); - qint8x16_t gamma_vec = vdupq_n_qs8(sqcvt_qs8_f32(1, fixed_point_position)); - qint8x16_t beta_vec = vdupq_n_qs8(sqcvt_qs8_f32(0, fixed_point_position)); - qint8x16_t denominator = vdupq_n_qs8(0); - const qint8x16_t epsilon_vec = vdupq_n_qs8(sqcvt_qs8_f32(_epsilon, fixed_point_position)); - execute_window_loop(window, [&](const Coordinates & id) - { - if(slice != id.z()) - { - // Conctruct vectors - mean_vec = vdupq_n_qs8(*(input_mean + id.z())); - var_vec = vdupq_n_qs8(*(input_var + id.z())); - if(input_gamma != nullptr) - { - gamma_vec = vdupq_n_qs8(*(input_gamma + id.z())); - } - if(input_beta != nullptr) - { - beta_vec = vdupq_n_qs8(*(input_beta + id.z())); - } - - // Calculate denominator - denominator = vqinvsqrtq_qs8(vqaddq_qs8(var_vec, epsilon_vec), fixed_point_position); - slice = id.z(); - } - - // Calculate x bar and store results - const qint8x16_t numerator = vqsubq_qs8(vld1q_qs8(reinterpret_cast<const qint8_t *>(input.ptr())), mean_vec); - const qint8x16_t x_bar = vqmulq_qs8(numerator, denominator, fixed_point_position); - vst1q_qs8(reinterpret_cast<qint8_t *>(output.ptr()), vqmlaq_qs8(beta_vec, x_bar, gamma_vec, fixed_point_position)); - }, - input, output); -} - -template <bool fused_activation> -void NEBatchNormalizationLayerKernel::batch_normalization_qs16(const Window &window) -{ - static_assert(!fused_activation, "Activation is not supported for QS16"); - - Iterator input(_input, window); - Iterator output(_output, window); - - // Hold information about the current feature map we are iterating. - // Only compute denominator and NEON vectors once per feature map. - int slice = -1; - - const int fixed_point_position = _input->info()->fixed_point_position(); - const auto input_mean = reinterpret_cast<const qint16_t *>(_mean->ptr_to_element(Coordinates(0, 0))); - const auto input_var = reinterpret_cast<const qint16_t *>(_var->ptr_to_element(Coordinates(0, 0))); - const auto input_gamma = (_gamma != nullptr) ? reinterpret_cast<const qint16_t *>(_gamma->ptr_to_element(Coordinates(0, 0))) : nullptr; - const auto input_beta = (_beta != nullptr) ? reinterpret_cast<const qint16_t *>(_beta->ptr_to_element(Coordinates(0, 0))) : nullptr; - - qint16x8_t mean_vec = vdupq_n_qs16(0); - qint16x8_t var_vec = vdupq_n_qs16(0); - qint16x8_t gamma_vec = vdupq_n_qs16(sqcvt_qs16_f32(1, fixed_point_position)); - qint16x8_t beta_vec = vdupq_n_qs16(sqcvt_qs16_f32(0, fixed_point_position)); - qint16x8_t denominator = vdupq_n_qs16(0); - const qint16x8_t epsilon_vec = vdupq_n_qs16(sqcvt_qs16_f32(_epsilon, fixed_point_position)); - execute_window_loop(window, [&](const Coordinates & id) - { - if(slice != id.z()) - { - // Conctruct vectors - mean_vec = vdupq_n_qs16(*(input_mean + id.z())); - var_vec = vdupq_n_qs16(*(input_var + id.z())); - if(input_gamma != nullptr) - { - gamma_vec = vdupq_n_qs16(*(input_gamma + id.z())); - } - if(input_beta != nullptr) - { - beta_vec = vdupq_n_qs16(*(input_beta + id.z())); - } - - // Calculate denominator - denominator = vqinvsqrtq_qs16(vqaddq_qs16(var_vec, epsilon_vec), fixed_point_position); - slice = id.z(); - } - - // Calculate x bar and store results - const qint16x8_t numerator = vqsubq_qs16(vld1q_qs16(reinterpret_cast<const qint16_t *>(input.ptr())), mean_vec); - const qint16x8_t x_bar = vqmulq_qs16(numerator, denominator, fixed_point_position); - vst1q_qs16(reinterpret_cast<qint16_t *>(output.ptr()), vqmlaq_qs16(beta_vec, x_bar, gamma_vec, fixed_point_position)); - }, - input, output); -} - -template <bool fused_activation> void NEBatchNormalizationLayerKernel::batch_normalization_fp16_nchw(const Window &window) { static_assert(!fused_activation, "Activation is not supported for FP16"); @@ -406,12 +296,6 @@ void NEBatchNormalizationLayerKernel::configure_non_fused() const bool is_nhwc = _input->info()->data_layout() == DataLayout::NHWC; switch(_input->info()->data_type()) { - case DataType::QS8: - _func = &NEBatchNormalizationLayerKernel::batch_normalization_qs8<false>; - break; - case DataType::QS16: - _func = &NEBatchNormalizationLayerKernel::batch_normalization_qs16<false>; - break; case DataType::F16: _func = (is_nhwc) ? &NEBatchNormalizationLayerKernel::batch_normalization_fp16_nhwc<false> : &NEBatchNormalizationLayerKernel::batch_normalization_fp16_nchw<false>; break; diff --git a/src/core/NEON/kernels/NECol2ImKernel.cpp b/src/core/NEON/kernels/NECol2ImKernel.cpp index 9fda65feb4..d09d174e4f 100644 --- a/src/core/NEON/kernels/NECol2ImKernel.cpp +++ b/src/core/NEON/kernels/NECol2ImKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -50,8 +50,8 @@ TensorShape get_output_shape(const ITensorInfo *input, const Size2D &convolved_d Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const Size2D &convolved_dims) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::QASYMM8, - DataType::U16, DataType::S16, DataType::QS16, + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8, + DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32); @@ -60,7 +60,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), get_output_shape(input, convolved_dims)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; diff --git a/src/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.cpp b/src/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.cpp index b3746bddf2..e581f221a3 100644 --- a/src/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.cpp +++ b/src/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.cpp @@ -65,7 +65,7 @@ void NEConvertFullyConnectedWeightsKernel::configure(const ITensor *input, ITens Status NEConvertFullyConnectedWeightsKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::QASYMM8, DataType::U16, DataType::S16, DataType::QS16, DataType::U32, DataType::S32, + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8, DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::QS32, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); diff --git a/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp b/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp index 891a03c5cc..38443ca4a8 100644 --- a/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp +++ b/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -41,10 +41,6 @@ using namespace arm_compute; namespace { // Overloads of 128-bit vector loads -uint8x16_t loadq(const uint8_t *ptr) -{ - return vld1q_u8(ptr); -} uint16x8_t loadq(const uint16_t *ptr) { return vld1q_u16(ptr); @@ -54,10 +50,6 @@ uint32x4_t loadq(const uint32_t *ptr) return vld1q_u32(ptr); } // Overloads of 128-bit vector stores -void storeq(uint8_t *ptr, uint8x16_t val) -{ - return vst1q_u8(ptr, val); -} void storeq(uint16_t *ptr, uint16x8_t val) { return vst1q_u16(ptr, val); @@ -107,9 +99,8 @@ BorderSize NEDepthConcatenateLayerKernel::border_size() const void NEDepthConcatenateLayerKernel::configure(const ITensor *input, unsigned int depth_offset, ITensor *output) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, output); ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) + depth_offset > output->info()->dimension(2)); ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) > output->info()->dimension(0)); ARM_COMPUTE_ERROR_ON(input->info()->dimension(1) > output->info()->dimension(1)); @@ -129,10 +120,6 @@ void NEDepthConcatenateLayerKernel::configure(const ITensor *input, unsigned int switch(input->info()->data_type()) { - case DataType::QS8: - _func = &depth_concat<uint8_t>; - break; - case DataType::QS16: case DataType::F16: _func = &depth_concat<uint16_t>; break; diff --git a/src/core/NEON/kernels/NEDepthConvertLayerKernel.cpp b/src/core/NEON/kernels/NEDepthConvertLayerKernel.cpp index c29cb57513..8280b52fcb 100644 --- a/src/core/NEON/kernels/NEDepthConvertLayerKernel.cpp +++ b/src/core/NEON/kernels/NEDepthConvertLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016, 2017 ARM Limited. + * Copyright (c) 2016-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -40,13 +40,13 @@ class Coordinates; } // namespace arm_compute NEDepthConvertLayerKernel::NEDepthConvertLayerKernel() - : _input(nullptr), _output(nullptr), _policy(), _shift(0), _fixed_point_position_input(0), _fixed_point_position_output(0) + : _input(nullptr), _output(nullptr), _policy(), _shift(0) { } void NEDepthConvertLayerKernel::configure(ITensor *input, ITensor *output, ConvertPolicy policy, uint32_t shift) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::QS8, DataType::S16, DataType::U16, DataType::QS16, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S16, DataType::U16); _input = input; _output = input; @@ -58,48 +58,26 @@ void NEDepthConvertLayerKernel::configure(ITensor *input, ITensor *output, Conve // Auto initialize output shape if not initialized (We can only auto-configure the shape, datatype must be given) set_shape_if_empty(*output->info(), input->info()->tensor_shape()); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::QS8, DataType::S16, DataType::U16, DataType::QS16, DataType::U32, DataType::S32, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16, DataType::U16, DataType::U32, DataType::S32, DataType::F32); ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); // Set output _output = output; } - // Set initial fixed point position of input and output - _fixed_point_position_input = input->info()->fixed_point_position(); - _fixed_point_position_output = _output->info()->fixed_point_position(); - - // Set the fixed point position to the output tensor if needed - if(is_data_type_fixed_point(input->info()->data_type()) && is_data_type_fixed_point(_output->info()->data_type())) - { - // If in-place set the fixed point position of the output tensor to be equal to shift - _fixed_point_position_output = (_input == _output) ? static_cast<int>(_shift) : _fixed_point_position_output; - // Set fixed point position to output tensor - _output->info()->set_fixed_point_position(_fixed_point_position_output); - } - - ARM_COMPUTE_ERROR_ON(shift >= 8 && (!is_data_type_fixed_point(input->info()->data_type()) && !is_data_type_fixed_point(output->info()->data_type()))); + ARM_COMPUTE_ERROR_ON(shift >= 8); ARM_COMPUTE_ERROR_ON(input == output && (data_size_from_type(input->info()->data_type()) != data_size_from_type(output->info()->data_type()))); ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::U8 && (output->info()->data_type() != DataType::S16 && output->info()->data_type() != DataType::U16 && output->info()->data_type() != DataType::S32), "Only data_types supported [in] U8 -> [out] U16, S16, S32"); - ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::QS8 && (output->info()->data_type() != DataType::QS8 && output->info()->data_type() != DataType::F32), - "Only data_types supported [in] QS8 -> [out] QS8, F32"); - ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::U16 && (output->info()->data_type() != DataType::U8 && output->info()->data_type() != DataType::U32), "Only data_types supported [in] U16 -> [out] U8, U32"); ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::S16 && (output->info()->data_type() != DataType::U8 && output->info()->data_type() != DataType::S32), "Only data_types supported [in] S16 -> [out] U8, S32"); - ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::QS16 && (output->info()->data_type() != DataType::QS16 && output->info()->data_type() != DataType::F32), - "Only data_types supported [in] QS16 -> [out] QS16, F32"); - - ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::F32 && (output->info()->data_type() != DataType::QS8 && output->info()->data_type() != DataType::QS16), - "Only data_types supported [in] F32 -> [out] QS8, QS16"); - constexpr unsigned int num_elems_processed_per_iteration = 16; // Configure kernel window @@ -132,8 +110,6 @@ void NEDepthConvertLayerKernel::run(const Window &window, const ThreadInfo &info Iterator input(_input, window); Iterator output(_output, window); - bool in_place = (_input == _output); - switch(_input->info()->data_type()) { case DataType::U8: @@ -212,49 +188,6 @@ void NEDepthConvertLayerKernel::run(const Window &window, const ThreadInfo &info } break; } - case DataType::QS8: - { - switch(_output->info()->data_type()) - { - case DataType::QS8: - { - const int relative_shift = _fixed_point_position_output - _fixed_point_position_input; - /* Fixed point position conversion QS8 -> QS8 */ - if(relative_shift != 0 || !in_place) - { - const auto relative_shift_vec = vdupq_n_qs8(relative_shift); - execute_window_loop(window, [&](const Coordinates & id) - { - const qint8x16_t texels_qs8 = vld1q_qs8(reinterpret_cast<const qint8_t *>(input.ptr())); - vst1q_qs8(reinterpret_cast<qint8_t *>(output.ptr()), vqrshlq_s8(texels_qs8, relative_shift_vec)); - }, - input, output); - } - break; - } - case DataType::F32: - { - /* Up-conversion QS8 -> F32 */ - execute_window_loop(window, [&](const Coordinates & id) - { - const qint8x16_t texels_qs8 = vld1q_qs8(reinterpret_cast<const qint8_t *>(input.ptr())); - - float32x4x2_t texels_low = vcvt_f32_qs8(vget_low_s8(texels_qs8), _fixed_point_position_input); - float32x4x2_t texels_high = vcvt_f32_qs8(vget_high_s8(texels_qs8), _fixed_point_position_input); - - vst1q_f32(reinterpret_cast<float *>(output.ptr()), texels_low.val[0]); - vst1q_f32(reinterpret_cast<float *>(output.ptr()) + 4, texels_low.val[1]); - vst1q_f32(reinterpret_cast<float *>(output.ptr()) + 8, texels_high.val[0]); - vst1q_f32(reinterpret_cast<float *>(output.ptr()) + 12, texels_high.val[1]); - }, - input, output); - break; - } - default: - ARM_COMPUTE_ERROR("Output data type not supported"); - } - break; - } case DataType::S16: { switch(_output->info()->data_type()) @@ -408,116 +341,6 @@ void NEDepthConvertLayerKernel::run(const Window &window, const ThreadInfo &info } break; } - case DataType::QS16: - { - switch(_output->info()->data_type()) - { - case DataType::QS16: - { - const int relative_shift = _fixed_point_position_output - _fixed_point_position_input; - /* Fixed point position conversion QS16 -> QS16 */ - if(relative_shift != 0 || !in_place) - { - const auto relative_shift_vec = vdupq_n_qs16(relative_shift); - execute_window_loop(window, [&](const Coordinates & id) - { - const qint16x8x2_t texels_qs16 = - { - { - vld1q_qs16(reinterpret_cast<qint16_t *>(input.ptr())), - vld1q_qs16(reinterpret_cast<qint16_t *>(input.ptr()) + 8) - } - }; - vst1q_qs16(reinterpret_cast<qint16_t *>(output.ptr()), vqrshlq_s16(texels_qs16.val[0], relative_shift_vec)); - vst1q_qs16(reinterpret_cast<qint16_t *>(output.ptr()) + 8, vqrshlq_s16(texels_qs16.val[1], relative_shift_vec)); - }, - input, output); - } - break; - } - case DataType::F32: - { - /* Up-conversion QS16 -> F32 */ - execute_window_loop(window, [&](const Coordinates & id) - { - const int16x8x2_t texels_qs16 = - { - { - vld1q_s16(reinterpret_cast<qint16_t *>(input.ptr())), - vld1q_s16(reinterpret_cast<qint16_t *>(input.ptr()) + 8) - } - }; - - vst1q_f32(reinterpret_cast<float *>(output.ptr()), vcvt_f32_qs16(vget_low_s16(texels_qs16.val[0]), _fixed_point_position_input)); - vst1q_f32(reinterpret_cast<float *>(output.ptr()) + 4, vcvt_f32_qs16(vget_high_s16(texels_qs16.val[0]), _fixed_point_position_input)); - vst1q_f32(reinterpret_cast<float *>(output.ptr()) + 8, vcvt_f32_qs16(vget_low_s16(texels_qs16.val[1]), _fixed_point_position_input)); - vst1q_f32(reinterpret_cast<float *>(output.ptr()) + 12, vcvt_f32_qs16(vget_high_s16(texels_qs16.val[1]), _fixed_point_position_input)); - }, - input, output); - break; - } - default: - ARM_COMPUTE_ERROR("Output data type not supported"); - } - break; - } - case DataType::F32: - { - switch(_output->info()->data_type()) - { - case DataType::QS8: - { - /* Down-conversion F32 -> QS8 */ - execute_window_loop(window, [&](const Coordinates & id) - { - const float32x4x4_t texels_f32 = - { - { - vld1q_f32(reinterpret_cast<const float *>(input.ptr())), - vld1q_f32(reinterpret_cast<const float *>(input.ptr()) + 4), - vld1q_f32(reinterpret_cast<const float *>(input.ptr()) + 8), - vld1q_f32(reinterpret_cast<const float *>(input.ptr()) + 12) - } - }; - - const qint8x16_t texels_s8 = vqcvtq_qs8_f32(texels_f32, _fixed_point_position_output); - - vst1q_s8(reinterpret_cast<int8_t *>(output.ptr()), texels_s8); - }, - input, output); - break; - } - case DataType::QS16: - { - /* Down-conversion F32 -> QS16 */ - execute_window_loop(window, [&](const Coordinates & id) - { - const float32x4x2_t texels_f32_1 = - { - { - vld1q_f32(reinterpret_cast<const float *>(input.ptr())), - vld1q_f32(reinterpret_cast<const float *>(input.ptr()) + 4), - } - }; - const float32x4x2_t texels_f32_2 = - { - { - vld1q_f32(reinterpret_cast<const float *>(input.ptr()) + 8), - vld1q_f32(reinterpret_cast<const float *>(input.ptr()) + 12) - } - }; - - vst1q_s16(reinterpret_cast<qint16_t *>(output.ptr()), vqcvtq_qs16_f32(texels_f32_1, _fixed_point_position_output)); - vst1q_s16(reinterpret_cast<qint16_t *>(output.ptr()) + 8, vqcvtq_qs16_f32(texels_f32_2, _fixed_point_position_output)); - }, - input, output); - break; - } - default: - ARM_COMPUTE_ERROR("Output data type not supported"); - } - break; - } default: ARM_COMPUTE_ERROR("Not supported"); } diff --git a/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp b/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp index 8cdf175d8a..09728e2a8d 100644 --- a/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp +++ b/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp @@ -115,7 +115,7 @@ public: in_top += delta_input, in_mid += delta_input, in_low += delta_input, p_out += num_elems_written_per_iteration) { - auto vres = convolve_3x3<stridex>(in_top, in_mid, in_low, vw_r0, vw_r1, vw_r2, 0, input_offset); + auto vres = convolve_3x3<stridex>(in_top, in_mid, in_low, vw_r0, vw_r1, vw_r2, input_offset); store_results<stridex>(p_out, vres); } } diff --git a/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp b/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp index cfd8eacfdd..5b43e2b14f 100644 --- a/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp +++ b/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp @@ -122,7 +122,6 @@ void NEDepthwiseIm2ColKernel::configure(const ITensor *input, ITensor *output, c { ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); ARM_COMPUTE_ERROR_ON(is_data_type_quantized_asymmetric(input->info()->data_type()) && has_bias); ARM_COMPUTE_ERROR_ON((input->info()->dimension(2) * depth_multiplier) != output->info()->dimension(2)); ARM_COMPUTE_ERROR_ON(output->info()->dimension(0) != (kernel_dims.width * kernel_dims.height + ((has_bias) ? 1 : 0))); diff --git a/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp b/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp index 8960d8a8af..86a6d1c1a8 100644 --- a/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp +++ b/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp @@ -89,7 +89,6 @@ void NEDepthwiseVectorToTensorKernel::configure(const ITensor *input, ITensor *o ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); _input = input; _output = output; diff --git a/src/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.cpp b/src/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.cpp index 36b17bfc4c..47fcf12874 100644 --- a/src/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.cpp +++ b/src/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.cpp @@ -88,7 +88,6 @@ void NEDepthwiseWeightsReshapeKernel::configure(const ITensor *input, ITensor *o { ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); ARM_COMPUTE_ERROR_ON(is_data_type_quantized_asymmetric(input->info()->data_type()) && (biases != nullptr)); ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != output->info()->dimension(1)); ARM_COMPUTE_ERROR_ON(output->info()->dimension(0) != (input->info()->dimension(0) * input->info()->dimension(1) + ((biases != nullptr) ? 1 : 0))); @@ -96,7 +95,6 @@ void NEDepthwiseWeightsReshapeKernel::configure(const ITensor *input, ITensor *o if(biases != nullptr) { ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases); ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != