From 4bfc70e31766587c951204c93a127a486e007d0c Mon Sep 17 00:00:00 2001 From: Gunes Bayir Date: Fri, 10 Dec 2021 16:17:56 +0000 Subject: Add Gemm MMUL Reshaped Only Rhs Support for FP32/FP16 This patch introduces a GEMM routine that is optimized for Arm(R) Mali(TM)-G715 and Arm(R) Mali(TM)-G615 Resolves: COMPMID-5216 Signed-off-by: Gunes Bayir Change-Id: I2e5d7806f5904347185bb3e250f73d73d6669dba Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7914 Reviewed-by: SiCong Li Tested-by: Arm Jenkins Comments-Addressed: Arm Jenkins Benchmark: Arm Jenkins --- src/gpu/cl/ClKernelLibrary.cpp | 6 + ...GemmMatrixMultiplyReshapedOnlyRhsMMULKernel.cpp | 365 +++++++++++++++++++++ ...ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel.h | 89 +++++ src/gpu/cl/kernels/gemm/ClGemmHelpers.cpp | 19 +- src/gpu/cl/kernels/gemm/ClGemmHelpers.h | 17 +- .../ClGemmDefaultConfigReshapedRhsOnlyValhall.cpp | 42 ++- .../ClGemmDefaultConfigReshapedRhsOnlyValhall.h | 4 +- src/gpu/cl/operators/ClGemm.cpp | 134 ++++++++ src/gpu/cl/operators/ClGemm.h | 29 +- 9 files changed, 689 insertions(+), 16 deletions(-) create mode 100644 src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel.cpp create mode 100644 src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel.h (limited to 'src/gpu/cl') diff --git a/src/gpu/cl/ClKernelLibrary.cpp b/src/gpu/cl/ClKernelLibrary.cpp index 1bf7f2b3ac..52661d6d79 100644 --- a/src/gpu/cl/ClKernelLibrary.cpp +++ b/src/gpu/cl/ClKernelLibrary.cpp @@ -272,6 +272,8 @@ const std::map ClKernelLibrary::_kernel_program_map = { "gemm_mv", "common/gemv.cl" }, { "gemm_mv_quantized", "common/gemv.cl" }, { "gemm_mm_native", "common/gemm.cl" }, + { "gemm_mm_reshaped_only_rhs_nt_mmul", "common/gemm_reshaped_only_rhs_mmul.cl" }, + { "gemm_mm_reshaped_only_rhs_nt_mmul_texture", "common/gemm_reshaped_only_rhs_mmul.cl" }, { "gemm_mm_native_post_act_eltwise_op_act", "common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl" }, { "gemm_mm_reshaped_lhs_nt_rhs_t", "common/gemm.cl" }, { "gemm_mm_reshaped_lhs_nt_rhs_t_texture", "common/gemm.cl" }, @@ -582,6 +584,10 @@ const std::map ClKernelLibrary::_program_source_map = { "common/gemm.cl", #include "./cl_kernels/common/gemm.clembed" + }, + { + "common/gemm_reshaped_only_rhs_mmul.cl", +#include "./cl_kernels/common/gemm_reshaped_only_rhs_mmul.clembed" }, { "common/gemm_utils.cl", diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel.cpp b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel.cpp new file mode 100644 index 0000000000..fe46913517 --- /dev/null +++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel.cpp @@ -0,0 +1,365 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel.h" + +#include "arm_compute/core/CL/CLHelpers.h" +#include "arm_compute/core/CL/CLKernelLibrary.h" +#include "arm_compute/core/CL/ICLTensor.h" +#include "arm_compute/core/CL/OpenCL.h" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "src/core/CL/CLUtils.h" +#include "src/core/helpers/AutoConfiguration.h" +#include "src/core/helpers/WindowHelpers.h" +#include "src/core/utils/helpers/float_ops.h" +#include "src/gpu/cl/kernels/gemm/ClGemmHelpers.h" +#include "support/Cast.h" +#include "support/StringSupport.h" + +namespace arm_compute +{ +namespace opencl +{ +namespace kernels +{ +namespace +{ +using ElementsProcessed = Steps; + +// Block size dimensions for the MMUL extension +constexpr int mmul_m0 = 4; +constexpr int mmul_n0 = 4; +constexpr int mmul_k0 = 4; + +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) +{ + ARM_COMPUTE_UNUSED(alpha); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src0, src1, dst); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(!arm_matrix_multiply_supported(CLKernelLibrary::get().get_device()), "The extension cl_arm_matrix_multiply is not supported on the target platform"); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src0, 1, DataType::F16, DataType::F32); + 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(lhs_info.m0 < 1, "Only values greater than 0 are supported for m0"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.n0 != 1 && rhs_info.n0 != 2 && rhs_info.n0 != 3 && rhs_info.n0 != 4 && rhs_info.n0 != 8 && rhs_info.n0 != 16, "Only 1,2,3,4,8, and 16 are supported for n0"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG((rhs_info.k0 != 1 || lhs_info.k0 != 1), "Only 1 is supported for k0"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG((rhs_info.h0 != 4), "Only 4 is supported for h0"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.interleave != true, "Only true is supported for interleave with mmul extension enabled"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.transpose != false, "Only false is supported for transpose with mmul extension enabled"); + 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)); + + const unsigned int m = gemm_info.m; + const unsigned int n = gemm_info.n; + const unsigned int k = gemm_info.k; + + ARM_COMPUTE_UNUSED(m); + ARM_COMPUTE_UNUSED(n); + ARM_COMPUTE_UNUSED(k); + + ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(0) != k); + + // Validate the reinterpreted-as-3D-case + if(gemm_info.depth_output_gemm3d != 0) + { + ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(1) * src0->dimension(2) != m); + } + else + { + ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(1) != m); + } + + // Validate the gemm-batched case + if(src1->num_dimensions() > 2) + { + if(gemm_info.depth_output_gemm3d != 0) + { + ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(3) != src1->dimension(2)); + } + else + { + ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(2) != src1->dimension(2)); + } + } + + if(src2 != nullptr && !(helpers::float_ops::is_zero(beta))) + { + const unsigned int src2_dim0 = src2->dimension(0); + const unsigned int src2_dim1 = src2->dimension(1); + + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src2, src1); + if(gemm_info.broadcast_bias) + { + ARM_COMPUTE_RETURN_ERROR_ON_MSG((src2_dim1 != 1 || src2_dim0 != n), "Incorrect dimension of bias matrix which is to be broadcasted"); + } + else + { + ARM_COMPUTE_RETURN_ERROR_ON_MSG((src2_dim0 != n || src2_dim1 != m), "Incorrect dimension of bias matrix"); + } + } + + TensorShape tensor_shape1{ src1->tensor_shape() }; + tensor_shape1.set(0, n); + tensor_shape1.set(1, k); + + const TensorInfo tensor_info1 = src1->clone()->set_tensor_shape(tensor_shape1); + const TensorInfo tensor_info_reshaped1 = src1->clone()->set_tensor_shape(misc::shape_calculator::compute_rhs_reshaped_shape(tensor_info1, rhs_info)); + + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src1, &tensor_info_reshaped1); + + if(dst->total_size() != 0) + { + 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); + } + + return Status{}; +} + +std::pair validate_and_configure_window(ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, const GEMMLHSMatrixInfo &lhs_info, + const GEMMRHSMatrixInfo &rhs_info, + const GEMMKernelInfo &gemm_info) +{ + ARM_COMPUTE_UNUSED(src0, src1, src2); + 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(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info))); + + TensorInfo tmp_info(*dst); + + 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 + TensorShape tmp_shape(dst->tensor_shape()); + tmp_shape.collapse(2U, 1U); + tmp_info.set_tensor_shape(tmp_shape); + } + + Window win = calculate_max_window(tmp_info, Steps(1, 1)); + + // Collapse along the Z direction + // This collapse needs to be here in order to tune the Z dimension of LWS + const unsigned int dimension_to_collapse = std::min(static_cast(dst->num_dimensions()), 2u); + Window collapsed = win.collapse(win, dimension_to_collapse); + + // Reconfigure window size, one arm_matrix_multiply kernel needs 16 threads to finish. + Window::Dimension x_dimension = collapsed.x(); + Window::Dimension y_dimension = collapsed.y(); + + // Make M and N multiple of M0 and N0 respectively + const unsigned int ceil_to_multiple_n_n0 = ceil_to_multiple(x_dimension.end(), rhs_info.n0); + const unsigned int ceil_to_multiple_m_m0 = ceil_to_multiple(y_dimension.end(), lhs_info.m0); + + // Divide M and N by M0 and N0 respectively + const unsigned int n_div_n0 = ceil_to_multiple_n_n0 / rhs_info.n0; + const unsigned int m_div_m0 = ceil_to_multiple_m_m0 / lhs_info.m0; + + // Make n_div_n0 and m_div_m0 multiple of mmul_n0 and mmul_k0 respectively + const unsigned int ceil_to_multiple_n_div_n0_mmul_n0 = ceil_to_multiple(n_div_n0, mmul_n0); + const unsigned int ceil_to_multiple_m_div_m0_mmul_k0 = ceil_to_multiple(m_div_m0, mmul_k0); + + // Ensure