From a046e164b96a8441b2fa14ef578f7db46a0e97da Mon Sep 17 00:00:00 2001 From: Michele Di Giorgio Date: Tue, 8 Oct 2019 09:36:26 +0100 Subject: COMPMID-2600: Implement a new and generic depthwise convolution for CL QASYMM8 NHWC The NCHW case is supported at function level by permuting the inputs/outputs to NHWC. This patch also removes CLDirectConvolutionLayerOutputStageKernel which is deprecated and some kernels which were only used in the generic case of depthwise convolution. Change-Id: I91e0f02d0a2f4a4a352e08c248e648944137fe68 Signed-off-by: Michele Di Giorgio Reviewed-on: https://review.mlplatform.org/c/2056 Reviewed-by: Giorgio Arena Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Gian Marco Iodice --- src/core/CL/cl_kernels/activation_quant_helpers.h | 2 +- src/core/CL/cl_kernels/depthwise_convolution.cl | 187 +-------------------- .../cl_kernels/depthwise_convolution_quantized.cl | 165 +++++++++++++++++- .../direct_convolution_1x1_3x3_5x5_quantized.cl | 78 +-------- src/core/CL/cl_kernels/helpers.h | 22 +++ src/core/CL/cl_kernels/helpers_asymm.h | 2 + 6 files changed, 189 insertions(+), 267 deletions(-) (limited to 'src/core/CL/cl_kernels') diff --git a/src/core/CL/cl_kernels/activation_quant_helpers.h b/src/core/CL/cl_kernels/activation_quant_helpers.h index 402e7ac41f..0e4eb2b32e 100644 --- a/src/core/CL/cl_kernels/activation_quant_helpers.h +++ b/src/core/CL/cl_kernels/activation_quant_helpers.h @@ -41,7 +41,7 @@ inline TYPE relu_op(TYPE x) // Bounded RELU Activation inline TYPE brelu_op(TYPE x) { - return min((TYPE)A_VAL, max(CONST_0, x)); + return min((TYPE)A_VAL, max((TYPE)CONST_0, x)); } // Lower Upper Bounded RELU Activation inline TYPE lu_brelu_op(TYPE x) diff --git a/src/core/CL/cl_kernels/depthwise_convolution.cl b/src/core/CL/cl_kernels/depthwise_convolution.cl index 1b2f5cccaa..3a227282ff 100644 --- a/src/core/CL/cl_kernels/depthwise_convolution.cl +++ b/src/core/CL/cl_kernels/depthwise_convolution.cl @@ -782,173 +782,6 @@ __kernel void depthwise_convolution_reshape_weights( } #endif // defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DST_WIDTH) -#if defined(NCHW) -#define in_stride_x src_stride_x -#define in_stride_y src_stride_y -#define in_stride_z src_stride_z -#define out_stride_x dst_stride_x -#define out_stride_y dst_stride_y -#define out_stride_z dst_stride_z -#else //defined(NCHW) -#define in_stride_x src_stride_y -#define in_stride_y src_stride_z -#define in_stride_z src_stride_x -#define out_stride_x dst_stride_y -#define out_stride_y dst_stride_z -#define out_stride_z dst_stride_x -#endif //defined(NCHW) - -#if defined(SRC_WIDTH) && defined(DATA_TYPE) -/** This kernel reshapes each of the tensor's low three dimensions to single rows. - * - * @note Datatype and source width should be given as a preprocessor argument using -DDATA_TYPE=type and -DSRC_WIDTH=width. e.g. -DSRC_WIDTH=128 - * - * @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) - * @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 source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * 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. 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_offset_first_element_in_bytes The offset of the first element in the destination tensor - * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: F16/F32 - * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) - * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector - */ -__kernel void depthwise_convolution_reshape_weights_generic( - TENSOR3D_DECLARATION(src), - IMAGE_DECLARATION(dst) -#ifdef HAS_BIAS - , - VECTOR_DECLARATION(biases) -#endif /* HAS_BIAS */ -) -{ -#ifdef HAS_BIAS - Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); -#endif /* HAS_BIAS */ - - __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + get_global_id(1) * in_stride_y + get_global_id(2) * in_stride_z; - __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + get_global_id(1) * SRC_WIDTH * dst_stride_x + get_global_id(2) * dst_stride_y; - - for(int i = 