input->info()->dimension(2)); ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1); } diff --git a/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp b/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp index 4120e5f87a..47c895c594 100644 --- a/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp +++ b/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp @@ -54,7 +54,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, ITensorInfo *min_max) { // Output tensor auto initialization if not yet initialized - auto_init_if_empty(*output, input->tensor_shape(), 1, DataType::F32, 0); + auto_init_if_empty(*output, input->tensor_shape(), 1, DataType::F32); constexpr unsigned int num_elems_processed_per_iteration = 8; diff --git a/src/core/NEON/kernels/NEDirectConvolutionLayerKernel.cpp b/src/core/NEON/kernels/NEDirectConvolutionLayerKernel.cpp index 5eafdf0363..54a046846a 100644 --- a/src/core/NEON/kernels/NEDirectConvolutionLayerKernel.cpp +++ b/src/core/NEON/kernels/NEDirectConvolutionLayerKernel.cpp @@ -43,34 +43,6 @@ using namespace arm_compute::detail; namespace { -template <unsigned int stridex> -qint16x8_t internal_vld1q(const qint16_t *in); - -template <> -qint16x8_t internal_vld1q<1>(const qint16_t *in) -{ - return vld1q_qs16(in); -} - -template <> -qint16x8_t internal_vld1q<2>(const qint16_t *in) -{ - const int16x8x2_t tmp = vld2q_s16(in); - return tmp.val[0]; -} - -template <> -qint16x8_t internal_vld1q<3>(const qint16_t *in) -{ - const int16x8x3_t tmp = vld3q_s16(in); - return tmp.val[0]; -} - -inline qint16x8_t internal_vdupq_n(qint16_t v) -{ - return vdupq_n_qs16(v); -} - #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC template <unsigned int stridex> float16x8_t internal_vld1q(const float16_t *in); @@ -105,15 +77,13 @@ inline void internal_vst1q(float16_t *p, const float16x8_t &v) vst1q_f16(p, v); } -float16x8_t internal_vmull(const float16x8_t &x, const float16x8_t &y, int fixed_point_position) +float16x8_t internal_vmull(const float16x8_t &x, const float16x8_t &y) { - ARM_COMPUTE_UNUSED(fixed_point_position); return vmulq_f16(x, y); } -inline float16x8_t internal_vmlal(const float16x8_t &x, const float16x8_t &y, const float16x8_t &z, int fixed_point_position) +inline float16x8_t internal_vmlal(const float16x8_t &x, const float16x8_t &y, const float16x8_t &z) { - ARM_COMPUTE_UNUSED(fixed_point_position); return vaddq_f16(x, vmulq_f16(y, z)); } #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ @@ -151,107 +121,16 @@ inline void internal_vst1q(float *p, const float32x4_t &v) vst1q_f32(p, v); } -float32x4_t internal_vmull(const float32x4_t &x, const float32x4_t &y, int fixed_point_position) +float32x4_t internal_vmull(const float32x4_t &x, const float32x4_t &y) { - ARM_COMPUTE_UNUSED(fixed_point_position); return vmulq_f32(x, y); } -inline float32x4_t internal_vmlal(const float32x4_t &x, const float32x4_t &y, const float32x4_t &z, int fixed_point_position) +inline float32x4_t internal_vmlal(const float32x4_t &x, const float32x4_t &y, const float32x4_t &z) { - ARM_COMPUTE_UNUSED(fixed_point_position); return vmlaq_f32(x, y, z); } -template <unsigned int stridex> -qint8x8_t internal_vld1q(const qint8_t *in); - -template <> -qint8x8_t internal_vld1q<1>(const qint8_t *in) -{ - return vld1_qs8(in); -} - -template <> -qint8x8_t internal_vld1q<2>(const qint8_t *in) -{ - const qint8x8x2_t tmp = vld2_s8(in); - return tmp.val[0]; -} - -template <> -qint8x8_t internal_vld1q<3>(const qint8_t *in) -{ - const qint8x8x3_t tmp = vld3_s8(in); - return tmp.val[0]; -} - -inline qint8x8_t internal_vdupq_n(qint8_t v) -{ - return vdup_n_qs8(v); -} - -inline qint16x8_t internal_vmull(const qint8x8_t &x, const qint8x8_t &y, int fixed_point_position) -{ - return vmull_qs8(x, y, fixed_point_position); -} - -inline qint16x8_t internal_vmlal(const qint16x8_t &x, const qint8x8_t &y, const qint8x8_t &z, int fixed_point_position) -{ - return vqmlal_qs8(x, y, z, fixed_point_position); -} - -inline void internal_vst1q(qint16_t *p, const qint16x8_t &v) -{ - vst1q_qs16(p, v); -} - -inline void internal_vst1q(int32_t *p, const qint32x4x2_t &v) -{ - vst1q_s32(p, v.val[0]); - vst1q_s32(p + 4, v.val[1]); -} - -template <unsigned int stridex> -qint32x4x2_t internal_vld1q(const qint32_t *in); - -template <> -qint32x4x2_t internal_vld1q<1>(const qint32_t *in) -{ - const qint32x4x2_t r = - { - { - vld1q_s32(in), - vld1q_s32(in + 4) - } - }; - return r; -} - -inline qint32x4x2_t internal_vmull(const qint16x8_t &x, const qint16x8_t &y, int fixed_point_position) -{ - const qint32x4x2_t r = - { - { - vmull_qs16(vget_low_s16(x), vget_low_s16(y), fixed_point_position), - vmull_qs16(vget_high_s16(x), vget_high_s16(y), fixed_point_position), - } - }; - return r; -} - -inline qint32x4x2_t internal_vmlal(const qint32x4x2_t &x, const qint16x8_t &y, const qint16x8_t &z, int fixed_point_position) -{ - const qint32x4x2_t r = - { - { - vqmlal_qs16(x.val[0], vget_low_s16(y), vget_low_s16(z), fixed_point_position), - vqmlal_qs16(x.val[1], vget_high_s16(y), vget_high_s16(z), fixed_point_position) - } - }; - return r; -} - constexpr int small_tensor_size_optim = 8; inline bool run_optim_small_tensor_info(const ITensorInfo *t) { @@ -355,21 +234,20 @@ public: static void convolve(const Window &window, unsigned int num_elems_read_per_iteration, unsigned int num_elems_written_per_iteration, const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info) { - const int input_stride_x = input->info()->strides_in_bytes().x(); - const int input_stride_y = input->info()->strides_in_bytes().y(); - const int input_stride_z = input->info()->strides_in_bytes().z(); - const int output_stride_y = output->info()->strides_in_bytes().y(); - const int output_stride_z = output->info()->strides_in_bytes().z(); - const int kernel_stride_z = weights->info()->strides_in_bytes().z(); - const int kernel_stride_w = weights->info()->strides_in_bytes()[3]; - const int output_w = output->info()->dimension(0); - const int output_h = output->info()->dimension(1); - const int range_z = window.z().end() - window.z().start(); - const int kernel_depth = weights->info()->dimension(Window::DimZ); - const unsigned int conv_stride_y = std::get<1>(conv_info.stride()); - const unsigned int conv_pad_left = conv_info.pad_left(); - const unsigned int conv_pad_top = conv_info.pad_top(); - const int fixed_point_position = input->info()->fixed_point_position(); + const int input_stride_x = input->info()->strides_in_bytes().x(); + const int input_stride_y = input->info()->strides_in_bytes().y(); + const int input_stride_z = input->info()->strides_in_bytes().z(); + const int output_stride_y = output->info()->strides_in_bytes().y(); + const int output_stride_z = output->info()->strides_in_bytes().z(); + const int kernel_stride_z = weights->info()->strides_in_bytes().z(); + const int kernel_stride_w = weights->info()->strides_in_bytes()[3]; + const int output_w = output->info()->dimension(0); + const int output_h = output->info()->dimension(1); + const int range_z = window.z().end() - window.z().start(); + const int kernel_depth = weights->info()->dimension(Window::DimZ); + const unsigned int conv_stride_y = std::get<1>(conv_info.stride()); + const unsigned int conv_pad_left = conv_info.pad_left(); + const unsigned int conv_pad_top = conv_info.pad_top(); // setup output window for the iterator Window window_out = window; @@ -414,7 +292,7 @@ public: auto p_out = reinterpret_cast<T2 *>(p_out_base + oh * output_stride_y); for(int ow = 0; ow < output_w; ow += num_elems_written_per_iteration, in_val += num_elems_read_per_iteration, p_out += num_elems_written_per_iteration) { - internal_vst1q(p_out, internal_vmull(vk, internal_vld1q<stridex>(in_val), fixed_point_position)); + internal_vst1q(p_out, internal_vmull(vk, internal_vld1q<stridex>(in_val))); } } } @@ -431,7 +309,7 @@ public: auto p_out = reinterpret_cast<T2 *>(p_out_base + oh * output_stride_y); for(int ow = 0; ow < output_w; ow += num_elems_written_per_iteration, in_val += num_elems_read_per_iteration, p_out += num_elems_written_per_iteration) { - internal_vst1q(p_out, internal_vmlal(internal_vld1q<1>(p_out), vk, internal_vld1q<stridex>(in_val), fixed_point_position)); + internal_vst1q(p_out, internal_vmlal(internal_vld1q<1>(p_out), vk, internal_vld1q<stridex>(in_val))); } } } @@ -469,7 +347,7 @@ void accumulate_results<3>(float16_t *buffer, const float16x8x2_t &values) template <unsigned int stridex> float32x4x2_t convolve_5x5(const float *in_0, const float *in_1, const float *in_2, const float *in_3, const float *in_4, - const float *m0, const float *m1, const float *m2, const float *m3, const float *m4, int fixed_point_position); + const float *m0, const float *m1, const float *m2, const float *m3, const float *m4); inline float32x4x3_t load_matrix_hi(const float *const m0, const float *const m1, const float *const m2) { @@ -511,9 +389,8 @@ inline float32x4x3_t load_input(const float *const in) template <> inline float32x4x2_t convolve_5x5<1>(const float *in_0, const float *in_1, const float *in_2, const float *in_3, const float *in_4, - const float *m0, const float *m1, const float *m2, const float *m3, const float *m4, int fixed_point_position) + const float *m0, const float *m1, const float *m2, const float *m3, const float *m4) { - ARM_COMPUTE_UNUSED(fixed_point_position); const float32x4x3_t vin0 = load_input(in_0); const float32x4x3_t vin1 = load_input(in_1); const float32x4x3_t vin2 = load_input(in_2); @@ -601,10 +478,9 @@ inline float32x4x2_t convolve_5x5<1>(const float *in_0, const float *in_1, const template <> inline float32x4x2_t convolve_5x5<2>(const float *in_0, const float *in_1, const float *in_2, const float *in_3, const float *in_4, - const float *m0, const float *m1, const float *m2, const float *m3, const float *m4, int fixed_point_position) + const float *m0, const float *m1, const float *m2, const float *m3, const float *m4) { - ARM_COMPUTE_UNUSED(fixed_point_position); - float32x4x2_t out = convolve_5x5<1>(in_0, in_1, in_2, in_3, in_4, m0, m1, m2, m3, m4, fixed_point_position); + float32x4x2_t out = convolve_5x5<1>(in_0, in_1, in_2, in_3, in_4, m0, m1, m2, m3, m4); out.val[0] = vsetq_lane_f32(vgetq_lane_f32(out.val[0], 2), out.val[0], 1); out.val[0] = vsetq_lane_f32(vgetq_lane_f32(out.val[1], 0), out.val[0], 2); out.val[0] = vsetq_lane_f32(vgetq_lane_f32(out.val[1], 2), out.val[0], 3); @@ -613,9 +489,9 @@ inline float32x4x2_t convolve_5x5<2>(const float *in_0, const float *in_1, const template <> inline float32x4x2_t convolve_5x5<3>(const float *in_0, const float *in_1, const float *in_2, const float *in_3, const float *in_4, - const float *m0, const float *m1, const float *m2, const float *m3, const float *m4, int fixed_point_position) + const float *m0, const float *m1, const float *m2, const float *m3, const float *m4) { - float32x4x2_t out = convolve_5x5<1>(in_0, in_1, in_2, in_3, in_4, m0, m1, m2, m3, m4, fixed_point_position); + float32x4x2_t out = convolve_5x5<1>(in_0, in_1, in_2, in_3, in_4, m0, m1, m2, m3, m4); out.val[0] = vsetq_lane_f32(vgetq_lane_f32(out.val[0], 3), out.val[0], 1); return out; } @@ -642,28 +518,6 @@ void accumulate_results<3>(float *buffer, const float32x4x2_t &values) vst1_f32(buffer, vadd_f32(vld1_f32(buffer), vget_low_f32(values.val[0]))); } -template <unsigned int stridex> -void accumulate_results(qint16_t *buffer, const qint16x8x2_t &values); - -template <> -void accumulate_results<1>(qint16_t *buffer, const qint16x8x2_t &values) -{ - vst1q_qs16(buffer, vqaddq_qs16(vld1q_qs16(buffer), values.val[0])); - vst1q_qs16(buffer + 8, vqaddq_qs16(vld1q_qs16(buffer + 8), values.val[1])); -} - -template <> -void accumulate_results<2>(qint16_t *buffer, const qint16x8x2_t &values) -{ - vst1q_qs16(buffer, vqaddq_qs16(vld1q_qs16(buffer), values.val[0])); -} - -template <> -void accumulate_results<3>(qint16_t *buffer, const qint16x8x2_t &values) -{ - vst1_qs16(buffer, vqadd_qs16(vld1_qs16(buffer), vget_low_s16(values.val[0]))); -} - template <typename T1> class convolver_nhwc { @@ -745,7 +599,7 @@ public: const auto we_addr = reinterpret_cast<const T1 *>(we_addr_base1 + x * kernel_stride_x); const auto we_values = internal_vld1q<1>(we_addr); - out_values = internal_vmlal(out_values, in_values, we_values, 0); + out_values = internal_vmlal(out_values, in_values, we_values); } out_val += out_values[0]; @@ -784,24 +638,23 @@ public: const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info) { ARM_COMPUTE_UNUSED(num_elems_read_per_iteration); - const int input_stride_x = input->info()->strides_in_bytes().x(); - const int input_stride_y = input->info()->strides_in_bytes().y(); - const int input_stride_z = input->info()->strides_in_bytes().z(); - const int output_stride_y = output->info()->strides_in_bytes().y(); - const int output_stride_z = output->info()->strides_in_bytes().z(); - const int kernel_stride_x = weights->info()->strides_in_bytes().x(); - const int kernel_stride_y = weights->info()->strides_in_bytes().y(); - const int kernel_stride_z = weights->info()->strides_in_bytes().z(); - const int kernel_stride_w = weights->info()->strides_in_bytes()[3]; - const int output_w = output->info()->dimension(0); - const int output_h = output->info()->dimension(1); - const int num_planes_z = window.z().end() - window.z().start(); - const int delta_input = get_input_num_elems_processed<stridex>(num_elems_written_per_iteration); - const int kernel_depth = weights->info()->dimension(Window::DimZ); - const unsigned int conv_stride_y = std::get<1>(conv_info.stride()); - const unsigned int conv_pad_left = conv_info.pad_left(); - const unsigned int conv_pad_top = conv_info.pad_top(); - const int fixed_point_position = input->info()->fixed_point_position(); + const int input_stride_x = input->info()->strides_in_bytes().x(); + const int input_stride_y = input->info()->strides_in_bytes().y(); + const int input_stride_z = input->info()->strides_in_bytes().z(); + const int output_stride_y = output->info()->strides_in_bytes().y(); + const int output_stride_z = output->info()->strides_in_bytes().z(); + const int kernel_stride_x = weights->info()->strides_in_bytes().x(); + const int kernel_stride_y = weights->info()->strides_in_bytes().y(); + const int kernel_stride_z = weights->info()->strides_in_bytes().z(); + const int kernel_stride_w = weights->info()->strides_in_bytes()[3]; + const int output_w = output->info()->dimension(0); + const int output_h = output->info()->dimension(1); + const int num_planes_z = window.z().end() - window.z().start(); + const int delta_input = get_input_num_elems_processed<stridex>(num_elems_written_per_iteration); + const int kernel_depth = weights->info()->dimension(Window::DimZ); + const unsigned int conv_stride_y = std::get<1>(conv_info.stride()); + const unsigned int conv_pad_left = conv_info.pad_left(); + const unsigned int conv_pad_top = conv_info.pad_top(); // setup output window for the iterator Window window_out = window; @@ -864,7 +717,7 @@ public: for(int ow = 0; ow < output_w; ow += num_elems_written_per_iteration, in_top += delta_input, in_mid += delta_input, in_low += delta_input, p_out += num_elems_written_per_iteration) { - auto vres = convolve_3x3<stridex>(in_top, in_mid, in_low, vk_r0, vk_r1, vk_r2, fixed_point_position); + auto vres = convolve_3x3<stridex>(in_top, in_mid, in_low, vk_r0, vk_r1, vk_r2); store_results<stridex>(p_out, vres); } } @@ -889,7 +742,7 @@ public: for(int ow = 0; ow < output_w; ow += num_elems_written_per_iteration, in_top += delta_input, in_mid += delta_input, in_low += delta_input, p_out += num_elems_written_per_iteration) { - auto vres = convolve_3x3<stridex>(in_top, in_mid, in_low, vk_r0, vk_r1, vk_r2, fixed_point_position); + auto vres = convolve_3x3<stridex>(in_top, in_mid, in_low, vk_r0, vk_r1, vk_r2); accumulate_results<stridex>(p_out, vres); } } @@ -908,24 +761,23 @@ public: const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info) { ARM_COMPUTE_UNUSED(num_elems_read_per_iteration); - const int input_stride_x = input->info()->strides_in_bytes().x(); - const int input_stride_y = input->info()->strides_in_bytes().y(); - const int input_stride_z = input->info()->strides_in_bytes().z(); - const int output_stride_y = output->info()->strides_in_bytes().y(); - const int output_stride_z = output->info()->strides_in_bytes().z(); - const int kernel_stride_x = weights->info()->strides_in_bytes().x(); - const int kernel_stride_y = weights->info()->strides_in_bytes().y(); - const int kernel_stride_z = weights->info()->strides_in_bytes().z(); - const int kernel_stride_w = weights->info()->strides_in_bytes()[3]; - const int output_w = output->info()->dimension(0); - const int output_h = output->info()->dimension(1); - const int num_planes_z = window.z().end() - window.z().start(); - const int delta_input = get_input_num_elems_processed<stridex>(num_elems_written_per_iteration); - const int kernel_depth = weights->info()->dimension(Window::DimZ); - const unsigned int conv_stride_y = std::get<1>(conv_info.stride()); - const unsigned int conv_pad_left = conv_info.pad_left(); - const unsigned int conv_pad_top = conv_info.pad_top(); - const int fixed_point_position = input->info()->fixed_point_position(); + const int input_stride_x = input->info()->strides_in_bytes().x(); + const int input_stride_y = input->info()->strides_in_bytes().y(); + const int input_stride_z = input->info()->strides_in_bytes().z(); + const int output_stride_y = output->info()->strides_in_bytes().y(); + const int output_stride_z = output->info()->strides_in_bytes().z(); + const int kernel_stride_x = weights->info()->strides_in_bytes().x(); + const int kernel_stride_y = weights->info()->strides_in_bytes().y(); + const int kernel_stride_z = weights->info()->strides_in_bytes().z(); + const int kernel_stride_w = weights->info()->strides_in_bytes()[3]; + const int output_w = output->info()->dimension(0); + const int output_h = output->info()->dimension(1); + const int num_planes_z = window.z().end() - window.z().start(); + const int delta_input = get_input_num_elems_processed<stridex>(num_elems_written_per_iteration); + const int kernel_depth = weights->info()->dimension(Window::DimZ); + const unsigned int conv_stride_y = std::get<1>(conv_info.stride()); + const unsigned int conv_pad_left = conv_info.pad_left(); + const unsigned int conv_pad_top = conv_info.pad_top(); // setup output window for the iterator Window window_out = window; @@ -976,7 +828,7 @@ public: for(int ow = 0; ow < output_w; ow += num_elems_written_per_iteration, in_0 += delta_input, in_1 += delta_input, in_2 += delta_input, in_3 += delta_input, in_4 += delta_input, p_out += num_elems_written_per_iteration) { - auto vres = convolve_5x5<stridex>(in_0, in_1, in_2, in_3, in_4, ptr_k_r0, ptr_k_r1, ptr_k_r2, ptr_k_r3, ptr_k_r4, fixed_point_position); + auto vres = convolve_5x5<stridex>(in_0, in_1, in_2, in_3, in_4, ptr_k_r0, ptr_k_r1, ptr_k_r2, ptr_k_r3, ptr_k_r4); store_results<stridex>(p_out, vres); } } @@ -1001,7 +853,7 @@ public: for(int ow = 0; ow < output_w; ow += num_elems_written_per_iteration, in_0 += delta_input, in_1 += delta_input, in_2 += delta_input, in_3 += delta_input, in_4 += delta_input, p_out += num_elems_written_per_iteration) { - auto vres = convolve_5x5<stridex>(in_0, in_1, in_2, in_3, in_4, ptr_k_r0, ptr_k_r1, ptr_k_r2, ptr_k_r3, ptr_k_r4, fixed_point_position); + auto vres = convolve_5x5<stridex>(in_0, in_1, in_2, in_3, in_4, ptr_k_r0, ptr_k_r1, ptr_k_r2, ptr_k_r3, ptr_k_r4); accumulate_results<stridex>(p_out, vres); } } @@ -1120,7 +972,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::UNKNOWN); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); const DataLayout data_layout = input->data_layout(); @@ -1140,11 +992,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, TensorShape output_shape = misc::shape_calculator::compute_deep_convolution_shape(*input, *weights, conv_info); DataType data_type = input->data_type(); - if(is_data_type_fixed_point(data_type)) - { - // Promote data type in case of fixed point - data_type = ((data_type == DataType::QS8) ? DataType::QS16 : DataType::QS32); - } ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); ARM_COMPUTE_RETURN_ERROR_ON(output->data_type() != data_type); @@ -1180,11 +1027,9 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen { #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - case DataType::QS8: - case DataType::QS16: num_elems_written_per_iteration = 8; break; +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ case DataType::F32: if(run_optim_small_tensor_info(input)) { @@ -1215,13 +1060,11 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen break; #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - case DataType::QS8: - case DataType::QS16: num_weight_elems_read_per_row = 8 + kernel_size - 1; num_elems_read_per_iteration = 24; num_elems_written_per_iteration = 32 >> conv_stride_x; break; +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ default: ARM_COMPUTE_ERROR("Data type not supported."); break; @@ -1315,14 +1158,8 @@ void NEDirectConvolutionLayerKernel::configure(const ITensor *input, const ITens DataType data_type = input->info()->data_type(); - if(is_data_type_fixed_point(data_type)) - { - // Promote data type in case of fixed point - data_type = ((data_type == DataType::QS8) ? DataType::QS16 : DataType::QS32); - } - // Output auto inizialitation if not yet initialized - auto_init_if_empty(*output->info(), output_shape, 1, data_type, input->info()->fixed_point_position()); + auto_init_if_empty(*output->info(), output_shape, 1, data_type); // Perform validation step ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), output->info(), conv_info)); @@ -1371,12 +1208,6 @@ void NEDirectConvolutionLayerKernel::run(const Window &window, const ThreadInfo { switch(_input->info()->data_type()) { - case DataType::QS8: - convolve_1x1<qint8_t, qint16_t>(window, _num_elems_read_per_iteration, _num_elems_written_per_iteration, _input, _weights, _output, _conv_info); - break; - case DataType::QS16: - convolve_1x1<qint16_t, qint32_t>(window, _num_elems_read_per_iteration, _num_elems_written_per_iteration, _input, _weights, _output, _conv_info); - break; case DataType::F32: convolve_1x1<float, float>(window, _num_elems_read_per_iteration, _num_elems_written_per_iteration, _input, _weights, _output, _conv_info); break; @@ -1395,9 +1226,6 @@ void NEDirectConvolutionLayerKernel::run(const Window &window, const ThreadInfo { switch(_input->info()->data_type()) { - case DataType::QS8: - convolve_3x3<qint8_t, qint16_t>(window, _num_elems_read_per_iteration, _num_elems_written_per_iteration, _input, _weights, _output, _conv_info); - break; case DataType::F32: convolve_3x3<float, float>(window, _num_elems_read_per_iteration, _num_elems_written_per_iteration, _input, _weights, _output, _conv_info); break; diff --git a/src/core/NEON/kernels/NEDirectConvolutionLayerOutputStageKernel.cpp b/src/core/NEON/kernels/NEDirectConvolutionLayerOutputStageKernel.cpp index 37a3804289..e4cd4d0465 100644 --- a/src/core/NEON/kernels/NEDirectConvolutionLayerOutputStageKernel.cpp +++ b/src/core/NEON/kernels/NEDirectConvolutionLayerOutputStageKernel.cpp @@ -45,22 +45,15 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, con { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::UNKNOWN); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, - DataType::QS16, DataType::F16, + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, + DataType::F16, DataType::QS32, DataType::S32, DataType::F32); if(bias != nullptr) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::QS32, DataType::S32, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::F16, DataType::QS32, DataType::S32, DataType::F32); - if(is_data_type_fixed_point(input->data_type())) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::QS8 && bias->data_type() != DataType::QS8, "Wrong data type for bias"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::QS16 && bias->data_type() != DataType::QS8, "Wrong data type for bias"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::QS32 && bias->data_type() != DataType::QS16, "Wrong data type for bias"); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, bias); - } - else if(is_data_type_quantized_asymmetric(input->data_type())) + if(is_data_type_quantized_asymmetric(input->data_type())) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32); } @@ -80,17 +73,10 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, con // Checks performed when output is configured if((output != nullptr) && (output->total_size() != 0)) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); - if(is_data_type_fixed_point(input->data_type())) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::QS8 && output->data_type() != DataType::QS8, "Wrong data type for output"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::QS16 && output->data_type() != DataType::QS8, "Wrong data type for output"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::QS32 && output->data_type() != DataType::QS16, "Wrong data type for output"); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, output); - } - else if(is_data_type_quantized_asymmetric(output->data_type())) + if(is_data_type_quantized_asymmetric(output->data_type())) { ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::S32 && output->data_type() != DataType::QASYMM8, "Wrong data type for bias"); } @@ -168,81 +154,24 @@ inline float32x4_t internal_vld1q(const float *in) { return vld1q_f32(in); } -inline qint8x16_t internal_vld1q(const qint8_t *in) -{ - return vld1q_qs8(in); -} -inline qint16x8_t internal_vld1q(const qint16_t *in) -{ - return vld1q_qs16(in); -} -inline qint32x4_t internal_vld1q(const qint32_t *in) -{ - return vld1q_s32(in); -} // Internal store inline void internal_vst1q(float *p, const float32x4_t &v) { vst1q_f32(p, v); } -inline void internal_vst1q(qint8_t *p, const qint8x16_t &v) -{ - vst1q_qs8(p, v); -} -inline void internal_vst1q(qint8_t *p, const qint16x8_t &v) -{ - vst1_qs8(p, vqmovn_s16(v)); -} -inline void internal_vst1q(qint16_t *p, const qint16x8_t &v) -{ - vst1q_qs16(p, v); -} -inline void internal_vst1q(qint32_t *p, const qint32x4_t &v) -{ - vst1q_s32(p, v); -} - -inline void internal_vst1q(qint16_t *p, const qint32x4_t &v) -{ - vst1_qs16(p, vqmovn_qs32(v)); -} // Internal vdup inline float32x4_t internal_vdupq_n(float v) { return vdupq_n_f32(v); } -inline qint8x16_t internal_vdupq_n(qint8_t v) -{ - return vdupq_n_qs8(v); -} -inline qint16x8_t internal_vdupq_n(qint16_t v) -{ - return vdupq_n_qs16(v); -} -inline qint32x4_t internal_vdupq_n(qint32_t v) -{ - return vdupq_n_qs32(v); -} // Internal vadd inline float32x4_t internal_vqaddq(const float32x4_t &x, const float32x4_t &y) { return vaddq_f32(x, y); } -inline qint8x16_t internal_vqaddq(const qint8x16_t &x, const qint8x16_t &y) -{ - return vqaddq_qs8(x, y); -} -inline qint16x8_t internal_vqaddq(const qint16x8_t &x, const qint16x8_t &y) -{ - return vqaddq_qs16(x, y); -} -inline qint32x4_t internal_vqaddq(const qint32x4_t &x, const qint32x4_t &y) -{ - return vqaddq_qs32(x, y); -} #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC inline float16x8_t internal_vld1q(const float16_t *in) @@ -494,39 +423,6 @@ void NEDirectConvolutionLayerOutputStageKernel::configure(ITensor *input, const { switch(input->info()->data_type()) { - case DataType::QS8: - { - if(bias == nullptr) - { - _func = (output == nullptr) ? &output_stage<qint8_t, qint8_t, true, false> : &output_stage<qint8_t, qint8_t, false, false>; - } - else - { - _func = (output == nullptr) ? &output_stage<qint8_t, qint8_t, true, true> : &output_stage<qint8_t, qint8_t, false, true>; - } - break; - } - case DataType::QS16: - { - if(bias != nullptr && bias->info()->data_type() == DataType::QS8) - { - _func = (output == nullptr) ? &output_stage<qint16_t, qint8_t, true, true> : &output_stage<qint16_t, qint8_t, false, true>; - } - else if(bias == nullptr) - { - _func = (output == nullptr) ? &output_stage<qint16_t, qint8_t, true, false> : &output_stage<qint16_t, qint8_t, false, false>; - } - else - { - ARM_COMPUTE_ERROR("Not implemented"); - } - break; - } - case DataType::QS32: - { - _func = (output == nullptr) ? &output_stage<qint32_t, qint16_t, true, true> : &output_stage<qint32_t, qint16_t, false, true>; - break; - } case DataType::S32: { _func = (bias == nullptr) ? &output_stage<int32_t, uint8_t, false, false> : &output_stage<int32_t, uint8_t, false, true>; diff --git a/src/core/NEON/kernels/NEFillBorderKernel.cpp b/src/core/NEON/kernels/NEFillBorderKernel.cpp index 747b8b1bfe..3d08cafa93 100644 --- a/src/core/NEON/kernels/NEFillBorderKernel.cpp +++ b/src/core/NEON/kernels/NEFillBorderKernel.cpp @@ -105,8 +105,8 @@ NEFillBorderKernel::NEFillBorderKernel() void NEFillBorderKernel::configure(ITensor *tensor, BorderSize border_size, BorderMode border_mode, const PixelValue &constant_border_value) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(tensor, 1, DataType::U8, DataType::QS8, DataType::QASYMM8, - DataType::QS16, DataType::U16, DataType::S16, + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(tensor, 1, DataType::U8, DataType::QASYMM8, + DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32); @@ -147,7 +147,6 @@ void NEFillBorderKernel::run(const Window &window, const ThreadInfo &info) case DataType::U8: fill_constant_value_single_channel<uint8_t>(window); break; - case DataType::QS8: case DataType::S8: fill_constant_value_single_channel<int8_t>(window); break; @@ -155,7 +154,6 @@ void NEFillBorderKernel::run(const Window &window, const ThreadInfo &info) fill_constant_value_single_channel<uint16_t>(window); break; case DataType::S16: - case DataType::QS16: fill_constant_value_single_channel<int16_t>(window); break; case DataType::U32: @@ -192,7 +190,6 @@ void NEFillBorderKernel::run(const Window &window, const ThreadInfo &info) case DataType::U8: fill_replicate_single_channel<uint8_t>(window); break; - case DataType::QS8: case DataType::S8: fill_replicate_single_channel<int8_t>(window); break; @@ -200,7 +197,6 @@ void NEFillBorderKernel::run(const Window &window, const ThreadInfo &info) fill_replicate_single_channel<uint16_t>(window); break; case DataType::S16: - case DataType::QS16: fill_replicate_single_channel<int16_t>(window); break; case DataType::U32: diff --git a/src/core/NEON/kernels/NEFloorKernel.cpp b/src/core/NEON/kernels/NEFloorKernel.cpp index 72b652d5dc..872ac2661e 100644 --- a/src/core/NEON/kernels/NEFloorKernel.cpp +++ b/src/core/NEON/kernels/NEFloorKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -40,7 +40,7 @@ void NEFloorKernel::configure(const ITensor *input, ITensor *output) ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); // Auto initialize output - auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type(), input->info()->fixed_point_position()); + auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type()); ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); diff --git a/src/core/NEON/kernels/NEGEMMInterleave4x4Kernel.cpp b/src/core/NEON/kernels/NEGEMMInterleave4x4Kernel.cpp index 12755a45f8..6519a39b9c 100644 --- a/src/core/NEON/kernels/NEGEMMInterleave4x4Kernel.cpp +++ b/src/core/NEON/kernels/NEGEMMInterleave4x4Kernel.cpp @@ -44,11 +44,10 @@ namespace { Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::U8, DataType::S8, - DataType::QS16, DataType::U16, DataType::S16, DataType::U32, DataType::S32, + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::U8, DataType::S8, + DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); if(output->total_size() != 0) { @@ -57,7 +56,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output) output_shape.set(1, std::ceil(input->dimension(1) / 4.0f)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; diff --git a/src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp b/src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp index cab3c7a58f..421a6f0ef9 100644 --- a/src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp +++ b/src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp @@ -43,9 +43,8 @@ namespace { inline Status validate_arguments(const ITensorInfo *accum, const ITensorInfo *biases) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(accum, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(accum, 1, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(biases, accum); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(biases, accum); ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != accum->dimension(0)); @@ -161,33 +160,6 @@ void NEGEMMMatrixAccumulateBiasesKernel::run(const Window &window, const ThreadI break; } #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - case DataType::QS8: - { - execute_window_loop(window, [&](const Coordinates & id) - { - const qint8x16_t accum = vld1q_qs8(reinterpret_cast<const qint8_t *>(in0_out.ptr())); - const qint8x16_t biases = vld1q_qs8(reinterpret_cast<const qint8_t *>(in1.ptr())); - - vst1q_qs8(reinterpret_cast<qint8_t *>(in0_out.ptr()), vqaddq_qs8(accum, biases)); - }, - in0_out, in1); - break; - } - case DataType::QS16: - { - execute_window_loop(window, [&](const Coordinates & id) - { - qint16x8x2_t accum = vld2q_s16(reinterpret_cast<const qint16_t *>(in0_out.ptr())); - const qint16x8x2_t biases = vld2q_s16(reinterpret_cast<const qint16_t *>(in1.ptr())); - - accum.val[0] = vqaddq_qs16(accum.val[0], biases.val[0]); - accum.val[1] = vqaddq_qs16(accum.val[1], biases.val[1]); - - vst2q_s16(reinterpret_cast<qint16_t *>(in0_out.ptr()), accum); - }, - in0_out, in1); - break; - } default: ARM_COMPUTE_ERROR("Data type not supported"); break; diff --git a/src/core/NEON/kernels/NEGEMMMatrixAdditionKernel.cpp b/src/core/NEON/kernels/NEGEMMMatrixAdditionKernel.cpp index dfba74355b..d02504329a 100644 --- a/src/core/NEON/kernels/NEGEMMMatrixAdditionKernel.cpp +++ b/src/core/NEON/kernels/NEGEMMMatrixAdditionKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016, 2017 ARM Limited. + * Copyright (c) 2016-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -91,54 +91,6 @@ void matrix_addition_f16(const ITensor *input, ITensor *output, const Window &wi } #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ -void matrix_addition_qs8(const ITensor *input, ITensor *output, const Window &window, float beta) -{ - const int fixed_point_position = input->info()->fixed_point_position(); - const qint8x16_t beta_qs8 = vdupq_n_qs8(sqcvt_qs8_f32(beta, fixed_point_position)); - - Iterator in(input, window); - Iterator out(output, window); - - execute_window_loop(window, [&](const Coordinates & id) - { - const auto in_ptr = reinterpret_cast<const qint8_t *>(in.ptr()); - const auto out_ptr = reinterpret_cast<qint8_t *>(out.ptr()); - - qint8x16_t alpha_ab = vld1q_qs8(out_ptr); - const qint8x16_t c = vld1q_qs8(in_ptr); - - // Multiply matrix C by its weight and accumulate - alpha_ab = vqmlaq_qs8(alpha_ab, c, beta_qs8, fixed_point_position); - - vst1q_qs8(out_ptr, alpha_ab); - }, - in, out); -} - -void matrix_addition_qs16(const ITensor *input, ITensor *output, const Window &window, float beta) -{ - const int fixed_point_position = input->info()->fixed_point_position(); - const qint16x8_t beta_qs16 = vdupq_n_qs16(sqcvt_qs16_f32(beta, fixed_point_position)); - - Iterator in(input, window); - Iterator out(output, window); - - execute_window_loop(window, [&](const Coordinates & id) - { - const auto in_ptr = reinterpret_cast<const qint16_t *>(in.ptr()); - const auto out_ptr = reinterpret_cast<qint16_t *>(out.ptr()); - - qint16x8x2_t alpha_ab = vld2q_s16(out_ptr); - const qint16x8x2_t c = vld2q_s16(in_ptr); - - // Multiply matrix C by its weight and accumulate - alpha_ab.val[0] = vqmlaq_qs16(alpha_ab.val[0], c.val[0], beta_qs16, fixed_point_position); - alpha_ab.val[1] = vqmlaq_qs16(alpha_ab.val[1], c.val[1], beta_qs16, fixed_point_position); - - vst2q_s16(out_ptr, alpha_ab); - }, - in, out); -} } // namespace NEGEMMMatrixAdditionKernel::NEGEMMMatrixAdditionKernel() @@ -148,10 +100,9 @@ NEGEMMMatrixAdditionKernel::NEGEMMMatrixAdditionKernel() void NEGEMMMatrixAdditionKernel::configure(const ITensor *input, ITensor *output, float beta) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F16, DataType::F32); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != output->info()->dimension(0)); ARM_COMPUTE_ERROR_ON(input->info()->dimension(1) != output->info()->dimension(1)); @@ -160,12 +111,6 @@ void NEGEMMMatrixAdditionKernel::configure(const ITensor *input, ITensor *output case DataType::F32: _func = &matrix_addition_f32; break; - case DataType::QS8: - _func = &matrix_addition_qs8; - break; - case DataType::QS16: - _func = &matrix_addition_qs16; - break; case DataType::F16: #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC _func = &matrix_addition_f16; diff --git a/src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp b/src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp index 69b052a9bd..196398a2de 100644 --- a/src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp +++ b/src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp @@ -356,263 +356,6 @@ void vector_matrix_multiply_f32(const ITensor *input0, const ITensor *input1, IT } template <bool multiply_alpha> -void vector_matrix_multiply_qs8(const ITensor *input0, const ITensor *input1, ITensor *output, const Window &window, const ThreadInfo &info, float alpha) -{ - const auto width_matrix_b = static_cast<int>(output->info()->dimension(0)); - const auto in_b_stride = static_cast<int>(input1->info()->strides_in_bytes()[1] / data_size_from_type(input1->info()->data_type())); - const auto num_elems_vec_a = static_cast<int>(input0->info()->dimension(0)); - const int fixed_point_position = input0->info()->fixed_point_position(); - - // The implementation computes 32 elements per iteration - const int window_start_x = 32 * info.thread_id; - const int window_step_x = 32 * info.num_threads; - // Make sure (window_end_x - window_start_x) is a multiple of window_step_x - const int window_end_x = ceil_to_multiple(width_matrix_b - window_start_x, window_step_x) + window_start_x; - - Window win_out(window); - win_out.set(Window::DimX, Window::Dimension(window_start_x, window_end_x, window_step_x)); - win_out.set(Window::DimY, Window::Dimension(0, 1, 1)); - - Window win_a(window); - win_a.set(Window::DimX, Window::Dimension(0, 0, 0)); - win_a.set(Window::DimY, Window::Dimension(0, 0, 0)); - - Window win_b; - // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 - // This scenario can happen when the the matrix multiplication is used to perform a convolution operation - if(input1->info()->num_dimensions() >= 3) - { - win_b = window; - } - win_b.set(Window::DimX, Window::Dimension(window_start_x, window_end_x, window_step_x)); - win_b.set(Window::DimY, Window::Dimension(0, 1, 1)); - - Iterator ina(input0, win_a); - Iterator inb(input1, win_b); - Iterator out(output, win_out); - - execute_window_loop(win_out, [&](const Coordinates & id) - { - if(id.x() > width_matrix_b) - { - return; - } - - // Reset accumulators - qint16x8_t acc00_qs16 = vdupq_n_qs16(0); - qint16x8_t acc01_qs16 = vdupq_n_qs16(0); - qint16x8_t acc02_qs16 = vdupq_n_qs16(0); - qint16x8_t acc03_qs16 = vdupq_n_qs16(0); - - auto vec_a = reinterpret_cast<const qint8_t *>(ina.ptr()); - auto matrix_b = reinterpret_cast<const qint8_t *>(inb.ptr()); - - auto vec_a_end_addr = vec_a + num_elems_vec_a; - for(; vec_a <= (vec_a_end_addr - 2);) - { - const qint8x8_t a0 = vld1_dup_qs8(vec_a + 0); - const qint8x8_t a1 = vld1_dup_qs8(vec_a + 1); - - const qint8x8_t b00 = vld1_qs8(matrix_b + 0 + 0 * in_b_stride); - const qint8x8_t b01 = vld1_qs8(matrix_b + 8 + 0 * in_b_stride); - const qint8x8_t b02 = vld1_qs8(matrix_b + 16 + 0 * in_b_stride); - const qint8x8_t b03 = vld1_qs8(matrix_b + 24 + 0 * in_b_stride); - const qint8x8_t b10 = vld1_qs8(matrix_b + 0 + 1 * in_b_stride); - const qint8x8_t b11 = vld1_qs8(matrix_b + 8 + 1 * in_b_stride); - const qint8x8_t b12 = vld1_qs8(matrix_b + 16 + 1 * in_b_stride); - const qint8x8_t b13 = vld1_qs8(matrix_b + 24 + 1 * in_b_stride); - - // First accumulation - acc00_qs16 = vqmlal_qs8(acc00_qs16, b00, a0, fixed_point_position); - acc01_qs16 = vqmlal_qs8(acc01_qs16, b01, a0, fixed_point_position); - acc02_qs16 = vqmlal_qs8(acc02_qs16, b02, a0, fixed_point_position); - acc03_qs16 = vqmlal_qs8(acc03_qs16, b03, a0, fixed_point_position); - - // Second accumulation - acc00_qs16 = vqmlal_qs8(acc00_qs16, b10, a1, fixed_point_position); - acc01_qs16 = vqmlal_qs8(acc01_qs16, b11, a1, fixed_point_position); - acc02_qs16 = vqmlal_qs8(acc02_qs16, b12, a1, fixed_point_position); - acc03_qs16 = vqmlal_qs8(acc03_qs16, b13, a1, fixed_point_position); - - vec_a += 2; - matrix_b += 2 * in_b_stride; - } - - for(; vec_a < vec_a_end_addr;) - { - const qint8x8_t a0 = vld1_dup_qs8(vec_a); - - const qint8x8_t b00 = vld1_qs8(matrix_b + 0); - const qint8x8_t b01 = vld1_qs8(matrix_b + 8); - const qint8x8_t b02 = vld1_qs8(matrix_b + 16); - const qint8x8_t b03 = vld1_qs8(matrix_b + 24); - - acc00_qs16 = vqmlal_qs8(acc00_qs16, b00, a0, fixed_point_position); - acc01_qs16 = vqmlal_qs8(acc01_qs16, b01, a0, fixed_point_position); - acc02_qs16 = vqmlal_qs8(acc02_qs16, b02, a0, fixed_point_position); - acc03_qs16 = vqmlal_qs8(acc03_qs16, b03, a0, fixed_point_position); - - vec_a += 1; - matrix_b += in_b_stride; - } - - // Convert back to qint8x8_t and saturate - qint8x8_t acc00_qs8 = vqmovn_qs16(acc00_qs16); - qint8x8_t acc01_qs8 = vqmovn_qs16(acc01_qs16); - qint8x8_t acc02_qs8 = vqmovn_qs16(acc02_qs16); - qint8x8_t acc03_qs8 = vqmovn_qs16(acc03_qs16); - - // Multiply by the weight of the matrix product (alpha) - if(multiply_alpha) - { - const qint8x8_t alpha_qs8 = vdup_n_qs8(sqcvt_qs8_f32(alpha, fixed_point_position)); - acc00_qs8 = vqmul_qs8(acc00_qs8, alpha_qs8, fixed_point_position); - acc01_qs8 = vqmul_qs8(acc01_qs8, alpha_qs8, fixed_point_position); - acc02_qs8 = vqmul_qs8(acc02_qs8, alpha_qs8, fixed_point_position); - acc03_qs8 = vqmul_qs8(acc03_qs8, alpha_qs8, fixed_point_position); - } - - const auto mtx_out0 = reinterpret_cast<qint8_t *>(out.ptr()); - - // Store 8x4 output elements - vst1_qs8(mtx_out0 + 0, acc00_qs8); - vst1_qs8(mtx_out0 + 8, acc01_qs8); - vst1_qs8(mtx_out0 + 16, acc02_qs8); - vst1_qs8(mtx_out0 + 24, acc03_qs8); - }, - ina, inb, out); -} - -template <bool multiply_alpha> -void vector_matrix_multiply_qs16(const ITensor *input0, const ITensor *input1, ITensor *output, const Window &window, const ThreadInfo &info, float alpha) -{ - const auto width_matrix_b = static_cast<int>(output->info()->dimension(0)); - const auto in_b_stride = static_cast<int>(input1->info()->strides_in_bytes()[1] / data_size_from_type(input1->info()->data_type())); - const auto num_elems_vec_a = static_cast<int>(input0->info()->dimension(0)); - const int fixed_point_position = input0->info()->fixed_point_position(); - - // The implementation computes 16 elements per iteration - const int window_start_x = 16 * info.thread_id; - const int window_step_x = 16 * info.num_threads; - // Make sure (window_end_x - window_start_x) is a multiple of window_step_x - const int window_end_x = ceil_to_multiple(width_matrix_b - window_start_x, window_step_x) + window_start_x; - ARM_COMPUTE_ERROR_ON_MSG((window_end_x - window_start_x) % window_step_x, " (window_end_x - window_start_x) must be multiple of window_step_x"); - - Window win_out(window); - win_out.set(Window::DimX, Window::Dimension(window_start_x, window_end_x, window_step_x)); - win_out.set(Window::DimY, Window::Dimension(0, 1, 1)); - - Window win_a(window); - win_a.set(Window::DimX, Window::Dimension(0, 0, 0)); - win_a.set(Window::DimY, Window::Dimension(0, 0, 0)); - - Window win_b; - // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 - // This scenario can happen when the the matrix multiplication is used to perform a convolution operation - if(input1->info()->num_dimensions() >= 3) - { - win_b = window; - } - win_b.set(Window::DimX, Window::Dimension(window_start_x, window_end_x, window_step_x)); - win_b.set(Window::DimY, Window::Dimension(0, 1, 1)); - - Iterator ina(input0, win_a); - Iterator inb(input1, win_b); - Iterator out(output, win_out); - - execute_window_loop(win_out, [&](const Coordinates & id) - { - if(id.x() > width_matrix_b) - { - return; - } - - // Reset accumulators - qint32x4_t acc00_qs32 = vdupq_n_qs32(0); - qint32x4_t acc01_qs32 = vdupq_n_qs32(0); - qint32x4_t acc02_qs32 = vdupq_n_qs32(0); - qint32x4_t acc03_qs32 = vdupq_n_qs32(0); - - auto vec_a = reinterpret_cast<const qint16_t *>(ina.ptr()); - auto matrix_b = reinterpret_cast<const qint16_t *>(inb.ptr()); - - auto vec_a_end_addr = vec_a + num_elems_vec_a; - for(; vec_a <= (vec_a_end_addr - 2);) - { - const qint16x4_t a0 = vld1_dup_qs16(vec_a + 0); - const qint16x4_t a1 = vld1_dup_qs16(vec_a + 1); - - const qint16x4_t b00 = vld1_qs16(matrix_b + 0 + 0 * in_b_stride); - const qint16x4_t b01 = vld1_qs16(matrix_b + 4 + 0 * in_b_stride); - const qint16x4_t b02 = vld1_qs16(matrix_b + 8 + 0 * in_b_stride); - const qint16x4_t b03 = vld1_qs16(matrix_b + 12 + 0 * in_b_stride); - const qint16x4_t b10 = vld1_qs16(matrix_b + 0 + 1 * in_b_stride); - const qint16x4_t b11 = vld1_qs16(matrix_b + 4 + 1 * in_b_stride); - const qint16x4_t b12 = vld1_qs16(matrix_b + 8 + 1 * in_b_stride); - const qint16x4_t b13 = vld1_qs16(matrix_b + 12 + 1 * in_b_stride); - - // First accumulation - acc00_qs32 = vqmlal_qs16(acc00_qs32, b00, a0, fixed_point_position); - acc01_qs32 = vqmlal_qs16(acc01_qs32, b01, a0, fixed_point_position); - acc02_qs32 = vqmlal_qs16(acc02_qs32, b02, a0, fixed_point_position); - acc03_qs32 = vqmlal_qs16(acc03_qs32, b03, a0, fixed_point_position); - - // Second accumulation - acc00_qs32 = vqmlal_qs16(acc00_qs32, b10, a1, fixed_point_position); - acc01_qs32 = vqmlal_qs16(acc01_qs32, b11, a1, fixed_point_position); - acc02_qs32 = vqmlal_qs16(acc02_qs32, b12, a1, fixed_point_position); - acc03_qs32 = vqmlal_qs16(acc03_qs32, b13, a1, fixed_point_position); - - vec_a += 2; - matrix_b += 2 * in_b_stride; - } - - for(; vec_a < vec_a_end_addr;) - { - const qint16x4_t a0 = vld1_dup_qs16(vec_a); - - const qint16x4_t b00 = vld1_qs16(matrix_b + 0); - const qint16x4_t b01 = vld1_qs16(matrix_b + 4); - const qint16x4_t b02 = vld1_qs16(matrix_b + 8); - const qint16x4_t b03 = vld1_qs16(matrix_b + 12); - - acc00_qs32 = vqmlal_qs16(acc00_qs32, b00, a0, fixed_point_position); - acc01_qs32 = vqmlal_qs16(acc01_qs32, b01, a0, fixed_point_position); - acc02_qs32 = vqmlal_qs16(acc02_qs32, b02, a0, fixed_point_position); - acc03_qs32 = vqmlal_qs16(acc03_qs32, b03, a0, fixed_point_position); - - vec_a += 1; - matrix_b += in_b_stride; - } - - // Convert back to qint16x4_t and saturate - qint16x4_t acc00_qs16 = vqmovn_qs32(acc00_qs32); - qint16x4_t acc01_qs16 = vqmovn_qs32(acc01_qs32); - qint16x4_t acc02_qs16 = vqmovn_qs32(acc02_qs32); - qint16x4_t acc03_qs16 = vqmovn_qs32(acc03_qs32); - - // Multiply by the weight of the matrix product (alpha) - if(multiply_alpha) - { - const qint16x4_t alpha_qs16 = vdup_n_qs16(sqcvt_qs16_f32(alpha, fixed_point_position)); - acc00_qs16 = vqmul_qs16(acc00_qs16, alpha_qs16, fixed_point_position); - acc01_qs16 = vqmul_qs16(acc01_qs16, alpha_qs16, fixed_point_position); - acc02_qs16 = vqmul_qs16(acc02_qs16, alpha_qs16, fixed_point_position); - acc03_qs16 = vqmul_qs16(acc03_qs16, alpha_qs16, fixed_point_position); - } - - const auto mtx_out0 = reinterpret_cast<qint16_t *>(out.ptr()); - - // Store 16x4 output elements - vst1_qs16(mtx_out0 + 0, acc00_qs16); - vst1_qs16(mtx_out0 + 4, acc01_qs16); - vst1_qs16(mtx_out0 + 8, acc02_qs16); - vst1_qs16(mtx_out0 + 12, acc03_qs16); - }, - ina, inb, out); -} - -template <bool multiply_alpha> void matrix_matrix_multiply_f32(const ITensor *input0, const ITensor *input1, ITensor *output, const Window &window, float alpha) { const size_t in_b_stride = input1->info()->strides_in_bytes()[1] / data_size_from_type(input1->info()->data_type()); @@ -1063,361 +806,12 @@ void matrix_matrix_multiply_f16(const ITensor *input0, const ITensor *input1, IT #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ } -template <bool multiply_alpha> -void matrix_matrix_multiply_qs8(const ITensor *input0, const ITensor *input1, ITensor *output, const Window &window, float alpha) -{ - const size_t in_b_stride = input1->info()->strides_in_bytes()[1] / data_size_from_type(input1->info()->data_type()); - const size_t out_stride1 = output->info()->strides_in_bytes()[1] / data_size_from_type(output->info()->data_type()); - const size_t out_stride2 = out_stride1 * 2; - const size_t out_stride3 = out_stride1 * 3; - const int num_elems_matrix_b_x = input1->info()->dimension(0); - const int fixed_point_position = input0->info()->fixed_point_position(); - const qint8x8_t alpha_qs8 = vdup_n_qs8(sqcvt_qs8_f32(alpha, fixed_point_position)); - ARM_COMPUTE_UNUSED(alpha_qs8); - - // Set step_x and step_y for matrix A. Scale by a factor of 4 the Y range as the input interleaved matrix A has 4 times less the rows of the output matrix - Window win_a(window); - win_a.set(Window::DimX, Window::Dimension(0, 0, 0)); - win_a.set(Window::DimY, Window::Dimension(window.y().start() / 4, std::max(window.y().end() / 4, 1), 1)); - - Window win_b; - // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 - // This scenario can happen when the the matrix multiplication is used to perform a convolution operation - if(input1->info()->num_dimensions() >= 