diff options
Diffstat (limited to 'src/gpu/cl/kernels')
6 files changed, 59 insertions, 194 deletions
diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp b/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp index 5fea097ae3..b8997dfc7f 100644 --- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp +++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp @@ -23,7 +23,6 @@ */ #include "src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h" -#include "arm_compute/core/utils/ActivationFunctionUtils.h" #include "arm_compute/core/CL/CLHelpers.h" #include "arm_compute/core/CL/CLKernelLibrary.h" #include "arm_compute/core/CL/ICLTensor.h" @@ -31,11 +30,11 @@ #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/ActivationFunctionUtils.h" #include "arm_compute/core/utils/StringUtils.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "src/core/AccessWindowStatic.h" #include "src/core/CL/CLUtils.h" -#include "src/core/experimental/PostOpUtils.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" #include "src/core/utils/helpers/float_ops.h" @@ -52,25 +51,6 @@ namespace { using ElementsProcessed = Steps; -const auto post_op_utils = experimental::PostOpCLKernelUtils( -{ - // PostOp sequence -> {Kernel Postfix, PostOp Slots} - { {}, { "", {} } }, - { { experimental::PostOpType::Activation }, { "", { 1 } } }, - - { { experimental::PostOpType::Eltwise_Add }, { "_post_act_eltwise_op_act", { 2 } } }, - { { experimental::PostOpType::Eltwise_PRelu }, { "_post_act_eltwise_op_act", { 2 } } }, - - { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_Add }, { "_post_act_eltwise_op_act", { 1, 2 } } }, - { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_PRelu }, { "_post_act_eltwise_op_act", { 1, 2 } } }, - - { { experimental::PostOpType::Eltwise_Add, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 2, 3 } } }, - { { experimental::PostOpType::Eltwise_PRelu, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 2, 3 } } }, - - { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_Add, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 1, 2, 3 } } }, - { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_PRelu, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 1, 2, 3 } } } -}); - Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info) @@ -90,7 +70,6 @@ Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, cons "Bias addition only supported with broadcast mode in case the input or dst has to be reinterpreted as 3D"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.fp_mixed_precision, "Mixed precision not supported"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.export_to_cl_image, "Export to CLImage not supported for GEMM native"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(!post_op_utils.is_post_op_sequence_supported(gemm_info.post_ops), "The sequence of Post Ops is not supported"); const unsigned int m = gemm_info.m; const unsigned int n = gemm_info.n; @@ -133,7 +112,6 @@ Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, cons const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_dst); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, dst); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(!post_op_utils.are_post_op_shapes_compliant(dst, gemm_info.post_ops), "The Post Op shapes are not compliant"); } return Status{}; @@ -240,7 +218,6 @@ void ClGemmMatrixMultiplyNativeKernel::configure(const CLCompileContext &compile _reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0; _use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device()); _add_bias = src2 != nullptr; - _num_post_op_args = gemm_info.post_ops.total_num_arguments(); // In case both input and dst have to be reinterpreted as 3D tensors, // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false. @@ -298,20 +275,11 @@ void ClGemmMatrixMultiplyNativeKernel::configure(const CLCompileContext &compile build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0)); build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0)); build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0)); - // If post_ops are used, then we disable the use of gemm_info.activation_info - if(gemm_info.post_ops.size() > 0) - { - post_op_utils.set_post_ops_cl_build_options(build_opts, gemm_info.post_ops); - } - else - { - build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation()))); - build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a())); - build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b())); - } + build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation()))); + build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a())); + build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b())); std::string