x_dimension is multiple of MMUL block size (mmul_n0 * mmul_k0) + x_dimension.set_end(ceil_to_multiple_n_div_n0_mmul_n0 * mmul_k0); + y_dimension.set_end(ceil_to_multiple_m_div_m0_mmul_k0 / mmul_k0); + + collapsed.set(Window::DimX, x_dimension); + collapsed.set(Window::DimY, y_dimension); + + return std::make_pair(Status{}, collapsed); +} +} // namespace + +ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel::ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel() +{ + _type = CLKernelType::GEMM; +} + +void ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, float alpha, + float beta, + const GEMMLHSMatrixInfo &lhs_info, + const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst); + + // dst tensor auto initialization if not yet initialized + auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info))); + + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info)); + + auto padding_info = get_padding_info({ src0, src1, src2, dst }); + _add_bias = src2 != nullptr; + _export_to_cl_image = rhs_info.export_to_cl_image; + + // Configure kernel window + auto win_config = validate_and_configure_window(src0, src1, src2, dst, lhs_info, rhs_info, gemm_info); + ARM_COMPUTE_ERROR_THROW_ON(win_config.first); + + IClKernel::configure_internal(win_config.second); + + _m = gemm_info.m; + _n = gemm_info.n; + _k = gemm_info.k; + + const unsigned int m0_leftover = _m % lhs_info.m0; + const unsigned int n0_leftover = _n % rhs_info.n0; + + // Create build options + CLBuildOptions build_opts; + build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src0->data_type())); + build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha)); + build_opts.add_option_if(src2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta)); + build_opts.add_option_if(helpers::float_ops::is_one(beta), "-DUNIT_BETA"); + build_opts.add_option_if(gemm_info.broadcast_bias, "-DBROADCAST_BIAS"); + build_opts.add_option_if(src0->data_type() == DataType::F16, "-DHALF_PRECISION"); + build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0)); + build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0)); + build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0)); + build_opts.add_option("-DM0_LEFTOVER=" + support::cpp11::to_string(m0_leftover)); + build_opts.add_option("-DN0_LEFTOVER=" + support::cpp11::to_string(n0_leftover)); + build_opts.add_option("-DMMUL_M0=" + support::cpp11::to_string(mmul_m0)); + build_opts.add_option("-DMMUL_N0=" + support::cpp11::to_string(mmul_n0)); + build_opts.add_option("-DMMUL_K0=" + support::cpp11::to_string(mmul_k0)); + build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation()))); + build_opts.add_option("-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a())); + build_opts.add_option("-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b())); + + std::string kernel_name("gemm_mm_reshaped_only_rhs_nt_mmul"); + kernel_name += rhs_info.export_to_cl_image ? "_texture" : ""; + + // A macro guard to compile ONLY the kernel of interest + build_opts.add_option("-D" + upper_string(kernel_name)); + + // Create kernel + _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); + + // Set config_id for enabling LWS tuning + _config_id = kernel_name; + _config_id += "_"; + _config_id += (_add_bias ? "add_bias_" : ""); + _config_id += (gemm_info.broadcast_bias ? "broadcast_bias_" : ""); + _config_id += (gemm_info.activation_info.enabled() ? "fused_activation_" : ""); + _config_id += lower_string(string_from_data_type(src0->data_type())); + _config_id += "_"; + _config_id += support::cpp11::to_string(_m); + _config_id += "_"; + _config_id += support::cpp11::to_string(_n); + _config_id += "_"; + _config_id += support::cpp11::to_string(_k); + _config_id += "_"; + _config_id += support::cpp11::to_string(lhs_info.m0); + _config_id += "_"; + _config_id += support::cpp11::to_string(rhs_info.n0); + + ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); +} + +Status ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel::validate(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) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src0->clone().get(), + src1->clone().get(), + src2 != nullptr ? src2->clone().get() : nullptr, + dst->clone().get(), + lhs_info, + rhs_info, + gemm_info) + .first); + + return