0; i < SRC_WIDTH; ++i, input_ptr += in_stride_x) - { - *((__global DATA_TYPE *)(output_ptr + i * dst_stride_x)) = *((__global DATA_TYPE *)input_ptr); - } - -#if defined(HAS_BIAS) - if(get_global_id(1) == 0) - { - *((__global DATA_TYPE *)(output_ptr + SRC_WIDTH * get_global_size(1) * dst_stride_x)) = *((__global DATA_TYPE *)(biases.ptr + get_global_id(2) * biases_stride_x)); - } -#endif // defined(HAS_BIAS) -} -#endif //defined(SRC_WIDTH) && defined(DATA_TYPE) - -#if defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP) && defined(PAD_RIGHT) && defined(PAD_BOTTOM) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(DATA_TYPE) && defined(PAD_VALUE) && defined(DEPTH_MULTIPLIER) && defined(DILATION_X) && defined(DILATION_Y) -/** This kernel performs a reshaping of the input tensor to a tensor used to perform depthwise convolution using vector to matrix multiplication. - * - * @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 - * @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 - * - * @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) - * @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 source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_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 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 - */ -__kernel void depthwise_im2col(TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst)) -{ - Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); - - const int src_pixel_linear = get_global_id(1) * STRIDE_X; - const int full_length = SRC_WIDTH + PAD_LEFT + PAD_RIGHT; - const int max_initial_x = STRIDE_X * (((full_length - (KERNEL_WIDTH + (KERNEL_WIDTH - 1) * (DILATION_X - 1))) / STRIDE_X) + 1); - - const int src_x = -PAD_LEFT + src_pixel_linear % max_initial_x; - const int src_y = -PAD_TOP + src_pixel_linear / max_initial_x * STRIDE_Y; - const int src_z = get_global_id(2) / DEPTH_MULTIPLIER; - - __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + src_z * in_stride_z; - __global DATA_TYPE *output_ptr = ((__global DATA_TYPE *)(dst.ptr)); - - for(int y = src_y; y < src_y + KERNEL_HEIGHT + (KERNEL_HEIGHT - 1) * (DILATION_Y - 1); y += DILATION_Y) - { - for(int x = src_x; x < src_x + KERNEL_WIDTH + (KERNEL_WIDTH - 1) * (DILATION_X - 1); x += DILATION_X, ++output_ptr) - { - if(x < 0 || x >= SRC_WIDTH || y < 0 || y >= SRC_HEIGHT) - { - *output_ptr = PAD_VALUE; - } - else - { - *output_ptr = *((__global DATA_TYPE *)(input_ptr + x * in_stride_x + y * in_stride_y)); - } - } - } -#if defined(HAS_BIAS) - *output_ptr = (DATA_TYPE)(1); -#endif // defined(HAS_BIAS) -} - -#endif //defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP) && defined(PAD_RIGHT) && defined(PAD_BOTTOM) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(SRC_WIDTH) && defined(DATA_TYPE) && defined(PAD_VALUE) && defined(DEPTH_MULTIPLIER) - -#if defined(CONV_WIDTH) && defined(CONV_HEIGHT) && defined(DATA_TYPE) - -/** This kernel performs a reshaping of the output of the depthwise generic convolution. - * - * @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: 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[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 - */ -__kernel void depthwise_vector_to_tensor( - VECTOR_DECLARATION(src), - TENSOR3D_DECLARATION(dst)) -{ - Vector src = CONVERT_TO_VECTOR_STRUCT(src); - - const int patch_size = CONV_WIDTH * CONV_HEIGHT; - const int id0 = get_global_id(0); - const int z = id0 / patch_size; - const int index2D = id0 - z * patch_size; - - __global uchar *out_ptr = dst_ptr + dst_offset_first_element_in_bytes + index2D % CONV_WIDTH * out_stride_x + index2D / CONV_WIDTH * out_stride_y + z * out_stride_z; - *((__global DATA_TYPE *)out_ptr) = *((__global DATA_TYPE *)src.ptr); -} - -#endif //defined(CONV_WIDTH) && defined(CONV_HEIGHT) && defined(DATA_TYPE) - #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(DEPTH_MULTIPLIER) && defined(DST_CHANNELS) && defined(IS_F16) #if defined(CONV_STRIDE_X) #if CONV_STRIDE_X == 1 @@ -1478,7 +1311,7 @@ __kernel void