Apply clang-format on repository
Code is formatted as per a revised clang format configuration
file(not part of this delivery). Version 14.0.6 is used.
Exclusion List:
- files with .cl extension
- files that are not strictly C/C++ (e.g. Android.bp, Sconscript ...)
And the following directories
- compute_kernel_writer/validation/
- tests/
- include/
- src/core/NEON/kernels/convolution/
- src/core/NEON/kernels/arm_gemm/
- src/core/NEON/kernels/arm_conv/
- data/
There will be a follow up for formatting of .cl files and the
files under tests/ and compute_kernel_writer/validation/.
Signed-off-by: Felix Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>
Change-Id: Ib7eb1fcf4e7537b9feaefcfc15098a804a3fde0a
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10391
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
diff --git a/src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyNativeKernel.cpp b/src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyNativeKernel.cpp
index accafee..e0d925d 100644
--- a/src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyNativeKernel.cpp
+++ b/src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyNativeKernel.cpp
@@ -29,14 +29,13 @@
#include "arm_compute/core/CL/OpenCL.h"
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include "arm_compute/core/utils/StringUtils.h"
#include "arm_compute/core/Validate.h"
-#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include "src/core/AccessWindowStatic.h"
#include "src/core/helpers/AutoConfiguration.h"
#include "src/core/helpers/WindowHelpers.h"
-
#include "support/Cast.h"
#include "support/StringSupport.h"
@@ -50,26 +49,35 @@
{
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);
- if(src0->data_type() == DataType::QASYMM8)
+ if (src0->data_type() == DataType::QASYMM8)
{
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, src1);
}
else
{
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src1, 1, DataType::QASYMM8, DataType::QSYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8_PER_CHANNEL);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src1, 1, DataType::QASYMM8, DataType::QSYMM8,
+ DataType::QASYMM8_SIGNED, DataType::QSYMM8_PER_CHANNEL);
}
- 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.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 < 1 || 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();
@@ -83,7 +91,7 @@
ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(0) != static_cast<unsigned int>(k));
ARM_COMPUTE_RETURN_ERROR_ON(src1->dimension(0) != static_cast<unsigned int>(n));
ARM_COMPUTE_RETURN_ERROR_ON(src1->dimension(1) != static_cast<unsigned int>(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) != static_cast<unsigned int>(m));
}
@@ -92,9 +100,10 @@
ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(1) != static_cast<unsigned int>(m));
}
- 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_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::S32);
}
@@ -102,8 +111,13 @@
return Status{};
}
-std::pair<Status, Window> validate_and_configure_window(const ITensorInfo *src0, 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,
+ 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];
@@ -115,17 +129,19 @@
// In case both input and dst have to be reinterpreted as 3D tensors,
// force reinterpret_dst_as_3d to be false.
- if(reinterpret_input_as_3d == reinterpret_dst_as_3d)
+ if (reinterpret_input_as_3d == reinterpret_dst_as_3d)
{
reinterpret_dst_as_3d = false;
}
// 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)).set_data_type(DataType::S32));
+ auto_init_if_empty(*dst, src0->clone()
+ ->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info))
+ .set_data_type(DataType::S32));
TensorInfo tmp_info(*dst);
- if(reinterpret_dst_as_3d)
+ if (reinterpret_dst_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
@@ -138,12 +154,12 @@
num_elems_processed_per_iteration_x = rhs_info.n0;
num_elems_processed_per_iteration_y = lhs_info.m0;
- win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+ win =
+ calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
// RHS matrix still needs padding on the X
- AccessWindowStatic src1_access(src1, 0, 0,
- ceil_to_multiple(src1->dimension(0), num_elems_processed_per_iteration_x),
- src1->dimension(1));
+ AccessWindowStatic src1_access(
+ src1, 0, 0, ceil_to_multiple(src1->dimension(0), num_elems_processed_per_iteration_x), src1->dimension(1));
window_changed = update_window_and_padding(win, src1_access); // window used by the execute_window_loop
@@ -153,7 +169,8 @@
const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(dst->num_dimensions()), 2u);
collapsed = win.collapse(win, dimension_to_collapse);
- Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+ Status err =
+ (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
return std::make_pair(err, collapsed);
}
} // namespace
@@ -163,8 +180,13 @@
_type = CLKernelType::GEMM;
}
-void ClGemmLowpMatrixMultiplyNativeKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *dst,
- const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info)
+void ClGemmLowpMatrixMultiplyNativeKernel::configure(const CLCompileContext &compile_context,
+ const ITensorInfo *src0,
+ ITensorInfo *src1,
+ ITensorInfo *dst,
+ const GEMMLHSMatrixInfo &lhs_info,
+ const GEMMRHSMatrixInfo &rhs_info,
+ const GEMMReshapeInfo &gemm_info)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
@@ -175,11 +197,11 @@
_use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
// We still need padding on the X dimension for the RHS matrix
- auto padding_info = get_padding_info({ src0, dst });
+ auto padding_info = get_padding_info({src0, dst});
// In case both input and dst have to be reinterpreted as 3D tensors,
// force reinterpret_input_as_3d and reinterpret_dst_as_3d to be false.
- if(_reinterpret_input_as_3d == _reinterpret_output_as_3d)
+ if (_reinterpret_input_as_3d == _reinterpret_output_as_3d)
{
_reinterpret_input_as_3d = false;
_reinterpret_output_as_3d = false;
@@ -192,7 +214,8 @@
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);
@@ -212,8 +235,10 @@
CLBuildOptions build_opts;
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(dst->dimension(1)));
- build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(dst->dimension(2)));
+ build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d,
+ "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(dst->dimension(1)));
+ build_opts.add_option_if(_reinterpret_input_as_3d || _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(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
build_opts.add_option("-DM=" + support::cpp11::to_string(src0->dimension(1)));
@@ -258,19 +283,19 @@
ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
}
-Status ClGemmLowpMatrixMultiplyNativeKernel::validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst, const GEMMLHSMatrixInfo &lhs_info,
- const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info)
+Status ClGemmLowpMatrixMultiplyNativeKernel::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{};
}
@@ -280,11 +305,13 @@
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);
@@ -296,7 +323,7 @@
slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));
- if(_reinterpret_input_as_3d)
+ if (_reinterpret_input_as_3d)
{
// Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor
const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 3;
@@ -304,10 +331,10 @@
_kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
}
- if(_reinterpret_output_as_3d)
+ if (_reinterpret_output_as_3d)
{
// Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor
- const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 3 + (_reinterpret_input_as_3d ? 1 : 0);
+ const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 3 + (_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));
}
@@ -317,7 +344,7 @@
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;
}
@@ -330,8 +357,7 @@
_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