From c4f582ee0dbdd7579f751277d1f3676c8db91ae5 Mon Sep 17 00:00:00 2001 From: Isabella Gottardi Date: Thu, 11 Oct 2018 19:14:55 +0100 Subject: COMPMID-1451: Reverting changes for CLGEMM and CLGEMMLowp previuosly done (384496) Mirroring CLGEMM behaviour to CLGEMMLowp Change-Id: I308b54e2c0de131a5322b77e83e7454db498d692 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/153175 Reviewed-by: Gian Marco Iodice Tested-by: bsgcomp --- .../CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp | 12 ++--- src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp | 8 +--- src/runtime/CL/functions/CLGEMM.cpp | 20 ++------ .../CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp | 55 +++++++++------------- 4 files changed, 33 insertions(+), 62 deletions(-) (limited to 'src') diff --git a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp index 56f318d6a8..99e184050e 100644 --- a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp +++ b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp @@ -57,6 +57,7 @@ using ElementsProcessed = Steps; Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info) { + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the matrix A must be <= 4"); @@ -87,7 +88,7 @@ Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const TensorInfo tensor_info0 = input0->clone()->set_tensor_shape(tensor_shape0); const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1); - const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(compute_interleaved_shape(tensor_info0, mult_interleave4x4_height, reshape_info.reinterpret_input_as_3d())); + const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(compute_interleaved_shape(tensor_info0, mult_interleave4x4_height)); const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(compute_transpose1xW_with_element_size_shape(tensor_info1, mult_transpose1xW_width)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input0, &tensor_info_reshaped0); @@ -124,11 +125,8 @@ std::pair validate_and_configure_window(ITensorInfo *input0, ITe reinterpret_output_as_3d = false; } - GEMMReshapeInfo reshape_info_to_use = GEMMReshapeInfo(reshape_info.m(), reshape_info.n(), reshape_info.k(), reshape_info.mult_transpose1xW_width(), reshape_info.mult_interleave4x4_height(), - reinterpret_output_as_3d ? reshape_info.depth_output_gemm3d() : 1, reinterpret_input_as_3d); - // Output tensor auto inizialitation if not yet initialized - auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, is_interleaved_transposed, reshape_info_to_use))); + auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, is_interleaved_transposed, reshape_info))); TensorInfo tmp_info(*output); @@ -145,7 +143,7 @@ std::pair validate_and_configure_window(ITensorInfo *input0, ITe if(is_interleaved_transposed) { // reinterpret_input_as_3d is not supported if is_interleaved_transposed is set - ARM_COMPUTE_ERROR_ON(reinterpret_input_as_3d); + ARM_COMPUTE_ERROR_ON(reshape_info.reinterpret_input_as_3d()); // Configure kernel window num_elems_processed_per_iteration_x = 4; @@ -198,7 +196,7 @@ std::pair validate_and_configure_window(ITensorInfo *input0, ITe Coordinates coord; coord.set_num_dimensions(output->num_dimensions()); - output_access.set_valid_region(win, ValidRegion(coord, output->tensor_shape())); + output_access.set_valid_region(win_out, ValidRegion(coord, output->tensor_shape())); } // Collapse along the Z direction diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp index c8bcb37b9c..715edae606 100644 --- a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp +++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp @@ -120,12 +120,8 @@ inline std::pair validate_and_configure_window(ITensorInfo *inpu reinterpret_output_as_3d = false; } - const GEMMReshapeInfo reshape_info_to_use(reshape_info.m(), reshape_info.n(), reshape_info.k(), reshape_info.mult_transpose1xW_width(), - reshape_info.mult_interleave4x4_height(), reinterpret_output_as_3d ? reshape_info.depth_output_gemm3d() : 1, reinterpret_input_as_3d); - // Output tensor auto inizialitation if not yet initialized - auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, is_interleaved_transposed, - reshape_info_to_use))); + auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, is_interleaved_transposed, reshape_info))); TensorInfo tmp_info(*output); @@ -141,7 +137,7 @@ inline std::pair validate_and_configure_window(ITensorInfo *inpu if(is_interleaved_transposed) { // reinterpret_input_as_3d is not supported if is_interleaved_transposed is set - ARM_COMPUTE_ERROR_ON(reinterpret_input_as_3d); + ARM_COMPUTE_ERROR_ON(reshape_info.reinterpret_input_as_3d()); // Configure kernel