diff options
| author | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2017-08-15 11:45:22 +0100 |
|---|---|---|
| committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:35:24 +0000 |
| commit | edfa9f463bed084f8b0953557202b2a1e56da817 (patch) | |
| tree | 5d1e92926d112fde05dcbc61324d96f73f692390 | |
| parent | dc460f13ee65e27b2a428e44c2d80afb1f516a99 (diff) | |
| download | ComputeLibrary-edfa9f463bed084f8b0953557202b2a1e56da817.tar.gz | |
COMPMID-477 - Optimized batched case in CLConvolutionLayer
Change-Id: I4ef18f49f1da0cb816aaa0762466b940792c15ed
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/84162
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
| -rw-r--r-- | arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h | 18 | ||||
| -rw-r--r-- | arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h | 42 | ||||
| -rw-r--r-- | arm_compute/runtime/CL/functions/CLGEMM.h | 2 | ||||
| -rw-r--r-- | src/core/CL/CLKernelLibrary.cpp | 15 | ||||
| -rw-r--r-- | src/core/CL/cl_kernels/gemm.cl | 568 | ||||
| -rw-r--r-- | src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp | 4 | ||||
| -rw-r--r-- | src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp | 96 | ||||
| -rw-r--r-- | src/runtime/CL/functions/CLConvolutionLayer.cpp | 5 | ||||
| -rw-r--r-- | src/runtime/CL/functions/CLFullyConnectedLayer.cpp | 239 | ||||
| -rw-r--r-- | src/runtime/CL/functions/CLGEMM.cpp | 30 | ||||
| -rw-r--r-- | tests/model_objects/AlexNet.h | 4 | ||||
| -rw-r--r-- | tests/networks_new/AlexNetNetwork.h | 4 | ||||
| -rw-r--r-- | tests/validation_new/CL/FullyConnectedLayer.cpp | 14 | ||||
| -rw-r--r-- | tests/validation_new/NEON/FullyConnectedLayer.cpp | 4 | ||||
| -rw-r--r-- | tests/validation_new/fixtures/FullyConnectedLayerFixture.h | 15 |
15 files changed, 549 insertions, 511 deletions
diff --git a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h index dec63e0679..a768a19914 100644 --- a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h +++ b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h @@ -30,10 +30,10 @@ namespace arm_compute { class ICLTensor; -/** OpenCL kernel to multiply two input matrices "A" and "B" or to multiply a vector "A" by a matrix "B". All elements of the output matrix/vector will be multiplied by alpha +/** OpenCL kernel to multiply two input matrices "A" and "B" . All elements of the output matrix will be multiplied by alpha * - * @note If the output tensor is a matrix, the implementation assumes that the input tensors @p input0 and @p input1 are both matrices and reshaped respectively with @ref CLGEMMInterleave4x4Kernel" and @ref CLGEMMTranspose1xWKernel - * @note If the output tensor is a vector and the data type is F32, the implementation assumes that the first input tensor @p input0 is a vector and the second input tensor @p input1 a matrix. The implementation also assumes that both tensors have not been reshaped + * @note If the input tensors @p input0 and @p input1 have been reshaped respectively with @ref CLGEMMInterleave4x4Kernel" and @ref CLGEMMTranspose1xWKernel, + * the flag @p is_interleaved_transposed must be set to true * * @attention The second input tensor must have at least 2 dimensions (matrix) * @@ -53,13 +53,13 @@ public: CLGEMMMatrixMultiplyKernel &operator=(CLGEMMMatrixMultiplyKernel &&) = default; /** Initialise the kernel's input, output and alpha * - * @param[in] input0 Input tensor containing the interleaved Matrix A or the vector A. Data types supported: QS8/QS16/F16/F32 - * @param[in] input1 Input tensor containing the transposed Matrix B if the first input tensor A is not a vector. - * If the output tensor is a vector, input1 must contain the matrix B not reshaped. Data type supported: same as @p input0 - * @param[out] output Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0 - * @param[in] alpha Weight of the matrix product + * @param[in] input0 Input tensor containing the Matrix A. Data types supported: QS8/QS16/F16/F32 + * @param[in] input1 Input tensor containing the Matrix B. Data type supported: same as @p input0 + * @param[out] output Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0 + * @param[in] alpha Weight of the matrix product + * @param[in] is_interleaved_transposed (Optional) True if input0 and input1 have been reshaped respectively using @ref CLGEMMInterleave4x4Kernel and @ref CLGEMMTranspose1xWKernel */ - void configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha); + void configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha, bool is_interleaved_transposed = true); // Inherited methods overridden: void run(const Window &window, cl::CommandQueue &queue) override; diff --git a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h index a29f68fcf1..e076f51b26 100644 --- a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h +++ b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h @@ -24,12 +24,10 @@ #ifndef __ARM_COMPUTE_CLFULLYCONNECTEDLAYER_H__ #define __ARM_COMPUTE_CLFULLYCONNECTEDLAYER_H__ -#include "arm_compute/runtime/IFunction.h" +#include "arm_compute/runtime/CL/ICLSimpleFunction.h" -#include "arm_compute/core/CL/kernels/CLGEMMInterleave4x4Kernel.h" #include "arm_compute/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.h" #include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h" -#include "arm_compute/core/CL/kernels/CLGEMMTranspose1xWKernel.h" #include "arm_compute/core/CL/kernels/CLIm2ColKernel.h" #include "arm_compute/core/CL/kernels/CLTransposeKernel.h" #include "arm_compute/runtime/CL/CLTensor.h" @@ -38,41 +36,25 @@ namespace arm_compute { /** Basic function to reshape the weights of Fully Connected layer with OpenCL. This function calls the following kernels: * - * -# @ref CLTransposeKernel (if @p transpose_weights is set to true) - * -# @ref CLGEMMTranspose1xWKernel (if @p is_batched_fc_layer is set to true) + * -# @ref CLTransposeKernel * * @note The fully connected layer accepts "weights" tensors only with 2 dimensions. */ -class CLFullyConnectedLayerReshapeWeights : public IFunction +class CLFullyConnectedLayerReshapeWeights : public ICLSimpleFunction { public: - /** Constructor */ - CLFullyConnectedLayerReshapeWeights(); /** Set the input and output tensors. * - * @param[in] input Weights tensor. The weights must be 2 dimensional. Data types supported: QS8/QS16/F16/F32. - * @param[out] output Destination tensor. Data type supported: Same as @p input. - * @param[in] transpose_weights True if the weights must be transposed. Data types supported: Same as @p weights. - * @param[in] is_batched_fc_layer True if it is a batched fully connected layer + * @param[in] input Weights tensor. The weights must be 2 dimensional. Data types supported: QS8/QS16/F16/F32. + * @param[out] output Destination tensor which stores the transposed input tensor. Data type supported: Same as @p input. */ - void configure(const ICLTensor *input, ICLTensor *output, bool transpose_weights, bool is_batched_fc_layer); - - // Inherited methods overridden: - void run() override; - -private: - CLTransposeKernel _transpose_kernel; - CLGEMMTranspose1xWKernel _transpose1xW_kernel; - CLTensor _transpose_output; - bool _transpose_weights; - bool _is_batched_fc_layer; + void configure(const ICLTensor *input, ICLTensor *output); }; /** Basic function to compute a Fully Connected layer on OpenCL. This function calls the following OpenCL kernels: * * -# @ref CLIm2ColKernel (called when the input comes from a convolutional layer) - * -# @ref CLFullyConnectedLayerReshapeWeights (if @p are_weights_reshaped is set to false) (called once) - * -# @ref CLGEMMInterleave4x4Kernel (called if we have a multi-batch input) + * -# @ref CLFullyConnectedLayerReshapeWeights (if @p are_weights_reshaped is set to false and transpose_weights is set to true ) (called once) * -# @ref CLGEMMMatrixMultiplyKernel * -# @ref CLGEMMMatrixAccumulateBiasesKernel (if @p biases is not equal to nullptr) * @@ -85,7 +67,7 @@ public: CLFullyConnectedLayer(); /** Set the input and output tensors. * - * @param[in] input Source tensor. Data type supported: QS8/F16/F32. + * @param[in] input Source tensor. Data type supported: QS8/QS16/F16/F32. * @param[in] weights Weights tensor. The weights must be 2 dimensional. Data type supported: Same as @p input * @param[in] biases Bias tensor. It can be nullptr. Data type supported:Same as @p input. * @param[out] output Destination tensor. Data type supported: Same as @p input. @@ -98,17 +80,17 @@ public: void run() override; private: + void configure_fc_fc(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output); + void