kernel_name("gemm_mm_native"); - post_op_utils.set_post_ops_cl_kernel_name(kernel_name, gemm_info.post_ops); // A macro guard to compile ONLY the kernel of interest build_opts.add_option("-D" + upper_string(kernel_name)); @@ -396,11 +364,11 @@ void ClGemmMatrixMultiplyNativeKernel::run_op(ITensorPack &tensors, const Window unsigned int idx0; if(_add_bias) { - idx0 = (4 + _num_post_op_args) * num_arguments_per_2D_tensor() + (7 + _num_post_op_args); + idx0 = 4 * num_arguments_per_2D_tensor() + 7; } else { - idx0 = (3 + _num_post_op_args) * num_arguments_per_2D_tensor() + (6 + _num_post_op_args); + idx0 = 3 * num_arguments_per_2D_tensor() + 6; } const unsigned int total_cross_plane_pad = src0->info()->padding().top + src0->info()->padding().bottom; _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad)); @@ -412,11 +380,11 @@ void ClGemmMatrixMultiplyNativeKernel::run_op(ITensorPack &tensors, const Window unsigned int idx0; if(_add_bias) { - idx0 = (4 + _num_post_op_args) * num_arguments_per_2D_tensor() + 7 + (_reinterpret_input_as_3d ? 1 : 0) + _num_post_op_args; + idx0 = 4 * num_arguments_per_2D_tensor() + 7 + (_reinterpret_input_as_3d ? 1 : 0); } else { - idx0 = (3 + _num_post_op_args) * num_arguments_per_2D_tensor() + 6 + (_reinterpret_input_as_3d ? 1 : 0) + _num_post_op_args; + idx0 = 3 * num_arguments_per_2D_tensor() + 6 + (_reinterpret_input_as_3d ? 1 : 0); } const unsigned int total_cross_plane_pad = dst->info()->padding().top + dst->info()->padding().bottom; _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad)); @@ -440,12 +408,7 @@ void ClGemmMatrixMultiplyNativeKernel::run_op(ITensorPack &tensors, const Window add_2D_tensor_argument(idx, src2, slice); } add_2D_tensor_argument(idx, dst, slice); - // post op argument buffers - for(size_t i = 0; i < _num_post_op_args; ++i) - { - const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i))); - add_2D_tensor_argument(idx, post_op_arg, slice); - } + _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src0->info()->strides_in_bytes()[2])); _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src1->info()->strides_in_bytes()[2])); if(_add_bias) @@ -453,12 +416,6 @@ void ClGemmMatrixMultiplyNativeKernel::run_op(ITensorPack &tensors, const Window _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src2->info()->strides_in_bytes()[2])); } _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[2])); - // post op argument stride_z - for(size_t i = 0; i < _num_post_op_args; ++i) - { - const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i))); - _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(post_op_arg->info()->strides_in_bytes()[2])); - } // Pass m, n and k at runtime _kernel.setArg<cl_int>(idx++, _m); diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h b/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h index e478df727a..80f8355932 100644 --- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h +++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_CL_GEMM_MATRIXMULTIPLY_NATIVE_KERNEL_H -#define ARM_COMPUTE_CL_GEMM_MATRIXMULTIPLY_NATIVE_KERNEL_H +#ifndef ACL_SRC_GPU_CL_KERNELS_CLGEMMMATRIXMULTIPLYNATIVEKERNEL_H +#define ACL_SRC_GPU_CL_KERNELS_CLGEMMMATRIXMULTIPLYNATIVEKERNEL_H #include "arm_compute/core/KernelDescriptors.h" #include "src/core/common/Macros.h" @@ -76,17 +76,16 @@ public: void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override; private: - bool _slide_matrix_b{ true }; - bool _reinterpret_input_as_3d{ false }; - bool _reinterpret_output_as_3d{ false }; - bool _use_dummy_work_items{ false }; - bool _add_bias{ false }; - signed int _m{ 1 }; - signed int _n{ 1 }; - signed int _k{ 1 }; - unsigned int _num_post_op_args{ 0 }; // (EXPERIMENTAL_POST_OPS) total number of post op arguments + bool _slide_matrix_b{ true }; + bool _reinterpret_input_as_3d{ false }; + bool _reinterpret_output_as_3d{ false }; + bool _use_dummy_work_items{ false }; + bool _add_bias{ false }; + signed int _m{ 1 }; + signed int _n{ 1 }; + signed int _k{ 1 }; }; } // namespace kernels } // namespace opencl } // namespace arm_compute -#endif /* ARM_COMPUTE_CL_GEMM_MATRIXMULTIPLY_NATIVE_KERNEL_H */ +#endif // ACL_SRC_GPU_CL_KERNELS_CLGEMMMATRIXMULTIPLYNATIVEKERNEL_H diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.cpp b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.cpp index f14a6f1900..d72d29ea1e 100644 --- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.cpp +++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.cpp @@ -23,7 +23,6 @@ */ #include "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.h" -#include "arm_compute/core/utils/ActivationFunctionUtils.h" #include "arm_compute/core/CL/CLHelpers.h" #include "arm_compute/core/CL/CLKernelLibrary.h" #include "arm_compute/core/CL/ICLTensor.h" @@ -31,11 +30,11 @@ #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/ActivationFunctionUtils.h" #include "arm_compute/core/utils/StringUtils.