diff options
Diffstat (limited to 'src/gpu/cl/kernels/gemm/ClGemmHelpers.cpp')
| -rw-r--r-- | src/gpu/cl/kernels/gemm/ClGemmHelpers.cpp | 79 |
1 files changed, 55 insertions, 24 deletions
diff --git a/src/gpu/cl/kernels/gemm/ClGemmHelpers.cpp b/src/gpu/cl/kernels/gemm/ClGemmHelpers.cpp index 9350bf74bb..b5ebac3b49 100644 --- a/src/gpu/cl/kernels/gemm/ClGemmHelpers.cpp +++ b/src/gpu/cl/kernels/gemm/ClGemmHelpers.cpp @@ -39,14 +39,24 @@ namespace kernels { namespace gemm { -std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_lhs_rhs_info(unsigned int m, unsigned int n, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0, - bool lhs_interleave, bool rhs_interleave, bool lhs_transpose, bool rhs_transpose, bool export_to_cl_image) +std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_lhs_rhs_info(unsigned int m, + unsigned int n, + unsigned int m0, + unsigned int n0, + unsigned int k0, + unsigned int v0, + unsigned int h0, + bool lhs_interleave, + bool rhs_interleave, + bool lhs_transpose, + bool rhs_transpose, + bool export_to_cl_image) { ARM_COMPUTE_ERROR_ON(m0 == 0 || n0 == 0); ARM_COMPUTE_ERROR_ON(v0 == 0); v0 = std::max(std::min(static_cast<int>(m / m0), static_cast<int>(v0)), static_cast<int>(1)); - if(h0 == 0) + if (h0 == 0) { // When h0 is 0, we should take the maximum H0 possible h0 = std::max(n / n0, 1U); @@ -62,17 +72,22 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_lhs_rhs_info(unsigned return std::make_pair(lhs_info, rhs_info); } -std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> select_lhs_rhs_info(std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> info_img, - std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> info_buf, - unsigned int n, unsigned int k, unsigned int b, DataType data_type) +std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> +select_lhs_rhs_info(std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> info_img, + std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> info_buf, + unsigned int n, + unsigned int k, + unsigned int b, + DataType data_type) { - ARM_COMPUTE_ERROR_ON_MSG(info_buf.second.export_to_cl_image == true, "The fallback GeMM configuration cannot have export_to_cl_image = true"); + ARM_COMPUTE_ERROR_ON_MSG(info_buf.second.export_to_cl_image == true, + "The fallback GeMM configuration cannot have export_to_cl_image = true"); const TensorInfo tensor_rhs_info(TensorShape(n, k, b), 1, data_type); const TensorShape shape = misc::shape_calculator::compute_rhs_reshaped_shape(tensor_rhs_info, info_img.second); const TensorInfo tensor_reshaped_info(shape, 1, data_type); - if(bool(validate_image2d_support_on_rhs(tensor_reshaped_info, info_img.second))) + if (bool(validate_image2d_support_on_rhs(tensor_reshaped_info, info_img.second))) { return info_img; } @@ -90,42 +105,56 @@ void update_padding_for_cl_image(ITensorInfo *tensor) const unsigned int pixel_alignment = get_cl_image_pitch_alignment(CLKernelLibrary::get().get_device()); ARM_COMPUTE_ERROR_ON_MSG(pixel_alignment == 0, "Cannot retrieve cl_image pitch alignment"); - if(pixel_alignment == 0) + if (pixel_alignment == 0) { return; } const unsigned int row_pitch_alignment = pixel_alignment * num_floats_per_pixel; - const unsigned int round_up_width = ((stride_y_in_elements + row_pitch_alignment - 1) / row_pitch_alignment) * row_pitch_alignment; - const unsigned int padding = round_up_width - stride_y_in_elements; + const unsigned int round_up_width = + ((stride_y_in_elements + row_pitch_alignment - 1) / row_pitch_alignment) * row_pitch_alignment; + const unsigned int padding = round_up_width - stride_y_in_elements; tensor->extend_padding(PaddingSize(0, tensor->padding().right + padding, 0, 0)); } Status validate_image2d_support_on_rhs(const ITensorInfo &tensor_reshaped_info, const GEMMRHSMatrixInfo &rhs_info) { - if(rhs_info.export_to_cl_image) + if (rhs_info.export_to_cl_image) { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.n0 == 2) || (rhs_info.n0 == 3)) && rhs_info.transpose == false, "Export to cl_image only supported with n0 = 4, 8 or 16"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.k0 == 2) || (rhs_info.k0 == 3)) && rhs_info.transpose == true, "Export to cl_image only supported with k0 = 4, 8 or 16"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.n0 == 2) || (rhs_info.n0 == 