configure_conv_fc(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output); + CLIm2ColKernel _im2col_kernel; CLFullyConnectedLayerReshapeWeights _reshape_weights_kernel; - CLGEMMInterleave4x4Kernel _interleave4x4_kernel; CLGEMMMatrixMultiplyKernel _mm_kernel; CLGEMMMatrixAccumulateBiasesKernel _accumulate_biases_kernel; CLTensor _im2col_output; - CLTensor _interleave4x4_output; CLTensor _reshape_weights_output; bool _are_weights_reshaped; - bool _is_batched_fc_layer; - bool _linearize_input; + bool _is_fc_after_conv; bool _accumulate_biases; }; } diff --git a/arm_compute/runtime/CL/functions/CLGEMM.h b/arm_compute/runtime/CL/functions/CLGEMM.h index 9207efd68f..9b887305cb 100644 --- a/arm_compute/runtime/CL/functions/CLGEMM.h +++ b/arm_compute/runtime/CL/functions/CLGEMM.h @@ -76,7 +76,7 @@ private: CLGEMMMatrixAdditionKernel _ma_kernel; CLTensor _tmp_a; CLTensor _tmp_b; - bool _run_vector_matrix_multiplication; + bool _is_interleaved_transposed; bool _run_addition; }; } diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp index 019f3ea132..2589bd12b5 100644 --- a/src/core/CL/CLKernelLibrary.cpp +++ b/src/core/CL/CLKernelLibrary.cpp @@ -168,16 +168,15 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map = { "gemm_ma_f32", "gemm.cl" }, { "gemm_ma_qs8", "gemm.cl" }, { "gemm_ma_qs16", "gemm.cl" }, - { "gemm_mm_u8", "gemm.cl" }, - { "gemm_mm_f16", "gemm.cl" }, - { "gemm_mm_f32_midgard", "gemm.cl" }, - { "gemm_mm_f32_bifrost", "gemm.cl" }, + { "gemm_mm_interleaved_transposed_u8", "gemm.cl" }, + { "gemm_mm_interleaved_transposed_f16", "gemm.cl" }, + { "gemm_mm_interleaved_transposed_f32_midgard", "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_qs8", "gemm.cl" }, { "gemm_mm_qs16", "gemm.cl" }, - { "gemm_vm_f16", "gemm.cl" }, - { "gemm_vm_f32", "gemm.cl" }, - { "gemm_vm_qs8", "gemm.cl" }, - { "gemm_vm_qs16", "gemm.cl" }, { "gemm_lc_vm_f32", "gemm.cl" }, { "gemm_transpose1x16", "gemm.cl" }, { "gemm_transpose1x8", "gemm.cl" }, diff --git a/src/core/CL/cl_kernels/gemm.cl b/src/core/CL/cl_kernels/gemm.cl index 00c73e7be0..35a2e4704f 100644 --- a/src/core/CL/cl_kernels/gemm.cl +++ b/src/core/CL/cl_kernels/gemm.cl @@ -48,10 +48,10 @@ __kernel void gemm_transpose1x4(IMAGE_DECLARATION(src), uint x = get_global_id(0); uint y = get_global_id(1); - /* Compute address for Matrix B - source */ + // Compute address for Matrix B - source Image src = CONVERT_TO_IMAGE_STRUCT(src); - /* Compute address for Matrix B transposed - destination. X and Y are swapped */ + // Compute address for Matrix B transposed - destination. X and Y are swapped uint dst_addr_in_bytes = y * 16 + ((x * dst_stride_y + dst_offset_first_element_in_bytes)); uint4 b0 = vload4(0, (__global uint *)src.ptr); @@ -288,11 +288,11 @@ __kernel void gemm_accumulate_biases( } #endif /* DATA_TYPE */ -#ifdef WIDTH_MATRIX_B +#ifdef COLS_B /** This OpenCL kernel computes the matrix multiplication between matrix A (src0) and matrix B (src1) * Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_8bit and @ref gemm_transpose1x16 before running the matrix multiplication * - * @attention The width of matrix B and the alpha's value need to be passed at compile time using -DWIDTH_MATRIX_B + * @attention The width of matrix B and the alpha's value need to be passed at compile time using -DCOLS_B * * @param[in] src0_ptr Pointer to the source matrix. Supported formats: U8 * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) @@ -318,14 +318,14 @@ __kernel void gemm_accumulate_biases( * @param[in] c_mult_int Multiplied with each element of the matrix C. * @param[in] shift Number of bits to shift right the result. */ -__kernel void gemm_mm_u8(IMAGE_DECLARATION(src0), - IMAGE_DECLARATION(src1), - IMAGE_DECLARATION(dst), - int a_offset, - int b_offset, - int c_offset, - int c_mult_int, - int shift) +__kernel void gemm_mm_interleaved_transposed_u8(IMAGE_DECLARATION(src0), + IMAGE_DECLARATION(src1), + IMAGE_DECLARATION(dst), + int a_offset, + int b_offset, + int c_offset, + int c_mult_int, + int shift) { /* src_addr.s0 = address of matrix A */ /* src_addr.s1 = address of matrix B */ @@ -338,7 +338,7 @@ __kernel void gemm_mm_u8(IMAGE_DECLARATION(src0), src_addr = src_addr + ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); /* Compute end row address for matrix B */ - int end_row_mtx_b = src_addr.s1 + WIDTH_MATRIX_B; + int end_row_mtx_b = src_addr.s1 + COLS_B; /* Reset accumulators */ int16 c00 = 0.0f; @@ -392,13 +392,13 @@ __kernel void gemm_mm_u8(IMAGE_DECLARATION(src0), vstore16(convert_uchar16_sat(c20), 0, (__global uchar *)(offset(&dst, 0, 2))); vstore16(convert_uchar16_sat(c30), 0, (__global uchar *)(offset(&dst, 0, 3))); } -#endif /* WIDTH_MATRIX_B */ +#endif /* COLS_B */ -#if defined(WIDTH_MATRIX_B) && defined(ALPHA) +#if defined(COLS_B) && defined(ALPHA) /** This OpenCL kernel is optimised for Midgard. It computes the matrix multiplication between matrix A (src0) and matrix B (src1) * Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_32bit and @ref gemm_transpose1x4 before running the matrix multiplication * - * @attention The width of matrix B and the alpha's value need to be passed at compile time using -DWIDTH_MATRIX_B and -DALPHA + * @attention The width of matrix B and the alpha's value need to be passed at compile time using -DCOLS_B and -DALPHA * * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F32 * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) @@ -419,9 +419,9 @@ __kernel void gemm_mm_u8(IMAGE_DECLARATION(src0), * @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 */ -__kernel void gemm_mm_f32_midgard(IMAGE_DECLARATION(src0), - IMAGE_DECLARATION(src1), - IMAGE_DECLARATION(dst)) +__kernel void gemm_mm_interleaved_transposed_f32_midgard(IMAGE_DECLARATION(src0), + IMAGE_DECLARATION(src1), + IMAGE_DECLARATION(dst)) { /* src_addr.s0 = address of matrix A */ /* src_addr.s1 = address of matrix B */ @@ -437,7 +437,7 @@ __kernel void gemm_mm_f32_midgard(IMAGE_DECLARATION(src0), src_addr = src_addr >> 2; /* Compute end row address for matrix B */ - int end_row_mtx_b = src_addr.s1 + WIDTH_MATRIX_B; + int end_row_mtx_b = src_addr.s1 + COLS_B; /* Reset accumulators */ float4 c00 = 0.0f; @@ -497,7 +497,7 @@ __kernel void gemm_mm_f32_midgard(IMAGE_DECLARATION(src0), /** This OpenCL kernel is optimised for Bifrost. It computes the matrix multiplication between matrix A (src0) and matrix B (src1) * Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_32bit and @ref gemm_transpose1x4 before running the matrix multiplication * - * @attention The width of matrix B and the alpha's value need to be passed at compile time using -DWIDTH_MATRIX_B and -DALPHA + * @attention The width of matrix B and the alpha's value need to be passed at compile time using -DCOLS_B and -DALPHA * * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F32 * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) @@ -518,9 +518,9 @@ __kernel void gemm_mm_f32_midgard(IMAGE_DECLARATION(src0), * @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 */ -__kernel void gemm_mm_f32_bifrost(IMAGE_DECLARATION(src0), - IMAGE_DECLARATION(src1), - IMAGE_DECLARATION(dst)) +__kernel void gemm_mm_interleaved_transposed_f32_bifrost(IMAGE_DECLARATION(src0), + IMAGE_DECLARATION(src1), + IMAGE_DECLARATION(dst)) { // src_addr_a = address of matrix A // src_addr_b = address of matrix B @@ -528,7 +528,7 @@ __kernel void gemm_mm_f32_bifrost(IMAGE_DECLARATION(src0), __global float *src_addr_b = (__global float *)(src1_ptr + get_global_id(0) * src1_stride_y + src1_offset_first_element_in_bytes); // Compute end row address for matrix B - __global float *src_end_addr_b = src_addr_b + WIDTH_MATRIX_B; + __global float *src_end_addr_b = src_addr_b + COLS_B; // Reset accumulators float c00 = 0.0f; @@ -707,7 +707,7 @@ __kernel void gemm_mm_f32_bifrost(IMAGE_DECLARATION(src0), /** This OpenCL kernel computes the matrix multiplication between matrix A (src0) and matrix B (src1) * Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_16bit and @ref gemm_transpose1x8 before running the matrix multiplication * - * @attention The width of matrix B and the alpha's value need to be passed at compile time using -DWIDTH_MATRIX_B and -DALPHA + * @attention The width of matrix B and the alpha's value need to be passed at compile time using -DCOLS_B and -DALPHA * * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F16 * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) @@ -728,9 +728,9 @@ __kernel void gemm_mm_f32_bifrost(IMAGE_DECLARATION(src0), * @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 */ -__kernel void gemm_mm_f16(IMAGE_DECLARATION(src0), - IMAGE_DECLARATION(src1), - IMAGE_DECLARATION(dst)) +__kernel