window num_elems_processed_per_iteration_x = max_cl_vector_width / data_size_from_type(data_type); diff --git a/src/runtime/CL/functions/CLGEMM.cpp b/src/runtime/CL/functions/CLGEMM.cpp index 0c82e6d788..6adbdc0cb6 100644 --- a/src/runtime/CL/functions/CLGEMM.cpp +++ b/src/runtime/CL/functions/CLGEMM.cpp @@ -133,10 +133,7 @@ void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor * if(_is_interleaved_transposed) { reinterpret_input_as_3d = false; - } - if(_is_interleaved_transposed) - { matrix_a = &_tmp_a; matrix_b = &_tmp_b; @@ -200,8 +197,7 @@ Status CLGEMM::validate(const ITensorInfo *a, const ITensorInfo *b, const ITenso // If we pass the matrix A and matrix B reshaped to CLGEMMMatrixMultiplyKernel, we need to pass m, n, k, mult_transpose1xW_width and mult_interleave4x4_height to CLGEMMReshapeInfo // in order to know how the matrices have been reshaped bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); - bool reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 1); - const int m = a->dimension(1); + const int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); const int n = b->dimension(0); const int k = a->dimension(0); int mult_transpose1xW_width = 1; @@ -217,21 +213,13 @@ Status CLGEMM::validate(const ITensorInfo *a, const ITensorInfo *b, const ITenso // Check if we need to reshape the matrix A and matrix B const bool run_interleave_transpose = is_interleaved_transposed(m, n, k, a->data_type(), reshape_b_only_on_first_run, gpu_target); - // In case both input and output have to be reinterpreted as 3D tensors, - // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false. - if(reinterpret_input_as_3d == reinterpret_output_as_3d) - { - reinterpret_input_as_3d = false; - reinterpret_output_as_3d = false; - } - // if _is_interleaved_transposed is set, force reinterpret_input_as_3d to be false as the output of CLGEMMInterleaveKernel will be 2D if(run_interleave_transpose) { reinterpret_input_as_3d = false; } - const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height, reinterpret_output_as_3d ? depth_output_gemm3d : 1, reinterpret_input_as_3d); + const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height, depth_output_gemm3d, reinterpret_input_as_3d); if(run_interleave_transpose) { @@ -239,8 +227,8 @@ Status CLGEMM::validate(const ITensorInfo *a, const ITensorInfo *b, const ITenso matrix_b_info = &tmp_b_info; // Validate interleave kernel - auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_interleaved_shape(*a, mult_interleave4x4_height, reinterpret_input_as_3d))); - ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMInterleave4x4Kernel::validate(a, &tmp_a_info, mult_interleave4x4_height, reinterpret_input_as_3d)); + auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_interleaved_shape(*a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d()))); + ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMInterleave4x4Kernel::validate(a, &tmp_a_info, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d())); // Validate transpose kernel auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_transpose1xW_with_element_size_shape(*b, mult_transpose1xW_width))); diff --git a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp index 7aeade1e3e..509b668bc9 100644 --- a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp +++ b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp @@ -108,22 +108,13 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor // If we pass the matrix A and matrix B reshaped to CLGEMMMatrixMultiplyKernel, we need to pass m, n, k, mult_transpose1xW_width and mult_interleave4x4_height to CLGEMMReshapeInfo // in order to know how the matrices have been reshaped bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); - bool reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 1); - int m = reinterpret_input_as_3d ? (a->info()->dimension(1) * a->info()->dimension(2)) : a->info()->dimension(1); + const int m = reinterpret_input_as_3d ? (a->info()->dimension(1) * a->info()->dimension(2)) : a->info()->dimension(1); const int n = b->info()->dimension(0); const int k = a->info()->dimension(0); const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); constexpr int mult_transpose1xW_width = 1; constexpr int mult_interleave4x4_height = 1; - // In case both input and output have to be reinterpreted as 3D tensors, - // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false. - if(reinterpret_input_as_3d == reinterpret_output_as_3d) - { - reinterpret_input_as_3d = false; - reinterpret_output_as_3d = false; - } - // Check if we need to reshape the matrix A and matrix B _is_interleaved_transposed = is_interleaved_transposed(m, n, k, _reshape_b_only_on_first_run, gpu_target); @@ -131,7 +122,6 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor { // if _is_interleaved_transposed is set, force reinterpret_input_as_3d to be false as the output of CLGEMMInterleaveKernel will be 2D reinterpret_input_as_3d = false; - m = a->info()->dimension(1); matrix_a = &_tmp_a; matrix_b = &_tmp_b; @@ -143,7 +133,7 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor } // Configure interleave kernel - _mtx_a_reshape_kernel.configure(a, &_tmp_a, mult_interleave4x4_height, reinterpret_input_as_3d); + _mtx_a_reshape_kernel.configure(a, &_tmp_a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d()); // Configure transpose kernel _mtx_b_reshape_kernel.configure(b, &_tmp_b, mult_transpose1xW_width); @@ -151,7 +141,7 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor // Configure matrix multiply kernel _mm_kernel.configure(matrix_a, matrix_b, output, _is_interleaved_transposed, GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height, - reinterpret_output_as_3d ? depth_output_gemm3d : 1, reinterpret_input_as_3d)); + depth_output_gemm3d, reinterpret_input_as_3d)); // Initialize matrix B reduction kernel only if _a_offset is not equal to 0 if(_a_offset != 0) @@ -213,23 +203,20 @@ Status CLGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso int32_t a_offset = a->quantization_info().offset; int32_t b_offset = b->quantization_info().offset; + const ITensorInfo *matrix_a_info = a; + const ITensorInfo *matrix_b_info = b; + + TensorInfo tmp_a_info{}; + TensorInfo tmp_b_info{}; + bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); - bool reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 1); - int m = a->dimension(1); + const int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); const int n = b->dimension(0); const int k = a->dimension(0); constexpr int mult_transpose1xW_width = 1; constexpr int mult_interleave4x4_height = 1; const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); - // In case both input and output have to be reinterpreted as 3D tensors, - // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false. - if(reinterpret_input_as_3d == reinterpret_output_as_3d) - { - reinterpret_input_as_3d = false; - reinterpret_output_as_3d = false; - } - bool reshape_matrices = is_interleaved_transposed(m, n, k, gemm_info.reshape_b_only_on_first_run(), CLScheduler::get().target()); // if reshape_matrices is set, force reinterpret_input_as_3d to be false as the output of CLGEMMInterleaveKernel will be 2D @@ -238,22 +225,24 @@ Status CLGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso reinterpret_input_as_3d = false; } - const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height, reinterpret_output_as_3d ? depth_output_gemm3d : 1, reinterpret_input_as_3d); + const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height, depth_output_gemm3d, reinterpret_input_as_3d); if(reshape_matrices) { - TensorInfo info_a(compute_interleaved_shape(*a, mult_interleave4x4_height, reinterpret_input_as_3d), 1, a->data_type()); - TensorInfo info_b(compute_transpose1xW_with_element_size_shape(*b, mult_transpose1xW_width), 1, b->data_type()); + matrix_a_info = &tmp_a_info; + matrix_b_info = &tmp_b_info; - ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMInterleave4x4Kernel::validate(a, &info_a, mult_interleave4x4_height, reinterpret_input_as_3d)); - ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMTranspose1xWKernel::validate(b, &info_b, mult_transpose1xW_width)); - ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyKernel::validate(&info_a, &info_b, output, reshape_matrices, reshape_info)); - } - else - { - ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyKernel::validate(a, b, output, reshape_matrices, reshape_info)); + // Validate interleave kernel + auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_interleaved_shape(*a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d()))); + ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMInterleave4x4Kernel::validate(a, &tmp_a_info, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d())); + + // Validate transpose kernel + auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_transpose1xW_with_element_size_shape(*b, mult_transpose1xW_width))); } + // Validate matrix multiply + ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyKernel::validate(matrix_a_info, matrix_b_info, output, reshape_matrices, reshape_info)); + TensorInfo info_vector_sum_col, info_vector_sum_row; // Validate matrix B reduction kernel only if _a_offset is not equal to 0 -- cgit v1.2.1