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "src/core/CL/CLUtils.h" #include "src/core/CL/CLValidate.h" -#include "src/core/experimental/PostOpUtils.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" #include "src/core/utils/helpers/float_ops.h" @@ -53,25 +52,6 @@ namespace { using ElementsProcessed = Steps; -const auto post_op_utils = experimental::PostOpCLKernelUtils( -{ - // PostOp sequence -> {Kernel Postfix, PostOp Slots} - { {}, { "", {} } }, - { { experimental::PostOpType::Activation }, { "", { 1 } } }, - - { { experimental::PostOpType::Eltwise_Add }, { "_post_act_eltwise_op_act", { 2 } } }, - { { experimental::PostOpType::Eltwise_PRelu }, { "_post_act_eltwise_op_act", { 2 } } }, - - { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_Add }, { "_post_act_eltwise_op_act", { 1, 2 } } }, - { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_PRelu }, { "_post_act_eltwise_op_act", { 1, 2 } } }, - - { { experimental::PostOpType::Eltwise_Add, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 2, 3 } } }, - { { experimental::PostOpType::Eltwise_PRelu, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 2, 3 } } }, - - { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_Add, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 1, 2, 3 } } }, - { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_PRelu, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 1, 2, 3 } } } -}); - Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info) @@ -95,7 +75,6 @@ Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, cons "Bias addition only supported with broadcast mode in case the input or dst has to be reinterpreted as 3D"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.fp_mixed_precision && (src0->data_type() == DataType::F32), "Mixed precision only supported for F16 data type"); ARM_COMPUTE_RETURN_ON_ERROR(gemm::validate_image2d_support_on_rhs(*src1, rhs_info)); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(!post_op_utils.is_post_op_sequence_supported(gemm_info.post_ops), "The sequence of Post Ops is not supported"); const unsigned int m = gemm_info.m; const unsigned int n = gemm_info.n; @@ -139,7 +118,6 @@ Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, cons const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_dst); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, dst); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(!post_op_utils.are_post_op_shapes_compliant(dst, gemm_info.post_ops), "The Post Op shapes are not compliant"); } return Status{}; @@ -202,7 +180,6 @@ void ClGemmMatrixMultiplyReshapedKernel::configure(const CLCompileContext &compi _use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device()); _add_bias = src2 != nullptr; _export_to_cl_image = rhs_info.export_to_cl_image; - _num_post_op_args = gemm_info.post_ops.total_num_arguments(); // Check if we need to slide the matrix B const unsigned int num_dimensions_src0 = src0->num_dimensions(); @@ -260,23 +237,14 @@ void ClGemmMatrixMultiplyReshapedKernel::configure(const CLCompileContext &compi build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0)); build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0)); build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0)); - // If post_ops are used, then we disable the use of gemm_info.activation_info - if(gemm_info.post_ops.size() > 0) - { - post_op_utils.set_post_ops_cl_build_options(build_opts, gemm_info.post_ops); - } - else - { - build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation()))); - build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a())); - build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b())); - } + build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation()))); + build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a())); + build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b())); std::string kernel_name("gemm_mm_reshaped_"); kernel_name += lhs_info.transpose ? "lhs_t_" : "lhs_nt_"; kernel_name += rhs_info.transpose ? "rhs_t" : "rhs_nt"; kernel_name += rhs_info.export_to_cl_image ? "_texture" : ""; - post_op_utils.set_post_ops_cl_kernel_name(kernel_name, gemm_info.post_ops); // A macro guard to compile ONLY the kernel of interest build_opts.add_option("-D" + upper_string(kernel_name)); @@ -395,13 +363,6 @@ void ClGemmMatrixMultiplyReshapedKernel::run_op(ITensorPack &tensors, const Wind // dst buffer add_2D_tensor_argument(idx, dst, slice); - // post op argument buffers - for(size_t i = 0; i < _num_post_op_args; ++i) - { - const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i))); - add_2D_tensor_argument(idx, post_op_arg, slice); - } - // LHS stride_z _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src0->info()->strides_in_bytes()[2])); @@ -417,12 +378,6 @@ void ClGemmMatrixMultiplyReshapedKernel::run_op(ITensorPack &tensors, const Wind // dst stride_z _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[2])); - // post op argument stride_z - for(size_t i = 0; i < _num_post_op_args; ++i) - { - const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i))); - _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(post_op_arg->info()->strides_in_bytes()[2])); - } // Cross-plan padding (if _reinterpret_output_as_3d = true) if(_reinterpret_output_as_3d) { diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.h b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.h index 2d668b91a3..8d25412a40 100644 --- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.h +++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_CL_GEMM_MATRIXMULTIPLY_RESHAPED_KERNEL_H -#define ARM_COMPUTE_CL_GEMM_MATRIXMULTIPLY_RESHAPED_KERNEL_H +#ifndef ACL_SRC_GPU_CL_KERNELS_CLGEMMMATRIXMULTIPLYRESHAPEDKERNEL_H +#define ACL_SRC_GPU_CL_KERNELS_CLGEMMMATRIXMULTIPLYRESHAPEDKERNEL_H #include "src/core/common/Macros.h" #include "src/gpu/cl/ClCompileContext.h" @@ -100,17 +100,16 @@ public: void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override; private: - bool _slide_matrix_b{ true }; - bool _reinterpret_output_as_3d{ false }; - bool _use_dummy_work_items{ false }; - bool _add_bias{ false }; - bool _export_to_cl_image{ false }; - signed int _m{ 1 }; - signed int _n{ 1 }; - signed int _k{ 1 }; - unsigned int _num_post_op_args{ 0 }; // (EXPERIMENTAL_POST_OPS) total number of post op arguments + bool _slide_matrix_b{ true }; + bool _reinterpret_output_as_3d{ false }; + bool _use_dummy_work_items{ false }; + bool _add_bias{ false }; + bool _export_to_cl_image{ false }; + signed int _m{ 1 }; + signed int _n{ 1 }; + signed int _k{ 1 }; }; } // namespace kernels } // namespace opencl } // namespace arm_compute -#endif /* ARM_COMPUTE_CL_GEMM_MATRIXMULTIPLY_RESHAPED_KERNEL_H */
\ No newline at end of file +#endif // ACL_SRC_GPU_CL_KERNELS_CLGEMMMATRIXMULTIPLYRESHAPEDKERNEL_H diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp index f780538f53..b34c17cda8 100644 --- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp +++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp @@ -23,13 +23,12 @@ */ #include "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h" -#include "arm_compute/core/utils/ActivationFunctionUtils.h" #include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/core/utils/ActivationFunctionUtils.h" #include "arm_compute/core/utils/StringUtils.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "src/core/CL/CLUtils.h" #include "src/core/CL/CLValidate.h" -#include "src/core/experimental/PostOpUtils.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" #include "src/core/utils/helpers/float_ops.h" @@ -47,25 +46,6 @@ namespace { using ElementsProcessed = Steps; -const auto post_op_utils = experimental::PostOpCLKernelUtils( -{ - // PostOp sequence -> {Kernel Postfix, PostOp Slots} - { {}, { "", {} } }, - { { experimental::PostOpType::Activation }, { "", { 1 } } }, - - { { experimental::PostOpType::Eltwise_Add }, { "_post_act_eltwise_op_act", { 2 } } }, - { { experimental::PostOpType::Eltwise_PRelu }, { "_post_act_eltwise_op_act", { 2 } } }, - - { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_Add }, { "_post_act_eltwise_op_act", { 1, 2 } } }, - { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_PRelu }, { "_post_act_eltwise_op_act", { 1, 2 } } }, - - { { experimental::PostOpType::Eltwise_Add, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 2, 3 } } }, - { { experimental::PostOpType::Eltwise_PRelu, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 2, 3 } } }, - - { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_Add, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 1, 2, 3 } } }, - { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_PRelu, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 1, 2, 3 } } } -}); - Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info) { @@ -86,7 +66,6 @@ Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, cons "Bias addition only supported with broadcast mode in case the input or dst has to be reinterpreted as 3D"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.fp_mixed_precision, "Mixed precision not supported"); ARM_COMPUTE_RETURN_ON_ERROR(gemm::validate_image2d_support_on_rhs(*src1, rhs_info)); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(!post_op_utils.is_post_op_sequence_supported(gemm_info.post_ops), "The