void gemm_mm_interleaved_transposed_f16(IMAGE_DECLARATION(src0), + IMAGE_DECLARATION(src1), + IMAGE_DECLARATION(dst)) { /* src_addr.s0 = address of matrix A */ /* src_addr.s1 = address of matrix B */ @@ -746,7 +746,7 @@ __kernel void gemm_mm_f16(IMAGE_DECLARATION(src0), src_addr = src_addr >> 1; /* Compute end row address for matrix B */ - int end_row_mtx_b = src_addr.s1 + WIDTH_MATRIX_B; + int end_row_mtx_b = src_addr.s1 + COLS_B; /* Reset accumulators */ half8 c00 = 0.0f; @@ -807,7 +807,7 @@ __kernel void gemm_mm_f16(IMAGE_DECLARATION(src0), /** 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 * - * @attention The width of matrix B, the alpha's value and fixed point position need to be passed at compile time using -DWIDTH_MATRIX_B -DALPHA and -DFIXED_POINT_POSITION + * @attention The width of matrix B, the alpha's value and fixed point position need to be passed at compile time using -DCOLS_B -DALPHA and -DFIXED_POINT_POSITION * * @note: ALPHA must be passed in 8 bit fixed point format * @@ -830,9 +830,9 @@ __kernel void gemm_mm_f16(IMAGE_DECLARATION(src0), * @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 */ -__kernel void gemm_mm_qs8(IMAGE_DECLARATION(src0), - IMAGE_DECLARATION(src1), - IMAGE_DECLARATION(dst)) +__kernel void gemm_mm_interleaved_transposed_qs8(IMAGE_DECLARATION(src0), + IMAGE_DECLARATION(src1), + IMAGE_DECLARATION(dst)) { /* src_addr.s0 = address of matrix A */ /* src_addr.s1 = address of matrix B */ @@ -845,7 +845,7 @@ __kernel void gemm_mm_qs8(IMAGE_DECLARATION(src0), src_addr = src_addr + ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); /* Compute end row address for matrix B */ - int end_row_mtx_b = src_addr.s1 + WIDTH_MATRIX_B; + int end_row_mtx_b = src_addr.s1 + COLS_B; /* Reset accumulators */ short8 c00 = 0.0f; @@ -899,7 +899,7 @@ __kernel void gemm_mm_qs8(IMAGE_DECLARATION(src0), /** 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 * - * @attention The width of matrix B, the alpha's value and fixed point position need to be passed at compile time using -DWIDTH_MATRIX_B -DALPHA and -DFIXED_POINT_POSITION + * @attention The width of matrix B, the alpha's value and fixed point position need to be passed at compile time using -DCOLS_B -DALPHA and -DFIXED_POINT_POSITION * * @note: ALPHA must be passed in 16 bit fixed point format * @@ -922,9 +922,9 @@ __kernel void gemm_mm_qs8(IMAGE_DECLARATION(src0), * @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 */ -__kernel void gemm_mm_qs16(IMAGE_DECLARATION(src0), - IMAGE_DECLARATION(src1), - IMAGE_DECLARATION(dst)) +__kernel void gemm_mm_interleaved_transposed_qs16(IMAGE_DECLARATION(src0), + IMAGE_DECLARATION(src1), + IMAGE_DECLARATION(dst)) { /* src_addr.s0 = address of matrix A */ /* src_addr.s1 = address of matrix B */ @@ -940,7 +940,7 @@ __kernel void gemm_mm_qs16(IMAGE_DECLARATION(src0), src_addr = src_addr >> 1; /* Compute end row address for matrix B */ - int end_row_mtx_b = src_addr.s1 + WIDTH_MATRIX_B; + int end_row_mtx_b = src_addr.s1 + COLS_B; /* Reset accumulators */ int8 c00 = 0.0f; @@ -983,14 +983,17 @@ __kernel void gemm_mm_qs16(IMAGE_DECLARATION(src0), } #endif // defined(FIXED_POINT_POSITION) -#ifdef WIDTH_VECTOR_A -/** This OpenCL kernel computes the vector by matrix multiplication between the vector A (src0) and matrix B (src1) - * - * @attention The width of vector A, the width of matrix B and the alpha's value need to be passed at compile time using -DWIDTH_VECTOR_A -DWIDTH_MATRIX_B and -DALPHA +#if defined(COLS_A) && defined(NUM_ELEMS_PROCESSED_PER_THREAD_X) && (NUM_ELEMS_PROCESSED_PER_THREAD_Y) +#if defined(DATA_TYPE) +#define VECTOR_TYPE VEC_DATA_TYPE(DATA_TYPE, NUM_ELEMS_PROCESSED_PER_THREAD_X) +/** 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 * - * @attention The input vector A and matrix B must not be reshaped + * @note This OpenCL kernel works with floating point data types (F16/F32) + * @note The floating point data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) + * @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 width of matrix A and the alpha's value need to be passed at compile time using -DCOLS_A and -DALPHA * - * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F32 + * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F16/F32 * @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) @@ -1009,127 +1012,136 @@ __kernel void gemm_mm_qs16(IMAGE_DECLARATION(src0), * @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 */ -__kernel void gemm_vm_f32(IMAGE_DECLARATION(src0), - IMAGE_DECLARATION(src1), - IMAGE_DECLARATION(dst)) +__kernel void gemm_mm_floating_point(IMAGE_DECLARATION(src0), + IMAGE_DECLARATION(src1), + IMAGE_DECLARATION(dst)) { - int idx = get_global_id(0) * 4; + int idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X; - /* Compute the address for the vector A and matrix B */ + // 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)); - src_addr.s1 += idx * sizeof(float); - - int end_row_vec_a = src_addr.s0 + (WIDTH_VECTOR_A * sizeof(float)); - - float4 acc = 0.0f; - for(; src_addr.s0 <= (end_row_vec_a - 2 * sizeof(float)); src_addr += (int2)(2 * sizeof(float), 2 * src1_stride_y)) - { - float2 a0 = vload2(0, (__global float *)(src0_ptr + src_addr.s0)); - float4 b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1)); - float4 b1 = vload4(0, (__global float *)(src1_ptr + src_addr.s1 + src1_stride_y)); - - acc += b0 * (float4)a0.s0; - acc += b1 * (float4)a0.s1; - } - - for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(sizeof(float), src1_stride_y)) - { - float a0 = *((__global float *)(src0_ptr + src_addr.s0)); - float4 b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1)); - - acc += b0 * (float4)a0; - } - - /* Compute destination address */ - Image dst = CONVERT_TO_IMAGE_STRUCT(dst); - - /* Multiply by the weight of vector-matrix product */ - acc = acc * (float4)ALPHA; - - vstore4(acc, 0, (__global float *)(offset(&dst, 0, 0))); -} - -/** This OpenCL kernel computes the vector by matrix multiplication between the vector A (src0) and matrix B (src1) - * - * @attention The width of vector A, the width of matrix B and the alpha's value need to be passed at compile time using -DWIDTH_VECTOR_A -DWIDTH_MATRIX_B and -DALPHA - * - * @attention The input vector A and matrix B must not be reshaped - * - * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F16 - * @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 - */ -__kernel void gemm_vm_f16(IMAGE_DECLARATION(src0), - IMAGE_DECLARATION(src1), - IMAGE_DECLARATION(dst)) -{ - int idx = get_global_id(0) * 8; + // Update address for the matrix A + src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y; - /* Compute the address for the vector A and matrix B */ - int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); - src_addr.s1 += idx * sizeof(half); + // Update address for the matrix B + src_addr.s1 += idx * sizeof(DATA_TYPE); - int end_row_vec_a = src_addr.s0 + (WIDTH_VECTOR_A * sizeof(half)); + int end_row_vec_a = src_addr.s0 + (COLS_A * sizeof(DATA_TYPE)); - half8 acc = 0.0f; + VECTOR_TYPE acc0 = 0.0f; +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 + VECTOR_TYPE acc1 = 0.0f; +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 + VECTOR_TYPE acc2 = 0.0f; +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 + VECTOR_TYPE acc3 = 0.0f; +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - for(; src_addr.s0 <= (end_row_vec_a - 4 * sizeof(half)); src_addr += (int2)(4 * sizeof(half), 4 * src1_stride_y)) + for(; src_addr.s0 <= (end_row_vec_a - 2 * sizeof(DATA_TYPE)); src_addr += (int2)(2 * sizeof(DATA_TYPE), 2 * src1_stride_y)) { - half4 a0 = vload4(0, (__global half *)(src0_ptr + src_addr.s0)); - half8 b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1 + 0 * src1_stride_y)); - half8 b1 = vload8(0, (__global half *)(src1_ptr + src_addr.s1 + 1 * src1_stride_y)); - half8 b2 = vload8(0, (__global half *)(src1_ptr + src_addr.s1 + 2 * src1_stride_y)); - half8 b3 = vload8(0, (__global half *)(src1_ptr + src_addr.s1 + 3 * src1_stride_y)); - - acc += b0 * (half8)a0.s0; - acc += b1 * (half8)a0.s1; - acc += b2 * (half8)a0.s2; - acc += b3 * (half8)a0.s3; + // Load values from matrix A + VEC_DATA_TYPE(DATA_TYPE, 2) + a0 = vload2(0, (__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 + VEC_DATA_TYPE(DATA_TYPE, 2) + a1 = vload2(0, (__global DATA_TYPE *)(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 + VEC_DATA_TYPE(DATA_TYPE, 2) + a2 = vload2(0, (__global DATA_TYPE *)(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 + VEC_DATA_TYPE(DATA_TYPE, 