COMPMID-2110: Enable CLGEMMLowpMatrixMultiplyReshapeOnlyRHSKernel in CLGEMMLowp
Change-Id: Ic32c803c3e2a067de10a7e46c85c962a970957b6
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/969
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
diff --git a/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h b/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
index 67b2282..a07101c 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
@@ -25,11 +25,10 @@
#define __ARM_COMPUTE_CLGEMMLOWPMATRIXMULTIPLYCORE_H__
#include "arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.h"
-#include "arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.h"
#include "arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.h"
#include "arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.h"
#include "arm_compute/core/CL/kernels/CLGEMMLowpReductionKernel.h"
-#include "arm_compute/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.h"
#include "arm_compute/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
#include "arm_compute/runtime/CL/CLMemoryGroup.h"
#include "arm_compute/runtime/CL/CLTensor.h"
@@ -42,10 +41,9 @@
/** Basic function to execute GEMMLowpMatrixMultiplyCore on OpenCL. This function calls the following OpenCL kernels:
*
- * -# @ref CLGEMMReshapeLHSMatrixKernel (if the output tensor is a matrix)
* -# @ref CLGEMMReshapeRHSMatrixKernel (if the output tensor is a matrix)
- * -# @ref CLGEMMLowpMatrixMultiplyKernel (if the input matrix is a vector or for Midgard architectures)
- * -# @ref CLGEMMLowpMatrixMultiplyReshapedKernel (if the input matrix is not a vector and if the GPU architecture is not Midgard)
+ * -# @ref CLGEMMLowpMatrixMultiplyKernel (if the parameter "reshape_b_only_on_first_run" of GEMMInfo is FALSE)
+ * -# @ref CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel (if the parameter "reshape_b_only_on_first_run" of GEMMInfo is TRUE)
* -# @ref CLGEMMLowpMatrixAReductionKernel (if the offset of matrix B is not 0)
* -# @ref CLGEMMLowpMatrixBReductionKernel (if the offset of matrix A is not 0)
* -# @ref CLGEMMLowpOffsetContributionKernel (if gemm_info.gemmlowp_output_stage == NONE)
@@ -103,8 +101,7 @@
private:
CLMemoryGroup _memory_group;
CLGEMMLowpMatrixMultiplyKernel _mm_kernel;
- CLGEMMLowpMatrixMultiplyReshapedKernel _mm_reshaped_kernel;
- CLGEMMReshapeLHSMatrixKernel _mtx_a_reshape_kernel;
+ CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel _mm_reshaped_only_rhs_kernel;
CLGEMMReshapeRHSMatrixKernel _mtx_b_reshape_kernel;
CLGEMMLowpMatrixAReductionKernel _mtx_a_reduction_kernel;
CLGEMMLowpMatrixBReductionKernel _mtx_b_reduction_kernel;
@@ -112,7 +109,6 @@
CLGEMMLowpOffsetContributionOutputStageKernel _offset_contribution_output_stage_kernel;
CLTensor _vector_sum_col;
CLTensor _vector_sum_row;
- CLTensor _tmp_a;
CLTensor _tmp_b;
CLTensor _mm_result_s32;
const ICLTensor *_original_b;
diff --git a/src/core/CL/cl_kernels/gemmlowp.cl b/src/core/CL/cl_kernels/gemmlowp.cl
index cf377e1..033b4b4 100644
--- a/src/core/CL/cl_kernels/gemmlowp.cl
+++ b/src/core/CL/cl_kernels/gemmlowp.cl
@@ -4006,4 +4006,4 @@
// Store the result
vstore4(res, 0, dst_addr);
}
-#endif // defined(REAL_MULTIPLIER) && defined(OUTPUT_OFFSET)
+#endif // defined(REAL_MULTIPLIER) && defined(OUTPUT_OFFSET)
\ No newline at end of file
diff --git a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp
index f696f0b..483bab8 100644
--- a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp
+++ b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp
@@ -133,21 +133,13 @@
{
if(m == 1)
{
- if(n > 2048)
- {
- const unsigned int h0 = std::max(n / 4, static_cast<unsigned int>(1));
- return configure_lhs_rhs_info(m, n, 1, 4, 16, 1, h0, false, true, false, true);
- }
- else
- {
- const unsigned int h0 = std::max(n / 2, static_cast<unsigned int>(1));
- return configure_lhs_rhs_info(m, n, 1, 2, 16, 1, h0, false, true, false, true);
- }
+ const unsigned int h0 = std::max(n / 2, static_cast<unsigned int>(1));
+ return configure_lhs_rhs_info(m, n, 1, 2, 4, 1, h0, false, true, false, true);