sequence of Post Ops is not supported"); const unsigned int m = gemm_info.m; const unsigned int n = gemm_info.n; @@ -132,7 +111,6 @@ Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, cons const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_dst); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, dst); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(!post_op_utils.are_post_op_shapes_compliant(dst, gemm_info.post_ops), "The Post Op shapes are not compliant"); } return Status{}; @@ -203,7 +181,6 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext _add_bias = src2 != nullptr; _export_to_cl_image = rhs_info.export_to_cl_image; _has_pad_y = gemm_info.has_pad_y; - _num_post_op_args = gemm_info.post_ops.total_num_arguments(); auto padding_info = get_padding_info({ src0, src1, src2, dst }); @@ -270,22 +247,14 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(h_gemm_3d)); build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(d_gemm_3d)); } - // If post_ops are used, then we disable the use of gemm_info.activation_info - if(gemm_info.post_ops.size() > 0) - { - post_op_utils.set_post_ops_cl_build_options(build_opts, gemm_info.post_ops); - } - else - { - build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation()))); - build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a())); - build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b())); - } + + build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation()))); + build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a())); + build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b())); std::string kernel_name("gemm_mm_reshaped_only_rhs_"); kernel_name += rhs_info.transpose ? "t" : "nt"; kernel_name += rhs_info.export_to_cl_image ? "_texture" : ""; - post_op_utils.set_post_ops_cl_kernel_name(kernel_name, gemm_info.post_ops); // A macro guard to compile ONLY the kernel of interest build_opts.add_option("-D" + upper_string(kernel_name)); @@ -411,13 +380,6 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::run_op(ITensorPack &tensors, con // dst buffer add_2D_tensor_argument(idx, dst, slice); - // post op argument buffers - for(size_t i = 0; i < _num_post_op_args; ++i) - { - const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i))); - add_2D_tensor_argument(idx, post_op_arg, slice); - } - // LHS stride_z _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src0->info()->strides_in_bytes()[lhs_idx_batch_size])); @@ -432,12 +394,6 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::run_op(ITensorPack &tensors, con // dst stride_z _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[out_idx_batch_size])); - // post op argument stride_z - for(size_t i = 0; i < _num_post_op_args; ++i) - { - const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i))); - _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(post_op_arg->info()->strides_in_bytes()[2])); - } // Cross-plan padding (if _reinterpret_input_as_3d = true) if(_reinterpret_input_as_3d && _has_pad_y) diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h index 00cdb299ce..471160c94b 100644 --- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h +++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2022 Arm Limited. + * Copyright (c) 2019-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_CL_GEMM_MATRIXMULTIPLY_RESHAPED_ONLY_RHS_KERNEL_H -#define ARM_COMPUTE_CL_GEMM_MATRIXMULTIPLY_RESHAPED_ONLY_RHS_KERNEL_H +#ifndef ACL_SRC_GPU_CL_KERNELS_CLGEMMMATRIXMULTIPLYRESHAPEDONLYRHSKERNEL_H +#define ACL_SRC_GPU_CL_KERNELS_CLGEMMMATRIXMULTIPLYRESHAPEDONLYRHSKERNEL_H #include "src/core/common/Macros.h" #include "src/gpu/cl/ClCompileContext.h" @@ -90,19 +90,18 @@ public: void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override; private: - bool _slide_matrix_b{ true }; - bool _reinterpret_input_as_3d{ false }; - bool _reinterpret_output_as_3d{ false }; - bool _use_dummy_work_items{ false }; - bool _add_bias{ false }; - bool _export_to_cl_image{ false }; - bool _has_pad_y{ false }; - signed int _m{ 1 }; - signed int _n{ 1 }; - signed int _k{ 1 }; - unsigned int _num_post_op_args{ 0 }; // (EXPERIMENTAL_POST_OPS) total number of post op arguments + bool _slide_matrix_b{ true }; + bool _reinterpret_input_as_3d{ false }; + bool _reinterpret_output_as_3d{ false }; + bool _use_dummy_work_items{ false }; + bool _add_bias{ false }; + bool _export_to_cl_image{ false }; + bool _has_pad_y{ false }; + signed int _m{ 1 }; + signed int _n{ 1 }; + signed int _k{ 1 }; }; } // namespace kernels } // namespace opencl } // namespace arm_compute -#endif /* ARM_COMPUTE_CL_GEMM_MATRIXMULTIPLY_RESHAPED_ONLY_RHS_KERNEL_H */ +#endif // ACL_SRC_GPU_CL_KERNELS_CLGEMMMATRIXMULTIPLYRESHAPEDONLYRHSKERNEL_H |