2) + a3 = vload2(0, (__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 + // Load values from matrix B + VECTOR_TYPE b0 = VLOAD(NUM_ELEMS_PROCESSED_PER_THREAD_X)(0, (__global DATA_TYPE *)(src1_ptr + src_addr.s1)); + VECTOR_TYPE b1 = VLOAD(NUM_ELEMS_PROCESSED_PER_THREAD_X)(0, (__global DATA_TYPE *)(src1_ptr + src_addr.s1 + src1_stride_y)); + + // Accumulate + acc0 += b0 * (VECTOR_TYPE)a0.s0; + acc0 += b1 * (VECTOR_TYPE)a0.s1; +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 + acc1 += b0 * (VECTOR_TYPE)a1.s0; + acc1 += b1 * (VECTOR_TYPE)a1.s1; +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 + acc2 += b0 * (VECTOR_TYPE)a2.s0; + acc2 += b1 * (VECTOR_TYPE)a2.s1; +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 + acc3 += b0 * (VECTOR_TYPE)a3.s0; + acc3 += b1 * (VECTOR_TYPE)a3.s1; +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 } - for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(sizeof(half), src1_stride_y)) + for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(sizeof(DATA_TYPE), src1_stride_y)) { - half a0 = *((__global half *)(src0_ptr + src_addr.s0)); - half8 b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1)); - - acc += b0 * (half8)a0; + // Load values from matrix A + DATA_TYPE a0 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 + DATA_TYPE a1 = *((__global DATA_TYPE *)(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 + DATA_TYPE a2 = *((__global DATA_TYPE *)(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 + DATA_TYPE a3 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 + // Load values from matrix B + VECTOR_TYPE b0 = VLOAD(NUM_ELEMS_PROCESSED_PER_THREAD_X)(0, (__global DATA_TYPE *)(src1_ptr + src_addr.s1)); + + // Accumulate + acc0 += b0 * (VECTOR_TYPE)a0; +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 + acc1 += b0 * (VECTOR_TYPE)a1; +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 + acc2 += b0 * (VECTOR_TYPE)a2; +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 + acc3 += b0 * (VECTOR_TYPE)a3; +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 } - /* Compute destination address */ + // Compute destination address Image dst = CONVERT_TO_IMAGE_STRUCT(dst); - /* Multiply by the weight of vector-matrix product */ - acc = acc * (half8)ALPHA; - - vstore8(acc, 0, (__global half *)(offset(&dst, 0, 0))); + // Multiply by the weight of matrix-matrix product and store the result + acc0 = acc0 * (VECTOR_TYPE)ALPHA; + VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X) + (acc0, 0, (__global DATA_TYPE *)(offset(&dst, 0, 0))); +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 + acc1 = acc1 * (VECTOR_TYPE)ALPHA; + VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X) + (acc1, 0, (__global DATA_TYPE *)(offset(&dst, 0, 1))); +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 + acc2 = acc2 * (VECTOR_TYPE)ALPHA; + VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X) + (acc2, 0, (__global DATA_TYPE *)(offset(&dst, 0, 2))); +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 + acc3 = acc3 * (VECTOR_TYPE)ALPHA; + VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X) + (acc3, 0, (__global DATA_TYPE *)(offset(&dst, 0, 3))); +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 } +#endif // defined(DATA_TYPE) #ifdef FIXED_POINT_POSITION -/** This OpenCL kernel computes the vector by matrix multiplication between the vector A (src0) and matrix B (src1) in 8 bit fixed point - * - * @attention The width of vector A, the width of matrix B, the alpha's value and the fixed point position need to be passed at compile time using -DWIDTH_VECTOR_A -DWIDTH_MATRIX_B, -DALPHA and -DFIXED_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 * - * @attention The input vector A and matrix B must not be 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 width of matrix A, 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 alpha value must be passed in 8 bit fixed point format using -DALPHA * - * @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_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) @@ -1148,72 +1160,143 @@ __kernel void gemm_vm_f16(IMAGE_DECLARATION(src0), * @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 */ -__kernel void gemm_vm_qs8(IMAGE_DECLARATION(src0), +__kernel void gemm_mm_qs8(IMAGE_DECLARATION(src0), IMAGE_DECLARATION(src1), IMAGE_DECLARATION(dst)) { - int idx = get_global_id(0) * 16; + int idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X; - /* Compute the address for the vector A and matrix B */ + // 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)); - src_addr.s1 += idx; - - int end_row_vec_a = src_addr.s0 + WIDTH_VECTOR_A; - - short8 acc0 = 0; - short8 acc1 = 0; - /* This for loop performs 4 accumulations per iteration */ - for(; src_addr.s0 <= (end_row_vec_a - 4); src_addr += (int2)(4, 4 * src1_stride_y)) + // 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); + + 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)) { - char4 a0 = vload4(0, (__global char *)(src0_ptr + src_addr.s0)); + 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)); - char16 b2 = vload16(0, (__global char *)(src1_ptr + src_addr.s1 + 2 * src1_stride_y)); - char16 b3 = vload16(0, (__global char *)(src1_ptr + src_addr.s1 + 3 * src1_stride_y)); - - acc0 = mlal_sat_qs8x8(acc0, (char8)a0.s0, b0.s01234567, FIXED_POINT_POSITION); - acc0 = mlal_sat_qs8x8(acc0, (char8)a0.s1, b1.s01234567, FIXED_POINT_POSITION); - acc0 = mlal_sat_qs8x8(acc0, (char8)a0.s2, b2.s01234567, FIXED_POINT_POSITION); - acc0 = mlal_sat_qs8x8(acc0, (char8)a0.s3, b3.s01234567, FIXED_POINT_POSITION); - - acc1 = mlal_sat_qs8x8(acc1, (char8)a0.s0, b0.s89ABCDEF, FIXED_POINT_POSITION); - acc1 = mlal_sat_qs8x8(acc1, (char8)a0.s1, b1.s89ABCDEF, FIXED_POINT_POSITION); - acc1 = mlal_sat_qs8x8(acc1, (char8)a0.s2, b2.s89ABCDEF, FIXED_POINT_POSITION); - acc1 = mlal_sat_qs8x8(acc1, (char8)a0.s3, b3.s89ABCDEF, FIXED_POINT_POSITION); + + 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 */ + // 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)); + 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)); - acc0 = mlal_sat_qs8x8(acc0, (char8)a0, b0.s01234567, FIXED_POINT_POSITION); - acc1 = mlal_sat_qs8x8(acc1, (char8)a0, b0.s89ABCDEF, FIXED_POINT_POSITION); + 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 */ + // Compute destination address Image dst = CONVERT_TO_IMAGE_STRUCT(dst); - /* Multiply by the weight of matrix product */ - char16 acc_qs8 = convert_char16_sat((short16)(acc0, acc1)); - + // Multiply by the weight of matrix product and store the result + char16 acc_qs8; + acc_qs8 = convert_char16_sat((short16)(acc00, acc01)); acc_qs8 = mul_sat_qs8x16(acc_qs8, (char16)ALPHA, FIXED_POINT_POSITION); - - /* Store 16 values */ vstore16(acc_qs8, 0, (__global char *)(offset(&dst, 0, 0))); +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 + acc_qs8 = convert_char16_sat((short16)(acc10, acc11)); + acc_qs8 = mul_sat_qs8x16(acc_qs8, (char16)ALPHA, FIXED_POINT_POSITION); + vstore16(acc_qs8, 0, (__global char *)(offset(&dst, 0, 1))); +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 + acc_qs8 = convert_char16_sat((short16)(acc20, acc21)); + acc_qs8 = mul_sat_qs8x16(acc_qs8, (char16)ALPHA, FIXED_POINT_POSITION); + vstore16(acc_qs8, 0, (__global char *)(offset(&dst, 0, 2))); +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 + acc_qs8 = convert_char16_sat((short16)(acc30, acc31)); + acc_qs8 = mul_sat_qs8x16(acc_qs8, (char16)ALPHA, FIXED_POINT_POSITION); + vstore16(acc_qs8, 0, (__global char *)(offset(&dst, 0, 3))); +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 } -/** This OpenCL kernel computes the vector by matrix multiplication between the vector A (src0) and matrix B (src1) in 16 bit fixed point +/** 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 * - * @attention The width of vector A, the width of matrix B, the alpha's value and the fixed point position need to be passed at compile time using -DWIDTH_VECTOR_A -DWIDTH_MATRIX_B, -DALPHA and -DFIXED_POINT_POSITION + * @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 width of matrix A, 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 alpha value must be passed in 16 bit fixed point format using -DALPHA * - * @attention The input vector A and matrix B must not be reshaped - * - * @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_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) @@ -1232,59 +1315,120 @@ __kernel void gemm_vm_qs8(IMAGE_DECLARATION(src0), * @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 */ -__kernel void gemm_vm_qs16(IMAGE_DECLARATION(src0), +__kernel