}
else
{
const unsigned int h0 = std::max(n / 4, static_cast<unsigned int>(1));
- return configure_lhs_rhs_info(m, n, 4, 1, 16, 1, h0, false, true, false, true);
+ return configure_lhs_rhs_info(m, n, 2, 2, 16, 1, h0, false, true, false, true);
}
}
}
diff --git a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp
index b1b0a16..eca2416 100644
--- a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp
@@ -152,7 +152,7 @@
window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop
update_window_and_padding(win_out, output_access); // window used to update the padding requirements of output tensor
- output_access.set_valid_region(win_out, ValidRegion(Coordinates(0, 0), output->tensor_shape()));
+ output_access.set_valid_region(win_out, ValidRegion(Coordinates(), output->tensor_shape()));
// Collapse along the Z direction
// This collapse needs to be here in order to tune the Z dimension of LWS
diff --git a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
index c447cb8..cd53708 100644
--- a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
@@ -24,7 +24,7 @@
#include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h"
#include "arm_compute/core/CL/ICLTensor.h"
-#include "arm_compute/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfiguration.h"
+#include "arm_compute/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfiguration.h"
#include "arm_compute/core/Error.h"
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/TensorInfo.h"
@@ -40,17 +40,16 @@
namespace
{
-inline bool is_gemm_reshaped(unsigned int m, bool reshape_b_only_on_first_run, GPUTarget gpu_target)
+inline bool is_gemm_reshaped(bool reshape_b_only_on_first_run, GPUTarget gpu_target)
{
- return (get_arch_from_target(gpu_target) != GPUTarget::MIDGARD) && (m > 1) && (reshape_b_only_on_first_run);
+ return (get_arch_from_target(gpu_target) != GPUTarget::MIDGARD) && (reshape_b_only_on_first_run);
}
} // namespace
CLGEMMLowpMatrixMultiplyCore::CLGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemoryManager> memory_manager)
: _memory_group(std::move(memory_manager)),
_mm_kernel(),
- _mm_reshaped_kernel(),
- _mtx_a_reshape_kernel(),
+ _mm_reshaped_only_rhs_kernel(),
_mtx_b_reshape_kernel(),
_mtx_a_reduction_kernel(),
_mtx_b_reduction_kernel(),
@@ -58,7 +57,6 @@
_offset_contribution_output_stage_kernel(),
_vector_sum_col(),
_vector_sum_row(),
- _tmp_a(),
_tmp_b(),
_mm_result_s32(),
_original_b(nullptr),
@@ -86,7 +84,6 @@
const GPUTarget gpu_target = CLScheduler::get().target();
// Set the target for the kernels
- _mtx_a_reshape_kernel.set_target(gpu_target);
_mm_kernel.set_target(gpu_target);
const ICLTensor *matrix_a = a;
@@ -105,27 +102,19 @@
const int depth_output_gemm3d = gemm_info.depth_output_gemm3d();
// Check if we need to reshape the matrix A and matrix B
- _is_gemm_reshaped = is_gemm_reshaped(m, _reshape_b_only_on_first_run, gpu_target);
+ _is_gemm_reshaped = is_gemm_reshaped(_reshape_b_only_on_first_run, gpu_target);
if(_is_gemm_reshaped)
{
- // if _is_interleaved_transposed is set, force reinterpret_input_as_3d to be false as the output of CLGEMMInterleaveKernel will be 2D
- reinterpret_input_as_3d = false;
-
- matrix_a = &_tmp_a;
matrix_b = &_tmp_b;
- _memory_group.manage(&_tmp_a);
if(!_reshape_b_only_on_first_run)
{
_memory_group.manage(&_tmp_b);
}
// Pick up the GEMM configuration
- std::tie(lhs_info, rhs_info) = CLGEMMReshapedKernelConfigurationFactory::create(gpu_target)->configure(m, n, k, batch_size, DataType::QASYMM8);
-
- // Configure reshape LHS kernel
- _mtx_a_reshape_kernel.configure(a, &_tmp_a, lhs_info, gemm_info.reinterpret_input_as_3d());
+ std::tie(lhs_info, rhs_info) = CLGEMMReshapedOnlyRHSKernelConfigurationFactory::create(gpu_target)->configure(m, n, k, batch_size, DataType::QASYMM8);
// Configure reshape RHS kernel
_mtx_b_reshape_kernel.configure(b, &_tmp_b, rhs_info);
@@ -166,7 +155,7 @@
if(_is_gemm_reshaped)
{
// Configure and tune matrix multiply kernel