3) - { - win_b = window; - } - // Set step_x and step_y for matrix B. Scale by a factor of 16 the X range as the input transposed matrix A has 16 times less the cols of the output matrix - // The step along the x direction is 2 times the in_b_stride because for each iteration we compute 2 blocks of size 16x4 - win_b.set(Window::DimX, Window::Dimension(window.x().start() / 16, window.x().end() / 16, 2 * in_b_stride)); - win_b.set(Window::DimY, Window::Dimension(0, 0, 0)); - - Iterator ina(input0, win_a); - Iterator inb(input1, win_b); - Iterator out(output, window); - - // The implementation assumes that the matrix A and Matrix B have been reshaped respectively with NEGEMMInterleave4x4 and NEGEMMTranspose1xW - // The reshaping of the matrices helps to have a cache friendly implementation and helps to avoid the data re-arrangements needed for computing 16x4 elements per iteration - // All the values needed for computing a single 32x4 block will be read from consecutive memory positions - execute_window_loop(window, [&](const Coordinates & id) - { - auto mtx_a0 = reinterpret_cast<const qint8_t *>(ina.ptr()); - auto mtx_b0 = reinterpret_cast<const qint8_t *>(inb.ptr()); - auto mtx_b1 = mtx_b0 + in_b_stride; - - qint16x8_t acc00_qs16 = vdupq_n_qs16(0); - qint16x8_t acc10_qs16 = vdupq_n_qs16(0); - qint16x8_t acc20_qs16 = vdupq_n_qs16(0); - qint16x8_t acc30_qs16 = vdupq_n_qs16(0); - - qint16x8_t acc01_qs16 = vdupq_n_qs16(0); - qint16x8_t acc11_qs16 = vdupq_n_qs16(0); - qint16x8_t acc21_qs16 = vdupq_n_qs16(0); - qint16x8_t acc31_qs16 = vdupq_n_qs16(0); - - qint16x8_t acc02_qs16 = vdupq_n_qs16(0); - qint16x8_t acc12_qs16 = vdupq_n_qs16(0); - qint16x8_t acc22_qs16 = vdupq_n_qs16(0); - qint16x8_t acc32_qs16 = vdupq_n_qs16(0); - - qint16x8_t acc03_qs16 = vdupq_n_qs16(0); - qint16x8_t acc13_qs16 = vdupq_n_qs16(0); - qint16x8_t acc23_qs16 = vdupq_n_qs16(0); - qint16x8_t acc33_qs16 = vdupq_n_qs16(0); - - int k = 0; - // This for loop performs 2 accumulations - for(; k <= (num_elems_matrix_b_x - 32); k += 32) - { - const qint8x8_t a0 = vld1_dup_qs8(mtx_a0 + 0); - const qint8x8_t a1 = vld1_dup_qs8(mtx_a0 + 1); - const qint8x8_t a2 = vld1_dup_qs8(mtx_a0 + 2); - const qint8x8_t a3 = vld1_dup_qs8(mtx_a0 + 3); - const qint8x8_t a4 = vld1_dup_qs8(mtx_a0 + 4); - const qint8x8_t a5 = vld1_dup_qs8(mtx_a0 + 5); - const qint8x8_t a6 = vld1_dup_qs8(mtx_a0 + 6); - const qint8x8_t a7 = vld1_dup_qs8(mtx_a0 + 7); - - const qint8x8_t b00 = vld1_qs8(mtx_b0 + 0); - const qint8x8_t b01 = vld1_qs8(mtx_b0 + 8); - const qint8x8_t b10 = vld1_qs8(mtx_b1 + 0); - const qint8x8_t b11 = vld1_qs8(mtx_b1 + 8); - - // First accumulation - acc00_qs16 = vqmlal_qs8(acc00_qs16, b00, a0, fixed_point_position); - acc10_qs16 = vqmlal_qs8(acc10_qs16, b00, a1, fixed_point_position); - acc20_qs16 = vqmlal_qs8(acc20_qs16, b00, a2, fixed_point_position); - acc30_qs16 = vqmlal_qs8(acc30_qs16, b00, a3, fixed_point_position); - acc02_qs16 = vqmlal_qs8(acc02_qs16, b10, a0, fixed_point_position); - acc12_qs16 = vqmlal_qs8(acc12_qs16, b10, a1, fixed_point_position); - acc22_qs16 = vqmlal_qs8(acc22_qs16, b10, a2, fixed_point_position); - acc32_qs16 = vqmlal_qs8(acc32_qs16, b10, a3, fixed_point_position); - - const qint8x8_t b02 = vld1_qs8(mtx_b0 + 16); - const qint8x8_t b03 = vld1_qs8(mtx_b0 + 24); - const qint8x8_t b12 = vld1_qs8(mtx_b1 + 16); - const qint8x8_t b13 = vld1_qs8(mtx_b1 + 24); - - acc01_qs16 = vqmlal_qs8(acc01_qs16, b01, a0, fixed_point_position); - acc11_qs16 = vqmlal_qs8(acc11_qs16, b01, a1, fixed_point_position); - acc21_qs16 = vqmlal_qs8(acc21_qs16, b01, a2, fixed_point_position); - acc31_qs16 = vqmlal_qs8(acc31_qs16, b01, a3, fixed_point_position); - acc03_qs16 = vqmlal_qs8(acc03_qs16, b11, a0, fixed_point_position); - acc13_qs16 = vqmlal_qs8(acc13_qs16, b11, a1, fixed_point_position); - acc23_qs16 = vqmlal_qs8(acc23_qs16, b11, a2, fixed_point_position); - acc33_qs16 = vqmlal_qs8(acc33_qs16, b11, a3, fixed_point_position); - -#if __arm__ - asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_a0))); - asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b0))); - asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b1))); -#endif /* __arm__ */ - - // Second accumulation - acc00_qs16 = vqmlal_qs8(acc00_qs16, b02, a4, fixed_point_position); - acc10_qs16 = vqmlal_qs8(acc10_qs16, b02, a5, fixed_point_position); - acc20_qs16 = vqmlal_qs8(acc20_qs16, b02, a6, fixed_point_position); - acc30_qs16 = vqmlal_qs8(acc30_qs16, b02, a7, fixed_point_position); - acc01_qs16 = vqmlal_qs8(acc01_qs16, b03, a4, fixed_point_position); - acc11_qs16 = vqmlal_qs8(acc11_qs16, b03, a5, fixed_point_position); - acc21_qs16 = vqmlal_qs8(acc21_qs16, b03, a6, fixed_point_position); - acc31_qs16 = vqmlal_qs8(acc31_qs16, b03, a7, fixed_point_position); - acc02_qs16 = vqmlal_qs8(acc02_qs16, b12, a4, fixed_point_position); - acc12_qs16 = vqmlal_qs8(acc12_qs16, b12, a5, fixed_point_position); - acc22_qs16 = vqmlal_qs8(acc22_qs16, b12, a6, fixed_point_position); - acc32_qs16 = vqmlal_qs8(acc32_qs16, b12, a7, fixed_point_position); - acc03_qs16 = vqmlal_qs8(acc03_qs16, b13, a4, fixed_point_position); - acc13_qs16 = vqmlal_qs8(acc13_qs16, b13, a5, fixed_point_position); - acc23_qs16 = vqmlal_qs8(acc23_qs16, b13, a6, fixed_point_position); - acc33_qs16 = vqmlal_qs8(acc33_qs16, b13, a7, fixed_point_position); - - mtx_a0 += 8; - mtx_b0 += 32; - mtx_b1 += 32; - } - - // This for loop performs the left over accumulations - for(; k < num_elems_matrix_b_x; k += 16) - { - const qint8x8_t a0 = vld1_dup_qs8(mtx_a0 + 0); - const qint8x8_t a1 = vld1_dup_qs8(mtx_a0 + 1); - const qint8x8_t a2 = vld1_dup_qs8(mtx_a0 + 2); - const qint8x8_t a3 = vld1_dup_qs8(mtx_a0 + 3); - - const qint8x8_t b00 = vld1_qs8(mtx_b0 + 0); - const qint8x8_t b01 = vld1_qs8(mtx_b0 + 8); - const qint8x8_t b10 = vld1_qs8(mtx_b1 + 0); - const qint8x8_t b11 = vld1_qs8(mtx_b1 + 8); - - acc00_qs16 = vqmlal_qs8(acc00_qs16, b00, a0, fixed_point_position); - acc10_qs16 = vqmlal_qs8(acc10_qs16, b00, a1, fixed_point_position); - acc20_qs16 = vqmlal_qs8(acc20_qs16, b00, a2, fixed_point_position); - acc30_qs16 = vqmlal_qs8(acc30_qs16, b00, a3, fixed_point_position); - acc01_qs16 = vqmlal_qs8(acc01_qs16, b01, a0, fixed_point_position); - acc11_qs16 = vqmlal_qs8(acc11_qs16, b01, a1, fixed_point_position); - acc21_qs16 = vqmlal_qs8(acc21_qs16, b01, a2, fixed_point_position); - acc31_qs16 = vqmlal_qs8(acc31_qs16, b01, a3, fixed_point_position); - acc02_qs16 = vqmlal_qs8(acc02_qs16, b10, a0, fixed_point_position); - acc12_qs16 = vqmlal_qs8(acc12_qs16, b10, a1, fixed_point_position); - acc22_qs16 = vqmlal_qs8(acc22_qs16, b10, a2, fixed_point_position); - acc32_qs16 = vqmlal_qs8(acc32_qs16, b10, a3, fixed_point_position); - acc03_qs16 = vqmlal_qs8(acc03_qs16, b11, a0, fixed_point_position); - acc13_qs16 = vqmlal_qs8(acc13_qs16, b11, a1, fixed_point_position); - acc23_qs16 = vqmlal_qs8(acc23_qs16, b11, a2, fixed_point_position); - acc33_qs16 = vqmlal_qs8(acc33_qs16, b11, a3, fixed_point_position); - - mtx_a0 += 4; - mtx_b0 += 16; - mtx_b1 += 16; - } - - // Convert back to qint8x8_t and saturate - qint8x8_t acc00_qs8 = vqmovn_qs16(acc00_qs16); - qint8x8_t acc10_qs8 = vqmovn_qs16(acc10_qs16); - qint8x8_t acc20_qs8 = vqmovn_qs16(acc20_qs16); - qint8x8_t acc30_qs8 = vqmovn_qs16(acc30_qs16); - - qint8x8_t acc01_qs8 = vqmovn_qs16(acc01_qs16); - qint8x8_t acc11_qs8 = vqmovn_qs16(acc11_qs16); - qint8x8_t acc21_qs8 = vqmovn_qs16(acc21_qs16); - qint8x8_t acc31_qs8 = vqmovn_qs16(acc31_qs16); - - qint8x8_t acc02_qs8 = vqmovn_qs16(acc02_qs16); - qint8x8_t acc12_qs8 = vqmovn_qs16(acc12_qs16); - qint8x8_t acc22_qs8 = vqmovn_qs16(acc22_qs16); - qint8x8_t acc32_qs8 = vqmovn_qs16(acc32_qs16); - - qint8x8_t acc03_qs8 = vqmovn_qs16(acc03_qs16); - qint8x8_t acc13_qs8 = vqmovn_qs16(acc13_qs16); - qint8x8_t acc23_qs8 = vqmovn_qs16(acc23_qs16); - qint8x8_t acc33_qs8 = vqmovn_qs16(acc33_qs16); - - // Multiply by the weight of the matrix product (alpha) - if(multiply_alpha) - { - acc00_qs8 = vqmul_qs8(acc00_qs8, alpha_qs8, fixed_point_position); - acc10_qs8 = vqmul_qs8(acc10_qs8, alpha_qs8, fixed_point_position); - acc20_qs8 = vqmul_qs8(acc20_qs8, alpha_qs8, fixed_point_position); - acc30_qs8 = vqmul_qs8(acc30_qs8, alpha_qs8, fixed_point_position); - acc01_qs8 = vqmul_qs8(acc01_qs8, alpha_qs8, fixed_point_position); - acc11_qs8 = vqmul_qs8(acc11_qs8, alpha_qs8, fixed_point_position); - acc21_qs8 = vqmul_qs8(acc21_qs8, alpha_qs8, fixed_point_position); - acc31_qs8 = vqmul_qs8(acc31_qs8, alpha_qs8, fixed_point_position); - acc02_qs8 = vqmul_qs8(acc02_qs8, alpha_qs8, fixed_point_position); - acc12_qs8 = vqmul_qs8(acc12_qs8, alpha_qs8, fixed_point_position); - acc22_qs8 = vqmul_qs8(acc22_qs8, alpha_qs8, fixed_point_position); - acc32_qs8 = vqmul_qs8(acc32_qs8, alpha_qs8, fixed_point_position); - acc03_qs8 = vqmul_qs8(acc03_qs8, alpha_qs8, fixed_point_position); - acc13_qs8 = vqmul_qs8(acc13_qs8, alpha_qs8, fixed_point_position); - acc23_qs8 = vqmul_qs8(acc23_qs8, alpha_qs8, fixed_point_position); - acc33_qs8 = vqmul_qs8(acc33_qs8, alpha_qs8, fixed_point_position); - } - - const auto mtx_out0 = reinterpret_cast<qint8_t *>(out.ptr()); - - // Store 32x4 output elements - vst1_qs8(mtx_out0 + 0, acc00_qs8); - vst1_qs8(mtx_out0 + 8, acc01_qs8); - vst1_qs8(mtx_out0 + 16, acc02_qs8); - vst1_qs8(mtx_out0 + 24, acc03_qs8); - vst1_qs8(mtx_out0 + out_stride1 + 0, acc10_qs8); - vst1_qs8(mtx_out0 + out_stride1 + 8, acc11_qs8); - vst1_qs8(mtx_out0 + out_stride1 + 16, acc12_qs8); - vst1_qs8(mtx_out0 + out_stride1 + 24, acc13_qs8); - vst1_qs8(mtx_out0 + out_stride2 + 0, acc20_qs8); - vst1_qs8(mtx_out0 + out_stride2 + 8, acc21_qs8); - vst1_qs8(mtx_out0 + out_stride2 + 16, acc22_qs8); - vst1_qs8(mtx_out0 + out_stride2 + 24, acc23_qs8); - vst1_qs8(mtx_out0 + out_stride3 + 0, acc30_qs8); - vst1_qs8(mtx_out0 + out_stride3 + 8, acc31_qs8); - vst1_qs8(mtx_out0 + out_stride3 + 16, acc32_qs8); - vst1_qs8(mtx_out0 + out_stride3 + 24, acc33_qs8); - }, - ina, inb, out); -} - -template <bool multiply_alpha> -void matrix_matrix_multiply_qs16(const ITensor *input0, const ITensor *input1, ITensor *output, const Window &window, float alpha) -{ - const size_t in_b_stride = input1->info()->strides_in_bytes()[1] / data_size_from_type(input1->info()->data_type()); - const size_t out_stride1 = output->info()->strides_in_bytes()[1] / data_size_from_type(output->info()->data_type()); - const size_t out_stride2 = out_stride1 * 2; - const size_t out_stride3 = out_stride1 * 3; - const int num_elems_matrix_b_x = input1->info()->dimension(0); - const int fixed_point_position = input0->info()->fixed_point_position(); - const qint16x4_t alpha_qs16 = vdup_n_qs16(sqcvt_qs16_f32(alpha, fixed_point_position)); - ARM_COMPUTE_UNUSED(alpha_qs16); - - // Set step_x and step_y for matrix A. Scale by a factor of 4 the Y range as the input interleaved matrix A has 4 times less the rows of the output matrix - Window win_a(window); - win_a.set(Window::DimX, Window::Dimension(0, 0, 0)); - win_a.set(Window::DimY, Window::Dimension(window.y().start() / 4, std::max(window.y().end() / 4, 1), 1)); - - Window win_b; - // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 - // This scenario can happen when the the matrix multiplication is used to perform a convolution operation - if(input1->info()->num_dimensions() >= 3) - { - win_b = window; - } - // Set step_x and step_y for matrix B. Scale by a factor of 16 the X range as the input transposed matrix A has 16 times less the cols of the output matrix - win_b.set(Window::DimX, Window::Dimension(window.x().start() / 8, window.x().end() / 8, in_b_stride)); - win_b.set(Window::DimY, Window::Dimension(0, 0, 0)); - - Iterator ina(input0, win_a); - Iterator inb(input1, win_b); - Iterator out(output, window); - - // The implementation assumes that the matrix A and Matrix B have been reshaped respectively with NEGEMMInterleave4x4 and NEGEMMTranspose1xW - // The reshaping of the matrices helps to have a cache friendly implementation and helps to avoid the data re-arrangements needed for computing 8x4 elements per iteration - // All the values needed for computing a single 8x4 block will be read from consecutive memory positions - execute_window_loop(window, [&](const Coordinates & id) - { - auto mtx_a0 = reinterpret_cast<const qint16_t *>(ina.ptr()); - auto mtx_b0 = reinterpret_cast<const qint16_t *>(inb.ptr()); - auto mtx_b1 = mtx_b0 + in_b_stride; - - qint32x4_t acc00_qs32 = vdupq_n_qs32(0); - qint32x4_t acc10_qs32 = vdupq_n_qs32(0); - qint32x4_t acc20_qs32 = vdupq_n_qs32(0); - qint32x4_t acc30_qs32 = vdupq_n_qs32(0); - - qint32x4_t acc01_qs32 = vdupq_n_qs32(0); - qint32x4_t acc11_qs32 = vdupq_n_qs32(0); - qint32x4_t acc21_qs32 = vdupq_n_qs32(0); - qint32x4_t acc31_qs32 = vdupq_n_qs32(0); - - // This for loop performs 1 accumulation - for(int k = 0; k <= (num_elems_matrix_b_x - 8); k += 8) - { - const qint16x4_t a0 = vld1_dup_qs16(mtx_a0 + 0); - const qint16x4_t a1 = vld1_dup_qs16(mtx_a0 + 1); - const qint16x4_t a2 = vld1_dup_qs16(mtx_a0 + 2); - const qint16x4_t a3 = vld1_dup_qs16(mtx_a0 + 3); - - const qint16x4_t b00 = vld1_qs16(mtx_b0 + 0); - const qint16x4_t b01 = vld1_qs16(mtx_b0 + 4); - - acc00_qs32 = vqmlal_qs16(acc00_qs32, b00, a0, fixed_point_position); - acc10_qs32 = vqmlal_qs16(acc10_qs32, b00, a1, fixed_point_position); - acc20_qs32 = vqmlal_qs16(acc20_qs32, b00, a2, fixed_point_position); - acc30_qs32 = vqmlal_qs16(acc30_qs32, b00, a3, fixed_point_position); - acc01_qs32 = vqmlal_qs16(acc01_qs32, b01, a0, fixed_point_position); - acc11_qs32 = vqmlal_qs16(acc11_qs32, b01, a1, fixed_point_position); - acc21_qs32 = vqmlal_qs16(acc21_qs32, b01, a2, fixed_point_position); - acc31_qs32 = vqmlal_qs16(acc31_qs32, b01, a3, fixed_point_position); - - mtx_a0 += 4; - mtx_b0 += 8; - mtx_b1 += 8; - } - - // Convert back to qint16x4_t and saturate - qint16x4_t acc00_qs16 = vqmovn_qs32(acc00_qs32); - qint16x4_t acc10_qs16 = vqmovn_qs32(acc10_qs32); - qint16x4_t acc20_qs16 = vqmovn_qs32(acc20_qs32); - qint16x4_t acc30_qs16 = vqmovn_qs32(acc30_qs32); - - qint16x4_t acc01_qs16 = vqmovn_qs32(acc01_qs32); - qint16x4_t acc11_qs16 = vqmovn_qs32(acc11_qs32); - qint16x4_t acc21_qs16 = vqmovn_qs32(acc21_qs32); - qint16x4_t acc31_qs16 = vqmovn_qs32(acc31_qs32); - - // Multiply by the weight of the matrix product (alpha) - if(multiply_alpha) - { - acc00_qs16 = vqmul_qs16(acc00_qs16, alpha_qs16, fixed_point_position); - acc10_qs16 = vqmul_qs16(acc10_qs16, alpha_qs16, fixed_point_position); - acc20_qs16 = vqmul_qs16(acc20_qs16, alpha_qs16, fixed_point_position); - acc30_qs16 = vqmul_qs16(acc30_qs16, alpha_qs16, fixed_point_position); - acc01_qs16 = vqmul_qs16(acc01_qs16, alpha_qs16, fixed_point_position); - acc11_qs16 = vqmul_qs16(acc11_qs16, alpha_qs16, fixed_point_position); - acc21_qs16 = vqmul_qs16(acc21_qs16, alpha_qs16, fixed_point_position); - acc31_qs16 = vqmul_qs16(acc31_qs16, alpha_qs16, fixed_point_position); - } - - const auto mtx_out0 = reinterpret_cast<qint16_t *>(out.ptr()); - - // Store 8x4 output elements - vst1_qs16(mtx_out0 + 0, acc00_qs16); - vst1_qs16(mtx_out0 + 4, acc01_qs16); - vst1_qs16(mtx_out0 + out_stride1 + 0, acc10_qs16); - vst1_qs16(mtx_out0 + out_stride1 + 4, acc11_qs16); - vst1_qs16(mtx_out0 + out_stride2 + 0, acc20_qs16); - vst1_qs16(mtx_out0 + out_stride2 + 4, acc21_qs16); - vst1_qs16(mtx_out0 + out_stride3 + 0, acc30_qs16); - vst1_qs16(mtx_out0 + out_stride3 + 4, acc31_qs16); - }, - ina, inb, out); -} - inline Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, bool is_interleaved, const GEMMReshapeInfo &reshape_info) { ARM_COMPUTE_UNUSED(alpha); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32, DataType::QS8, DataType::QS16); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output); if(!is_interleaved) { @@ -1428,7 +822,6 @@ inline Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *i ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(0) != output->dimension(0)); ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) != output->dimension(1)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input0, output); } } else @@ -1467,7 +860,6 @@ inline Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *i } ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(1) != static_cast<size_t>(m)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input0, output); } } @@ -1492,16 +884,6 @@ inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *inpu num_elems_processed_per_iteration_x = 16; break; } - case DataType::QS8: - { - num_elems_processed_per_iteration_x = 32; - break; - } - case DataType::QS16: - { - num_elems_processed_per_iteration_x = 16; - break; - } #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: { @@ -1539,16 +921,6 @@ inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *inpu num_elems_processed_per_iteration_x = 8; break; } - case DataType::QS8: - { - num_elems_processed_per_iteration_x = 32; - break; - } - case DataType::QS16: - { - num_elems_processed_per_iteration_x = 8; - break; - } #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: { @@ -1638,18 +1010,6 @@ void NEGEMMMatrixMultiplyKernel::run(const Window &window, const ThreadInfo &inf vector_matrix_multiply_f32<false>(_input0, _input1, _output, window, info, _alpha); break; } - case DataType::QS8: - { - multiply_alpha ? vector_matrix_multiply_qs8<true>(_input0, _input1, _output, window, info, _alpha) : - vector_matrix_multiply_qs8<false>(_input0, _input1, _output, window, info, _alpha); - break; - } - case DataType::QS16: - { - multiply_alpha ? vector_matrix_multiply_qs16<true>(_input0, _input1, _output, window, info, _alpha) : - vector_matrix_multiply_qs16<false>(_input0, _input1, _output, window, info, _alpha); - break; - } #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: { @@ -1675,18 +1035,6 @@ void NEGEMMMatrixMultiplyKernel::run(const Window &window, const ThreadInfo &inf matrix_matrix_multiply_f32<false>(_input0, _input1, _output, window, _alpha); break; } - case DataType::QS8: - { - multiply_alpha ? matrix_matrix_multiply_qs8<true>(_input0, _input1, _output, window, _alpha) : - matrix_matrix_multiply_qs8<false>(_input0, _input1, _output, window, _alpha); - break; - } - case DataType::QS16: - { - multiply_alpha ? matrix_matrix_multiply_qs16<true>(_input0, _input1, _output, window, _alpha) : - matrix_matrix_multiply_qs16<false>(_input0, _input1, _output, window, _alpha); - break; - } #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: { diff --git a/src/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.cpp b/src/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.cpp index c1e975e77e..8588f43edf 100644 --- a/src/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.cpp +++ b/src/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.cpp @@ -177,7 +177,6 @@ void NEGEMMMatrixVectorMultiplyKernel::configure(const ITensor *input0, const IT { ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8, DataType::F32); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output); ARM_COMPUTE_ERROR_ON(is_data_type_quantized_asymmetric(input0->info()->data_type()) && (output->info()->data_type() != DataType::S32)); ARM_COMPUTE_ERROR_ON(input0->info()->dimension(2) != input1->info()->dimension(1)); diff --git a/src/core/NEON/kernels/NEGEMMTranspose1xWKernel.cpp b/src/core/NEON/kernels/NEGEMMTranspose1xWKernel.cpp index 5d6163d583..4517f46139 100644 --- a/src/core/NEON/kernels/NEGEMMTranspose1xWKernel.cpp +++ b/src/core/NEON/kernels/NEGEMMTranspose1xWKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016, 2018 ARM Limited. + * Copyright (c) 2016-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -54,17 +54,15 @@ TensorShape get_output_shape(const ITensorInfo *input) Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::U8, DataType::S8, - DataType::QS16, DataType::U16, DataType::S16, DataType::U32, DataType::S32, + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::U8, DataType::S8, + DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); if(output->total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), get_output_shape(input)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; @@ -102,7 +100,7 @@ void NEGEMMTranspose1xWKernel::configure(const ITensor *input, ITensor *output) ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); // Output tensor auto inizialitation if not yet initialized - auto_init_if_empty(*output->info(), get_output_shape(input->info()), 1, input->info()->data_type(), input->info()->fixed_point_position()); + auto_init_if_empty(*output->info(), get_output_shape(input->info()), 1, input->info()->data_type()); // Perform validate step ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info())); diff --git a/src/core/NEON/kernels/NEIm2ColKernel.cpp b/src/core/NEON/kernels/NEIm2ColKernel.cpp index 86e3fd7a84..f03bc49ed3 100644 --- a/src/core/NEON/kernels/NEIm2ColKernel.cpp +++ b/src/core/NEON/kernels/NEIm2ColKernel.cpp @@ -24,7 +24,6 @@ #include "arm_compute/core/NEON/kernels/NEIm2ColKernel.h" #include "arm_compute/core/Error.h" -#include "arm_compute/core/FixedPoint.