void gemm_mm_qs16(IMAGE_DECLARATION(src0), IMAGE_DECLARATION(src1), IMAGE_DECLARATION(dst)) { - int idx = get_global_id(0) * 8; + int idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X; - /* Compute the address for the vector A and matrix B */ + // 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); - int end_row_vec_a = src_addr.s0 + (WIDTH_VECTOR_A * sizeof(short)); + int end_row_vec_a = src_addr.s0 + (COLS_A * sizeof(short)); - /* Reset accumulator */ int8 acc0 = 0; - - /* This for loop performs 4 accumulations per iteration */ - for(; src_addr.s0 <= (end_row_vec_a - 4 * sizeof(short)); src_addr += (int2)(4 * sizeof(short), 4 * src1_stride_y)) +#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 * sizeof(short)); src_addr += (int2)(2 * sizeof(short), 2 * src1_stride_y)) { - short4 a0 = vload4(0, (__global short *)(src0_ptr + src_addr.s0)); + 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)); - short8 b2 = vload8(0, (__global short *)(src1_ptr + src_addr.s1 + 2 * src1_stride_y)); - short8 b3 = vload8(0, (__global short *)(src1_ptr + src_addr.s1 + 3 * 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); - acc0 = mlal_sat_qs16x8(acc0, (short8)a0.s2, b2, FIXED_POINT_POSITION); - acc0 = mlal_sat_qs16x8(acc0, (short8)a0.s3, b3, 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 */ + // 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)); + 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 */ + // Compute destination address Image dst = CONVERT_TO_IMAGE_STRUCT(dst); - /* Multiply by the weight of matrix product */ - short8 acc_qs16 = convert_short8_sat(acc0); - + // Multiply by the weight of matrix product and store the result + short8 acc_qs16; + acc_qs16 = convert_short8_sat(acc0); acc_qs16 = mul_sat_qs16x8(acc_qs16, (short8)ALPHA, FIXED_POINT_POSITION); - - /* Store 8 values */ vstore8(acc_qs16, 0, (__global short *)(offset(&dst, 0, 0))); +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 + acc_qs16 = convert_short8_sat(acc1); + acc_qs16 = mul_sat_qs16x8(acc_qs16, (short8)ALPHA, FIXED_POINT_POSITION); + vstore8(acc_qs16, 0, (__global short *)(offset(&dst, 0, 1))); +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 + acc_qs16 = convert_short8_sat(acc2); + acc_qs16 = mul_sat_qs16x8(acc_qs16, (short8)ALPHA, FIXED_POINT_POSITION); + vstore8(acc_qs16, 0, (__global short *)(offset(&dst, 0, 2))); +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 +#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 + acc_qs16 = convert_short8_sat(acc3); + acc_qs16 = mul_sat_qs16x8(acc_qs16, (short8)ALPHA, FIXED_POINT_POSITION); + vstore8(acc_qs16, 0, (__global short *)(offset(&dst, 0, 3))); +#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 } -#endif /* defined(FIXED_POINT_POSITION) */ -#endif /* defined(WIDTH_VECTOR_A) */ -#endif /* defined(WIDTH_MATRIX_B) && defined(ALPHA) */ +#endif // defined(FIXED_POINT_POSITION) +#endif // defined(COLS_A) && defined(NUM_ELEMS_PROCESSED_PER_THREAD_X) && (NUM_ELEMS_PROCESSED_PER_THREAD_Y) +#endif // defined(COLS_B) && defined(ALPHA) #ifdef BETA /** This OpenCL kernel performs the in-place matrix addition between 2 matrices taking into account that the second matrix might be weighted by a scalar value beta: @@ -1508,4 +1652,4 @@ __kernel void gemm_lc_vm_f32(IMAGE_DECLARATION(src0), vstore4(acc, 0, (__global float *)(offset(&dst, 0, 0))); } -#endif /* WIDTH_VECTOR_A */ +#endif /* WIDTH_VECTOR_A */
\ No newline at end of file diff --git a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp index ce68c1f9cd..ef572cfc7e 100644 --- a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp +++ b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp @@ -64,8 +64,8 @@ void CLGEMMLowpMatrixMultiplyKernel::configure(const ICLTensor *input0, const IC _output = output; // Create kernel and set static arguments - std::set<std::string> build_opts = { ("-DWIDTH_MATRIX_B=" + support::cpp11::to_string(input1->info()->dimension(0))) }; - _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemm_mm_u8", build_opts)); + std::set<std::string> build_opts = { ("-DCOLS_B=" + support::cpp11::to_string(input1->info()->dimension(0))) }; + _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemm_mm_interleaved_transposed_u8", build_opts)); unsigned int idx = 3 * num_arguments_per_2D_tensor(); //Skip the input and output parameters _kernel.setArg<int32_t>(idx++, a_offset); _kernel.setArg<int32_t>(idx++, b_offset); diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp index 39526a23e1..684e3232d5 100644 --- a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp +++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp @@ -48,13 +48,13 @@ CLGEMMMatrixMultiplyKernel::CLGEMMMatrixMultiplyKernel() { } -void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha) +void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha, bool is_interleaved_transposed) { ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output); ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output); - if(output->info()->dimension(1) == 1) + if(!is_interleaved_transposed) { ARM_COMPUTE_ERROR_ON(input0->info()->dimension(0) != input1->info()->dimension(1)); } @@ -72,79 +72,89 @@ void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTen _lws_hint = cl::NDRange(8, 8); } - std::ostringstream mm_arguments; - mm_arguments << "-DWIDTH_MATRIX_B=" << input1->info()->dimension(0) << " "; + std::set<std::string> build_opts; + build_opts.emplace(("-DCOLS_A=" + support::cpp11::to_string(input0->info()->dimension(0)))); + build_opts.emplace(("-DCOLS_B=" + support::cpp11::to_string(input1->info()->dimension(0)))); + if(is_data_type_fixed_point(input0->info()->data_type())) { - mm_arguments << "-DALPHA=" << (input0->info()->data_type() == DataType::QS8 ? - sqcvt_qs8_f32(alpha, input0->info()->fixed_point_position()) : - sqcvt_qs16_f32(alpha, input0->info()->fixed_point_position())) - << " "; - mm_arguments << "-DFIXED_POINT_POSITION=" << input0->info()->fixed_point_position() << " "; + build_opts.emplace(("-DALPHA=" + support::cpp11::to_string((input0->info()->data_type() == DataType::QS8 ? + sqcvt_qs8_f32(alpha, input0->info()->fixed_point_position()) : + sqcvt_qs16_f32(alpha, input0->info()->fixed_point_position()))))); + + build_opts.emplace(("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input0->info()->fixed_point_position()))); } else { - mm_arguments << "-DALPHA=" << alpha << " "; + build_opts.emplace(("-DALPHA=" + float_to_string_with_full_precision(alpha))); } - std::set<std::string> build_opts; - // Check if the output tensor is a vector. If so,the kernel runs the vector-matrix multiplication - if(output->info()->dimension(1) == 1) + if(is_interleaved_transposed) { - mm_arguments << "-DWIDTH_VECTOR_A=" << input0->info()->dimension(0) << " "; - build_opts.emplace(mm_arguments.str()); - // Create kernel std::string data_type_name = lower_string(string_from_data_type(input0->info()->data_type())); - _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(("gemm_vm_" + data_type_name), build_opts)); + + if(data_type_name == "f32") + { + GPUTarget arch_target = get_arch_from_target(get_target()); + _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemm_mm_interleaved_transposed_f32_" + string_from_target(arch_target), build_opts)); + } + else + { + _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemm_mm_interleaved_transposed_" + data_type_name, build_opts)); + } // Configure window kernel - const unsigned int num_elems_processed_per_iteration_x = max_cl_vector_width / data_size_from_type(input0->info()->data_type()); + const unsigned int num_elems_processed_per_iteration_x = max_cl_vector_width / data_size_from_type(input0->info()->data_type()); + constexpr unsigned int num_elems_processed_per_iteration_y = 4; - Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x)); + Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); - AccessWindowStatic input0_access(input0->info(), 0, 0, input0->info()->tensor_shape().x(), 1); - AccessWindowHorizontal input1_access(input1->info(), 0, num_elems_processed_per_iteration_x); - AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration_x); + AccessWindowRectangle input0_access(input0->info(), 0, 0, num_elems_processed_per_iteration_y, 1, 1.f, 0.25f); + AccessWindowTranspose input1_access(input1->info(), 0, 0, num_elems_processed_per_iteration_x, 1, 0.f, 0.25f); + AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); update_window_and_padding(win, input0_access, input1_access, output_access); - Coordinates