Status{}; +} + +void ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) +{ + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); + + const auto src0 = utils::cast::polymorphic_downcast(tensors.get_const_tensor(TensorType::ACL_SRC_0)); + const auto src1 = utils::cast::polymorphic_downcast(tensors.get_const_tensor(TensorType::ACL_SRC_1)); + const auto src2 = utils::cast::polymorphic_downcast(tensors.get_const_tensor(TensorType::ACL_SRC_2)); + auto dst = utils::cast::polymorphic_downcast(tensors.get_tensor(TensorType::ACL_DST)); + + ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst); + ARM_COMPUTE_ERROR_ON(_add_bias && src2 == nullptr); + + 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); + } + + cl::Image2D src1_image2d; + + if(_export_to_cl_image) + { + const TensorShape shape2d(src1->info()->dimension(0) / 4, src1->info()->dimension(1) * src1->info()->dimension(2)); + const size_t image_row_pitch = src1->info()->strides_in_bytes()[1]; + + src1_image2d = create_image2d_from_buffer(CLKernelLibrary::get().context(), src1->cl_buffer(), shape2d, src1->info()->data_type(), image_row_pitch); + } + + Window slice = window.first_slice_window_3D(); + + do + { + unsigned int idx = 0; + + add_3d_tensor_nhw_argument(idx, src0); + if(_export_to_cl_image) + { + _kernel.setArg(idx++, src1_image2d); + } + add_3d_tensor_nhw_argument(idx, src1); + + // Bias buffer (_add_bias == true) + if(_add_bias) + { + add_3d_tensor_nhw_argument(idx, src2); + } + // dst buffer + add_3d_tensor_nhw_argument(idx, dst); + + // Pass m, n and k at runtime as signed ints, to ensure results of any subtractions they could be operand in, would still be signed. + _kernel.setArg(idx++, _m); + _kernel.setArg(idx++, _n); + _kernel.setArg(idx++, _k); + + // LWS_x should be multiple of 16 at least. (32, 2) has been chosen to have more work-items on a single core + // LWS also enforces the order of execution of the workitems which improves cache utilization + enqueue(queue, *this, slice, cl::NDRange(32, 2), false); + } + while(window.slide_window_slice_3D(slice)); +} +} // namespace kernels +} // namespace opencl +} // namespace arm_compute diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel.h b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel.h new file mode 100644 index 0000000000..59612fcf5d --- /dev/null +++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel.h @@ -0,0 +1,89 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * 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_MMUL_KERNEL_H +#define ARM_COMPUTE_CL_GEMM_MATRIXMULTIPLY_RESHAPED_ONLY_RHS_MMUL_KERNEL_H + +#include "arm_compute/core/KernelDescriptors.h" +#include "src/core/common/Macros.h" +#include "src/gpu/cl/ClCompileContext.h" +#include "src/gpu/cl/IClKernel.h" + +namespace arm_compute +{ +namespace opencl +{ +namespace kernels +{ +/** OpenCL kernel to multiply matrices using MMUL when only the input matrix RHS (src1) has been reshaped */ +class ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel : public IClKernel +{ +public: + ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel(); + ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel); + /** Initialize the kernel's input and dst. + * + * @param[in] compile_context The compile context to be used. + * @param[in] src0 Input tensor for the LHS matrix. Data type supported: F16/F32. + * @param[in] src1 Input tensor containing the RHS reshaped matrix. Data type supported: same as @p src0. + * @param[in] src2 Input tensor containing the bias matrix. Data type supported: same as @p src0. + * @param[out] dst dst tensor info. Data type supported: same as @p src0 + * @param[in] alpha Weight of the matrix product + * @param[in] beta Weight of the matrix bias + * @param[in] lhs_info LHS matrix information used to retrieve the number of rows and accumulations to be processed by each thread. Only the following values are supported: + * lhs_info.m0 > 0 + * lhs_info.k0: 1 + * @param[in] rhs_info RHS matrix information used to retrieve the number of columns and accumulations to be processed by each thread. Only the following values are supported: + * rhs_info.n0: 1,2,3,4,8,16 + * rhs_info.k0: same of lhs_info.k0 + * rhs_info.transpose: false + * @param[in] gemm_info