depthwise_convolution_3x3_stridex2_stridey2_bifrost_f16( //3x3 Convolution of elements starting in 0th row pixels0 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(src_addr, src.stride_x, src.stride_y, 0, weights_addr, weights_stride_y); //3x3 Convolution of elements starting in 2nd row - pixels1 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(src_addr, src.stride_x, src.stride_y, 2, weights_addr, weights_stride_y); + pixels1 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(src_addr, src.stride_x, src.stride_y, 2, weights_addr, weights_stride_y); #endif /* DILATION_X==1 && DILATION_Y==1 */ #ifdef HAS_BIAS @@ -1556,23 +1389,17 @@ __kernel void dwc_MxN_native_fp_nhwc( int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y int b = get_global_id(2) / (int)DST_DEPTH; // batch #else // defined(DST_DEPTH) - int z = get_global_id(2); // spatial coordinate y + int z = get_global_id(2); // spatial coordinate y #endif // defined(DST_DEPTH) - __global uchar *s_addr = src_ptr + src_offset_first_element_in_bytes + - x * sizeof(DATA_TYPE) * (int)N0; + __global uchar *s_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * (int)N0; - __global uchar *d_addr = dst_ptr + dst_offset_first_element_in_bytes + - x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0 + - y * dst_stride_y + - z * dst_stride_z; + __global uchar *d_addr = dst_ptr + dst_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0 + y * dst_stride_y + z * dst_stride_z; - __global uchar *w_addr = weights_ptr + weights_offset_first_element_in_bytes + - x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0; + __global uchar *w_addr = weights_ptr + weights_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0; #if defined(HAS_BIAS) - __global uchar *b_addr = biases_ptr + biases_offset_first_element_in_bytes + - x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0; + __global uchar *b_addr = biases_ptr + biases_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0; #endif // defined(HAS_BIAS) #if defined(DST_DEPTH) @@ -1611,7 +1438,7 @@ __kernel void dwc_MxN_native_fp_nhwc( #if GPU_ARCH == GPU_ARCH_MIDGARD res += i * w; #else // GPU_ARCH == GPU_ARCH_MIDGARD - res = fma(i, w, res); + res = fma(i, w, res); #endif // GPU_ARCH == GPU_ARCH_MIDGARD } x_coord_tmp += DILATION_X; diff --git a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl index 8f2e441693..10872d460a 100644 --- a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl +++ b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl @@ -24,19 +24,30 @@ #include "helpers_asymm.h" -#if defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && ((defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)) || defined(REAL_MULTIPLIER)) - -#if defined(ACTIVATION_TYPE) && defined(CONST_0) -#define DATA_TYPE uchar #ifndef VEC_SIZE +#if defined(N0) +#define VEC_SIZE N0 +#else /* defined(N0) */ #define VEC_SIZE 8 +#endif /* defined(N0) */ #endif /* VEC_SIZE */ + +#if defined(ACTIVATION_TYPE) && defined(CONST_0) +#define DATA_TYPE uchar #include "activation_layer_quant.cl" #define ACTIVATION_FUNC(x) PERFORM_ACTIVATION_QUANT(ACTIVATION_TYPE, x) #else /* defined(ACTIVATION_TYPE) && defined(CONST_0) */ #define ACTIVATION_FUNC(x) (x) #endif /* defined(ACTIVATION_TYPE) && defined(CONST_0) */ +#define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE) +#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE) +#define VEC_UCHAR VEC_DATA_TYPE(uchar, VEC_SIZE) +#define VEC_USHORT VEC_DATA_TYPE(ushort, VEC_SIZE) +#define VEC_SHORT VEC_DATA_TYPE(short, VEC_SIZE) + +#if defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && ((defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)) || defined(REAL_MULTIPLIER)) + #if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) #if defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8) #define ARM_DOT(x, y, val) val = arm_dot_acc((x), (y), val); @@ -635,11 +646,6 @@ __kernel void dwc_3x3_native_qasymm8_dot8_nchw( #define asymm_mult_by_quant_multiplier_less_than_one(x, y, z) ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(x, y, z, VEC_SIZE) -#define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE) -#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE) -#define VEC_UCHAR VEC_DATA_TYPE(uchar, VEC_SIZE) -#define VEC_USHORT VEC_DATA_TYPE(ushort, VEC_SIZE) - #define MULTIPLY_ADD(x, y, acc) acc += CONVERT(CONVERT(x, VEC_USHORT) * CONVERT(y, VEC_USHORT), VEC_INT) #if WEIGHTS_OFFSET != 0 @@ -1375,3 +1381,144 @@ __kernel void dwc_3x3_reshaped_qasymm8_dot8_stride1_nhwc( #endif // defined(VEC_SIZE) && defined(SRC_DIM_1) && defined(SRC_DIM_2) && defined(CONV_PAD_TOP) && defined(CONV_PAD_LEFT) #endif // defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && ((defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)) || defined(REAL_MULTIPLIER)) + +#if defined(SRC_DIM1) && defined(SRC_DIM2) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(N0) && defined(DILATION_X) && defined(DILATION_Y) && defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && defined(CONV_PAD_LEFT) && defined(CONV_PAD_TOP) && defined(INPUT_OFFSET) && defined(WEIGHTS_OFFSET) && defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT) +/** This function computes the depthwise convolution for NHWC data layout. This kernel assumes that the weights tensor is NOT reshaped + * + * @note The number of elements processed must be passed at compile time using -DN0 (e.g. -DN0=2) + * @note The depth multiplier must be passed at compile time using -DDEPTH_MULTIPLIER (e.g. -DDEPTH_MULTIPLIER=1) + * @note The first dimension of the input tensor must be passed at compile time using -DSRC_DIM1 (e.g. -DSRC_DIM1=112) + * @note The second dimension of the input tensor must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM2=80) + * @note The kernel width must be passed at compile time using -DKERNEL_WIDTH (e.g. -DKERNEL_WIDTH=5) + * @note The kernel height must be passed at compile time using -DKERNEL_HEIGHT (e.g. -DKERNEL_HEIGHT=5) + * @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1) + * @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1) + * @note The convolution stride along the width must be passed at compile time using -DCONV_STRIDE_X (e.g. -DCONV_STRIDE_Y=X) + * @note The convolution stride along the height must be passed at compile time using -DCONV_STRIDE_Y (e.g. -DCONV_STRIDE_Y=1) + * @note It is possible to select the activation function to apply using -DACTIVATION_TYPE e.g. -DACTIVATION_TYPE=relu + * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8 + * @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_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_y * number of elements along Z processed per workitem(in bytes) + * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) + * @param[in] src_step_w src_stride_w * number of elements along W 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] dst_ptr Pointer to the destination tensor. Supported data types: same as 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 Y processed per workitem(in bytes) + * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes) + * @param[in] dst_step_w dst_stride_w * number of elements along W 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] weights_ptr Pointer to the weights tensor. Supported data types: QASYMM8 + * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) + * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) + * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) + * @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes) + * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor + * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: same as src_ptr + * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) + * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector + */ +__kernel void dwc_MxN_native_quantized8_nhwc( + TENSOR4D_DECLARATION(src), + TENSOR4D_DECLARATION(dst), + TENSOR3D_DECLARATION(weights), +#if defined(HAS_BIAS) + VECTOR_DECLARATION(biases) +#endif // defined(HAS_BIAS) +) +{ + int x = get_global_id(0); // channels + int y = get_global_id(1); // spatial coordinate x +#if defined(DST_DEPTH) + int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y + int b = get_global_id(2) / (int)DST_DEPTH; // batch +#else // defined(DST_DEPTH) + int z = get_global_id(2); // spatial coordinate y +#endif // defined(DST_DEPTH) + + __global uchar *s_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(uchar) * (int)N0; + + __global uchar *d_addr = dst_ptr + dst_offset_first_element_in_bytes + x * sizeof(uchar) * (int)DEPTH_MULTIPLIER * (int)N0 + y * dst_stride_y + z * dst_stride_z; + + __global uchar *w_addr = weights_ptr + weights_offset_first_element_in_bytes + x * sizeof(uchar) * (int)DEPTH_MULTIPLIER * (int)N0; + +#if defined(HAS_BIAS) + __global uchar *b_addr = biases_ptr + biases_offset_first_element_in_bytes + x * sizeof(int) * (int)DEPTH_MULTIPLIER * (int)N0; +#endif // defined(HAS_BIAS) + +#if defined(DST_DEPTH) + s_addr += b * src_stride_w; + d_addr += b * dst_stride_w; +#endif // defined(DST_DEPTH) + +#if DEPTH_MULTIPLIER > 1 + for(int d = 0; d < (int)DEPTH_MULTIPLIER; ++d) + { +#endif // DEPTH_MULTIPLIER > 1 + // Each work-item computes N0x1x1 elements + VEC_SHORT res = 0; + + int x_coord = y * CONV_STRIDE_X - (int)CONV_PAD_LEFT; + int y_coord = z * CONV_STRIDE_Y - (int)CONV_PAD_TOP; + + for(int yk = 0; yk < KERNEL_HEIGHT; ++yk) + { + if(y_coord >= 0 && y_coord < SRC_DIM2) + { + int x_coord_tmp = x_coord; + + for(int xk = 0; xk < KERNEL_WIDTH; ++xk) + { + if(x_coord_tmp >= 0 && x_coord_tmp < SRC_DIM1) + { + int s_offset = x_coord_tmp * (int)src_stride_y + y_coord * (int)src_stride_z; + int w_offset = xk * weights_stride_y + yk * weights_stride_z; + + // Load input and weights values + VEC_SHORT i = CONVERT(VLOAD(N0)(0, (__global uchar *)(s_addr + s_offset)), VEC_SHORT); + VEC_SHORT w = CONVERT(VLOAD(N0)(0, (__global uchar *)(w_addr + w_offset)), VEC_SHORT); + + res += (i + (VEC_SHORT)INPUT_OFFSET) * (w + (VEC_SHORT)WEIGHTS_OFFSET); + } + x_coord_tmp += DILATION_X; + } + } + y_coord += DILATION_Y; + } + +#if defined(HAS_BIAS) + VEC_SHORT bias = CONVERT(VLOAD(N0)(0, (__global int *)(b_addr)), VEC_SHORT); + res += bias; +#endif // defined(HAS_BIAS) + + res = CONVERT(ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(CONVERT(res, VEC_INT), OUTPUT_MULTIPLIER, OUTPUT_SHIFT, N0), VEC_SHORT); + res += (VEC_SHORT)OUTPUT_OFFSET; + + VEC_UCHAR res1 = CONVERT_SAT(res, VEC_UCHAR); + + VSTORE(N0) + (ACTIVATION_FUNC(res1), 0, (__global uchar *)(d_addr)); + +#if DEPTH_MULTIPLIER > 1 + w_addr += sizeof(uchar); + d_addr += sizeof(uchar); +#if defined(HAS_BIAS) + b_addr += sizeof(int); +#endif // defined(HAS_BIAS) + } +#endif // DEPTH_MULTIPLIER > 1 +} +#endif // defined(SRC_DIM1) && defined(SRC_DIM2) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defiend(N0) && defined(DILATION_X) && defined(DILATION_Y) && defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && defined(CONV_PAD_LEFT) && defined(CONV_PAD_TOP) && defined(INPUT_OFFSET) && defined(WEIGHTS_OFFSET) && defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT) diff --git a/src/core/CL/cl_kernels/direct_convolution_1x1_3x3_5x5_quantized.cl b/src/core/CL/cl_kernels/direct_convolution_1x1_3x3_5x5_quantized.cl index 83da76785b..5ad9afb23c 100644 --- a/src/core/CL/cl_kernels/direct_convolution_1x1_3x3_5x5_quantized.cl +++ b/src/core/CL/cl_kernels/direct_convolution_1x1_3x3_5x5_quantized.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -247,79 +247,3 @@ __kernel void direct_convolution_1x1_3x3_5x5_quantized( vstore8(convert_uchar8_sat(pixels0), 0, (__global uchar *)dst.ptr); } #endif // defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) - -#if defined(VEC_SIZE) - -#define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE) -#define CONVERT_SAT_UCHAR_STR(x, size) (convert_uchar##size##_sat((x))) -#define CONVERT_SAT_UCHAR(x, size) CONVERT_SAT_UCHAR_STR(x, size) - -/** This function computes the output stage of a depthwise convolution. - * - * @param[in] src_ptr Pointer to the