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/ITensor.h" #include "arm_compute/core/Size2D.h" @@ -47,9 +46,8 @@ namespace Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, bool is_fully_connected, bool is_flatten, const Size2D &dilation) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, output); ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::QASYMM8 && has_bias); ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1)); @@ -90,7 +88,6 @@ inline void linearize_volume(const uint8_t *const in_ptr, int input_stride_x, int input_stride_y, int input_stride_z, - int fixed_point_position, int pad_value, int dilation_x, int dilation_y) @@ -171,18 +168,7 @@ inline void linearize_volume(const uint8_t *const in_ptr, // Append 1 if the convolution layer has biases if(has_bias) { - if(std::is_same<T, qint8_t>::value) - { - *out_ptr = sqcvt_qs8_f32(1.0f, fixed_point_position); - } - else if(std::is_same<T, qint16_t>::value) - { - *out_ptr = sqcvt_qs16_f32(1.0f, fixed_point_position); - } - else - { - *out_ptr = static_cast<T>(1); - } + *out_ptr = static_cast<T>(1); } } } // namespace @@ -251,7 +237,6 @@ void NEIm2ColKernel::run_generic(const Window &window) input_stride_x, input_stride_y, input_stride_z, - _input->info()->fixed_point_position(), offset, _dilation.x(), _dilation.y()); @@ -294,18 +279,7 @@ void NEIm2ColKernel::run_reduced(const Window &window) // Add bias if(_has_bias) { - if(std::is_same<T, qint8_t>::value) - { - *(reinterpret_cast<T *>(out_ptr) + out_width - 1) = sqcvt_qs8_f32(1.0f, _input->info()->fixed_point_position()); - } - else if(std::is_same<T, qint16_t>::value) - { - *(reinterpret_cast<T *>(out_ptr) + out_width - 1) = sqcvt_qs16_f32(1.0f, _input->info()->fixed_point_position()); - } - else - { - *(reinterpret_cast<T *>(out_ptr) + out_width - 1) = static_cast<T>(1); - } + *(reinterpret_cast<T *>(out_ptr) + out_width - 1) = static_cast<T>(1); } } while(in_window.slide_window_slice_3D(in_slice) && out_window.slide_window_slice_1D(out_slice)); @@ -366,12 +340,6 @@ void NEIm2ColKernel::configure(const ITensor *input, ITensor *output, const Size _func = &NEIm2ColKernel::run_reduced<float16_t>; break; #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - case DataType::QS8: - _func = &NEIm2ColKernel::run_reduced<qint8_t>; - break; - case DataType::QS16: - _func = &NEIm2ColKernel::run_reduced<qint16_t>; - break; case DataType::QASYMM8: _func = &NEIm2ColKernel::run_reduced<qasymm8_t>; break; @@ -392,12 +360,6 @@ void NEIm2ColKernel::configure(const ITensor *input, ITensor *output, const Size _func = (!conv_info.has_padding()) ? &NEIm2ColKernel::run_generic<float16_t, false> : &NEIm2ColKernel::run_generic<float16_t, true>; break; #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - case DataType::QS8: - _func = (!conv_info.has_padding()) ? &NEIm2ColKernel::run_generic<qint8_t, false> : &NEIm2ColKernel::run_generic<qint8_t, true>; - break; - case DataType::QS16: - _func = (!conv_info.has_padding()) ? &NEIm2ColKernel::run_generic<qint16_t, false> : &NEIm2ColKernel::run_generic<qint16_t, true>; - break; case DataType::QASYMM8: _func = (!conv_info.has_padding()) ? &NEIm2ColKernel::run_generic<qasymm8_t, false> : &NEIm2ColKernel::run_generic<qasymm8_t, true>; break; diff --git a/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp b/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp index 91776d8100..ed037832af 100644 --- a/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp +++ b/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp @@ -103,7 +103,7 @@ std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITe Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); // Output auto initialization if not yet initialized - auto_init_if_empty(*output, input->tensor_shape(), 1, input->data_type(), input->fixed_point_position()); + auto_init_if_empty(*output, input->tensor_shape(), 1, input->data_type()); AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration); AccessWindowHorizontal sum_access(sum, 0, num_elems_processed_per_iteration_sum); diff --git a/src/core/NEON/kernels/NEMinMaxLayerKernel.cpp b/src/core/NEON/kernels/NEMinMaxLayerKernel.cpp index 434f4eb3e9..d93dc09ff9 100644 --- a/src/core/NEON/kernels/NEMinMaxLayerKernel.cpp +++ b/src/core/NEON/kernels/NEMinMaxLayerKernel.cpp @@ -68,7 +68,7 @@ std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITe TensorShape output_shape = compute_min_max_shape(input); // Output auto initialization if not yet initialized - auto_init_if_empty(*output, output_shape, 1, input->data_type(), input->fixed_point_position()); + auto_init_if_empty(*output, output_shape, 1, input->data_type()); constexpr unsigned int num_elems_processed_per_iteration = 1; diff --git a/src/core/NEON/kernels/NENormalizationLayerKernel.cpp b/src/core/NEON/kernels/NENormalizationLayerKernel.cpp index 776cb27d7a..253a93f196 100644 --- a/src/core/NEON/kernels/NENormalizationLayerKernel.cpp +++ b/src/core/NEON/kernels/NENormalizationLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -39,26 +39,17 @@ namespace Status validate_arguments(const ITensorInfo *input, const ITensorInfo *input_squared, const ITensorInfo *output, const NormalizationLayerInfo &norm_info) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, input_squared, output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, input_squared); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, input_squared); ARM_COMPUTE_RETURN_ERROR_ON_MSG(!(norm_info.norm_size() % 2), "Normalization size should be odd"); - if(is_data_type_fixed_point(input->data_type())) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, input_squared); - ARM_COMPUTE_RETURN_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.beta(), input); - ARM_COMPUTE_RETURN_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.kappa(), input); - ARM_COMPUTE_RETURN_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.scale_coeff(), input); - } - // Checks performed when output is configured if(output->total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; @@ -162,44 +153,6 @@ void NENormalizationLayerKernel::configure(const ITensor *input, const ITensor * } break; } - case DataType::QS8: - { - switch(norm_info.type()) - { - case NormType::IN_MAP_1D: - _func = &NENormalizationLayerKernel::normalize_fixed_point<DataType::QS8, 0, false>; - break; - case NormType::IN_MAP_2D: - // Normalize over X and Y - _func = &NENormalizationLayerKernel::normalize_fixed_point<DataType::QS8, 0, true>; - break; - case NormType::CROSS_MAP: - _func = &NENormalizationLayerKernel::normalize_fixed_point<DataType::QS8, 2, false>; - break; - default: - break; - } - break; - } - case DataType::QS16: - { - switch(norm_info.type()) - { - case NormType::IN_MAP_1D: - _func = &NENormalizationLayerKernel::normalize_fixed_point<DataType::QS16, 0, false>; - break; - case NormType::IN_MAP_2D: - // Normalize over X and Y - _func = &NENormalizationLayerKernel::normalize_fixed_point<DataType::QS16, 0, true>; - break; - case NormType::CROSS_MAP: - _func = &NENormalizationLayerKernel::normalize_fixed_point<DataType::QS16, 2, false>; - break; - default: - break; - } - break; - } default: ARM_COMPUTE_ERROR("NOT SUPPORTED!"); } @@ -306,105 +259,6 @@ void NENormalizationLayerKernel::normalize_float(const Window &window) } } -template <DataType dt, unsigned int dim, bool do_2D_norm> -void NENormalizationLayerKernel::normalize_fixed_point(const Window &window) -{ - Iterator input(_input, window); - Iterator input_squared(_input_squared, window); - Iterator output(_output, window); - - const int dim_y = 1; - const int radius = _norm_info.norm_size() / 2; - const int total_size = _input->info()->dimension(dim) - 1; - const int input_squared_stride = _input_squared->info()->strides_in_bytes()[dim]; - // We account padding across X only and we iterate over rows - const int min_left = (dim == 2) ? 0 : -static_cast<int>(border_size().left); - const int max_right = (dim == 2) ? total_size : total_size + border_size().left; - const int min_top = 0; - const int max_bottom = _input->info()->dimension(dim_y) - 1; - - const int fixed_point_position = _input->info()->fixed_point_position(); - - if(dt == DataType::QS8) - { - const qint8x16_t coeff_vec = vdupq_n_qs8_f32(_norm_info.scale_coeff(), fixed_point_position); - const qint8x16_t beta_vec = vdupq_n_qs8_f32(_norm_info.beta(), fixed_point_position); - const qint8x16_t kappa_vec = vdupq_n_qs8_f32(_norm_info.kappa(), fixed_point_position); - - execute_window_loop(window, [&](const Coordinates & id) - { - // Get range to normalize - const int current_row = do_2D_norm ? id[dim_y] : 0; - const int current_slice = id[dim]; - const int first_row = do_2D_norm ? std::max(current_row - radius, min_top) : 0; - const int last_row = do_2D_norm ? std::min(current_row + radius, max_bottom) : 0; - const int first_slice = std::max(current_slice - radius, min_left); - const int last_slice = std::min(current_slice + radius, max_right); - - // Accumulate 2D In-Map values - qint8x16_t accu = vdupq_n_qs8(0); - for(int j = first_row; j <= last_row; ++j) - { - // Compute row displacement - const int row = (j - current_row) * _input_squared->info()->strides_in_bytes()[dim_y]; - const uint8_t *const input_squared_ptr = input_squared.ptr() + row - (current_slice * input_squared_stride); - for(int i = first_slice; i <= last_slice; ++i) - { - accu = vqaddq_qs8(accu, vld1q_qs8(reinterpret_cast<const qint8_t *>(input_squared_ptr + i * input_squared_stride))); - } - } - - // Normalize - const qint8x16_t accu_scale = vqmlaq_qs8(kappa_vec, coeff_vec, accu, fixed_point_position); - const qint8x16_t normalized = vqpowq_qs8(accu_scale, beta_vec, fixed_point_position); - const qint8x16_t normalized_pixel = vdivq_qs8(vld1q_qs8(reinterpret_cast<const qint8_t *>(input.ptr())), normalized, fixed_point_position); - vst1q_qs8(reinterpret_cast<qint8_t *>(output.ptr()), normalized_pixel); - }, - input, input_squared, output); - } - else if(dt == DataType::QS16) - { - const qint16x8_t coeff_vec = vdupq_n_qs16_f32(_norm_info.scale_coeff(), fixed_point_position); - const qint16x8_t beta_vec = vdupq_n_qs16_f32(_norm_info.beta(), fixed_point_position); - const qint16x8_t kappa_vec = vdupq_n_qs16_f32(_norm_info.kappa(), fixed_point_position); - - execute_window_loop(window, [&](const Coordinates & id) - { - // Get range to normalize - const int current_row = do_2D_norm ? id[dim_y] : 0; - const int current_slice = id[dim]; - const int first_row = do_2D_norm ? std::max(current_row - radius, min_top) : 0; - const int last_row = do_2D_norm ? std::min(current_row + radius, max_bottom) : 0; - const int first_slice = std::max(current_slice - radius, min_left); - const int last_slice = std::min(current_slice + radius, max_right); - - // Accumulate 2D In-Map values - qint16x8_t accu = vdupq_n_qs16(0); - for(int j = first_row; j <= last_row; ++j) - { - // Compute row displacement - const int row = (j - current_row) * _input_squared->info()->strides_in_bytes()[dim_y]; - const uint8_t *const input_squared_ptr = input_squared.ptr() + row - (current_slice * input_squared_stride); - for(int i = first_slice; i <= last_slice; ++i) - { - accu = vqaddq_qs16(accu, vld1q_qs16(reinterpret_cast<const qint16_t *>(input_squared_ptr + i * input_squared_stride))); - } - } - - // Normalize - const qint16x8_t accu_scale = vqmlaq_qs16(kappa_vec, coeff_vec, accu, fixed_point_position); - const qint16x8_t normalized = vqpowq_qs16(accu_scale, beta_vec, fixed_point_position); - const qint16x8_t normalized_pixel = vdivq_qs16(vld1q_qs16(reinterpret_cast<const qint16_t *>(input.ptr())), normalized, fixed_point_position); - vst1q_qs16(reinterpret_cast<qint16_t *>(output.ptr()), normalized_pixel); - }, - input, input_squared, output); - } - else - { - ARM_COMPUTE_ERROR("Not supported"); - } -} - Status NENormalizationLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *input_squared, const ITensorInfo *output, const NormalizationLayerInfo norm_info) { ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, input_squared, output, norm_info)); diff --git a/src/core/NEON/kernels/NEPermuteKernel.cpp b/src/core/NEON/kernels/NEPermuteKernel.cpp index ae1d48cc69..e9bc8effc6 100644 --- a/src/core/NEON/kernels/NEPermuteKernel.cpp +++ b/src/core/NEON/kernels/NEPermuteKernel.cpp @@ -45,8 +45,8 @@ namespace { Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PermutationVector &perm) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::QASYMM8, - DataType::U16, DataType::S16, DataType::QS16, + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8, + DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG((perm.num_dimensions() == 3 && !(perm[0] == 2 && perm[1] == 0 && perm[2] == 1) && !(perm[0] == 1 && perm[1] == 2 && perm[2] == 0)), @@ -59,7 +59,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; diff --git a/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp b/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp index 193ca3799c..0ec7e823a1 100644 --- a/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp +++ b/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp @@ -61,9 +61,9 @@ inline Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *i ARM_COMPUTE_UNUSED(overflow_policy); ARM_COMPUTE_UNUSED(rounding_policy); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() == DataType::U8 && (input1->data_type() != DataType::U8 || input2->data_type() != DataType::U8), "Output can only be U8 if both inputs are U8"); @@ -71,14 +71,6 @@ inline Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *i ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output->tensor_shape(), 0), "Wrong shape for output"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible"); - if(is_data_type_fixed_point(input1->data_type()) || is_data_type_fixed_point(input2->data_type()) || is_data_type_fixed_point(output->data_type())) - { - // Check that all data types are the same and all fixed-point positions are the same - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input1, input2, output); - // Check if scale is representable in fixed-point with the provided settings - ARM_COMPUTE_RETURN_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(scale, input1); - } - if(std::abs(scale - scale255_constant) < 0.00001f) { ARM_COMPUTE_RETURN_ERROR_ON(rounding_policy != RoundingPolicy::TO_NEAREST_UP && rounding_policy != RoundingPolicy::TO_NEAREST_EVEN); @@ -120,11 +112,6 @@ inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *inpu { set_format_if_unknown(*output, Format::F16); } - else if(input1->data_type() == DataType::QS8 && input2->data_type() == DataType::QS8) - { - set_data_type_if_unknown(*output, DataType::QS8); - set_fixed_point_position_if_zero(*output, input1->fixed_point_position()); - } } // Configure kernel window @@ -220,105 +207,6 @@ void mul_U8_U8_U8_n(const void *__restrict input1_ptr, const void *__restrict in } template <bool is_scale255, bool is_sat> -void mul_QS8_QS8_QS8_n(const void *__restrict input1_ptr, const void *__restrict input2_ptr, void *__restrict output_ptr, int n, int fixed_point_position) -{ - const auto output = static_cast<qint8_t *__restrict>(output_ptr); - - const qint8x16_t ta1 = vld1q_qs8(static_cast<const qint8_t *__restrict>(input1_ptr)); - const qint8x16_t ta2 = vld1q_qs8(static_cast<const qint8_t *__restrict>(input2_ptr)); - - if(is_scale255) - { - qint16x8_t tmp1_high = vmovl_s8(vget_high_s8(ta1)); - qint16x8_t tmp1_low = vmovl_s8(vget_low_s8(ta1)); - const qint16x8_t tmp2_high = vmovl_s8(vget_high_s8(ta2)); - const qint16x8_t tmp2_low = vmovl_s8(vget_low_s8(ta2)); - - const float32x4x2_t scale255_f32 = - { - { - scale255_constant_f32q, - scale255_constant_f32q - } - }; - const qint16x8_t scale255 = vqcvtq_qs16_f32(scale255_f32, fixed_point_position); - - tmp1_high = vmulq_qs16(tmp1_high, tmp2_high, fixed_point_position); - tmp1_low = vmulq_qs16(tmp1_low, tmp2_low, fixed_point_position); - tmp1_high = vmulq_qs16(tmp1_high, scale255, fixed_point_position); - tmp1_low = vmulq_qs16(tmp1_low, scale255, fixed_point_position); - - if(is_sat) - { - vst1q_qs8(output, vcombine_s8(vqmovn_s16(tmp1_low), vqmovn_s16(tmp1_high))); - } - else - { - vst1q_qs8(output, vcombine_s8(vmovn_s16(tmp1_low), vmovn_s16(tmp1_high))); - } - } - else - { - const qint8x16_t vn = vdupq_n_s8(-n); - qint8x16_t res = ta2; - - if(is_sat) - { - res = vqshlq_s8(vqmulq_qs8(ta1, res, fixed_point_position), vn); - } - else - { - res = vshlq_s8(vmulq_qs8(ta1, res, fixed_point_position), vn); - } - vst1q_qs8(output, res); - } -} - -template <bool is_scale255, bool is_sat> -void mul_QS16_QS16_QS16_n(const void *__restrict input1_ptr, const void *__restrict input2_ptr, void *__restrict output_ptr, int n, int fixed_point_position) -{ - const qint16x8x2_t ta1 = vld2q_qs16(static_cast<const qint16_t *__restrict>(input1_ptr)); - qint16x8x2_t res = vld2q_qs16(static_cast<const qint16_t *__restrict>(input2_ptr)); - - if(is_scale255) - { - const float32x4x2_t scale255_f32 = - { - { - scale255_constant_f32q, - scale255_constant_f32q - } - }; - const qint16x8_t scale255 = vqcvtq_qs16_f32(scale255_f32, fixed_point_position); - if(is_sat) - { - res.val[0] = vqmulq_qs16(vqmulq_qs16(ta1.val[0], res.val[0], fixed_point_position), scale255, fixed_point_position); - res.val[1] = vqmulq_qs16(vqmulq_qs16(ta1.val[1], res.val[1], fixed_point_position), scale255, fixed_point_position); - } - else - { - res.val[0] = vmulq_qs16(vmulq_qs16(ta1.val[0], res.val[0], fixed_point_position), scale255, fixed_point_position); - res.val[1] = vmulq_qs16(vmulq_qs16(ta1.val[1], res.val[1], fixed_point_position), scale255, fixed_point_position); - } - } - else - { - const qint16x8_t vn = vdupq_n_s16(-n); - if(is_sat) - { - res.val[0] = vqshlq_s16(vqmulq_qs16(ta1.val[0], res.val[0], fixed_point_position), vn); - res.val[1] = vqshlq_s16(vqmulq_qs16(ta1.val[1], res.val[1], fixed_point_position), vn); - } - else - { - res.val[0] = vshlq_s16(vmulq_qs16(ta1.val[0], res.val[0], fixed_point_position), vn); - res.val[1] = vshlq_s16(vmulq_qs16(ta1.val[1], res.val[1], fixed_point_position), vn); - } - } - vst2q_s16(static_cast<qint16_t *__restrict>(output_ptr), res); -} - -template <bool is_scale255, bool is_sat> inline int16x8_t mul_S16_S16_S16_n_loop(const int16x8_t &input1, const int16x8_t &input2, int n) { int32x4_t tmp1_high = vmovl_s16(vget_high_s16(input1)); @@ -529,7 +417,7 @@ void mul_U8_S16_S16_n(const void *__restrict input1_ptr, const void *__restrict } // namespace NEPixelWiseMultiplicationKernel::NEPixelWiseMultiplicationKernel() - : _func_float(nullptr), _func_int(nullptr), _func_q_int(nullptr), _input1(nullptr), _input2(nullptr), _output(nullptr), _scale{ 0 }, _scale_exponent{ 0 } + : _func_float(nullptr), _func_int(nullptr), _input1(nullptr), _input2(nullptr), _output(nullptr), _scale{ 0 }, _scale_exponent{ 0 } { } @@ -550,7 +438,6 @@ void NEPixelWiseMultiplicationKernel::configure(const ITensor *input1, const ITe _scale = scale; _scale_exponent = 0; _func_int = nullptr; - _func_q_int = nullptr; _func_float = nullptr; bool is_scale_255 = false; @@ -630,28 +517,6 @@ void NEPixelWiseMultiplicationKernel::configure(const ITensor *input1, const ITe _func_int = is_sat ? &mul_U8_U8_S16_n<false, true> : &mul_U8_U8_S16_n<false, false>; } } - else if(DataType::QS8 == dt_input1 && DataType::QS8 == dt_input2 && DataType::QS8 == dt_output) - { - if(is_scale_255) - { - _func_q_int = is_sat ? &mul_QS8_QS8_QS8_n<true, true> : &mul_QS8_QS8_QS8_n<true, false>; - } - else - { - _func_q_int = is_sat ? &mul_QS8_QS8_QS8_n<false, true> : &mul_QS8_QS8_QS8_n<false, false>; - } - } - else if(DataType::QS16 == dt_input1 && DataType::QS16 == dt_input2 && DataType::QS16 == dt_output) - { - if(is_scale_255) - { - _func_q_int = is_sat ? &mul_QS16_QS16_QS16_n<true, true> : &mul_QS16_QS16_QS16_n<true, false>; - } - else - { - _func_q_int = is_sat ? &mul_QS16_QS16_QS16_n<false, true> : &mul_QS16_QS16_QS16_n<false, false>; - } - } else if(DataType::F16 == dt_input1 && DataType::F16 == dt_input2 && DataType::F16 == dt_output) { _func_float = &mul_F16_F16_F16_n<false, false>; @@ -724,17 +589,6 @@ void NEPixelWiseMultiplicationKernel::run(const Window &window, const ThreadInfo }, input1, input2, output); } - else if(_func_q_int != nullptr) - { - int fixed_point_position = _input1->info()->fixed_point_position(); - execute_window_loop(collapsed, [&](const Coordinates & id) - { - (*_func_q_int)(input1.ptr(), input2.ptr(), output.ptr(), _scale_exponent, fixed_point_position); - collapsed.slide_window_slice_3D(slice_input1); - collapsed.slide_window_slice_3D(slice_input2); - }, - input1, input2, output); - } else { ARM_COMPUTE_ERROR_ON(_func_float == nullptr); diff --git a/src/core/NEON/kernels/NEPoolingLayerKernel.cpp b/src/core/NEON/kernels/NEPoolingLayerKernel.cpp index 7877cf5cc0..e586b72d30 100644 --- a/src/core/NEON/kernels/NEPoolingLayerKernel.cpp +++ b/src/core/NEON/kernels/NEPoolingLayerKernel.cpp @@ -25,7 +25,6 @@ #include "arm_compute/core/AccessWindowStatic.h" #include "arm_compute/core/Error.h" -#include "arm_compute/core/FixedPoint.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/ITensor.h" #include "arm_compute/core/NEON/NEAsymm.h" @@ -79,32 +78,6 @@ inline float calculate_avg_scale(const Coordinates &id, const int pool_size_x, c return 1.f / ((end_y - start_y) * (end_x - start_x)); } -inline qint8_t calculate_avg_scale_q8(const Coordinates &id, int pool_size, int upper_bound_w, int upper_bound_h, - int pad_x, int pad_y, int stride_x, int stride_y, int fixed_point_position) -{ - static const std::array<qint8_t, 10> scale_values_q8 = - { { 0x0, 0x0, 0x40, 0x2A, 0x20, 0x19, 0x15, 0x12, 0x10, 0xE } }; - const int start_x = id.x() * stride_x - pad_x; - const int start_y = id.y() * stride_y - pad_y; - const int end_x = std::min(start_x + pool_size, upper_bound_w); - const int end_y = std::min(start_y + pool_size, upper_bound_h); - const int val = ((end_y - start_y) * (end_x - start_x)); - return sshr_qs8(scale_values_q8[val], (7 - fixed_point_position)); -} - -inline qint16_t calculate_avg_scale_q16(const Coordinates &id, int pool_size, int upper_bound_w, int upper_bound_h, - int pad_x, int pad_y, int stride_x, int stride_y, int fixed_point_position) -{ - static std::array<qint16_t, 10> scale_values_q16 = - { { 0x0, 0x0, 0x4000, 0x2AAB, 0x2000, 0x199A, 0x1555, 0x1249, 0x1000, 0xE38 } }; - const int start_x = id.x() * stride_x - pad_x; - const int start_y = id.y() * stride_y - pad_y; - const int end_x = std::min(start_x + pool_size, upper_bound_w); - const int end_y = std::min(start_y + pool_size, upper_bound_h); - const int val = ((end_y - start_y) * (end_x - start_x)); - return sshr_qs16(scale_values_q16[val], (15 - fixed_point_position)); -} - template <bool exclude_padding> inline void scale_vector_s16x8(uint16x8_t &v, const Coordinates &id, int id_offset, int step, const int pool_size, const int upper_bound_w, const int upper_bound_h, @@ -163,22 +136,18 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c int pool_stride_y = 0; PoolingType pool_type = pool_info.pool_type(); const PadStrideInfo pad_stride_info = pool_info.pad_stride_info(); - const bool exclude_padding = pool_info.exclude_padding(); std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride(); static const std::set<int> supported_pool_sizes = { 2, 3 }; - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON(pool_type == PoolingType::L2 && is_data_type_quantized(input->data_type())); ARM_COMPUTE_RETURN_ERROR_ON((supported_pool_sizes.find(pool_size_x) == supported_pool_sizes.end()) && ((input->data_type() != DataType::F32) && (input->data_type() != DataType::QASYMM8)) && (pool_type != PoolingType::MAX)); - ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_fixed_point(input->data_type()) && pool_stride_x > 2); - ARM_COMPUTE_RETURN_ERROR_ON(exclude_padding && is_data_type_fixed_point(input->data_type())); if(output->total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output); ARM_COMPUTE_RETURN_ERROR_ON((output->dimension(get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH)) != pooled_w) || (output->dimension(get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT)) != pooled_h)); @@ -236,22 +205,6 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen { switch(input->data_type()) { - case DataType::QS8: - num_elems_read_per_iteration = 16; - switch(pool_size_x) - { - case 2: - num_elems_horizontal_window = (pool_stride_x == 2) ? 8 : 16; - num_elems_processed_per_iteration = (pool_stride_x == 2) ? 8 : 15; - break; - case 3: - num_elems_horizontal_window = (pool_stride_x == 2) ? 8 : 16; - num_elems_processed_per_iteration = (pool_stride_x == 2) ? 7 : 14; - break; - default: - break; - } - break; case DataType::QASYMM8: if(is_nhwc) { @@ -274,22 +227,6 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen break; } break; - case DataType::QS16: - num_elems_read_per_iteration = 8; - switch(pool_size_x) - { - case 2: - num_elems_horizontal_window = (pool_stride_x == 2) ? 4 : 8; - num_elems_processed_per_iteration = (pool_stride_x == 2) ? 4 : 7; - break; - case 3: - num_elems_horizontal_window = (pool_stride_x == 2) ? 4 : 8; - num_elems_processed_per_iteration = (pool_stride_x == 2) ? 