GEMM information used to retrieve the original dimensions of the input matrices + */ + void configure(const ClCompileContext &compile_context, ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, float alpha, float beta, + const GEMMLHSMatrixInfo &lhs_info, + const GEMMRHSMatrixInfo &rhs_info, + const GEMMKernelInfo &gemm_info); + /** Static function to check if given info will lead to a valid configuration + * + * Similar to @ref ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel::configure() + * + * @return a status + */ + static Status validate(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); + + // Inherited methods overridden: + void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override; + +private: + 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_ONLY_RHS_MMUL_KERNEL_H */ diff --git a/src/gpu/cl/kernels/gemm/ClGemmHelpers.cpp b/src/gpu/cl/kernels/gemm/ClGemmHelpers.cpp index 1bf27ba277..67da06102d 100644 --- a/src/gpu/cl/kernels/gemm/ClGemmHelpers.cpp +++ b/src/gpu/cl/kernels/gemm/ClGemmHelpers.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -110,6 +110,23 @@ Status validate_image2d_support_on_rhs(const ITensorInfo &tensor_reshaped_info, 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) +{ + ARM_COMPUTE_UNUSED(n, k, b, data_type); + + const unsigned int mmul_k0 = 4; + best_m0 = 4; + best_n0 = 4; + + const unsigned int ceil_to_multiple_m_m0 = ceil_to_multiple(m, best_m0); + const unsigned int m_div_m0 = ceil_to_multiple_m_m0 / best_m0; + const unsigned int ceil_to_multiple_m_div_m0_mmul_k0 = ceil_to_multiple(m_div_m0, mmul_k0); + const unsigned int gws_y = ceil_to_multiple_m_div_m0_mmul_k0 / mmul_k0; + + return ((k % mmul_k0) == 0) && (gws_y > 4); +} } // namespace gemm } // namespace kernels } // namespace opencl diff --git a/src/gpu/cl/kernels/gemm/ClGemmHelpers.h b/src/gpu/cl/kernels/gemm/ClGemmHelpers.h index 3fce8c9173..bf1e8fce82 100644 --- a/src/gpu/cl/kernels/gemm/ClGemmHelpers.h +++ b/src/gpu/cl/kernels/gemm/ClGemmHelpers.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -88,6 +88,21 @@ void update_padding_for_cl_image(ITensorInfo *tensor); * @return Status reporting if we can use the image2d OpenCL object on the RHS reshaped matrix */ Status validate_image2d_support_on_rhs(const ITensorInfo &tensor_reshaped_info, const GEMMRHSMatrixInfo &rhs_info); + +/** Determine if the MMUL kernels should be preferred + * + * @param[in] m Number of rows of the LHS matrix + * @param[in] n Number of columns of the RHS matrix + * @param[in] k Number of columns of the LHS matrix, rows of the RHS matrix + * @param[in] b Batch size + * @param[in] data_type Data type FP32/FP16 + * @param[in, out] best_m0 Suggested M0 (number of rows of the output block) for the kernel + * @param[in, out] best_n0 Suggested N0 (number of columns of the output block) for the kernel + * + * @return true if MMUL kernel is preferred over kernels w/o MMUL, false otherwise + */ +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); } // namespace gemm } // namespace kernels } // namespace opencl diff --git a/src/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmDefaultConfigReshapedRhsOnlyValhall.cpp b/src/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmDefaultConfigReshapedRhsOnlyValhall.cpp index a82084a8df..97762980be 100644 --- a/src/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmDefaultConfigReshapedRhsOnlyValhall.cpp +++ b/src/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmDefaultConfigReshapedRhsOnlyValhall.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2020-2021 Arm Limited. + * Copyright (c) 2020-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -29,7 +29,9 @@ #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/TensorShape.