coord; - coord.set_num_dimensions(output->info()->num_dimensions()); - output_access.set_valid_region(win, ValidRegion(coord, output->info()->tensor_shape())); + output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape())); ICLKernel::configure(win); } - else + else // The input tensors have not been reshaped { - build_opts.emplace(mm_arguments.str()); + ARM_COMPUTE_ERROR_ON(input0->info()->dimension(0) != input1->info()->dimension(1)); - // Create kernel - std::string data_type_name = lower_string(string_from_data_type(input0->info()->data_type())); + // Special case for 1xN, 2xN, 3xN and 4xN input0 tensor + const unsigned int num_elems_processed_per_iteration_x = max_cl_vector_width / data_size_from_type(input0->info()->data_type()); + const unsigned int num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->info()->dimension(1)), 4); - if(data_type_name == "f32") + build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type()))); + build_opts.emplace(("-DNUM_ELEMS_PROCESSED_PER_THREAD_X=" + support::cpp11::to_string(num_elems_processed_per_iteration_x))); + build_opts.emplace(("-DNUM_ELEMS_PROCESSED_PER_THREAD_Y=" + support::cpp11::to_string(num_elems_processed_per_iteration_y))); + + // Create kernel + if(is_data_type_fixed_point(input0->info()->data_type())) { - GPUTarget arch_target = get_arch_from_target(get_target()); - _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemm_mm_f32_" + string_from_target(arch_target), build_opts)); + std::string kernel_name = "gemm_mm_" + lower_string(string_from_data_type(input0->info()->data_type())); + _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel((kernel_name), build_opts)); } else { - _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemm_mm_" + data_type_name, build_opts)); + std::string kernel_name = "gemm_mm_floating_point"; + _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel((kernel_name), build_opts)); } - // Configure window kernel - const unsigned int num_elems_processed_per_iteration_x = max_cl_vector_width / data_size_from_type(input0->info()->data_type()); - constexpr unsigned int num_elems_processed_per_iteration_y = 4; - Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); - AccessWindowRectangle input0_access(input0->info(), 0, 0, num_elems_processed_per_iteration_y, 1, 1.f, 0.25f); - AccessWindowTranspose input1_access(input1->info(), 0, 0, num_elems_processed_per_iteration_x, 1, 0.f, 0.25f); + AccessWindowStatic input0_access(input0->info(), 0, 0, input0->info()->dimension(0), ceil_to_multiple(input0->info()->dimension(1), num_elems_processed_per_iteration_y)); + AccessWindowStatic input1_access(input1->info(), 0, 0, ceil_to_multiple(input1->info()->dimension(0), num_elems_processed_per_iteration_x), input1->info()->dimension(1)); AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); update_window_and_padding(win, input0_access, input1_access, output_access); - output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape())); + Coordinates coord; + coord.set_num_dimensions(output->info()->num_dimensions()); + output_access.set_valid_region(win, ValidRegion(coord, output->info()->tensor_shape())); ICLKernel::configure(win); } @@ -157,9 +167,9 @@ void CLGEMMMatrixMultiplyKernel::run(const Window &window, cl::CommandQueue &que Window slice = window.first_slice_window_2D(); Window slice_matrix_b = slice; - slice_matrix_b.set(Window::DimX, Window::Dimension(0, _input1->info()->dimension(0), 1)); - slice_matrix_b.set(Window::DimY, Window::Dimension(0, _input1->info()->dimension(1), 1)); - slice_matrix_b.set(Window::DimZ, Window::Dimension(0, 1, 1)); + + slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1)); + slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1)); do { diff --git a/src/runtime/CL/functions/CLConvolutionLayer.cpp b/src/runtime/CL/functions/CLConvolutionLayer.cpp index b1b83985d0..0bbec94e78 100644 --- a/src/runtime/CL/functions/CLConvolutionLayer.cpp +++ b/src/runtime/CL/functions/CLConvolutionLayer.cpp @@ -197,9 +197,12 @@ void CLConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weig // Configure kernels _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _has_bias); + + // Configure matrix multiply if(_is_fully_connected_convolution) { - _mm_kernel.configure(&_input_im2col_reshaped, weights, &_gemm_output, 1.0f); + // The matrix A and Matrix B have not been reshaped + _mm_kernel.configure(&_input_im2col_reshaped, weights, &_gemm_output, 1.0f, false); } else { diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp index 66a858d3ed..f7cea551f6 100644 --- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp +++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp @@ -26,217 +26,127 @@ #include "arm_compute/core/Size2D.h" #include "arm_compute/core/Validate.h" #include "arm_compute/runtime/CL/CLScheduler.h" +#include "support/ToolchainSupport.h" #include <algorithm> -#include <cmath> -namespace arm_compute +using namespace arm_compute; + +void CLFullyConnectedLayerReshapeWeights::configure(const ICLTensor *input, ICLTensor *output) { -CLFullyConnectedLayerReshapeWeights::CLFullyConnectedLayerReshapeWeights() - : _transpose_kernel(), _transpose1xW_kernel(), _transpose_output(), _transpose_weights(false), _is_batched_fc_layer(false) + auto k = arm_compute::support::cpp14::make_unique<CLTransposeKernel>(); + k->configure(input, output); + _kernel = std::move(k); +} + +CLFullyConnectedLayer::CLFullyConnectedLayer() + : _im2col_kernel(), _reshape_weights_kernel(), _mm_kernel(), _accumulate_biases_kernel(), _im2col_output(), _reshape_weights_output(), _are_weights_reshaped(true), _is_fc_after_conv(true), + _accumulate_biases(false) { } -void CLFullyConnectedLayerReshapeWeights::configure(const ICLTensor *input, ICLTensor *output, bool transpose_weights, bool is_batched_fc_layer) +void CLFullyConnectedLayer::configure_conv_fc(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); - ARM_COMPUTE_ERROR_ON(input->info()->num_dimensions() > 2); - ARM_COMPUTE_ERROR_ON(output == nullptr); - ARM_COMPUTE_ERROR_ON(!transpose_weights && !is_batched_fc_layer); + ARM_COMPUTE_ERROR_ON((weights->info()->dimension(1) != (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2)))); - const DataType data_type = input->info()->data_type(); + const DataType dt = input->info()->data_type(); const int fixed_point_position = input->info()->fixed_point_position(); - _transpose_weights = transpose_weights; - _is_batched_fc_layer = is_batched_fc_layer; + // If the fully connected layer is called after a convolution layer, the input tensor must be linearized - // Check if we need to transpose the weights - if(_transpose_weights) - { - if(_is_batched_fc_layer) - { - // Initialize the output tensor for transpose - TensorShape shape_transposed(input->info()->dimension(1), input->info()->dimension(0)); - _transpose_output.allocator()->init(TensorInfo(shape_transposed, 1, data_type, fixed_point_position)); - _transpose_kernel.configure(input, &_transpose_output); + // Initialize output tensor for im2col + TensorShape shape_im2col; + shape_im2col.set(0, input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2)); + shape_im2col.set(1, input->info()->dimension(3)); + shape_im2col.set(2, input->info()->dimension(4)); + shape_im2col.set(3, input->info()->dimension(5)); + _im2col_output.allocator()->init(TensorInfo(shape_im2col, 1, dt, fixed_point_position)); - // Configure transpose 1xW kernel - _transpose1xW_kernel.configure(&_transpose_output, output); + // Configure im2col kernel + _im2col_kernel.configure(input, &_im2col_output, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false); - // Allocate temporary tensor used for transposing the weights - _transpose_output.allocator()->allocate(); - } - else - { - _transpose_kernel.configure(input, output); - } - } - else - { - if(_is_batched_fc_layer) - { - // Configure transpose 1xW kernel - _transpose1xW_kernel.configure(input, output); - } - else - { - ARM_COMPUTE_ERROR("Configuration transpose_weights=false & is_batched_fc_layer=false not supported"); - } - } -} - -void CLFullyConnectedLayerReshapeWeights::run() -{ - if(_transpose_weights) - { - CLScheduler::get().enqueue(_transpose_kernel, _is_batched_fc_layer); - } + // Configure matrix multiply kernel + _mm_kernel.configure(&_im2col_output, weights, output, 1.0f, false); - if(_is_batched_fc_layer) - { - CLScheduler::get().enqueue(_transpose1xW_kernel); - } + // Allocate the output tensor for im2col once all the configure methods have been called + _im2col_output.allocator()->allocate(); } -CLFullyConnectedLayer::CLFullyConnectedLayer() - : _im2col_kernel(), _reshape_weights_kernel(), _interleave4x4_kernel(), _mm_kernel(), _accumulate_biases_kernel(), _im2col_output(), _interleave4x4_output(), _reshape_weights_output(), - _are_weights_reshaped(false), _is_batched_fc_layer(false), _linearize_input(false), _accumulate_biases(false) +void CLFullyConnectedLayer::configure_fc_fc(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output) { + ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != weights->info()->dimension(1)); + + // Configure matrix multiply kernel + _mm_kernel.configure(input, weights, output, 1.0f, false); } void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose_weights, bool are_weights_reshaped) { - // With the Fully Connected layer we can have 4 different cases: - // 1) Convolution layer -> Fully Connected layer without batches - // 2) Fully Connected layer -> Fully Connected layer without batches - // 3) Convolution layer -> Fully Connected layer with batches - // 4) Fully Connected layer -> Fully Connected layer with batches - - // Expected shape before transpose and reshaping - // Input: In x B (In and B can be multi-dimensional) - // Weights: flat(In) x Out - // Biases: Out - // Output: Out x B (B can be multi-dimensional) - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, weights, output); + ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() != 2); - const DataType data_type = input->info()->data_type(); - const int fixed_point_position = input->info()->fixed_point_position(); - const int num_batch_dimensions = std::max(0, static_cast<int>(output->info()->tensor_shape().num_dimensions()) - 1); - const int num_input_dimensions = input->info()->tensor_shape().num_dimensions() - num_batch_dimensions; - const size_t linear_input_size = input->info()->tensor_shape().total_size_lower(num_input_dimensions); - - _linearize_input = input->info()->tensor_shape().x() != linear_input_size; - _are_weights_reshaped = are_weights_reshaped; - _accumulate_biases = biases != nullptr; - _is_batched_fc_layer = num_batch_dimensions > 0; - - // Check if number of batches match - ARM_COMPUTE_ERROR_ON(input->info()->tensor_shape().total_size_upper(num_input_dimensions) != output->info()->tensor_shape().total_size_upper(1)); - ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 2); + _are_weights_reshaped = transpose_weights ? are_weights_reshaped : true; + _is_fc_after_conv = true; + _accumulate_biases = false; - const size_t interleave_width = 16 / input->info()->element_size(); - const ICLTensor *weights_to_use = weights; - - if(!are_weights_reshaped && (transpose_weights || _is_batched_fc_layer)) + if(biases != nullptr) { - weights_to_use = &_reshape_weights_output; + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); + + _accumulate_biases = true; - TensorShape reshaped_weights_shape(weights->info()->tensor_shape()); + // Configure accumulate biases kernel + _accumulate_biases_kernel.configure(output, biases); + } - // Transpose weights if the user hasn't done it - if(transpose_weights) - { - const size_t shape_x = reshaped_weights_shape.x(); - reshaped_weights_shape.set(0, reshaped_weights_shape.y()); - reshaped_weights_shape.set(1, shape_x); - } + // With the Fully Connected layer we can have 4 different cases: + // 1) Convolution layer -> Fully Connected layer without batches + // 2) Fully Connected layer -> Fully Connected layer without batches + // 3) Convolution layer -> Fully Connected layer with batches + // 4) Fully Connected layer -> Fully Connected layer with batches - // If the we run multiple batches we need 1xW transpose, too. - if(_is_batched_fc_layer) - { - const float shape_x = reshaped_weights_shape.x(); - reshaped_weights_shape.set(0, reshaped_weights_shape.y() * interleave_width); - reshaped_weights_shape.set(1, static_cast<unsigned int>(std::ceil(shape_x / interleave_width))); - } + const ICLTensor *weights_to_use = weights; - _reshape_weights_output.allocator()->init(TensorInfo(reshaped_weights_shape, 1, data_type, fixed_point_position)); + if(!_are_weights_reshaped) + { + weights_to_use = &_reshape_weights_output; // Reshape the weights - _reshape_weights_kernel.configure(weights, &_reshape_weights_output, transpose_weights, _is_batched_fc_layer); + _reshape_weights_kernel.configure(weights, &_reshape_weights_output); } - // Check correct shape of weights - if(_is_batched_fc_layer) + // Check if we have a fully connected layer with batches + const bool is_batched_fc_layer = output->info()->dimension(1) > 1; + + if(is_batched_fc_layer) { - // Transpose + Transpose1xW - ARM_COMPUTE_ERROR_ON(weights_to_use->info()->tensor_shape().x() != linear_input_size * interleave_width); - ARM_COMPUTE_ERROR_ON(weights_to_use->info()->tensor_shape().y() != static_cast<unsigned int>(std::ceil(static_cast<float>(output->info()->tensor_shape().x()) / interleave_width))); + _is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(input->info()->tensor_shape().cbegin() + 3, + input->info()->tensor_shape().cend(), + output->info()->tensor_shape().cbegin() + 1)); } else { - // Transpose - ARM_COMPUTE_ERROR_ON(weights_to_use->info()->tensor_shape().x() != output->info()->tensor_shape().x()); - ARM_COMPUTE_ERROR_ON(weights_to_use->info()->tensor_shape().y() != linear_input_size); + _is_fc_after_conv = input->info()->num_dimensions() > 1; } - const ICLTensor *multiply_input = input; - - if(_linearize_input) + if(_is_fc_after_conv) { - TensorShape shape_im2col(input->info()->tensor_shape()); - shape_im2col.collapse(num_input_dimensions); - _im2col_output.allocator()->init(TensorInfo(shape_im2col, 1, data_type, fixed_point_position)); - - // Configure im2col kernel - _im2col_kernel.configure(input, &_im2col_output, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false); - - multiply_input = &_im2col_output; + // Fully Connected layer after a Convolution Layer without batches + configure_conv_fc(input, weights_to_use, output); } - - if(_is_batched_fc_layer) - { - TensorShape shape_interleaved(multiply_input->info()->tensor_shape()); - shape_interleaved.set(0, shape_interleaved.x() * 4); - shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f)); - _interleave4x4_output.allocator()->init(TensorInfo(shape_interleaved, 1, data_type, fixed_point_position)); - - // Configure interleave4x4 kernel - _interleave4x4_kernel.configure(multiply_input, &_interleave4x4_output); - - multiply_input = &_interleave4x4_output; - } - - // Configure matrix multiply kernel - _mm_kernel.configure(multiply_input, weights_to_use, output, 1.0f); - - if(_accumulate_biases) + else { - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); - ARM_COMPUTE_ERROR_ON(biases->info()->tensor_shape().x() != output->info()->tensor_shape().x()); - - // Configure accumulate biases kernel - _accumulate_biases_kernel.configure(output, biases); + // Fully Connected layer after a Fully Connected Layer without batches + configure_fc_fc(input, weights_to_use, output); } // Allocate the transpose tensor if the are_weights_reshaped flag is false and once all the configure methods have been called - if(!are_weights_reshaped && (transpose_weights || _is_batched_fc_layer)) + if(!_are_weights_reshaped) { // Allocate the tensor for the weights reshaped _reshape_weights_output.allocator()->allocate(); } - - if(_linearize_input) - { - _im2col_output.allocator()->allocate(); - } - - if(_is_batched_fc_layer) - { - _interleave4x4_output.allocator()->allocate(); - } } void CLFullyConnectedLayer::run() @@ -249,17 +159,11 @@ void CLFullyConnectedLayer::run() } // Linearize input if it comes from a convolutional layer - if(_linearize_input) + if(_is_fc_after_conv) { CLScheduler::get().enqueue(_im2col_kernel, false); } - // Interleave input - if(_is_batched_fc_layer) - { - CLScheduler::get().enqueue(_interleave4x4_kernel, false); - } - // Run matrix multiply CLScheduler::get().enqueue(_mm_kernel, !