3 : 6; - break; - default: - break; - } - break; #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: if(is_nhwc) @@ -462,64 +399,7 @@ void NEPoolingLayerKernel::configure(const ITensor *input, ITensor *output, cons const DataType data_type = input->info()->data_type(); const bool is_nchw = data_layout == DataLayout::NCHW; - // Select appropriate function - if(data_type == DataType::QS8) - { - if(_is_square) - { - switch(pool_size_x) - { - case 2: - switch(pool_type) - { - case PoolingType::AVG: - _func = &NEPoolingLayerKernel::pooling2_q8_nchw<PoolingType::AVG>; - break; - case PoolingType::MAX: - _func = &NEPoolingLayerKernel::pooling2_q8_nchw<PoolingType::MAX>; - break; - default: - ARM_COMPUTE_ERROR("Unsupported pooling type!"); - } - break; - case 3: - switch(pool_type) - { - case PoolingType::AVG: - _func = &NEPoolingLayerKernel::pooling3_q8_nchw<PoolingType::AVG>; - break; - case PoolingType::MAX: - _func = &NEPoolingLayerKernel::pooling3_q8_nchw<PoolingType::MAX>; - break; - default: - ARM_COMPUTE_ERROR("Unsupported pooling type!"); - } - break; - default: - switch(pool_type) - { - case PoolingType::MAX: - _func = &NEPoolingLayerKernel::poolingMxN_q8_nchw<PoolingType::MAX>; - break; - default: - ARM_COMPUTE_ERROR("Unsupported pooling type!"); - } - break; - } - } - else - { - switch(pool_type) - { - case PoolingType::MAX: - _func = &NEPoolingLayerKernel::poolingMxN_q8_nchw<PoolingType::MAX>; - break; - default: - ARM_COMPUTE_ERROR("Unsupported pooling type!"); - } - } - } - else if(data_type == DataType::QASYMM8) + if(data_type == DataType::QASYMM8) { if(pool_size_x == 2 && pool_stride_x < 3 && _is_square) { @@ -606,62 +486,6 @@ void NEPoolingLayerKernel::configure(const ITensor *input, ITensor *output, cons } } } - else if(data_type == DataType::QS16) - { - if(_is_square) - { - switch(pool_size_x) - { - case 2: - switch(pool_type) - { - case PoolingType::AVG: - _func = &NEPoolingLayerKernel::pooling2_q16_nchw<PoolingType::AVG>; - break; - case PoolingType::MAX: - _func = &NEPoolingLayerKernel::pooling2_q16_nchw<PoolingType::MAX>; - break; - default: - ARM_COMPUTE_ERROR("Unsupported pooling type!"); - } - break; - case 3: - switch(pool_type) - { - case PoolingType::AVG: - _func = &NEPoolingLayerKernel::pooling3_q16_nchw<PoolingType::AVG>; - break; - case PoolingType::MAX: - _func = &NEPoolingLayerKernel::pooling3_q16_nchw<PoolingType::MAX>; - break; - default: - ARM_COMPUTE_ERROR("Unsupported pooling type!"); - } - break; - default: - switch(pool_type) - { - case PoolingType::MAX: - _func = &NEPoolingLayerKernel::poolingMxN_q16_nchw<PoolingType::MAX>; - break; - default: - ARM_COMPUTE_ERROR("Unsupported pooling type!"); - } - break; - } - } - else - { - switch(pool_type) - { - case PoolingType::MAX: - _func = &NEPoolingLayerKernel::poolingMxN_q16_nchw<PoolingType::MAX>; - break; - default: - ARM_COMPUTE_ERROR("Unsupported pooling type!"); - } - } - } else if(data_type == DataType::F16) { if(_is_square) @@ -1022,71 +846,6 @@ void NEPoolingLayerKernel::configure(const ITensor *input, ITensor *output, cons INEKernel::configure(win_config.second); } -template <PoolingType pooling_type> -void NEPoolingLayerKernel::pooling2_q8_nchw(const Window &window_input, const Window &window) -{ - Iterator input(_input, window_input); - Iterator output(_output, window); - - const int fixed_point_position = _input->info()->fixed_point_position(); - constexpr int pool_size = 2; - int pool_stride_x = 0; - int pool_stride_y = 0; - const int pool_pad_right = _pool_info.pad_stride_info().pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info().pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info().pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info().pad_bottom(); - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); - const int upper_bound_w = _input->info()->dimension(0) + pool_pad_right; - const int upper_bound_h = _input->info()->dimension(1) + pool_pad_bottom; - - const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top))); - const uint8_t *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1)); - - execute_window_loop(window, [&](const Coordinates & id) - { - const auto top_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_top_ptr + input.offset())); - const auto bottom_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_bottom_ptr + input.offset())); - qint8x8_t lower_res = {}; - qint8x8_t upper_res = {}; - if(pooling_type == PoolingType::AVG) - { - // Calculate scale - const qint8_t scale = calculate_avg_scale_q8(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y, fixed_point_position); - const qint8x8_t scale_vec = vdup_n_qs8(scale); - - // Perform pooling - const qint8x16_t sum_data = vqaddq_qs8(top_data, bottom_data); - lower_res = vqmul_qs8(vpadd_s8(vget_low_s8(sum_data), vget_high_s8(sum_data)), scale_vec, fixed_point_position); - if(pool_stride_x == 1) - { - const qint8x16_t sum_data_shifted = vextq_s8(sum_data, sum_data, 1); - upper_res = vqmul_qs8(vpadd_s8(vget_low_s8(sum_data_shifted), vget_high_s8(sum_data_shifted)), scale_vec, fixed_point_position); - } - } - else - { - const qint8x16_t max_data = vmaxq_s8(top_data, bottom_data); - lower_res = vpmax_s8(vget_low_s8(max_data), vget_high_s8(max_data)); - if(pool_stride_x == 1) - { - const qint8x16_t max_data_shifted = vextq_s8(max_data, max_data, 1); - upper_res = vpmax_s8(vget_low_s8(max_data_shifted), vget_high_s8(max_data_shifted)); - } - } - if(pool_stride_x == 1) - { - const qint8x8x2_t res = { { lower_res, upper_res } }; - vst2_s8(reinterpret_cast<qint8_t *>(output.ptr()), res); - } - else - { - vst1_qs8(reinterpret_cast<qint8_t *>(output.ptr()), lower_res); - } - }, - input, output); -} - template <PoolingType pooling_type, bool exclude_padding> void NEPoolingLayerKernel::pooling2_qasymm8_nchw(const Window &window_input, const Window &window) { @@ -1201,71 +960,6 @@ void NEPoolingLayerKernel::pooling2_qasymm8_nchw(const Window &window_input, con input, output); } -template <PoolingType pooling_type> -void NEPoolingLayerKernel::pooling2_q16_nchw(const Window &window_input, const Window &window) -{ - Iterator input(_input, window_input); - Iterator output(_output, window); - - const int fixed_point_position = _input->info()->fixed_point_position(); - constexpr int pool_size = 2; - const int pool_pad_right = _pool_info.pad_stride_info().pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info().pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info().pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info().pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); - const int upper_bound_w = _input->info()->dimension(0) + pool_pad_right; - const int upper_bound_h = _input->info()->dimension(1) + pool_pad_bottom; - - const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top))); - const unsigned char *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1)); - - execute_window_loop(window, [&](const Coordinates & id) - { - const auto top_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_top_ptr + input.offset())); - const auto bottom_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_bottom_ptr + input.offset())); - qint16x4_t lower_res = {}; - qint16x4_t upper_res = {}; - if(pooling_type == PoolingType::AVG) - { - // Calculate scale - const qint16_t scale = calculate_avg_scale_q16(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y, fixed_point_position); - const qint16x4_t scale_vec = vdup_n_qs16(scale); - - // Perform pooling - const qint16x8_t sum_data = vqaddq_qs16(top_data, bottom_data); - lower_res = vqmul_qs16(vpadd_s16(vget_low_s16(sum_data), vget_high_s16(sum_data)), scale_vec, fixed_point_position); - if(pool_stride_x == 1) - { - const qint16x8_t sum_data_shifted = vextq_s16(sum_data, sum_data, 1); - upper_res = vqmul_qs16(vpadd_s16(vget_low_s16(sum_data_shifted), vget_high_s16(sum_data_shifted)), scale_vec, fixed_point_position); - } - } - else - { - const qint16x8_t max_data = vmaxq_s16(top_data, bottom_data); - lower_res = vpmax_s16(vget_low_s16(max_data), vget_high_s16(max_data)); - if(pool_stride_x == 1) - { - const qint16x8_t max_data_shifted = vextq_s16(max_data, max_data, 1); - upper_res = vpmax_s16(vget_low_s16(max_data_shifted), vget_high_s16(max_data_shifted)); - } - } - if(pool_stride_x == 1) - { - const qint16x4x2_t res = { { lower_res, upper_res } }; - vst2_s16(reinterpret_cast<qint16_t *>(output.ptr()), res); - } - else - { - vst1_qs16(reinterpret_cast<qint16_t *>(output.ptr()), lower_res); - } - }, - input, output); -} - template <PoolingType pooling_type, bool exclude_padding> void NEPoolingLayerKernel::pooling3_f16_nchw(const Window &window_input, const Window &window) { @@ -1461,82 +1155,6 @@ void NEPoolingLayerKernel::pooling2_f32_nchw(const Window &window_input, const W input, output); } -template <PoolingType pooling_type> -void NEPoolingLayerKernel::pooling3_q8_nchw(const Window &window_input, const Window &window) -{ - Iterator input(_input, window_input); - Iterator output(_output, window); - - const int fixed_point_position = _input->info()->fixed_point_position(); - constexpr int pool_size = 3; - const int pool_pad_right = _pool_info.pad_stride_info().pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info().pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info().pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info().pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); - const int upper_bound_w = _input->info()->dimension(0) + pool_pad_right; - const int upper_bound_h = _input->info()->dimension(1) + pool_pad_bottom; - - const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top))); - const uint8_t *const input_middle_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1)); - const uint8_t *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 2)); - - execute_window_loop(window, [&](const Coordinates & id) - { - const auto top_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_top_ptr + input.offset())); - const auto middle_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_middle_ptr + input.offset())); - const auto bottom_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_bottom_ptr + input.offset())); - qint8x8_t res = {}; - if(pooling_type == PoolingType::AVG) - { - // Calculate scale - const qint8_t scale = calculate_avg_scale_q8(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y, fixed_point_position); - - // Perform pooling for stride 2 - const qint8x16_t sum_data = vqaddq_qs8(vqaddq_qs8(top_data, bottom_data), middle_data); - const qint8x16_t sum_data2 = vextq_s8(sum_data, sum_data, 1); - const qint8x16_t sum_data3 = vextq_s8(sum_data, sum_data, 2); - const qint8x16_t final_sum = vqaddq_qs8(vqaddq_qs8(sum_data, sum_data2), sum_data3); - if(pool_stride_x == 2) - { - const qint8x8x2_t table = { { vget_low_s8(final_sum), vget_high_s8(final_sum) } }; - static const qint8x8_t lookup_val = { 0, 2, 4, 6, 8, 10, 12, 14 }; - const qint8x8_t scale_vec = vdup_n_qs8(scale); - res = vtbl2_s8(table, lookup_val); - res = vqmul_qs8(res, scale_vec, fixed_point_position); - vst1_qs8(reinterpret_cast<qint8_t *>(output.ptr()), res); - } - else - { - const qint8x16_t scale_vec = vdupq_n_qs8(scale); - vst1q_qs8(reinterpret_cast<qint8_t *>(output.ptr()), vqmulq_qs8(final_sum, scale_vec, fixed_point_position)); - } - } - else - { - const qint8x16_t max_data = vmaxq_s8(vmaxq_s8(top_data, bottom_data), middle_data); - const qint8x16_t max_data2 = vextq_s8(max_data, max_data, 1); - const qint8x16_t max_data3 = vextq_s8(max_data, max_data, 2); - const qint8x16_t final_max = vmaxq_s8(vmaxq_s8(max_data, max_data2), max_data3); - - if(pool_stride_x == 2) - { - const qint8x8x2_t table = { { vget_low_s8(final_max), vget_high_s8(final_max) } }; - static const qint8x8_t lookup_val = { 0, 2, 4, 6, 8, 10, 12, 14 }; - res = vtbl2_s8(table, lookup_val); - vst1_qs8(reinterpret_cast<qint8_t *>(output.ptr()), res); - } - else - { - vst1q_qs8(reinterpret_cast<qint8_t *>(output.ptr()), final_max); - } - } - }, - input, output); -} - template <PoolingType pooling_type, bool exclude_padding> void NEPoolingLayerKernel::pooling3_qasymm8_nchw(const Window &window_input, const Window &window) { @@ -1657,77 +1275,6 @@ void NEPoolingLayerKernel::pooling3_qasymm8_nchw(const Window &window_input, con input, output); } -template <PoolingType pooling_type> -void NEPoolingLayerKernel::pooling3_q16_nchw(const Window &window_input, const Window &window) -{ - Iterator input(_input, window_input); - Iterator output(_output, window); - - const int fixed_point_position = _input->info()->fixed_point_position(); - constexpr int pool_size = 3; - const int pool_pad_right = _pool_info.pad_stride_info().pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info().pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info().pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info().pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); - const int upper_bound_w = _input->info()->dimension(0) + pool_pad_right; - const int upper_bound_h = _input->info()->dimension(1) + pool_pad_bottom; - - const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top))); - const unsigned char *const input_middle_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1)); - const unsigned char *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 2)); - - execute_window_loop(window, [&](const Coordinates & id) - { - const auto top_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_top_ptr + input.offset())); - const auto middle_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_middle_ptr + input.offset())); - const auto bottom_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_bottom_ptr + input.offset())); - - if(pooling_type == PoolingType::AVG) - { - // Calculate scale - const qint16_t scale = calculate_avg_scale_q16(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y, fixed_point_position); - - // Perform pooling for stride 2 - const qint16x8_t sum_data = vqaddq_qs16(vqaddq_qs16(top_data, bottom_data), middle_data); - const qint16x8_t sum_data2 = vextq_s16(sum_data, sum_data, 1); - const qint16x8_t sum_data3 = vextq_s16(sum_data, sum_data, 2); - const qint16x8_t final_sum = vqaddq_qs16(vqaddq_qs16(sum_data, sum_data2), sum_data3); - if(pool_stride_x == 2) - { - const qint16x4_t tmp = { vgetq_lane_s16(final_sum, 0), vgetq_lane_s16(final_sum, 2), vgetq_lane_s16(final_sum, 4), vgetq_lane_s16(final_sum, 6) }; - const qint16x4_t scale_vec = vdup_n_qs16(scale); - vst1_qs16(reinterpret_cast<qint16_t *>(output.ptr()), vqmul_qs16(tmp, scale_vec, fixed_point_position)); - } - else - { - const qint16x8_t scale_vec = vdupq_n_qs16(scale); - vst1q_qs16(reinterpret_cast<qint16_t *>(output.ptr()), vqmulq_qs16(final_sum, scale_vec, fixed_point_position)); - } - } - else - { - const qint16x8_t max_data = vmaxq_s16(vmaxq_s16(top_data, bottom_data), middle_data); - const qint16x8_t max_data2 = vextq_s16(max_data, max_data, 1); - const qint16x8_t max_data3 = vextq_s16(max_data, max_data, 2); - const qint16x8_t final_max = vmaxq_s16(vmaxq_s16(max_data, max_data2), max_data3); - - if(pool_stride_x == 2) - { - const qint16x4_t tmp = { vgetq_lane_s16(final_max, 0), vgetq_lane_s16(final_max, 2), vgetq_lane_s16(final_max, 4), vgetq_lane_s16(final_max, 6) }; - vst1_qs16(reinterpret_cast<qint16_t *>(output.ptr()), tmp); - } - else - { - vst1q_qs16(reinterpret_cast<qint16_t *>(output.ptr()), final_max); - } - } - }, - input, output); -} - template <PoolingType pooling_type, bool exclude_padding> void NEPoolingLayerKernel::pooling3_f32_nchw(const Window &window_input, const Window &window) { @@ -1879,110 +1426,6 @@ void NEPoolingLayerKernel::pooling7_f32_nchw(const Window &window_input, const W input, output); } -template <PoolingType pooling_type> -void NEPoolingLayerKernel::poolingMxN_q8_nchw(const Window &window_input, const Window &window) -{ - Iterator input(_input, window_input); - Iterator output(_output, window); - - const int pool_size_x = _pool_info.is_global_pooling() ? _input->info()->tensor_shape().x() : _pool_info.pool_size().width; - const int pool_size_y = _pool_info.is_global_pooling() ? _input->info()->tensor_shape().y() : _pool_info.pool_size().height; - const int pool_pad_top = _pool_info.pad_stride_info().pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info().pad_left(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); - - execute_window_loop(window, [&](const Coordinates & id) - { - qint8x16_t vres = {}; - qint8_t res = {}; - - //PoolingType::MAX - for(int y = 0; y < pool_size_y; ++y) - { - int x = 0; - for(; x <= (pool_size_x - 16); x += 16) - { - const qint8x16_t data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input.ptr() + (x - pool_pad_left) * _input->info()->strides_in_bytes().x() + - (y - pool_pad_top) * _input->info()->strides_in_bytes().y())); - vres = vmaxq_s8(vres, data); - } - - // Leftover for loop - for(; x < pool_size_x; ++x) - { - qint8_t data = *(reinterpret_cast<const qint8_t *>(input.ptr() + (x - pool_pad_left) * _input->info()->strides_in_bytes().x() + (y - pool_pad_top) * _input->info()->strides_in_bytes().y())); - res = std::max(res, data); - } - } - //Reduce - const qint8x8_t half_vres = vpmax_s8(vget_low_s8(vres), vget_high_s8(vres)); - res = std::max(res, vget_lane_s8(half_vres, 0)); - res = std::max(res, vget_lane_s8(half_vres, 1)); - res = std::max(res, vget_lane_s8(half_vres, 2)); - res = std::max(res, vget_lane_s8(half_vres, 3)); - res = std::max(res, vget_lane_s8(half_vres, 4)); - res = std::max(res, vget_lane_s8(half_vres, 5)); - res = std::max(res, vget_lane_s8(half_vres, 6)); - res = std::max(res, vget_lane_s8(half_vres, 7)); - - // Store result - *(reinterpret_cast<qint8_t *>(output.ptr())) = res; - }, - input, output); -} - -template <PoolingType pooling_type> -void NEPoolingLayerKernel::poolingMxN_q16_nchw(const Window &window_input, const Window &window) -{ - Iterator input(_input, window_input); - Iterator output(_output, window); - - const int pool_size_x = _pool_info.is_global_pooling() ? _input->info()->tensor_shape().x() : _pool_info.pool_size().width; - const int pool_size_y = _pool_info.is_global_pooling() ? _input->info()->tensor_shape().y() : _pool_info.pool_size().height; - const int pool_pad_top = _pool_info.pad_stride_info().pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info().pad_left(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); - - execute_window_loop(window, [&](const Coordinates & id) - { - qint16x8_t vres = {}; - qint16_t res = {}; - - //PoolingType::MAX - for(int y = 0; y < pool_size_y; ++y) - { - int x = 0; - for(; x <= (pool_size_x - 8); x += 8) - { - const qint16x8_t data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input.ptr() + (x - pool_pad_left) * _input->info()->strides_in_bytes().x() + - (y - pool_pad_top) * _input->info()->strides_in_bytes().y())); - vres = vmaxq_s16(vres, data); - } - - // Leftover for loop - for(; x < pool_size_x; ++x) - { - qint16_t data = *(reinterpret_cast<const qint16_t *>(input.ptr() + (x - pool_pad_left) * _input->info()->strides_in_bytes().x() + (y - pool_pad_top) * _input->info()->strides_in_bytes().y())); - res = std::max(res, data); - } - } - //Reduce - const qint16x4_t half_vres = vpmax_s16(vget_low_s16(vres), vget_high_s16(vres)); - res = std::max(res, vget_lane_s16(half_vres, 0)); - res = std::max(res, vget_lane_s16(half_vres, 1)); - res = std::max(res, vget_lane_s16(half_vres, 2)); - res = std::max(res, vget_lane_s16(half_vres, 3)); - - // Store result - *(reinterpret_cast<qint16_t *>(output.ptr())) = res; - }, - input, output); -} - template <PoolingType pooling_type, bool exclude_padding> void NEPoolingLayerKernel::poolingMxN_f16_nchw(const Window &window_input, const Window &window) { @@ -2688,8 +2131,6 @@ void NEPoolingLayerKernel::run(const Window &window, const ThreadInfo &info) unsigned int window_x_inc = 0; switch(_input->info()->data_type()) { - case DataType::QS8: - case DataType::QS16: case DataType::F16: { window_x_inc = (pool_stride_x == 2) ? _num_elems_processed_per_iteration * 2 : _num_elems_processed_per_iteration; diff --git a/src/core/NEON/kernels/NEQuantizationLayerKernel.cpp b/src/core/NEON/kernels/NEQuantizationLayerKernel.cpp index ee23e76c5c..b49400ab7d 100644 --- a/src/core/NEON/kernels/NEQuantizationLayerKernel.cpp +++ b/src/core/NEON/kernels/NEQuantizationLayerKernel.cpp @@ -54,7 +54,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, ITensorInfo *min_max) { // Output tensor auto initialization if not yet initialized - auto_init_if_empty(*output, input->tensor_shape(), 1, DataType::U8, 0); + auto_init_if_empty(*output, input->tensor_shape(), 1, DataType::U8); constexpr unsigned int num_elems_processed_per_iteration = 8; diff --git a/src/core/NEON/kernels/NEROIPoolingLayerKernel.cpp b/src/core/NEON/kernels/NEROIPoolingLayerKernel.cpp index a209a523d3..4d908db77b 100644 --- a/src/core/NEON/kernels/NEROIPoolingLayerKernel.cpp +++ b/src/core/NEON/kernels/NEROIPoolingLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -51,7 +51,7 @@ void NEROIPoolingLayerKernel::configure(const ITensor *input, const IROIArray *r // Output auto inizialitation if not yet initialized TensorShape output_shape(pool_info.pooled_width(), pool_info.pooled_height(), input->info()->dimension(2), rois->num_values()); - auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position()); + auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type()); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); ARM_COMPUTE_ERROR_ON((output->info()->dimension(0) != pool_info.pooled_width()) || (output->info()->dimension(1) != pool_info.pooled_height())); diff --git a/src/core/NEON/kernels/NEReductionOperationKernel.cpp b/src/core/NEON/kernels/NEReductionOperationKernel.cpp index 30d42fa25f..30f21bbf33 100644 --- a/src/core/NEON/kernels/NEReductionOperationKernel.cpp +++ b/src/core/NEON/kernels/NEReductionOperationKernel.cpp @@ -134,7 +134,7 @@ std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITe const TensorShape output_shape = calculate_output_shape(input->tensor_shape(), axis); // Output auto initialization if not yet initialized - auto_init_if_empty(*output, output_shape, 1, input->data_type(), input->fixed_point_position()); + auto_init_if_empty(*output, output_shape, 1, input->data_type()); unsigned int num_elems_processed_per_iteration = 16 / data_size_from_type(input->data_type()); diff --git a/src/core/NEON/kernels/NEReshapeLayerKernel.cpp b/src/core/NEON/kernels/NEReshapeLayerKernel.cpp index 45ba68d9fa..d6f470445f 100644 --- a/src/core/NEON/kernels/NEReshapeLayerKernel.cpp +++ b/src/core/NEON/kernels/NEReshapeLayerKernel.cpp @@ -59,11 +59,10 @@ inline void reshape_tensor(const Window &window, const ITensor *input, ITensor * void NEReshapeLayerKernel::configure(const ITensor *input, ITensor *output) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8, DataType::QS8, DataType::U16, DataType::S16, DataType::QS16, + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8, DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32); ARM_COMPUTE_ERROR_ON_NULLPTR(output); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); ARM_COMPUTE_ERROR_ON(input->info()->tensor_shape().total_size() != output->info()->tensor_shape().total_size()); _input = input; @@ -94,12 +93,10 @@ void NEReshapeLayerKernel::run(const Window &window, const ThreadInfo &info) case DataType::U8: case DataType::S8: case DataType::QASYMM8: - case DataType::QS8: reshape_tensor<uint8_t>(window, _input, _output); break; case DataType::U16: case DataType::S16: - case DataType::QS16: case DataType::F16: reshape_tensor<uint16_t>(window, _input, _output); break; diff --git a/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp b/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp index d91efd267f..9946f002de 100644 --- a/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp +++ b/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp @@ -194,56 +194,7 @@ T sqadd(T a, T b); template <typename T> T sqsub(T a, T b); template <typename T> -T sqmul(T a, T b, int fixed_point_position); - -#define DECLARE_NEON_FUNCTIONS_FOR_FIXED_POINT(TYPET, TYPEU, TAGT, TAGU) \ - inline vec_8_byte_t<TYPET> vqsub(vec_8_byte_t<TYPET> a, vec_8_byte_t<TYPET> b) \ - { \ - return vqsub_##TAGT(a, b); \ - } \ - inline vec_8_byte_t<TYPEU> vqadd(vec_8_byte_t<TYPEU> a, vec_8_byte_t<TYPEU> b) \ - { \ - return vqadd_##TAGU(a, b); \ - } \ - inline vec_16_byte_t<TYPEU> vqadd(vec_16_byte_t<TYPEU> a, vec_16_byte_t<TYPEU> b) \ - { \ - return vqaddq_##TAGU(a, b); \ - } \ - inline vec_8_byte_t<TYPET> vqexp(vec_8_byte_t<TYPET> vec, int fixed_point_position) \ - { \ - return vqexp_q##TAGT(vec, fixed_point_position); \ - } \ - inline auto vmovl(vec_8_byte_t<TYPET> vec)->decltype(vmovl_##TAGT(vec)) \ - { \ - return vmovl_##TAGT(vec); \ - } \ - inline vec_16_byte_t<TYPET> vqrecip(vec_16_byte_t<TYPET> vec, int fixed_point_position) \ - { \ - return vqrecipq_q##TAGT(vec, fixed_point_position); \ - } \ - inline vec_16_byte_t<TYPET> vqmul(vec_16_byte_t<TYPET> a, vec_16_byte_t<TYPET> b, int fixed_point_position) \ - { \ - return vqmulq_q##TAGT(a, b, fixed_point_position); \ - } \ - template <> \ - inline TYPEU sqadd<TYPEU>(TYPEU a, TYPEU b) \ - { \ - return sqadd_q##TAGU(a, b); \ - } \ - inline TYPET sqexp(TYPET val, int fixed_point_position) \ - { \ - return sqexp_q##TAGT(val, fixed_point_position); \ - } \ - template <> \ - inline TYPET sqsub<TYPET>(TYPET a, TYPET b) \ - { \ - return sqsub_q##TAGT(a, b); \ - } \ - template <> \ - inline TYPET sqmul<TYPET>(TYPET a, TYPET b, int fixed_point_position) \ - { \ - return sqmul_q##TAGT(a, b, fixed_point_position); \ - } +T sqmul(T a, T b); #define DECLARE_NEON_FUNCTIONS_FOR_FLOAT(TYPE, TAG) \ inline vec_8_byte_t<TYPE> vadd(vec_8_byte_t<TYPE> a, vec_8_byte_t<TYPE> b) \ @@ -278,9 +229,6 @@ DECLARE_NEON_FUNCTIONS_FOR_TYPE(float16_t, f16) #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ DECLARE_NEON_FUNCTIONS_FOR_TYPE(float, f32) -DECLARE_NEON_FUNCTIONS_FOR_FIXED_POINT(int8_t, int16_t, s8, s16) -DECLARE_NEON_FUNCTIONS_FOR_FIXED_POINT(int16_t, int32_t, s16, s32) - #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC DECLARE_NEON_FUNCTIONS_FOR_FLOAT(float16_t, f16) #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ @@ -373,16 +321,15 @@ namespace Status validate_arguments_logits_1d_max(const ITensorInfo &input, const ITensorInfo &output) { #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input, 1, DataType::QASYMM8, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); #else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input, 1, DataType::QASYMM8, DataType::QS8, DataType::QS16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input, 1, DataType::QASYMM8, DataType::F32); #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ // Validate in case of configured output if(output.total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input, &output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(&input, &output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&input, &output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output.tensor_shape(), TensorShape(input.tensor_shape()).set(0, 1)); } @@ -395,7 +342,7 @@ std::pair<Status, Window> validate_and_configure_window_logits_1d_max(ITensorInf // Softmax across the x dimension const TensorShape output_shape = TensorShape(input.tensor_shape()).set(0, 1); // Output auto initialization if not yet initialized - auto_init_if_empty(output, output_shape, 1, input.data_type(), input.fixed_point_position(), input.quantization_info()); + auto_init_if_empty(output, output_shape, 1, input.data_type(), input.quantization_info()); // Configure kernel window const int input_width = input.valid_region().shape.x(); @@ -488,12 +435,6 @@ void NELogits1DMaxKernel::configure(const ITensor *input, ITensor *output) case DataType::QASYMM8: _func = &logits_1d_max<qasymm8_t>; break; - case DataType::QS8: - _func = &logits_1d_max<qint8_t>; - break; - case DataType::QS16: - _func = &logits_1d_max<qint16_t>; - break; #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: _func = &logits_1d_max<float16_t>; @@ -543,11 +484,12 @@ namespace Status validate_arguments_logits_softmax(const ITensorInfo &input, const ITensorInfo &max, const ITensorInfo &output, const float beta, const ITensorInfo &tmp) { + ARM_COMPUTE_UNUSED(beta); // Check input #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input, 1, DataType::QASYMM8, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); #else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input, 1, DataType::QASYMM8, DataType::QS8, DataType::QS16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input, 1, DataType::QASYMM8, DataType::F32); #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ const bool is_quantized_asymmetric = is_data_type_quantized_asymmetric(input.data_type()); @@ -555,7 +497,6 @@ Status validate_arguments_logits_softmax(const ITensorInfo &input, const ITensor // Check max ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input, &max); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(TensorShape(input.tensor_shape()).set(0, 1), max.tensor_shape()); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(&input, &max); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&input, &max); // Check output if configured @@ -564,19 +505,14 @@ Status validate_arguments_logits_softmax(const ITensorInfo &input, const ITensor const QuantizationInfo output_quantization = is_quantized_asymmetric ? QuantizationInfo(1.f / 256.f, 0) : output.quantization_info(); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input, &output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&input, &output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(&input, &output); ARM_COMPUTE_RETURN_ERROR_ON(output.quantization_info() != output_quantization); } - // Check beta - ARM_COMPUTE_RETURN_ERROR_ON((beta != 1.0f) && is_data_type_fixed_point(input.data_type())); - // Check tmp if configured if(tmp.total_size() != 0) { const DataType tmp_data_type = is_quantized_asymmetric ? DataType::F32 : input.data_type(); ARM_COMPUTE_RETURN_ERROR_ON(tmp.data_type() != tmp_data_type); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(&input, &tmp); // We could potentially reduce tmp memory if we could predict or make an assumption // on the maximum number of threads that will run in parallel. ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&input, &tmp); @@ -727,88 +663,6 @@ void logits_1d_softmax_qasymm8(const ITensor &in, const ITensor &max, void *cons in_it, max_it, out_it); } -template <typename T, typename U> -void logits_1d_softmax_fixed_point(const ITensor &in, const ITensor &max, void *const tmp, - ITensor &out, const float /*beta*/, const Window &window) -{ - const int start_x = in.info()->valid_region().anchor.x(); - const int input_width = in.info()->valid_region().shape.x(); - - const int fixed_point_position = in.info()->fixed_point_position(); - - Iterator in_it(&in, window); - Iterator max_it(&max, window); - Iterator out_it(&out, window); - - execute_window_loop(window, [&](const Coordinates &) - { - /* Get pointers */ - const auto in_ptr = reinterpret_cast<const T *>(in_it.ptr()) + start_x; - const auto out_ptr = reinterpret_cast<T *>(out_it.ptr()) + start_x; - const auto tmp_ptr = reinterpret_cast<T *>(tmp); - - vec_16_byte_t<T> vec_sum_inversed; - - /* Compute exponentials and sum */ - { - /* Get max value */ - const auto max_val = *reinterpret_cast<const T *>(max_it.ptr()); - const auto vec_max = vdup_n<vec_8_byte_t<T>>(max_val); - - /* Init sum to zero */ - auto vec_sum = vdup_n<vec_16_byte_t<U>>(0); - - /* Loop over row and compute exponentials and sum */ - int i = 0; - constexpr int vec_size = vec_size_of(vec_sum); - for(; i <= (input_width - vec_size); i += vec_size) - { - auto vec_elements = vld<vec_8_byte_t<T>>(in_ptr + i); - vec_elements = vqsub(vec_elements, vec_max); - vec_elements = vqexp(vec_elements, fixed_point_position); - vec_sum = vqadd(vec_sum, vmovl(vec_elements)); - vst(tmp_ptr + i, vec_elements); - } - /* Reduce sum */ - const vec_8_byte_t<U> sum_8_byte = vqadd(vget_high(vec_sum), vget_low(vec_sum)); - U sum = reduce_add(sqadd<U>, sum_8_byte); - - /* Run remaining elements */ - for(; i < input_width; ++i) - { - T element = sqexp(sqsub(in_ptr[i], max_val), fixed_point_position); - sum = sqadd<U>(sum, element); - tmp_ptr[i] = element; - } - - const auto qsum = utility::saturate_cast<T>(sum); - vec_sum_inversed = vqrecip(vdup_n<vec_16_byte_t<T>>(qsum), fixed_point_position); - } - - /* Normalize exponentials */ - { - /* Loop over row and compute softmax */ - int i = 0; - constexpr int vec_size = vec_size_of(vec_sum_inversed); - for(; i <= (input_width - vec_size); i += vec_size) - { - const auto vec_in = vld<vec_16_byte_t<T>>(tmp_ptr + i); - const vec_16_byte_t<T> normalized_value = vqmul(vec_in, vec_sum_inversed, fixed_point_position); - vst(out_ptr + i, normalized_value); - } - - const T sum_inversed = vget_lane<0>(vec_sum_inversed); - - /* Run remaining elements */ - for(; i < input_width; ++i) - { - out_ptr[i] = sqmul(tmp_ptr[i], sum_inversed, fixed_point_position); - } - } - }, - in_it, max_it, out_it); -} - template <typename T> void logits_1d_softmax_float(const ITensor &in, const ITensor &max, void *const tmp, ITensor &out, const float beta, const Window &window) @@ -908,12 +762,6 @@ void NELogits1DSoftmaxKernel::configure(const ITensor *input, const ITensor *max case DataType::QASYMM8: _func = &logits_1d_softmax_qasymm8; break; - case DataType::QS8: - _func = &logits_1d_softmax_fixed_point<qint8_t, qint16_t>; - break; - case DataType::QS16: - _func = &logits_1d_softmax_fixed_point<qint16_t, qint32_t>; - break; #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: _func = &logits_1d_softmax_float<float16_t>; diff --git a/src/core/NEON/kernels/NETransposeKernel.cpp b/src/core/NEON/kernels/NETransposeKernel.cpp index 92271378ff..2630159561 100644 --- a/src/core/NEON/kernels/NETransposeKernel.cpp +++ b/src/core/NEON/kernels/NETransposeKernel.cpp @@ -74,7 +74,7 @@ unsigned int num_elems_processed(size_t element_size) Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::QASYMM8, DataType::U16, DataType::S16, DataType::QS16, DataType::U32, DataType::S32, + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8, DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32); @@ -84,7 +84,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output) ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; diff --git a/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp b/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp index 3031a87637..f398409b26 100644 --- a/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp +++ b/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp @@ -105,14 +105,13 @@ TensorShape get_output_shape(const ITensorInfo *input, bool has_bias) Status validate_arguments(const ITensorInfo *input, const ITensorInfo *biases, const ITensorInfo *output) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); if(biases != nullptr) { ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input->data_type())); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases); ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 4) && (biases->num_dimensions() != 1)); ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 5) && (biases->num_dimensions() != 2)); ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 4) && (biases->dimension(0) != input->tensor_shape()[3])); @@ -124,7 +123,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *biases, c { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), get_output_shape(input, biases != nullptr)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); } return Status{}; diff --git a/src/core/TensorInfo.cpp b/src/core/TensorInfo.cpp index 676938a231..b77a47e3e1 100644 --- a/src/core/TensorInfo.cpp +++ b/src/core/TensorInfo.cpp @@ -33,8 +33,8 @@ using namespace arm_compute; TensorInfo::TensorInfo() - : _total_size(0), _fixed_point_position(0), _offset_first_element_in_bytes(0), _strides_in_bytes(), _num_channels(0), _tensor_shape(), _data_type(DataType::UNKNOWN), _format(Format::UNKNOWN), - _is_resizable{ true }, _valid_region{ Coordinates(), _tensor_shape }, _padding{ 0 }, _quantization_info(), _data_layout(DataLayout::NCHW) + : _total_size(0), _offset_first_element_in_bytes(0), _strides_in_bytes(), _num_channels(0), _tensor_shape(), _data_type(DataType::UNKNOWN), _format(Format::UNKNOWN), _is_resizable{ true }, + _valid_region{ Coordinates(), _tensor_shape }, _padding{ 0 }, _quantization_info(), _data_layout(DataLayout::NCHW) { } @@ -42,7 +42,6 @@ TensorInfo::TensorInfo(const ITensorInfo &info) : TensorInfo() { _total_size = info.total_size(); - _fixed_point_position = info.fixed_point_position(); _offset_first_element_in_bytes = info.offset_first_element_in_bytes(); _strides_in_bytes = info.strides_in_bytes(); _num_channels = info.num_channels(); @@ -72,22 +71,22 @@ TensorInfo::TensorInfo(const TensorShape &tensor_shape, Format format) init(tensor_shape, format); } -TensorInfo::TensorInfo(size_t num_channels, DataType data_type, size_t fixed_point_position) +TensorInfo::TensorInfo(size_t num_channels, DataType data_type) : TensorInfo() { - init(TensorShape(), num_channels, data_type, fixed_point_position); + init(TensorShape(), num_channels, data_type); } -TensorInfo::TensorInfo(const TensorShape &tensor_shape, size_t num_channels, DataType data_type, int fixed_point_position) +TensorInfo::TensorInfo(const TensorShape &tensor_shape, size_t num_channels, DataType data_type) : TensorInfo() { - init(tensor_shape, num_channels, data_type, fixed_point_position); + init(tensor_shape, num_channels, data_type); } TensorInfo::TensorInfo(const TensorShape &tensor_shape, size_t num_channels, DataType data_type, QuantizationInfo quantization_info) : TensorInfo() { - init(tensor_shape, num_channels, data_type, 0); + init(tensor_shape, num_channels, data_type); _quantization_info = quantization_info; } @@ -124,34 +123,28 @@ void TensorInfo::init(const TensorShape &tensor_shape, Format format, _format = format; } -void TensorInfo::init(size_t num_channels, DataType data_type, size_t fixed_point_position) +void TensorInfo::init(size_t num_channels, DataType data_type) { - init(TensorShape(), num_channels, data_type, fixed_point_position); + init(TensorShape(), num_channels, data_type); } -void TensorInfo::init(const TensorShape &tensor_shape, size_t num_channels, DataType data_type, int fixed_point_position) +void TensorInfo::init(const TensorShape &tensor_shape, size_t num_channels, DataType data_type) { ARM_COMPUTE_ERROR_ON(num_channels == 0); - ARM_COMPUTE_ERROR_ON(data_type == DataType::QS8 && (fixed_point_position < 1 || fixed_point_position > 6)); - ARM_COMPUTE_ERROR_ON(data_type == DataType::QS16 && (fixed_point_position < 1 || fixed_point_position > 14)); - _fixed_point_position = fixed_point_position; - _data_type = data_type; - _num_channels = num_channels; - _format = Format::UNKNOWN; + _data_type = data_type; + _num_channels = num_channels; + _format = Format::UNKNOWN; set_tensor_shape(tensor_shape); } void TensorInfo::init(const TensorShape &tensor_shape, size_t num_channels, DataType data_type, const Strides &strides_in_bytes, size_t offset_first_element_in_bytes, - size_t total_size_in_bytes, int fixed_point_position) + size_t total_size_in_bytes) { ARM_COMPUTE_ERROR_ON(num_channels == 0); - ARM_COMPUTE_ERROR_ON(data_type == DataType::QS8 && (fixed_point_position < 1 || fixed_point_position > 6)); - ARM_COMPUTE_ERROR_ON(data_type == DataType::QS16 && (fixed_point_position < 1 || fixed_point_position > 14)); - _fixed_point_position = fixed_point_position; _data_type = data_type; _num_channels = num_channels; _format = Format::UNKNOWN; @@ -188,17 +181,14 @@ size_t TensorInfo::init_auto_padding(const TensorShape &tensor_shape, Format for return total_size; } -size_t TensorInfo::init_auto_padding(const TensorShape &tensor_shape, size_t num_channels, DataType data_type, int fixed_point_position) +size_t TensorInfo::init_auto_padding(const TensorShape &tensor_shape, size_t num_channels, DataType data_type) { ARM_COMPUTE_ERROR_ON(num_channels == 0); - ARM_COMPUTE_ERROR_ON(data_type == DataType::QS8 && (fixed_point_position < 1 || fixed_point_position > 6)); - ARM_COMPUTE_ERROR_ON(data_type == DataType::QS16 && (fixed_point_position < 1 || fixed_point_position > 14)); - _fixed_point_position = fixed_point_position; - _data_type = data_type; - _num_channels = num_channels; - _format = Format::UNKNOWN; - _tensor_shape = tensor_shape; + _data_type = data_type; + _num_channels = num_channels; + _format = Format::UNKNOWN; + _tensor_shape = tensor_shape; _valid_region = ValidRegion{ Coordinates(), _tensor_shape }; @@ -371,14 +361,6 @@ ITensorInfo &TensorInfo::set_tensor_shape(const TensorShape &shape) return *this; } -ITensorInfo &TensorInfo::set_fixed_point_position(int fixed_point_position) -{ - ARM_COMPUTE_ERROR_ON(_data_type == DataType::QS8 && (fixed_point_position < 1 || fixed_point_position > 6)); - ARM_COMPUTE_ERROR_ON(_data_type == DataType::QS16 && (fixed_point_position < 1 || fixed_point_position > 14)); - _fixed_point_position = fixed_point_position; - return *this; -} - ITensorInfo &TensorInfo::set_quantization_info(const QuantizationInfo &quantization_info) { _quantization_info = quantization_info; diff --git a/src/core/Utils.cpp b/src/core/Utils.cpp index b1c59924a7..11bdbdafe0 100644 --- a/src/core/Utils.cpp +++ b/src/core/Utils.cpp @@ -24,8 +24,6 @@ #include "arm_compute/core/Utils.h" -#include "arm_compute/core/FixedPoint.h" - #include "support/ToolchainSupport.h" #include <algorithm> @@ -145,10 +143,8 @@ const std::string &arm_compute::string_from_data_type(DataType dt) { DataType::UNKNOWN, "UNKNOWN" }, { DataType::S8, "S8" }, { DataType::U8, "U8" }, - { DataType::QS8, "QS8" }, { DataType::S16, "S16" }, { DataType::U16, "U16" }, - { DataType::QS16, "QS16" }, { DataType::S32, "S32" }, { DataType::U32, "U32" }, { DataType::S64, "S64" }, @@ -353,14 +349,12 @@ void arm_compute::print_consecutive_elements(std::ostream &s, DataType dt, const case DataType::U8: print_consecutive_elements_impl<uint8_t>(s, ptr, n, stream_width, element_delim); break; - case DataType::QS8: case DataType::S8: print_consecutive_elements_impl<int8_t>(s, reinterpret_cast<const int8_t *>(ptr), n, stream_width, element_delim); break; case DataType::U16: print_consecutive_elements_impl<uint16_t>(s, reinterpret_cast<const uint16_t *>(ptr), n, stream_width, element_delim); break; - case DataType::QS16: case DataType::S16: print_consecutive_elements_impl<int16_t>(s, reinterpret_cast<const int16_t *>(ptr), n, stream_width, element_delim); break; @@ -388,12 +382,10 @@ int arm_compute::max_consecutive_elements_display_width(std::ostream &s, DataTyp case DataType::QASYMM8: case DataType::U8: return max_consecutive_elements_display_width_impl<uint8_t>(s, ptr, n); - case DataType::QS8: case DataType::S8: return max_consecutive_elements_display_width_impl<int8_t>(s, reinterpret_cast<const int8_t *>(ptr), n); case DataType::U16: return max_consecutive_elements_display_width_impl<uint16_t>(s, reinterpret_cast<const uint16_t *>(ptr), n); - case DataType::QS16: case DataType::S16: return max_consecutive_elements_display_width_impl<int16_t>(s, reinterpret_cast<const int16_t *>(ptr), n); case DataType::U32: diff --git a/src/runtime/CL/functions/CLDeconvolutionLayer.cpp b/src/runtime/CL/functions/CLDeconvolutionLayer.cpp index 4c1ea5b9a2..9f3dc78022 100644 --- a/src/runtime/CL/functions/CLDeconvolutionLayer.cpp +++ b/src/runtime/CL/functions/CLDeconvolutionLayer.cpp @@ -64,12 +64,10 @@ Status CLDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInf const TensorShape output_shape = deconvolution_output_shape(out_dims, input->tensor_shape(), weights->tensor_shape()); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output, weights); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output, weights); if(bias != nullptr) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, bias); } ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimX) != output_shape.x(), "Output's width is invalid."); @@ -100,7 +98,7 @@ void CLDeconvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const TensorShape output_shape = deconvolution_output_shape(out_dims, input->info()->tensor_shape(), weights->info()->tensor_shape()); // Output auto initialization if not yet initialized - auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position()); + auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type()); // Perform validation step ARM_COMPUTE_ERROR_THROW_ON(CLDeconvolutionLayer::validate(input->info(), weights->info(), bias == nullptr ? nullptr : bias->info(), output->info(), info, inner_border_right, inner_border_top)); @@ -116,7 +114,7 @@ void CLDeconvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const unsigned int out_y = input->info()->dimension(1) + (input->info()->dimension(1) - 1) * (stride_y - 1) + inner_border_top + 2 * info.pad().second; scale_out_shape.set(0, out_x); scale_out_shape.set(1, out_y); - TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type(), input->info()->fixed_point_position()); + TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type()); _scaled_output.allocator()->init(scale_out_info); _scale_f.configure(input, &_scaled_output, BorderSize(inner_border_top, inner_border_right), info); diff --git a/src/runtime/CL/functions/CLDepthConcatenateLayer.cpp b/src/runtime/CL/functions/CLDepthConcatenateLayer.cpp index 26d46a438c..0b26f55a29 100644 --- a/src/runtime/CL/functions/CLDepthConcatenateLayer.cpp +++ b/src/runtime/CL/functions/CLDepthConcatenateLayer.cpp @@ -55,7 +55,7 @@ void CLDepthConcatenateLayer::configure(std::vector<ICLTensor *> inputs_vector, TensorShape output_shape = calculate_depth_concatenate_shape(inputs_vector); // Output auto inizialitation if not yet initialized - auto_init_if_empty(*output->info(), output_shape, 1, inputs_vector[0]->info()->data_type(), inputs_vector[0]->info()->fixed_point_position()); + auto_init_if_empty(*output->info(), output_shape, 1, inputs_vector[0]->info()->data_type()); for(unsigned int i = 0; i < _num_inputs; i++) { diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp index 9248bc559b..273ef96a03 100644 --- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp +++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp @@ -232,7 +232,7 @@ Status CLFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorIn { ARM_COMPUTE_UNUSED(retain_internal_weights); ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 2); diff --git a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp index ace3379618..f1d2924c92 100644 --- a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp +++ b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp @@ -62,7 +62,7 @@ void CLConvolutionLayerReshapeWeights::configure(const ICLTensor *weights, const Status CLConvolutionLayerReshapeWeights::validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(weights); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); if(biases != nullptr) @@ -77,7 +77,6 @@ Status CLConvolutionLayerReshapeWeights::validate(const ITensorInfo *weights, co if((output != nullptr) && (output->total_size() != 0)) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, output); CLWeightsReshapeKernel::validate(weights, biases, output); } @@ -233,7 +232,7 @@ void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor * shape_im2col.set(1, conv_w * conv_h); // FIXME: input->clone() doesn't work with subtensors for grouped convolutions. - TensorInfo im2col_reshaped_info(shape_im2col, 1, data_type, input->info()->fixed_point_position()); + TensorInfo im2col_reshaped_info(shape_im2col, 1, data_type); im2col_reshaped_info.set_quantization_info(input->info()->quantization_info()); _im2col_output.allocator()->init(im2col_reshaped_info); _memory_group.manage(&_im2col_output); @@ -257,7 +256,7 @@ void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor * // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input. const DataType gemm_data_type = _is_quantized ? DataType::S32 : data_type; // FIXME: input->clone() doesn't work with subtensors for grouped convolutions. - TensorInfo info_gemm(shape_gemm, 1, gemm_data_type, input->info()->fixed_point_position()); + TensorInfo info_gemm(shape_gemm, 1, gemm_data_type); info_gemm.set_quantization_info(output->info()->quantization_info()); _gemm_output.allocator()->init(info_gemm); _memory_group.manage(&_gemm_output); @@ -326,10 +325,9 @@ Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights_info.are_reshaped(), "Weights already reshaped are not supported!"); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, weights); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::QASYMM8 && input->data_layout() == DataLayout::NHWC, "NHWC is unsupported for QASYMM8!"); @@ -369,7 +367,6 @@ Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); } - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases); ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(idx_kernels)); ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); } @@ -395,7 +392,7 @@ Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI // Output tensor auto inizialitation if not yet initialized ARM_COMPUTE_RETURN_ON_ERROR(CLConvolutionLayerReshapeWeights::validate(weights, is_quantized ? nullptr : biases, nullptr)); - weights_reshaped_info = TensorInfo(compute_weights_reshaped_shape(*weights, append_bias), 1, data_type, weights->fixed_point_position()); + weights_reshaped_info = TensorInfo(compute_weights_reshaped_shape(*weights, append_bias), 1, data_type); weights_to_use = &weights_reshaped_info; if(!skip_im2col) @@ -408,7 +405,7 @@ Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI } shape_im2col.set(0, mat_weights_rows); shape_im2col.set(1, conv_w * conv_h); - im2col_reshaped_info = TensorInfo(shape_im2col, 1, data_type, input->fixed_point_position()); + im2col_reshaped_info = TensorInfo(shape_im2col, 1, data_type); im2col_reshaped_info.set_quantization_info(input->quantization_info()); ARM_COMPUTE_RETURN_ON_ERROR(CLIm2ColKernel::validate(input, &im2col_reshaped_info, Size2D(kernel_width, kernel_height), conv_info, append_bias, dilation)); gemm_input_to_use = &im2col_reshaped_info; @@ -422,7 +419,7 @@ Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI shape_gemm.set(1, conv_w * conv_h); const DataType gemm_data_type = is_quantized ? DataType::S32 : data_type; // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input. - info_gemm = TensorInfo(shape_gemm, 1, gemm_data_type, input->fixed_point_position()); + info_gemm = TensorInfo(shape_gemm, 1, gemm_data_type); info_gemm.set_quantization_info(output->quantization_info()); gemm_output_to_use = &info_gemm; } @@ -436,7 +433,7 @@ Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); if(!is_nhwc) { - tmp_info = TensorInfo(gemm_output_to_use->tensor_shape(), 1, DataType::QASYMM8, input->fixed_point_position()); + tmp_info = TensorInfo(gemm_output_to_use->tensor_shape(), 1, DataType::QASYMM8); tmp_info.set_quantization_info(output->quantization_info()); gemm_output_staged_to_use = &tmp_info; } diff --git a/src/runtime/CL/functions/CLReductionOperation.cpp b/src/runtime/CL/functions/CLReductionOperation.cpp index 3a5133d91f..2a171c3969 100644 --- a/src/runtime/CL/functions/CLReductionOperation.cpp +++ b/src/runtime/CL/functions/CLReductionOperation.cpp @@ -71,7 +71,6 @@ Status CLReductionOperation::validate(const ITensorInfo *input, const ITensorInf sums_vector[i].set_data_type(input->data_type()); sums_vector[i].set_tensor_shape(shape); sums_vector[i].set_num_channels(input->num_channels()); - sums_vector[i].set_fixed_point_position(input->fixed_point_position()); } // Validate ReductionOperation only on first kernel @@ -105,7 +104,7 @@ void CLReductionOperation::configure(ICLTensor *input, ICLTensor *output, unsign for(unsigned int i = 0; i < _num_of_stages - 1; i++) { shape.set(0, ceil(shape.x() / 128.f)); - _sums_vector[i].allocator()->init(TensorInfo(shape, input->info()->num_channels(), input->info()->data_type(), input->info()->fixed_point_position())); + _sums_vector[i].allocator()->init(TensorInfo(shape, input->info()->num_channels(), input->info()->data_type())); } // Apply ReductionOperation only on first kernel diff --git a/src/runtime/CL/functions/CLWidthConcatenateLayer.cpp b/src/runtime/CL/functions/CLWidthConcatenateLayer.cpp index d5427819c3..5233ff4f52 100644 --- a/src/runtime/CL/functions/CLWidthConcatenateLayer.cpp +++ b/src/runtime/CL/functions/CLWidthConcatenateLayer.cpp @@ -48,7 +48,7 @@ Status CLWidthConcatenateLayer::validate(const std::vector<ITensorInfo *> &input // Output auto inizialitation if not yet initialized TensorInfo tmp_output_info = *output->clone(); TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_width_concatenate_shape(inputs_vector); - auto_init_if_empty(tmp_output_info, output_shape, 1, inputs_vector[0]->data_type(), inputs_vector[0]->fixed_point_position()); + auto_init_if_empty(tmp_output_info, output_shape, 1, inputs_vector[0]->data_type()); unsigned int width_offset = 0; for(const auto &input : inputs_vector) @@ -73,7 +73,7 @@ void CLWidthConcatenateLayer::configure(std::vector<ICLTensor *> inputs_vector, TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_width_concatenate_shape(inputs_vector); // Output auto inizialitation if not yet initialized - auto_init_if_empty(*output->info(), output_shape, 1, inputs_vector[0]->info()->data_type(), inputs_vector[0]->info()->fixed_point_position()); + auto_init_if_empty(*output->info(), output_shape, 1, inputs_vector[0]->info()->data_type()); ARM_COMPUTE_ERROR_THROW_ON(CLWidthConcatenateLayer::validate(inputs_vector_info, output->info())); unsigned int width_offset = 0; diff --git a/src/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.cpp index 67b2ae9d61..5cfd72f724 100644 --- a/src/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.cpp +++ b/src/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.cpp @@ -149,7 +149,7 @@ void GCConvolutionLayer::configure(const IGCTensor *input, const IGCTensor *weig shape_im2col.set(2, 1); // FIXME: input->clone() doesn't work with subtensors for grouped convolutions. - TensorInfo im2col_reshaped_info(shape_im2col, 1, dt, input->info()->fixed_point_position()); + TensorInfo im2col_reshaped_info(shape_im2col, 1, dt); _input_im2col_reshaped.allocator()->init(im2col_reshaped_info); _memory_group.manage(&_input_im2col_reshaped); @@ -160,7 +160,7 @@ void GCConvolutionLayer::configure(const IGCTensor *input, const IGCTensor *weig const DataType gemm_data_type = dt; // FIXME: input->clone() doesn't work with subtensors for grouped convolutions. - TensorInfo info_gemm(shape_gemm, 1, gemm_data_type, input->info()->fixed_point_position()); + TensorInfo info_gemm(shape_gemm, 1, gemm_data_type); _gemm_output.allocator()->init(info_gemm); _memory_group.manage(&_gemm_output); diff --git a/src/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.cpp index 1748a5952b..0c8769b38f 100644 --- a/src/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.cpp +++ b/src/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.cpp @@ -42,11 +42,11 @@ void GCSoftmaxLayer::configure(const IGCTensor *input, IGCTensor *output, float ARM_COMPUTE_ERROR_ON(beta != 1.0f); // Create intermediate tensors shapes - _tmp.allocator()->init(TensorInfo(input->info()->tensor_shape(), input->info()->num_channels(), input->info()->data_type(), input->info()->fixed_point_position())); + _tmp.allocator()->init(TensorInfo(input->info()->tensor_shape(), input->info()->num_channels(), input->info()->data_type())); TensorShape shape = input->info()->tensor_shape(); shape.set(0, 1); - TensorInfo tensor_info_max_sum(shape, input->info()->num_channels(), input->info()->data_type(), input->info()->fixed_point_position()); + TensorInfo tensor_info_max_sum(shape, input->info()->num_channels(), input->info()->data_type()); _max.allocator()->init(tensor_info_max_sum); _sum.allocator()->init(tensor_info_max_sum); diff --git a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp index 8051d6da0e..fda9f57499 100644 --- a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp @@ -63,18 +63,15 @@ Status NEDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInf info.pad().first, info.pad().second, inner_border_right, inner_border_top, stride_x, stride_y); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, bias); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights, bias); if(bias != nullptr) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, bias); } if(output->tensor_shape().total_size() > 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); const TensorShape output_shape = deconvolution_output_shape(out_dims, input->tensor_shape(), weights->tensor_shape()); @@ -117,8 +114,7 @@ void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, con // configure scale function // Init and allocate intermmidiate tensor for output, same size as input but the first two axis are the same as the output tensor - const TensorInfo scale_out_info(compute_deconvolution_shape(*input->info(), stride_x, stride_y, inner_border_right, inner_border_top, info), 1, input->info()->data_type(), - input->info()->fixed_point_position()); + const TensorInfo scale_out_info(compute_deconvolution_shape(*input->info(), stride_x, stride_y, inner_border_right, inner_border_top, info), 1, input->info()->data_type()); _scaled_output.allocator()->init(scale_out_info); // setup the function to convolve the upscaled output diff --git a/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp b/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp index 930f8d5a26..3d47ec2ac2 100644 --- a/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp +++ b/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp @@ -52,7 +52,7 @@ void NEDepthConcatenateLayer::configure(std::vector<ITensor *> inputs_vector, IT TensorShape output_shape = calculate_depth_concatenate_shape(inputs_vector); // Output auto inizialitation if not yet initialized - auto_init_if_empty(*output->info(), output_shape, 1, inputs_vector[0]->info()->data_type(), inputs_vector[0]->info()->fixed_point_position()); + auto_init_if_empty(*output->info(), output_shape, 1, inputs_vector[0]->info()->data_type()); unsigned int depth_offset = 0; for(unsigned int i = 0; i < _num_inputs; ++i) diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp index 83c3e217f3..1d65dde2a6 100644 --- a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp @@ -88,7 +88,7 @@ void NEDepthwiseConvolutionLayer3x3::configure(ITensor *input, const ITensor *we } else { - // Allocate the intermediate accumulator tensor in case of fixed point input + // Allocate the intermediate accumulator tensor in case of quantized input if(_is_quantized) { _accumulator.allocator()->init(TensorInfo(output->info()->tensor_shape(), 1, DataType::S32)); diff --git a/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp index 445864c2a9..40e40c8ffa 100644 --- a/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp @@ -34,7 +34,7 @@ using namespace arm_compute; NEDirectConvolutionLayer::NEDirectConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager) - : _memory_group(std::move(memory_manager)), _output_stage_kernel(), _conv_kernel(), _input_border_handler(), _activationlayer_function(), _accumulator(), _has_bias(false), _is_fixed_point(false), + : _memory_group(std::move(memory_manager)), _output_stage_kernel(), _conv_kernel(), _input_border_handler(), _activationlayer_function(), _accumulator(), _has_bias(false), _is_activationlayer_enabled(false), _dim_split(Window::DimZ) { } @@ -54,26 +54,10 @@ void NEDirectConvolutionLayer::configure(ITensor *input, const ITensor *weights, // Check if bias should be added in the convolution result _has_bias = (bias != nullptr); - // Allocate the intermediate accumulator tensor in case of fixed point input - _is_fixed_point = is_data_type_fixed_point(input->info()->data_type()); - if(_is_fixed_point) + _conv_kernel.configure(input, weights, output, conv_info); + if(_has_bias) { - const DataType promoted_dt = (input->info()->data_type() == DataType::QS8) ? DataType::QS16 : DataType::QS32; - _accumulator.allocator()->init(TensorInfo(output->info()->tensor_shape(), 1, promoted_dt, output->info()->fixed_point_position())); - _memory_group.manage(&_accumulator); - _conv_kernel.configure(input, weights, &_accumulator, conv_info); - - // When no bias is provided, we need to downscale the accumulator tensor - _output_stage_kernel.configure(&_accumulator, bias, output); - _accumulator.allocator()->allocate(); - } - else - { - _conv_kernel.configure(input, weights, output, conv_info); - if(_has_bias) - { - _output_stage_kernel.configure(output, bias); - } + _output_stage_kernel.configure(output, bias); } // Add zero padding XY @@ -92,12 +76,7 @@ Status NEDirectConvolutionLayer::validate(const ITensorInfo *input, const ITenso { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); - DataType data_type = output->data_type(); - if(is_data_type_fixed_point(data_type)) - { - // Promote data type in case of fixed point - data_type = ((data_type == DataType::QS8) ? DataType::QS16 : DataType::QS32); - } + DataType data_type = output->data_type(); TensorInfo accumulator(output->clone()->set_is_resizable(true).reset_padding().set_data_type(data_type)); // Validate Convolution kernel @@ -129,7 +108,7 @@ void NEDirectConvolutionLayer::run() _memory_group.acquire(); NEScheduler::get().schedule(&_conv_kernel, _dim_split); - if(_has_bias || _is_fixed_point) + if(_has_bias) { NEScheduler::get().schedule(&_output_stage_kernel, Window::DimY); } diff --git a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp index 5b9f182bcb..3126823e9c 100644 --- a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp +++ b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp @@ -83,7 +83,7 @@ void NEFullyConnectedLayerReshapeWeights::configure(const ITensor *input, ITenso Status NEFullyConnectedLayerReshapeWeights::validate(const ITensorInfo *input, const ITensorInfo *output, bool transpose_weights, bool is_batched_fc_layer) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 2); ARM_COMPUTE_RETURN_ERROR_ON_MSG(!transpose_weights && !is_batched_fc_layer, "Configuration transpose_weights=false & is_batched_fc_layer=false not supported"); @@ -233,9 +233,8 @@ void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weigh Status NEFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose_weights, bool are_weights_reshaped) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, weights, output); const int num_batch_dimensions = std::max(0, static_cast<int>(output->tensor_shape().num_dimensions()) - 1); const int num_input_dimensions = input->tensor_shape().num_dimensions() - num_batch_dimensions; diff --git a/src/runtime/NEON/functions/NEGEMM.cpp b/src/runtime/NEON/functions/NEGEMM.cpp index a98309d304..795ffc5d1c 100644 --- a/src/runtime/NEON/functions/NEGEMM.cpp +++ b/src/runtime/NEON/functions/NEGEMM.cpp @@ -46,7 +46,7 @@ NEGEMM::NEGEMM(std::shared_ptr<IMemoryManager> memory_manager) void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *d, float alpha, float beta, const GEMMInfo &gemm_info) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::F32, DataType::F16, DataType::QS8, DataType::QS16); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::F32, DataType::F16); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, d); ARM_COMPUTE_ERROR_ON_MSG(a->info()->dimension(0) != b->info()->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B"); ARM_COMPUTE_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported"); @@ -54,7 +54,7 @@ void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITe if(c != nullptr) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(c, 1, DataType::F32, DataType::F16, DataType::QS8, DataType::QS16); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(c, 1, DataType::F32, DataType::F16); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, c); ARM_COMPUTE_ERROR_ON_MSG(a->info()->dimension(1) != c->info()->dimension(1), "The C matrix must have the same number of rows as the matrix A"); ARM_COMPUTE_ERROR_ON_MSG(b->info()->dimension(0) != c->info()->dimension(0), "The C matrix must have the same number of columns as the matrix B"); @@ -103,8 +103,8 @@ void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITe shape_tmp_b.set(0, b->info()->dimension(1) * transpose_w); shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / static_cast<float>(transpose_w))); - TensorInfo info_a(shape_tmp_a, 1, a->info()->data_type(), a->info()->fixed_point_position()); - TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type(), a->info()->fixed_point_position()); + TensorInfo info_a(shape_tmp_a, 1, a->info()->data_type()); + TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type()); _tmp_a.allocator()->init(info_a); _tmp_b.allocator()->init(info_b); diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp index d4400b8864..94ef4e7b32 100644 --- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp @@ -90,12 +90,11 @@ void NEConvolutionLayerReshapeWeights::configure(const ITensor *weights, const I Status NEConvolutionLayerReshapeWeights::validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose1xW) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); if(!is_data_type_quantized_asymmetric(weights->data_type())) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, output); } // Check if bias are present, if yes they will be embedded to the weights matrix const bool append_bias = (biases != nullptr); @@ -104,7 +103,6 @@ Status NEConvolutionLayerReshapeWeights::validate(const ITensorInfo *weights, co { ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(weights->data_type())); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, biases); ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3)); ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); } @@ -173,9 +171,8 @@ Status validate_and_initialize_values(const ITensorInfo *input, const ITensorInf unsigned int &mat_weights_cols, unsigned int &mat_weights_rows, unsigned int &conv_w, unsigned int &conv_h, const Size2D &dilation) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, weights); DataLayout data_layout = input->data_layout(); @@ -201,7 +198,6 @@ Status validate_and_initialize_values(const ITensorInfo *input, const ITensorInf { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); } - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases); ARM_COMPUTE_RETURN_ERROR_ON(!weights_info.are_reshaped() && biases->dimension(0) != weights->dimension(3)); ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); } @@ -287,10 +283,9 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig ARM_COMPUTE_ERROR_THROW_ON(status); - _is_prepared = false; - _original_weights = weights; - const unsigned int fixed_point_position = input->info()->fixed_point_position(); - const ITensor *biases_to_use = (_append_bias) ? biases : nullptr; + _is_prepared = false; + _original_weights = weights; + const ITensor *biases_to_use = (_append_bias) ? biases : nullptr; bool run_optimised = dt == DataType::F32; @@ -300,7 +295,7 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig TensorShape reshaped_weights_shape{ mat_weights_cols, mat_weights_rows }; // Create tensor to store the reshaped weights - _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt, fixed_point_position)); + _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt)); _reshape_weights.configure(weights, biases, &_weights_reshaped, false /* 1xW transpose */); weights = &_weights_reshaped; } @@ -336,7 +331,7 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig } // Create tensor to store the reshaped weights - _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt, fixed_point_position)); + _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt)); _reshape_weights.configure(weights, biases_to_use, &_weights_reshaped, _is_interleaved /* 1xW transpose */); weights = &_weights_reshaped; } @@ -372,7 +367,7 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig shape_gemm.set(1, mat_input_rows); const DataType gemm_data_type = _is_quantized ? DataType::S32 : dt; // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input. - TensorInfo info_gemm(shape_gemm, 1, gemm_data_type, input->info()->fixed_point_position()); + TensorInfo info_gemm(shape_gemm, 1, gemm_data_type); info_gemm.set_quantization_info(output->info()->quantization_info()); _gemm_output.allocator()->init(info_gemm); _memory_group.manage(&_gemm_output); diff --git a/src/runtime/NEON/functions/NENormalizationLayer.cpp b/src/runtime/NEON/functions/NENormalizationLayer.cpp index af98ac1f17..f00114f930 100644 --- a/src/runtime/NEON/functions/NENormalizationLayer.cpp +++ b/src/runtime/NEON/functions/NENormalizationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -41,7 +41,7 @@ void NENormalizationLayer::configure(const ITensor *input, ITensor *output, cons { ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); - TensorInfo tensor_info(input->info()->tensor_shape(), 1, input->info()->data_type(), input->info()->fixed_point_position()); + TensorInfo tensor_info(input->info()->tensor_shape(), 1, input->info()->data_type()); _input_squared.allocator()->init(tensor_info); // Manage intermediate buffers |