Issue COMPMID-1835: Remove CLGEMMInterleave4x4Kernel and replace with CLGEMMReshapeLHSMatrixKernel
Change-Id: Id6a1bd78f9b1698b64a004e4adebc41002b15745
Reviewed-on: https://review.mlplatform.org/496
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
diff --git a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp
index 66fafe4..2c072a8 100644
--- a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -72,16 +72,24 @@
else
{
GEMMRHSMatrixInfo rhs_info;
+ GEMMLHSMatrixInfo lhs_info;
const int m = reshape_info.m();
const int n = reshape_info.n();
const int k = reshape_info.k();
const int mult_transpose1xW_width = reshape_info.mult_transpose1xW_width();
const int mult_interleave4x4_height = reshape_info.mult_interleave4x4_height();
- rhs_info.n0 = 16 / input1->element_size();
- rhs_info.k0 = 1;
- rhs_info.h0 = mult_transpose1xW_width;
- rhs_info.interleave = false;
- rhs_info.transpose = false;
+ const bool unroll_block = dot8_supported(CLKernelLibrary::get().get_device());
+
+ rhs_info.n0 = 16 / input1->element_size();
+ rhs_info.k0 = 1;
+ rhs_info.h0 = mult_transpose1xW_width;
+ rhs_info.interleave = false;
+ rhs_info.transpose = false;
+ lhs_info.m0 = 4;
+ lhs_info.k0 = 4;
+ lhs_info.v0 = mult_interleave4x4_height;
+ lhs_info.interleave = true;
+ lhs_info.transpose = unroll_block;
TensorShape tensor_shape0{ input0->tensor_shape() };
tensor_shape0.set(0, k);
@@ -94,7 +102,7 @@
const TensorInfo tensor_info0 = input0->clone()->set_tensor_shape(tensor_shape0);
const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1);
- const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(compute_interleaved_shape(tensor_info0, mult_interleave4x4_height));
+ const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(compute_lhs_reshaped_shape(tensor_info0, lhs_info));
const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(compute_rhs_reshaped_shape(tensor_info1, rhs_info));
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input0, &tensor_info_reshaped0);
diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
index 69455cf..89fe7a4 100644
--- a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -40,7 +40,8 @@
#include <set>
#include <string>
-using namespace arm_compute;
+namespace arm_compute
+{
using namespace arm_compute::misc::shape_calculator;
namespace
@@ -67,6 +68,7 @@
else
{
GEMMRHSMatrixInfo rhs_info;
+ GEMMLHSMatrixInfo lhs_info;
const int m = reshape_info.m();
const int n = reshape_info.n();
const int k = reshape_info.k();
@@ -77,6 +79,11 @@
rhs_info.h0 = mult_transpose1xW_width;
rhs_info.interleave = false;
rhs_info.transpose = false;
+ lhs_info.m0 = 4;
+ lhs_info.k0 = 4;
+ lhs_info.v0 = mult_interleave4x4_height;
+ lhs_info.interleave = true;
+ lhs_info.transpose = true;
TensorShape tensor_shape0{ input0->tensor_shape() };
tensor_shape0.set(0, k);
@@ -89,7 +96,7 @@
const TensorInfo tensor_info0 = input0->clone()->set_tensor_shape(tensor_shape0);
const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1);
- const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(compute_interleaved_shape(tensor_info0, mult_interleave4x4_height));
+ const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(compute_lhs_reshaped_shape(tensor_info0, lhs_info));
const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(compute_rhs_reshaped_shape(tensor_info1, rhs_info));
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input0, &tensor_info_reshaped0);
@@ -439,3 +446,4 @@
}
while(window.slide_window_slice_3D(slice));
}
+} // namespace arm_compute
diff --git a/src/runtime/CL/functions/CLGEMM.cpp b/src/runtime/CL/functions/CLGEMM.cpp
index 9048b85..a3612f3 100644
--- a/src/runtime/CL/functions/CLGEMM.cpp
+++ b/src/runtime/CL/functions/CLGEMM.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -35,7 +35,8 @@
#include "arm_compute/runtime/CL/CLScheduler.h"
#include "arm_compute/runtime/ITensorAllocator.h"
-using namespace arm_compute;
+namespace arm_compute
+{
using namespace arm_compute::misc::shape_calculator;
namespace
@@ -117,7 +118,6 @@
CLGEMM::CLGEMM(std::shared_ptr<IMemoryManager> memory_manager)
: _memory_group(std::move(memory_manager)),
- _interleave_kernel(),
_mm_kernel(),
_ma_kernel(),
_reshape_lhs_kernel(),
@@ -153,7 +153,7 @@
const GPUTarget gpu_target = CLScheduler::get().target();
// Set the target for the kernels
- _interleave_kernel.set_target(gpu_target);
+ _reshape_lhs_kernel.set_target(gpu_target);
_mm_kernel.set_target(gpu_target);
// Arguments used by GEMMReshapeInfo
@@ -180,6 +180,13 @@
rhs_info.interleave = false;
rhs_info.transpose = false;
+ GEMMLHSMatrixInfo lhs_info;
+ lhs_info.m0 = 4;
+ lhs_info.k0 = 4;
+ lhs_info.v0 = mult_interleave4x4_height;
+ lhs_info.interleave = true;
+ lhs_info.transpose = true;
+