- _mm_reshaped_kernel.configure(matrix_a, matrix_b, &_mm_result_s32, lhs_info, rhs_info, GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d));
+ _mm_reshaped_only_rhs_kernel.configure(matrix_a, matrix_b, &_mm_result_s32, lhs_info, rhs_info, GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d));
}
else
{
@@ -185,7 +174,7 @@
if(_is_gemm_reshaped)
{
// Configure and tune matrix multiply kernel
- _mm_reshaped_kernel.configure(matrix_a, matrix_b, output, lhs_info, rhs_info, GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d));
+ _mm_reshaped_only_rhs_kernel.configure(matrix_a, matrix_b, output, lhs_info, rhs_info, GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d));
}
else
{
@@ -200,7 +189,6 @@
// Allocate tensors
if(_is_gemm_reshaped)
{
- _tmp_a.allocator()->allocate();
if(!_reshape_b_only_on_first_run)
{
_tmp_b.allocator()->allocate();
@@ -231,7 +219,6 @@
const ITensorInfo *matrix_a_info = a;
const ITensorInfo *matrix_b_info = b;
- TensorInfo tmp_a_info{};
TensorInfo tmp_b_info{};
GEMMRHSMatrixInfo rhs_info;
GEMMLHSMatrixInfo lhs_info;
@@ -246,27 +233,16 @@
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();
- bool reshape_matrices = is_gemm_reshaped(m, gemm_info.reshape_b_only_on_first_run(), CLScheduler::get().target());
-
- // if reshape_matrices is set, force reinterpret_input_as_3d to be false as the output of CLGEMMInterleaveKernel will be 2D
- if(reshape_matrices)
- {
- reinterpret_input_as_3d = false;
- }
+ bool reshape_matrix_b = is_gemm_reshaped(gemm_info.reshape_b_only_on_first_run(), CLScheduler::get().target());
const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d);
- if(reshape_matrices)
+ if(reshape_matrix_b)
{
- matrix_a_info = &tmp_a_info;
matrix_b_info = &tmp_b_info;
// Pick up the GEMM configuration
- std::tie(lhs_info, rhs_info) = CLGEMMReshapedKernelConfigurationFactory::create(gpu_target)->configure(m, n, k, batch_size, DataType::QASYMM8);
-
- // Validate reshape LHS kernel
- auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_lhs_reshaped_shape(*a, lhs_info, gemm_info.reinterpret_input_as_3d())));
- ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMReshapeLHSMatrixKernel::validate(a, &tmp_a_info, lhs_info, gemm_info.reinterpret_input_as_3d()));
+ std::tie(lhs_info, rhs_info) = CLGEMMReshapedOnlyRHSKernelConfigurationFactory::create(gpu_target)->configure(m, n, k, batch_size, DataType::QASYMM8);
// Validate reshape RHS kernel
auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info)));
@@ -297,13 +273,13 @@
{
TensorInfo mm_result_s32_info{};
- if(reshape_matrices)
+ if(reshape_matrix_b)
{
// Output tensor auto inizialitation if not yet initialized
auto_init_if_empty(mm_result_s32_info, a->clone()->set_tensor_shape(compute_mm_shape(*matrix_a_info, *matrix_b_info, reshape_info)).set_data_type(DataType::S32));
// Validate matrix multiply
- ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyReshapedKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info, lhs_info, rhs_info, reshape_info));
+ ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info, lhs_info, rhs_info, reshape_info));
}
else
{
@@ -324,10 +300,10 @@
}
else
{
- if(reshape_matrices)
+ if(reshape_matrix_b)
{
// Validate matrix multiply
- ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyReshapedKernel::validate(matrix_a_info, matrix_b_info, output, lhs_info, rhs_info, reshape_info));
+ ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::validate(matrix_a_info, matrix_b_info, output, lhs_info, rhs_info, reshape_info));
}
else
{
@@ -356,9 +332,6 @@
if(_is_gemm_reshaped)
{
- // Run reshape matrix A
- CLScheduler::get().enqueue(_mtx_a_reshape_kernel, false);
-
if(!_reshape_b_only_on_first_run)
{
// Run reshape matrix B
@@ -375,7 +348,7 @@
// Run matrix multiply
if(_is_gemm_reshaped)
{
- CLScheduler::get().enqueue(_mm_reshaped_kernel, false);
+ CLScheduler::get().enqueue(_mm_reshaped_only_rhs_kernel, false);
}
else
{
@@ -425,4 +398,4 @@
_is_prepared = true;
}
}
-} // namespace arm_compute
+} // namespace arm_compute
\ No newline at end of file