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" + #include "src/gpu/cl/kernels/gemm/ClGemmHelpers.h" +#include "src/runtime/CL/gemm/CLGEMMDefaultTypeValhall.h" #include @@ -61,6 +63,10 @@ std::pair ClGemmDefaultConfigReshapedRhsOn &ClGemmDefaultConfigReshapedRhsOnlyValhall::configure_G78_f16, &ClGemmDefaultConfigReshapedRhsOnlyValhall::configure_G77_u8); + CLGEMMConfigArray configs_G715(&ClGemmDefaultConfigReshapedRhsOnlyValhall::configure_G715_f32, + &ClGemmDefaultConfigReshapedRhsOnlyValhall::configure_G715_f16, + &ClGemmDefaultConfigReshapedRhsOnlyValhall::configure_G77_u8); + ConfigurationFunctionExecutorPtr func = nullptr; switch(_target) @@ -68,6 +74,10 @@ std::pair ClGemmDefaultConfigReshapedRhsOn case GPUTarget::G78: func = configs_G78.get_function(data_type); break; + case GPUTarget::G715: + case GPUTarget::G615: + func = configs_G715.get_function(data_type); + break; case GPUTarget::G77: default: func = configs_G77.get_function(data_type); @@ -564,6 +574,36 @@ std::pair ClGemmDefaultConfigReshapedRhsOn } } } + +std::pair ClGemmDefaultConfigReshapedRhsOnlyValhall::configure_G715_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b) +{ + unsigned int best_m0; + unsigned int best_n0; + + if(is_mmul_kernel_preferred(m, n, k, b, DataType::F32, best_m0, best_n0)) + { + return configure_lhs_rhs_info(m, n, best_m0, best_n0, 1, 1, 4, false, true, false, false, true); + } + else + { + return configure_G77_f32(m, n, k, b); + } +} + +std::pair ClGemmDefaultConfigReshapedRhsOnlyValhall::configure_G715_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b) +{ + unsigned int best_m0; + unsigned int best_n0; + + if(is_mmul_kernel_preferred(m, n, k, b, DataType::F16, best_m0, best_n0)) + { + return configure_lhs_rhs_info(m, n, best_m0, best_n0, 1, 1, 4, false, true, false, false, true); + } + else + { + return configure_G78_f16(m, n, k, b); + } +} } // namespace gemm } // namespace kernels } // namespace opencl diff --git a/src/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmDefaultConfigReshapedRhsOnlyValhall.h b/src/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmDefaultConfigReshapedRhsOnlyValhall.h index c5e80a7ddc..0ec068fffd 100644 --- a/src/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmDefaultConfigReshapedRhsOnlyValhall.h +++ b/src/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmDefaultConfigReshapedRhsOnlyValhall.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2020-2021 Arm Limited. + * Copyright (c) 2020-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -53,6 +53,8 @@ private: std::pair configure_G78_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b); std::pair configure_G78_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b); std::pair configure_G77_u8(unsigned int m, unsigned int n, unsigned int k, unsigned int b); + std::pair configure_G715_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b); + std::pair configure_G715_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b); }; } // namespace gemm } // namespace kernels diff --git a/src/gpu/cl/operators/ClGemm.cpp b/src/gpu/cl/operators/ClGemm.cpp index 88f6b79b56..4db39a635d 100644 --- a/src/gpu/cl/operators/ClGemm.cpp +++ b/src/gpu/cl/operators/ClGemm.cpp @@ -191,6 +191,7 @@ ClGemm::ClGemm() _mm_native_kernel(std::make_unique()), _mm_reshaped_kernel(std::make_unique()), _mm_reshaped_only_rhs_kernel(std::make_unique()), + _mm_reshaped_only_rhs_mmul_kernel(std::make_unique()), _tmp_a(), _tmp_b(), _reshape_b_only_on_first_run(false), @@ -324,6 +325,53 @@ void ClGemm::configure_reshaped_only_rhs(const CLCompileContext &compile_context _aux_mem[RhsReshape] = MemoryInfo(offset_int_vec(RhsReshape), _reshape_b_only_on_first_run ? MemoryLifetime::Persistent : MemoryLifetime::Temporary, _tmp_b.total_size()); } +void ClGemm::configure_reshaped_only_rhs_mmul(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, + const GEMMInfo &gemm_info) +{ + DataType data_type = a->data_type(); + bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); + const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); + const unsigned int n = b->dimension(0); + const unsigned int k = a->dimension(0); + const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); + const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); + const GPUTarget gpu_target = CLScheduler::get().target(); + bool broadcast_bias = gemm_info.broadcast_bias(); + + GEMMKernelInfo kernel_info; + kernel_info.m = m; + kernel_info.n = n; + kernel_info.k = k; + kernel_info.depth_output_gemm3d = depth_output_gemm3d; + kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d; + kernel_info.broadcast_bias = broadcast_bias; + kernel_info.activation_info = gemm_info.activation_info(); + kernel_info.post_ops = gemm_info.post_ops(); + + // Set the target for