// Check if we need to reshape the matrix A and matrix B
_is_interleaved_transposed = is_interleaved_transposed(m, n, k, a->info()->data_type(), _reshape_b_only_on_first_run, gpu_target);
@@ -219,8 +226,7 @@
else
{
// Configure interleave kernel
- _interleave_kernel.configure(a, &_tmp_a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d());
-
+ _reshape_lhs_kernel.configure(a, &_tmp_a, lhs_info, gemm_info.reinterpret_input_as_3d());
// Configure transpose kernel
_reshape_rhs_kernel.configure(b, &_tmp_b, rhs_info);
}
@@ -296,6 +302,13 @@
rhs_info.interleave = false;
rhs_info.transpose = false;
+ GEMMLHSMatrixInfo lhs_info;
+ lhs_info.m0 = 4;
+ lhs_info.k0 = 4;
+ lhs_info.v0 = mult_interleave4x4_height;
+ lhs_info.interleave = true;
+ lhs_info.transpose = true;
+
// Check if we need to reshape the matrix A and matrix B
const bool run_interleave_transpose = is_interleaved_transposed(m, n, k, a->data_type(), reshape_b_only_on_first_run, gpu_target);
@@ -335,8 +348,8 @@
else
{
// Validate interleave kernel
- auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_interleaved_shape(*a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d())));
- ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMInterleave4x4Kernel::validate(a, &tmp_a_info, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d()));
+ 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()));
// Validate transpose kernel
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));
@@ -367,14 +380,7 @@
if(_is_interleaved_transposed)
{
// Run interleave kernel
- if(_is_G76_path)
- {
- CLScheduler::get().enqueue(_reshape_lhs_kernel, false);
- }
- else
- {
- CLScheduler::get().enqueue(_interleave_kernel, false);
- }
+ CLScheduler::get().enqueue(_reshape_lhs_kernel, false);
if(!_reshape_b_only_on_first_run)
{
@@ -417,3 +423,4 @@
_is_prepared = true;
}
}
+} // namespace arm_compute
diff --git a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
index cf20bc6..edb3107 100644
--- a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -32,7 +32,8 @@
#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include "arm_compute/runtime/CL/CLScheduler.h"
-using namespace arm_compute;
+namespace arm_compute
+{
using namespace arm_compute::misc::shape_calculator;
namespace
@@ -109,6 +110,7 @@
const ICLTensor *matrix_a = a;
const ICLTensor *matrix_b = b;
GEMMRHSMatrixInfo rhs_info;
+ GEMMLHSMatrixInfo lhs_info;
// Arguments used by GEMMReshapeInfo
// If we pass the matrix A and matrix B reshaped to CLGEMMMatrixMultiplyKernel, we need to pass m, n, k, mult_transpose1xW_width and mult_interleave4x4_height to CLGEMMReshapeInfo
@@ -126,6 +128,11 @@
rhs_info.h0 = mult_transpose1xW_width;
rhs_info.interleave = false;
rhs_info.transpose = false;
+ lhs_info.m0 = 4;
+ lhs_info.k0 = 4;
+ lhs_info.v0 = mult_interleave4x4_height;
+ lhs_info.interleave = true;
+ lhs_info.transpose = unroll_block;
// Check if we need to reshape the matrix A and matrix B
_is_interleaved_transposed = is_interleaved_transposed(m, n, k, _reshape_b_only_on_first_run, gpu_target);
@@ -145,7 +152,7 @@
}
// Configure interleave kernel
- _mtx_a_reshape_kernel.configure(a, &_tmp_a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d(), unroll_block);
+ _mtx_a_reshape_kernel.configure(a, &_tmp_a, lhs_info, gemm_info.reinterpret_input_as_3d());
// Configure transpose kernel
_mtx_b_reshape_kernel.configure(b, &_tmp_b, rhs_info);
@@ -242,8 +249,10 @@
TensorInfo tmp_a_info{};
TensorInfo tmp_b_info{};
GEMMRHSMatrixInfo rhs_info;
+ GEMMLHSMatrixInfo lhs_info;
bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
+ const bool unroll_block = dot8_supported(CLKernelLibrary::get().get_device());
const int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1);
const int n = b->dimension(0);
const int k = a->dimension(0);
@@ -255,6 +264,11 @@
rhs_info.h0 = mult_transpose1xW_width;
rhs_info.interleave = false;
rhs_info.transpose = false;
+ lhs_info.m0 = 4;
+ lhs_info.k0 = 4;
+ lhs_info.v0 = mult_interleave4x4_height;
+ lhs_info.interleave = true;
+ lhs_info.transpose = unroll_block;
bool reshape_matrices = is_interleaved_transposed(m, n, k, gemm_info.reshape_b_only_on_first_run(), CLScheduler::get().target());
@@ -272,8 +286,8 @@
matrix_b_info = &tmp_b_info;
// Validate interleave kernel
- auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_interleaved_shape(*a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d())));
- ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMInterleave4x4Kernel::validate(a, &tmp_a_info, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d()));
+ 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()));
// Validate transpose kernel
@@ -408,3 +422,4 @@
_is_prepared = true;
}
}
+} // namespace arm_compute