diff options
Diffstat (limited to 'src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp')
| -rw-r--r-- | src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp | 166 |
1 files changed, 109 insertions, 57 deletions
diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp index b34c17cda8..1b19f1ec5b 100644 --- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp +++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp @@ -25,8 +25,9 @@ #include "arm_compute/core/CL/ICLTensor.h" #include "arm_compute/core/utils/ActivationFunctionUtils.h" -#include "arm_compute/core/utils/StringUtils.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/core/utils/StringUtils.h" + #include "src/core/CL/CLUtils.h" #include "src/core/CL/CLValidate.h" #include "src/core/helpers/AutoConfiguration.h" @@ -46,24 +47,36 @@ namespace { using ElementsProcessed = Steps; -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) +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_F16_UNSUPPORTED(src0); 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(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 || lhs_info.m0 > 8, "Only 1,2,3,4,5,6,7,8 are supported for m0"); ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.k0 > 16 || rhs_info.k0 < 2); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.k0 & (rhs_info.k0 - 1)) && rhs_info.k0 != 3), "Only 2,3,4,8,16 are supported for k0"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.k0 & (rhs_info.k0 - 1)) && rhs_info.k0 != 3), + "Only 2,3,4,8,16 are supported for k0"); ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.n0 > 16 || rhs_info.n0 < 2); - 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((gemm_info.reinterpret_input_as_3d || gemm_info.depth_output_gemm3d != 0) && (src2 != nullptr) - && (!gemm_info.broadcast_bias), - "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(((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( + (gemm_info.reinterpret_input_as_3d || gemm_info.depth_output_gemm3d != 0) && (src2 != nullptr) && + (!gemm_info.broadcast_bias), + "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)); @@ -71,19 +84,20 @@ Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, cons const unsigned int n = gemm_info.n; const unsigned int k = gemm_info.k; - TensorShape tensor_shape1{ src1->tensor_shape() }; + TensorShape tensor_shape1{src1->tensor_shape()}; tensor_shape1.set(0, n); tensor_shape1.set(1, k); - if(src2 != nullptr && !(helpers::float_ops::is_zero(beta))) + 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, src0); - if(gemm_info.broadcast_bias) + 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"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG((src2_dim1 != 1 || src2_dim0 != n), + "Incorrect dimension of bias matrix which is to be broadcasted"); } else { @@ -93,10 +107,11 @@ Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, cons 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)); + 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(src0->dimension(0) != k); - if(gemm_info.reinterpret_input_as_3d) + if (gemm_info.reinterpret_input_as_3d) { ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(1) * src0->dimension(2) != m); } @@ -106,9 +121,10 @@ Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, cons } 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(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info)); + 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); } @@ -116,8 +132,14 @@ Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, cons return Status{}; } -Window validate_and_configure_window(ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, const GEMMLHSMatrixInfo &lhs_info, - const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info, ElementsProcessed &num_elements_processed) +Window validate_and_configure_window(ITensorInfo *src0, + ITensorInfo *src1, + ITensorInfo *src2, + ITensorInfo *dst, + const GEMMLHSMatrixInfo &lhs_info, + const GEMMRHSMatrixInfo &rhs_info, + const GEMMKernelInfo &gemm_info, + ElementsProcessed &num_elements_processed) { ARM_COMPUTE_UNUSED(src0, src1, src2); unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0]; @@ -128,14 +150,14 @@ Window validate_and_configure_window(ITensorInfo *src0, ITensorInfo *src1, ITens // 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. // This approach should only be used when the input/dst tensors have pad on the y direction - if((reinterpret_input_as_3d == reinterpret_output_as_3d) && gemm_info.has_pad_y) + if ((reinterpret_input_as_3d == reinterpret_output_as_3d) && gemm_info.has_pad_y) { reinterpret_output_as_3d = false; } 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 @@ -148,7 +170,8 @@ Window validate_and_configure_window(ITensorInfo *src0, ITensorInfo *src1, ITens 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 @@ -164,14 +187,22 @@ ClGemmMatrixMultiplyReshapedOnlyRhsKernel::ClGemmMatrixMultiplyReshapedOnlyRhsKe _type = CLKernelType::GEMM; } -void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext &compile_context, - const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, ITensorInfo *dst, float alpha, float beta, - const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info) +void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext &compile_context, + const ITensorInfo *src0, + const ITensorInfo *src1, + const 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))); + 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)); @@ -182,11 +213,11 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext _export_to_cl_image = rhs_info.export_to_cl_image; _has_pad_y = gemm_info.has_pad_y; - auto padding_info = get_padding_info({ src0, src1, src2, dst }); + auto padding_info = get_padding_info({src0, src1, src2, dst}); // 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. - if((_reinterpret_input_as_3d == _reinterpret_output_as_3d) && _has_pad_y) + if ((_reinterpret_input_as_3d == _reinterpret_output_as_3d) && _has_pad_y) { _reinterpret_input_as_3d = false; _reinterpret_output_as_3d = false; @@ -199,8 +230,9 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext ElementsProcessed num_elements_processed{}; // Configure kernel window - Window win = 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, - num_elements_processed); + Window win = 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, num_elements_processed); ICLKernel::configure_internal(win); // If _reinterpret_input_as_3d = reinterpret_output_as_3d = true, @@ -225,7 +257,8 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext // 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(!(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"); @@ -240,17 +273,23 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext 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(_has_pad_y) + if (_has_pad_y) { build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D"); build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D"); - 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)); + 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)); } - 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"; @@ -294,28 +333,39 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); } -Status ClGemmMatrixMultiplyReshapedOnlyRhsKernel::validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta, +Status ClGemmMatrixMultiplyReshapedOnlyRhsKernel::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) + 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)); return Status{}; } -void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) +void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::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<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)); - const auto src2 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2)); - 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)); + const auto src2 = + utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2)); + auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(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) + 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); @@ -341,12 +391,14 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::run_op(ITensorPack &tensors, con cl::Image2D src1_image2d; - if(_export_to_cl_image) + if (_export_to_cl_image) { - const TensorShape shape2d(src1->info()->dimension(0) / 4, src1->info()->dimension(1) * src1->info()->dimension(2)); + 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, CLImage2DType::ReadOnly); + src1_image2d = create_image2d_from_buffer(CLKernelLibrary::get().context(), src1->cl_buffer(), shape2d, + src1->info()->data_type(), image_row_pitch, CLImage2DType::ReadOnly); } do @@ -354,7 +406,7 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::run_op(ITensorPack &tensors, con 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; } @@ -365,7 +417,7 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::run_op(ITensorPack &tensors, con add_2D_tensor_argument(idx, src0, slice); // RHS buffer or RHS OpenCL image (_export_to_cl_image == true) - if(_export_to_cl_image) + if (_export_to_cl_image) { _kernel.setArg(idx++, src1_image2d); } @@ -387,22 +439,23 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::run_op(ITensorPack &tensors, con _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src1->info()->strides_in_bytes()[rhs_idx_batch_size])); // Bias stride_z (if _add_bias == true) - if(_add_bias) + if (_add_bias) { - _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src2->info()->strides_in_bytes()[bia_idx_batch_size])); + _kernel.setArg<cl_uint>(idx++, + static_cast<unsigned int>(src2->info()->strides_in_bytes()[bia_idx_batch_size])); } // dst stride_z _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[out_idx_batch_size])); // Cross-plan padding (if _reinterpret_input_as_3d = true) - if(_reinterpret_input_as_3d && _has_pad_y) + if (_reinterpret_input_as_3d && _has_pad_y) { _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(total_cross_plane_pad_lhs)); } // Cross-plan padding (if reinterpret_output_as_3d = true) - if(_reinterpret_output_as_3d && _has_pad_y) + if (_reinterpret_output_as_3d && _has_pad_y) { _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(total_cross_plane_pad_out)); } @@ -413,8 +466,7 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::run_op(ITensorPack &tensors, con _kernel.setArg<cl_int>(idx++, _k); 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 |