the kernels + _mm_reshaped_only_rhs_mmul_kernel->set_target(gpu_target); + + GEMMLHSMatrixInfo lhs_info{}; + GEMMRHSMatrixInfo rhs_info{}; + + // Pick up the GEMM configuration + auto gemm_config = select_default_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size }); + lhs_info = gemm_config.lhs_info; + rhs_info = gemm_config.rhs_info; + // Force H0 to 4 in order to use the MMUL extension + rhs_info.h0 = 4; + + // Reshape Rhs matrix + _reshape_rhs_kernel->configure(compile_context, b, &_tmp_b, rhs_info); + + // Configure matrix multiply kernel with no y padding support + kernel_info.has_pad_y = false; + _mm_reshaped_only_rhs_mmul_kernel->configure(compile_context, a, &_tmp_b, c, output, alpha, beta, lhs_info, rhs_info, kernel_info); + + // Request memory for RHS reshape matrix + _aux_mem[RhsReshape] = MemoryInfo(offset_int_vec(RhsReshape), _reshape_b_only_on_first_run ? MemoryLifetime::Persistent : MemoryLifetime::Temporary, _tmp_b.total_size()); +} + Status ClGemm::validate_native(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info) { ARM_COMPUTE_UNUSED(alpha); @@ -458,6 +506,54 @@ Status ClGemm::validate_reshaped_only_rhs(const ITensorInfo *a, const ITensorInf return Status{}; } +Status ClGemm::validate_reshaped_only_rhs_mmul(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info) +{ + ARM_COMPUTE_UNUSED(alpha); + ARM_COMPUTE_UNUSED(output); + TensorInfo tmp_b_info{}; + + // Get the GPU target + const GPUTarget gpu_target = CLScheduler::get().target(); + const DataType data_type = a->data_type(); + bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); + const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); + const unsigned int n = b->dimension(0); + const unsigned int k = a->dimension(0); + const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); + const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); + const bool broadcast_bias = gemm_info.broadcast_bias(); + + GEMMKernelInfo kernel_info; + kernel_info.m = m; + kernel_info.n = n; + kernel_info.k = k; + kernel_info.depth_output_gemm3d = depth_output_gemm3d; + kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d; + kernel_info.broadcast_bias = broadcast_bias; + kernel_info.activation_info = gemm_info.activation_info(); + kernel_info.post_ops = gemm_info.post_ops(); + + GEMMLHSMatrixInfo lhs_info; + GEMMRHSMatrixInfo rhs_info; + + // Pick up the GEMM configuration + // NOTE: No need to validate mlgo configurations as they automatically fall back to default heuristics if validation fails + const auto gemm_config = select_default_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size }); + lhs_info = gemm_config.lhs_info; + rhs_info = gemm_config.rhs_info; + // Force H0 to 4 in order to use the MMUL extension + rhs_info.h0 = 4; + + auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info))); + ARM_COMPUTE_RETURN_ON_ERROR(ClGemmReshapeRhsMatrixKernel::validate(b, &tmp_b_info, rhs_info)); + + // Validate matrix multiply + kernel_info.has_pad_y = false; + ARM_COMPUTE_RETURN_ON_ERROR(ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel::validate(a, &tmp_b_info, c, output, alpha, beta, lhs_info, rhs_info, kernel_info)); + + return Status{}; +} + void ClGemm::configure(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info) { ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output); @@ -501,6 +597,11 @@ void ClGemm::configure(const CLCompileContext &compile_context, ITensorInfo *a, configure_reshaped_only_rhs(compile_context, a, b, c_to_use, output, alpha, beta, gemm_info); break; } + case CLGEMMKernelType::RESHAPED_ONLY_RHS_MMUL: + { + configure_reshaped_only_rhs_mmul(compile_context, a, b, c_to_use, output, alpha, beta, gemm_info); + break; + } default: { ARM_COMPUTE_ERROR("GEMMType not supported"); @@ -545,6 +646,11 @@ Status ClGemm::validate(const ITensorInfo *a, const ITensorInfo *b, const ITenso ARM_COMPUTE_RETURN_ON_ERROR(validate_reshaped_only_rhs(a, b, c_to_use, output, alpha, beta, gemm_info)); break; } + case CLGEMMKernelType::RESHAPED_ONLY_RHS_MMUL: + { + ARM_COMPUTE_RETURN_ON_ERROR(validate_reshaped_only_rhs_mmul(a, b, c_to_use, output, alpha, beta, gemm_info)); + break; + } default: { ARM_COMPUTE_RETURN_ERROR_MSG("GEMMType not supported"); @@ -627,6 +733,34 @@ void ClGemm::run(ITensorPack &tensors) } break; } + case CLGEMMKernelType::RESHAPED_ONLY_RHS_MMUL: + { + if(!