source image. Supported data types: QASYMM8 - * @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) - * @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 image - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: QASYMM8 - * @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 Y 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] bias_ptr (Optional) Pointer to the biases vector. Supported data types: S32 - * @param[in] bias_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) - * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector - * @param[in] output_offset Quantized offset of zero point of the output tensor data range - * @param[in] output_multiplier Output scale multiplier - * @param[in] output_shift Output scale divisor exponent - */ -__kernel void output_stage_quantized( - TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst), -#if defined(HAS_BIAS) - VECTOR_DECLARATION(bias), -#endif //defined(HAS_BIAS) - int output_offset, - int output_multiplier, - int output_shift) -{ - Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src); - Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst); -#if defined(HAS_BIAS) - Vector bias = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bias); -#endif //defined(HAS_BIAS) - - // Load input - VEC_INT vals = VLOAD(VEC_SIZE)(0, (__global int *)(src.ptr)); - -#if defined(HAS_BIAS) - // Load and add bias -#if defined(NCHW) - int bias_value = *((__global int *)(vector_offset(&bias, get_global_id(2)))); -#else // defined(NCHW) - VEC_INT bias_value = VLOAD(VEC_SIZE)(0, ((__global int *)(vector_offset(&bias, get_global_id(0) * VEC_SIZE)))); -#endif // defined(NCHW) - - vals += (VEC_INT)(bias_value); -#endif //defined(HAS_BIAS) - - vals = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(vals, output_multiplier, output_shift, VEC_SIZE); - vals = vals + output_offset; - - // Store result in dst - VSTORE(VEC_SIZE) - (CONVERT_SAT_UCHAR(vals, VEC_SIZE), 0, (__global uchar *)dst.ptr); -} - -#undef VEC_INT -#undef CONVERT_SAT_UCHAR_STR -#undef CONVERT_SAT_UCHAR - -#endif // defined(VEC_SIZE) diff --git a/src/core/CL/cl_kernels/helpers.h b/src/core/CL/cl_kernels/helpers.h index f7f208529a..8117c1e519 100644 --- a/src/core/CL/cl_kernels/helpers.h +++ b/src/core/CL/cl_kernels/helpers.h @@ -150,6 +150,28 @@ #define convert_half8_sat convert_half8 #define convert_half16_sat convert_half16 +#define convert_float1 convert_float +#define convert_half1 convert_half +#define convert_char1 convert_char +#define convert_uchar1 convert_uchar +#define convert_short1 convert_short +#define convert_ushort1 convert_ushort +#define convert_int1 convert_int +#define convert_uint1 convert_uint +#define convert_long1 convert_long +#define convert_ulong1 convert_ulong +#define convert_double1 convert_double + +#define convert_char1_sat convert_char_sat +#define convert_uchar1_sat convert_uchar_sat +#define convert_short1_sat convert_short_sat +#define convert_ushort1_sat convert_ushort_sat +#define convert_int1_sat convert_int_sat +#define convert_uint1_sat convert_uint_sat +#define convert_long1_sat convert_long_sat +#define convert_ulong1_sat convert_ulong_sat +#define convert_double1_sat convert_double_sat + #define VEC_DATA_TYPE_STR(type, size) type##size #define VEC_DATA_TYPE(type, size) VEC_DATA_TYPE_STR(type, size) diff --git a/src/core/CL/cl_kernels/helpers_asymm.h b/src/core/CL/cl_kernels/helpers_asymm.h index 53e6719cd7..57ecccc2b2 100644 --- a/src/core/CL/cl_kernels/helpers_asymm.h +++ b/src/core/CL/cl_kernels/helpers_asymm.h @@ -381,11 +381,13 @@ DEQUANTIZE_IMPL(uchar, 4) DEQUANTIZE_IMPL(ushort, 4) DEQUANTIZE_IMPL(short, 4) +ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(1) ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(2) ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(4) ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(8) ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(16) +ASYMM_MULT_IMPL(1) ASYMM_MULT_IMPL(2) ASYMM_MULT_IMPL(4) ASYMM_MULT_IMPL(8) -- cgit v1.2.1