diff options
Diffstat (limited to 'src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedKernel.cpp')
| -rw-r--r-- | src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedKernel.cpp | 108 |
1 files changed, 66 insertions, 42 deletions
diff --git a/src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedKernel.cpp b/src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedKernel.cpp index 15493f7ddc..ddbc809cdd 100644 --- a/src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedKernel.cpp +++ b/src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedKernel.cpp @@ -29,13 +29,12 @@ #include "arm_compute/core/CL/OpenCL.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Validate.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/core/utils/StringUtils.h" +#include "arm_compute/core/Validate.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" - #include "support/Cast.h" #include "support/StringSupport.h" @@ -51,45 +50,55 @@ namespace { using ElementsProcessed = Steps; -Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst, - const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info) +Status validate_arguments(const ITensorInfo *src0, + const ITensorInfo *src1, + const ITensorInfo *dst, + const GEMMLHSMatrixInfo &lhs_info, + const GEMMRHSMatrixInfo &rhs_info, + const GEMMReshapeInfo &gemm_info) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src0, src1, dst); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src0, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, src1); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(src0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(src1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(src0->num_dimensions() > 4, + "The number of dimensions for the LHS matrix must be <= 4"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(src1->num_dimensions() > 3, + "The number of dimensions for the RHS matrix must be <= 3"); ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.transpose); ARM_COMPUTE_RETURN_ERROR_ON(!rhs_info.transpose); ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 != rhs_info.k0); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(((lhs_info.k0 & (lhs_info.k0 - 1)) && lhs_info.k0 != 3), "Only 2,3,4,8,16 are supported for k0"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(((lhs_info.k0 & (lhs_info.k0 - 1)) && lhs_info.k0 != 3), + "Only 2,3,4,8,16 are supported for k0"); ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 > 16); ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 2 || lhs_info.m0 > 8); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.n0 & (rhs_info.n0 - 1)) && rhs_info.n0 != 3), "Only 2,3,4,8,16 are supported for n0"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.n0 & (rhs_info.n0 - 1)) && rhs_info.n0 != 3), + "Only 2,3,4,8,16 are supported for n0"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.export_to_cl_image, "Export to CLImage not supported for quantized GEMM"); const int m = gemm_info.m(); const int n = gemm_info.n(); const int k = gemm_info.k(); - TensorShape tensor_shape0{ src0->tensor_shape() }; + TensorShape tensor_shape0{src0->tensor_shape()}; tensor_shape0.set(0, k); tensor_shape0.set(1, m); - TensorShape tensor_shape1{ src1->tensor_shape() }; + TensorShape tensor_shape1{src1->tensor_shape()}; tensor_shape1.set(0, n); tensor_shape1.set(1, k); const TensorInfo tensor_info0 = src0->clone()->set_tensor_shape(tensor_shape0); const TensorInfo tensor_info1 = src1->clone()->set_tensor_shape(tensor_shape1); - const TensorInfo tensor_info_reshaped0 = src0->clone()->set_tensor_shape(compute_lhs_reshaped_shape(tensor_info0, lhs_info)); - const TensorInfo tensor_info_reshaped1 = src1->clone()->set_tensor_shape(compute_rhs_reshaped_shape(tensor_info1, rhs_info)); + const TensorInfo tensor_info_reshaped0 = + src0->clone()->set_tensor_shape(compute_lhs_reshaped_shape(tensor_info0, lhs_info)); + const TensorInfo tensor_info_reshaped1 = + src1->clone()->set_tensor_shape(compute_rhs_reshaped_shape(tensor_info1, rhs_info)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src0, &tensor_info_reshaped0); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src1, &tensor_info_reshaped1); - if(dst->total_size() != 0) + if (dst->total_size() != 0) { const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(compute_mm_shape(*src0, *src1, gemm_info)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_dst); @@ -99,19 +108,24 @@ Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, cons return Status{}; } -std::pair<Status, Window> validate_and_configure_window(const ITensorInfo *src0, const ITensorInfo *src1, ITensorInfo *dst, - const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info, - ElementsProcessed &num_elements_processed) +std::pair<Status, Window> validate_and_configure_window(const ITensorInfo *src0, + const ITensorInfo *src1, + ITensorInfo *dst, + const GEMMLHSMatrixInfo &lhs_info, + const GEMMRHSMatrixInfo &rhs_info, + const GEMMReshapeInfo &gemm_info, + ElementsProcessed &num_elements_processed) { unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0]; unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1]; bool reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0); // dst tensor auto initialization if not yet initialized - auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(compute_mm_shape(*src0, *src1, gemm_info)).set_data_type(DataType::S32)); + auto_init_if_empty( + *dst, src0->clone()->set_tensor_shape(compute_mm_shape(*src0, *src1, gemm_info)).set_data_type(DataType::S32)); TensorInfo tmp_info(*dst); - if(reinterpret_output_as_3d) + if (reinterpret_output_as_3d) { // Since the dst tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM, // the window needs to be constructed on the 2D collapsed version of the tensor @@ -123,7 +137,8 @@ std::pair<Status, Window> validate_and_configure_window(const ITensorInfo *src0, // Configure kernel window num_elems_processed_per_iteration_x = rhs_info.n0; num_elems_processed_per_iteration_y = lhs_info.m0; - Window win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); + Window win = + calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); // Collapse along the Z direction // This collapse needs to be here in order to tune the Z dimension of LWS @@ -140,8 +155,13 @@ ClGemmLowpMatrixMultiplyReshapedKernel::ClGemmLowpMatrixMultiplyReshapedKernel() _type = CLKernelType::GEMM; } -void ClGemmLowpMatrixMultiplyReshapedKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src0, const ITensorInfo *src1, ITensorInfo *dst, - const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info) +void ClGemmLowpMatrixMultiplyReshapedKernel::configure(const CLCompileContext &compile_context, + const ITensorInfo *src0, + const ITensorInfo *src1, + ITensorInfo *dst, + const GEMMLHSMatrixInfo &lhs_info, + const GEMMRHSMatrixInfo &rhs_info, + const GEMMReshapeInfo &gemm_info) { ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst); ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, dst, lhs_info, rhs_info, gemm_info)); @@ -154,11 +174,12 @@ void ClGemmLowpMatrixMultiplyReshapedKernel::configure(const CLCompileContext &c const unsigned int num_dimensionssrc0 = src0->num_dimensions(); _slide_matrix_b = (src1->num_dimensions() >= num_dimensionssrc0); - auto padding_info = get_padding_info({ src0, src1, dst }); + auto padding_info = get_padding_info({src0, src1, dst}); ElementsProcessed num_elements_processed{}; // Configure kernel window - auto win_config = validate_and_configure_window(src0, src1, dst, lhs_info, rhs_info, gemm_info, num_elements_processed); + auto win_config = + validate_and_configure_window(src0, src1, dst, lhs_info, rhs_info, gemm_info, num_elements_processed); ARM_COMPUTE_ERROR_THROW_ON(win_config.first); ICLKernel::configure_internal(win_config.second); @@ -171,8 +192,10 @@ void ClGemmLowpMatrixMultiplyReshapedKernel::configure(const CLCompileContext &c // Create build options CLBuildOptions build_opts; build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D"); - build_opts.add_option_if(_reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(dst->dimension(1))); - build_opts.add_option_if(_reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(dst->dimension(2))); + build_opts.add_option_if(_reinterpret_output_as_3d, + "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(dst->dimension(1))); + build_opts.add_option_if(_reinterpret_output_as_3d, + "-DDEPTH_GEMM3D=" + support::cpp11::to_string(dst->dimension(2))); build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(src1->dimension(2))); build_opts.add_option_if(lhs_info.interleave, "-DLHS_INTERLEAVE"); build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE"); @@ -230,19 +253,19 @@ void ClGemmLowpMatrixMultiplyReshapedKernel::configure(const CLCompileContext &c ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); } -Status ClGemmLowpMatrixMultiplyReshapedKernel::validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst, const GEMMLHSMatrixInfo &lhs_info, - const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info) +Status ClGemmLowpMatrixMultiplyReshapedKernel::validate(const ITensorInfo *src0, + const ITensorInfo *src1, + const ITensorInfo *dst, + const GEMMLHSMatrixInfo &lhs_info, + const GEMMRHSMatrixInfo &rhs_info, + const GEMMReshapeInfo &gemm_info) { ElementsProcessed num_elements_processed{}; ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src0, src1, dst, lhs_info, rhs_info, gemm_info)); - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src0->clone().get(), - src1->clone().get(), - dst->clone().get(), - lhs_info, - rhs_info, - gemm_info, + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src0->clone().get(), src1->clone().get(), + dst->clone().get(), lhs_info, rhs_info, gemm_info, num_elements_processed) - .first); + .first); return Status{}; } @@ -252,11 +275,13 @@ void ClGemmLowpMatrixMultiplyReshapedKernel::run_op(ITensorPack &tensors, const ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); - const auto src0 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0)); - const auto src1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1)); - auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); + const auto src0 = + utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0)); + const auto src1 = + utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1)); + auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); - if(src1->info()->num_dimensions() < 3) + if (src1->info()->num_dimensions() < 3) { // The stride_z for matrix B must be zero if we do not slice ARM_COMPUTE_ERROR_ON(src1->info()->strides_in_bytes()[3] != 0); @@ -268,7 +293,7 @@ void ClGemmLowpMatrixMultiplyReshapedKernel::run_op(ITensorPack &tensors, const slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1)); slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1)); - if(_reinterpret_output_as_3d) + if (_reinterpret_output_as_3d) { // Pass bottom paddings to the kernel if the dst has to be reinterpreted as 3D tensor const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 4; @@ -281,7 +306,7 @@ void ClGemmLowpMatrixMultiplyReshapedKernel::run_op(ITensorPack &tensors, const Window slice_b = slice; // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 // This scenario can happen when the matrix multiplication is used to perform a convolution operation - if(!_slide_matrix_b) + if (!_slide_matrix_b) { slice_b = slice_matrix_b; } @@ -295,8 +320,7 @@ void ClGemmLowpMatrixMultiplyReshapedKernel::run_op(ITensorPack &tensors, const _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src1->info()->strides_in_bytes()[2])); _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[2])); enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items); - } - while(window.slide_window_slice_3D(slice)); + } while (window.slide_window_slice_3D(slice)); } } // namespace kernels } // namespace opencl |