_reshape_b_only_on_first_run) + { + // Run transpose kernel + ITensorPack reshape_rhs_pack{ { ACL_SRC, rhs }, { ACL_DST, rhs_reshaped.get() } }; + CLScheduler::get().enqueue_op(*_reshape_rhs_kernel, reshape_rhs_pack, false); + } + // In case of RESHAPED_ONLY_RHS, we need to check the padding requirement + // Check if the lhs or dst tensors have padding + const unsigned int cross_plane_pad_lhs = lhs->info()->padding().top + lhs->info()->padding().bottom; + const unsigned int cross_plane_pad_dst = dst->info()->padding().top + dst->info()->padding().bottom; + bool has_pad_y = (cross_plane_pad_lhs != 0) || (cross_plane_pad_dst != 0); + + // Copy original tensor pack and overwrite rhs with reshaped counterpart + ITensorPack gemm_reshaped_onlyrhs_pack(tensors); + gemm_reshaped_onlyrhs_pack.add_const_tensor(ACL_SRC_1, rhs_reshaped.get()); + + if(has_pad_y) + { + ARM_COMPUTE_ERROR_ON(has_pad_y); + } + else + { + CLScheduler::get().enqueue_op(*_mm_reshaped_only_rhs_mmul_kernel, gemm_reshaped_onlyrhs_pack, true); + } + break; + } default: { ARM_COMPUTE_ERROR("GEMMType not supported"); diff --git a/src/gpu/cl/operators/ClGemm.h b/src/gpu/cl/operators/ClGemm.h index 3c0cad3ca4..aac463f0b8 100644 --- a/src/gpu/cl/operators/ClGemm.h +++ b/src/gpu/cl/operators/ClGemm.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2021 Arm Limited. + * Copyright (c) 2016-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -34,6 +34,7 @@ #include "src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h" #include "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.h" #include "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h" +#include "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel.h" #include "src/gpu/cl/kernels/ClGemmReshapeLhsMatrixKernel.h" #include "src/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h" @@ -50,6 +51,7 @@ namespace opencl * -# @ref kernels::ClGemmMatrixMultiplyNativeKernel (only if NATIVE is selected by the select_gemm_kernel method()) * -# @ref kernels::ClGemmMatrixMultiplyReshapedKernel (only if RESHAPED is selected by the select_gemm_kernel method()) * -# @ref kernels::ClGemmMatrixMultiplyReshapedOnlyRhsKernel (only if RESHAPED_ONLY_RHS is selected by the select_gemm_kernel method()) + * -# @ref kernels::ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel (only if RESHAPED_ONLY_RHS_MMUL is selected by the select_gemm_kernel method()) */ class ClGemm : public IClOperator { @@ -102,10 +104,12 @@ private: void configure_native(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info); void configure_reshaped(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info); void configure_reshaped_only_rhs(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info); + void configure_reshaped_only_rhs_mmul(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info); static Status validate_native(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info); static Status validate_reshaped(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info); static Status validate_reshaped_only_rhs(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info); + static Status validate_reshaped_only_rhs_mmul(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info); private: enum AuxTensorIdx @@ -116,17 +120,18 @@ private: }; private: - std::unique_ptr _reshape_lhs_kernel; - std::unique_ptr _reshape_rhs_kernel; - std::unique_ptr _mm_native_kernel; - std::unique_ptr _mm_reshaped_kernel; - std::unique_ptr _mm_reshaped_only_rhs_kernel; - TensorInfo _tmp_a; - TensorInfo _tmp_b; - bool _reshape_b_only_on_first_run; - CLGEMMKernelType _gemm_kernel_type; - bool _is_prepared; - experimental::MemoryRequirements _aux_mem{}; + std::unique_ptr _reshape_lhs_kernel; + std::unique_ptr _reshape_rhs_kernel; + std::unique_ptr _mm_native_kernel; + std::unique_ptr _mm_reshaped_kernel; + std::unique_ptr _mm_reshaped_only_rhs_kernel; + std::unique_ptr _mm_reshaped_only_rhs_mmul_kernel; + TensorInfo _tmp_a; + TensorInfo _tmp_b; + bool _reshape_b_only_on_first_run; + CLGEMMKernelType _gemm_kernel_type; + bool _is_prepared; + experimental::MemoryRequirements _aux_mem{}; }; } // namespace opencl } // namespace arm_compute -- cgit v1.2.1