3)) && rhs_info.transpose == false, + "Export to cl_image only supported with n0 = 4, 8 or 16"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.k0 == 2) || (rhs_info.k0 == 3)) && rhs_info.transpose == true, + "Export to cl_image only supported with k0 = 4, 8 or 16"); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(&tensor_reshaped_info, DataType::F32, DataType::F16); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(!image2d_from_buffer_supported(CLKernelLibrary::get().get_device()), "The extension cl_khr_image2d_from_buffer is not supported on the target platform"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(get_cl_image_pitch_alignment(CLKernelLibrary::get().get_device()) == 0, "Impossible to retrieve the cl_image pitch alignment"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG( + !image2d_from_buffer_supported(CLKernelLibrary::get().get_device()), + "The extension cl_khr_image2d_from_buffer is not supported on the target platform"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(get_cl_image_pitch_alignment(CLKernelLibrary::get().get_device()) == 0, + "Impossible to retrieve the cl_image pitch alignment"); // Check the width and height of the output tensor. // Since we cannot create a 3d image from a buffer, the third dimension is collapsed on the second dimension const size_t max_image_w = CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_IMAGE2D_MAX_WIDTH>(); const size_t max_image_h = CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_IMAGE2D_MAX_HEIGHT>(); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(tensor_reshaped_info.tensor_shape()[0] > max_image_w * 4, "Not supported width for cl_image"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(tensor_reshaped_info.tensor_shape()[1] * tensor_reshaped_info.tensor_shape()[2] > max_image_h, "Not supported height for cl_image"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(tensor_reshaped_info.tensor_shape()[0] > max_image_w * 4, + "Not supported width for cl_image"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG( + tensor_reshaped_info.tensor_shape()[1] * tensor_reshaped_info.tensor_shape()[2] > max_image_h, + "Not supported height for cl_image"); } return Status{}; } -bool is_mmul_kernel_preferred(const unsigned int m, const unsigned int n, const unsigned int k, const unsigned int b, - const DataType data_type, unsigned int &best_m0, unsigned int &best_n0) +bool is_mmul_kernel_preferred(const unsigned int m, + const unsigned int n, + const unsigned int k, + const unsigned int b, + const DataType data_type, + unsigned int &best_m0, + unsigned int &best_n0) { ARM_COMPUTE_UNUSED(n, k, b, data_type); @@ -141,7 +170,8 @@ bool is_mmul_kernel_preferred(const unsigned int m, const unsigned int n, const return ((k % mmul_k0) == 0) && (gws_y > 4); } -std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> find_lhs_rhs_info(const GeMMConfigsMatrix &configs, unsigned int m, unsigned int n, unsigned int k, unsigned int b) +std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> +find_lhs_rhs_info(const GeMMConfigsMatrix &configs, unsigned int m, unsigned int n, unsigned int k, unsigned int b) { size_t min_acc = std::numeric_limits<size_t>::max(); size_t min_idx = 0; @@ -150,12 +180,13 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> find_lhs_rhs_info(const GeMMConf const size_t num_rows = configs.size(); const size_t num_cols = configs[0].size(); - ARM_COMPUTE_ERROR_ON_MSG(num_cols != 14U, "The entry should have 14 integer values representing: M, N, K, B, M0, N0. K0, V0, H0, INT_LHS, INT_RHS, TRA_LHS, TRA_RHS, IMG_RHS"); + ARM_COMPUTE_ERROR_ON_MSG(num_cols != 14U, "The entry should have 14 integer values representing: M, N, K, B, M0, " + "N0. K0, V0, H0, INT_LHS, INT_RHS, TRA_LHS, TRA_RHS, IMG_RHS"); ARM_COMPUTE_UNUSED(num_cols); // Find nearest GeMM workload // Note: the workload does not depend on the K dimension - for(size_t y = 0; y < num_rows; ++y) + for (size_t y = 0; y < num_rows; ++y) { size_t mc0 = static_cast<size_t>(configs[y][0]); size_t nc0 = static_cast<size_t>(configs[y][1]); @@ -168,7 +199,7 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> find_lhs_rhs_info(const GeMMConf acc += (k - kc0) * (k - kc0); acc += (b - bc0) * (b - bc0); acc = std::sqrt(acc); - if(acc < min_acc) + if (acc < min_acc) { min_acc = acc; min_idx = y; |