_accumulate_biases); @@ -269,4 +173,3 @@ void CLFullyConnectedLayer::run() CLScheduler::get().enqueue(_accumulate_biases_kernel); } } -} // namespace arm_compute diff --git a/src/runtime/CL/functions/CLGEMM.cpp b/src/runtime/CL/functions/CLGEMM.cpp index e81d8a6b97..9867229a7c 100644 --- a/src/runtime/CL/functions/CLGEMM.cpp +++ b/src/runtime/CL/functions/CLGEMM.cpp @@ -39,7 +39,7 @@ using namespace arm_compute; CLGEMM::CLGEMM() - : _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _ma_kernel(), _tmp_a(), _tmp_b(), _run_vector_matrix_multiplication(false), _run_addition(false) + : _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _ma_kernel(), _tmp_a(), _tmp_b(), _is_interleaved_transposed(false), _run_addition(false) { } @@ -59,12 +59,16 @@ void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor * 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"); - _mm_kernel.set_target(CLScheduler::get().target()); + // If the input tensor has less than 16 rows, we run a special version of GEMM without reshaping the input tensors + _is_interleaved_transposed = a->info()->dimension(1) > 16; - // Check if the first input tensor is a vector. If so, all the kernels for reshaping the tensors can be skipped - if(a->info()->dimension(1) != 1) + const ICLTensor *matrix_a = a; + const ICLTensor *matrix_b = b; + + if(_is_interleaved_transposed) { - _run_vector_matrix_multiplication = false; + matrix_a = &_tmp_a; + matrix_b = &_tmp_b; TensorShape shape_tmp_a = a->info()->tensor_shape(); TensorShape shape_tmp_b = b->info()->tensor_shape(); @@ -89,19 +93,17 @@ void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor * _transpose_kernel.configure(b, &_tmp_b); // Configure matrix multiply kernel - _mm_kernel.configure(&_tmp_a, &_tmp_b, output, alpha); + _mm_kernel.set_target(CLScheduler::get().target()); + } + _mm_kernel.configure(matrix_a, matrix_b, output, alpha, _is_interleaved_transposed); + + if(_is_interleaved_transposed) + { // Allocate intermediate tensors _tmp_a.allocator()->allocate(); _tmp_b.allocator()->allocate(); } - else // The first input tensor is a vector - { - _run_vector_matrix_multiplication = true; - - // Configure the matrix multiply kernel - _mm_kernel.configure(a, b, output, alpha); - } // Configure matrix addition kernel if(beta != 0 && c != nullptr) @@ -113,7 +115,7 @@ void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor * void CLGEMM::run() { - if(!_run_vector_matrix_multiplication) + if(_is_interleaved_transposed) { // Run interleave kernel CLScheduler::get().enqueue(_interleave_kernel, false); diff --git a/tests/model_objects/AlexNet.h b/tests/model_objects/AlexNet.h index c9fd448d5d..45622e2118 100644 --- a/tests/model_objects/AlexNet.h +++ b/tests/model_objects/AlexNet.h @@ -24,6 +24,8 @@ #ifndef __ARM_COMPUTE_TEST_MODEL_OBJECTS_ALEXNET_H__ #define __ARM_COMPUTE_TEST_MODEL_OBJECTS_ALEXNET_H__ +#include "arm_compute/runtime/Tensor.h" + #include "tests/AssetsLibrary.h" #include "tests/Globals.h" #include "tests/Utils.h" @@ -149,7 +151,7 @@ public: b[6]->allocator()->init(TensorInfo(TensorShape(4096U), 1, dt, fixed_point_position)); b[7]->allocator()->init(TensorInfo(TensorShape(1000U), 1, dt, fixed_point_position)); - if(_batches > 1) + if(_batches > 1 && std::is_same<TensorType, Tensor>::value) { w[5]->allocator()->init(TensorInfo(TensorShape(9216U * dt_size, 4096U / dt_size), 1, dt, fixed_point_position)); w[6]->allocator()->init(TensorInfo(TensorShape(4096U * dt_size, 4096U / dt_size), 1, dt, fixed_point_position)); diff --git a/tests/networks_new/AlexNetNetwork.h b/tests/networks_new/AlexNetNetwork.h index 39c69daf60..b3a719671d 100644 --- a/tests/networks_new/AlexNetNetwork.h +++ b/tests/networks_new/AlexNetNetwork.h @@ -24,6 +24,8 @@ #ifndef __ARM_COMPUTE_TEST_MODEL_OBJECTS_ALEXNET_H__ #define __ARM_COMPUTE_TEST_MODEL_OBJECTS_ALEXNET_H__ +#include "arm_compute/runtime/Tensor.h" + #include "AssetsLibrary.h" #include "Globals.h" #include "Utils.h" @@ -153,7 +155,7 @@ public: b[6].allocator()->init(TensorInfo(TensorShape(4096U), 1, _data_type, _fixed_point_position)); b[7].allocator()->init(TensorInfo(TensorShape(1000U), 1, _data_type, _fixed_point_position)); - if(_batches > 1) + if(_batches > 1 && std::is_same<TensorType, Tensor>::value) { w[5].allocator()->init(TensorInfo(TensorShape(9216U * data_type_size, 4096U / data_type_size), 1, _data_type, _fixed_point_position)); w[6].allocator()->init(TensorInfo(TensorShape(4096U * data_type_size, 4096U / data_type_size), 1, _data_type, _fixed_point_position)); diff --git a/tests/validation_new/CL/FullyConnectedLayer.cpp b/tests/validation_new/CL/FullyConnectedLayer.cpp index 9bf3a75d88..e43997c47b 100644 --- a/tests/validation_new/CL/FullyConnectedLayer.cpp +++ b/tests/validation_new/CL/FullyConnectedLayer.cpp @@ -80,16 +80,6 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(frame const size_t shape_x = ws.x(); ws.set(0, ws.y()); ws.set(1, shape_x); - - // Weights have to be passed reshaped - // Transpose 1xW for batched version - if(!reshape_weights && dst_shape.y() > 1) - { - const float transpose_width = 16.0f / data_size_from_type(data_type); - const size_t shape_x = ws.x(); - ws.set(0, ws.y() * static_cast<unsigned int>(transpose_width)); - ws.set(1, static_cast<unsigned int>(std::ceil(shape_x / transpose_width))); - } } // Create tensors @@ -113,7 +103,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(frame } template <typename T> -using CLFullyConnectedLayerFixture = FullyConnectedLayerValidationFixture<CLTensor, CLAccessor, CLFullyConnectedLayer, T>; +using CLFullyConnectedLayerFixture = FullyConnectedLayerValidationFixture<CLTensor, CLAccessor, CLFullyConnectedLayer, T, false>; TEST_SUITE(Float) TEST_SUITE(FP16) @@ -150,7 +140,7 @@ TEST_SUITE_END() TEST_SUITE_END() template <typename T> -using CLFullyConnectedLayerFixedPointFixture = FullyConnectedLayerValidationFixedPointFixture<CLTensor, CLAccessor, CLFullyConnectedLayer, T>; +using CLFullyConnectedLayerFixedPointFixture = FullyConnectedLayerValidationFixedPointFixture<CLTensor, CLAccessor, CLFullyConnectedLayer, T, false>; TEST_SUITE(Quantized) TEST_SUITE(QS8) diff --git a/tests/validation_new/NEON/FullyConnectedLayer.cpp b/tests/validation_new/NEON/FullyConnectedLayer.cpp index 6eb18ebc6a..e859fb3872 100644 --- a/tests/validation_new/NEON/FullyConnectedLayer.cpp +++ b/tests/validation_new/NEON/FullyConnectedLayer.cpp @@ -117,7 +117,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(frame } template <typename T> -using NEFullyConnectedLayerFixture = FullyConnectedLayerValidationFixture<Tensor, Accessor, NEFullyConnectedLayer, T>; +using NEFullyConnectedLayerFixture = FullyConnectedLayerValidationFixture<Tensor, Accessor, NEFullyConnectedLayer, T, true>; TEST_SUITE(Float) #ifdef ARM_COMPUTE_ENABLE_FP16 @@ -156,7 +156,7 @@ TEST_SUITE_END() TEST_SUITE_END() template <typename T> -using NEFullyConnectedLayerFixedPointFixture = FullyConnectedLayerValidationFixedPointFixture<Tensor, Accessor, NEFullyConnectedLayer, T>; +using NEFullyConnectedLayerFixedPointFixture = FullyConnectedLayerValidationFixedPointFixture<Tensor, Accessor, NEFullyConnectedLayer, T, true>; TEST_SUITE(Quantized) TEST_SUITE(QS8) diff --git a/tests/validation_new/fixtures/FullyConnectedLayerFixture.h b/tests/validation_new/fixtures/FullyConnectedLayerFixture.h index eb4aad8952..0953b0b67e 100644 --- a/tests/validation_new/fixtures/FullyConnectedLayerFixture.h +++ b/tests/validation_new/fixtures/FullyConnectedLayerFixture.h @@ -76,7 +76,7 @@ RawTensor transpose(const RawTensor &src, int interleave = 1) } } // namespace -template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool run_interleave> class FullyConnectedLayerValidationFixedPointFixture : public framework::Fixture { public: @@ -131,7 +131,7 @@ protected: // Weights have to be passed reshaped // Transpose 1xW for batched version - if(!reshape_weights && output_shape.y() > 1) + if(!reshape_weights && output_shape.y() > 1 && run_interleave) { const int transpose_width = 16 / data_size_from_type(data_type); const float shape_x = reshaped_weights_shape.x(); @@ -182,7 +182,7 @@ protected: tmp = transpose(tmp); // Reshape weights for batched runs - if(!reshape_weights && output_shape.y() > 1) + if(!reshape_weights && output_shape.y() > 1 && run_interleave) { // Transpose with interleave const int interleave_size = 16 / tmp.element_size(); @@ -232,15 +232,16 @@ protected: DataType _data_type{}; }; -template <typename TensorType, typename AccessorType, typename FunctionType, typename T> -class FullyConnectedLayerValidationFixture : public FullyConnectedLayerValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T> +template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool run_interleave> +class FullyConnectedLayerValidationFixture : public FullyConnectedLayerValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T, run_interleave> { public: template <typename...> void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, bool transpose_weights, bool reshape_weights, DataType data_type) { - FullyConnectedLayerValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, transpose_weights, reshape_weights, data_type, - 0); + FullyConnectedLayerValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T, run_interleave>::setup(input_shape, weights_shape, bias_shape, output_shape, transpose_weights, + reshape_weights, data_type, + 0); } }; } // namespace validation |