COMPMID-1836: Remove CLGEMMTranspose1xWKernel and replace with CLGEMMReshapeRHSMatrixKernel
Change-Id: Ic5a4f32657a155380684dcd4b44fbb608ef40cb4
Reviewed-on: https://review.mlplatform.org/418
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/CL/CLHelpers.cpp b/src/core/CL/CLHelpers.cpp
index 924fb1d..18ef185 100644
--- a/src/core/CL/CLHelpers.cpp
+++ b/src/core/CL/CLHelpers.cpp
@@ -148,7 +148,7 @@
const GPUTarget gpu_target = get_target_from_name(device_name);
// SW_WORKAROUND: Workaround for DDK revision r14p0.to enable cl_arm_integer_dot_product_int8
- std::set<GPUTarget> sw_workaround_issue = {GPUTarget::G76};
+ std::set<GPUTarget> sw_workaround_issue = { GPUTarget::G76 };
return (device_supports_extension(device, "cl_arm_integer_dot_product_int8") || sw_workaround_issue.count(gpu_target) != 0);
}
@@ -255,5 +255,4 @@
return 1;
}
}
-
} // namespace arm_compute
diff --git a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp
index b2fb3e0..66fafe4 100644
--- a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp
@@ -71,11 +71,17 @@
}
else
{
- 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();
+ GEMMRHSMatrixInfo rhs_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;
TensorShape tensor_shape0{ input0->tensor_shape() };
tensor_shape0.set(0, k);
@@ -89,7 +95,7 @@
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_reshaped1 = input1->clone()->set_tensor_shape(compute_transpose1xW_with_element_size_shape(tensor_info1, mult_transpose1xW_width));
+ 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);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, &tensor_info_reshaped1);
diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
index c9ed776..69455cf 100644
--- a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
@@ -66,11 +66,17 @@
}
else
{
- 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();
+ GEMMRHSMatrixInfo rhs_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;
TensorShape tensor_shape0{ input0->tensor_shape() };
tensor_shape0.set(0, k);
@@ -84,7 +90,7 @@
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_reshaped1 = input1->clone()->set_tensor_shape(compute_transpose1xW_with_element_size_shape(tensor_info1, mult_transpose1xW_width));
+ 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);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, &tensor_info_reshaped1);
diff --git a/src/runtime/CL/functions/CLGEMM.cpp b/src/runtime/CL/functions/CLGEMM.cpp
index d0db876..9048b85 100644
--- a/src/runtime/CL/functions/CLGEMM.cpp
+++ b/src/runtime/CL/functions/CLGEMM.cpp
@@ -118,7 +118,6 @@
CLGEMM::CLGEMM(std::shared_ptr<IMemoryManager> memory_manager)
: _memory_group(std::move(memory_manager)),
_interleave_kernel(),
- _transpose_kernel(),
_mm_kernel(),
_ma_kernel(),
_reshape_lhs_kernel(),
@@ -174,13 +173,18 @@
mult_transpose1xW_width = 4;
mult_interleave4x4_height = 2;
}
+ GEMMRHSMatrixInfo rhs_info;
+ rhs_info.n0 = 16 / b->info()->element_size();
+ rhs_info.k0 = 1;
+ rhs_info.h0 = mult_transpose1xW_width;
+ rhs_info.interleave = false;
+ rhs_info.transpose = false;
// 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);
// Check if we can run the new reshaped GEMM
_is_G76_path = (gpu_target == GPUTarget::G76) && _is_interleaved_transposed && (data_type == DataType::F32);
- ;
// if _is_interleaved_transposed is set, force reinterpret_input_as_3d to be false as the output of CLGEMMInterleaveKernel will be 2D
if(_is_interleaved_transposed)
@@ -201,7 +205,6 @@
if(_is_G76_path)
{
GEMMLHSMatrixInfo lhs_info;
- GEMMRHSMatrixInfo rhs_info;
// Pick up the GEMM configuration based on M,N and K
select_gemm_configuration(m, n, lhs_info, rhs_info);
@@ -219,7 +222,7 @@
_interleave_kernel.configure(a, &_tmp_a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d());
// Configure transpose kernel
- _transpose_kernel.configure(b, &_tmp_b, mult_transpose1xW_width);
+ _reshape_rhs_kernel.configure(b, &_tmp_b, rhs_info);
}
}
@@ -286,6 +289,13 @@
mult_interleave4x4_height = 2;
}
+ GEMMRHSMatrixInfo rhs_info;
+ rhs_info.n0 = 16 / b->element_size();
+ rhs_info.k0 = 1;
+ rhs_info.h0 = mult_transpose1xW_width;
+ rhs_info.interleave = false;
+ rhs_info.transpose = false;
+
// 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);
@@ -308,7 +318,6 @@
if(is_G76_path)
{
GEMMLHSMatrixInfo lhs_info;
- GEMMRHSMatrixInfo rhs_info;
// Pick up the GEMM configuration based on M,N and K
select_gemm_configuration(m, n, lhs_info, rhs_info);
@@ -328,10 +337,9 @@
// 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()));
-
// Validate transpose kernel
- auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_transpose1xW_with_element_size_shape(*b, mult_transpose1xW_width)));
- ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMTranspose1xWKernel::validate(b, &tmp_b_info, mult_transpose1xW_width));
+ 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));
}
}
@@ -371,14 +379,7 @@
if(!_reshape_b_only_on_first_run)
{
// Run transpose kernel
- if(_is_G76_path)
- {
- CLScheduler::get().enqueue(_reshape_rhs_kernel, false);
- }
- else
- {
- CLScheduler::get().enqueue(_transpose_kernel, false);
- }
+ CLScheduler::get().enqueue(_reshape_rhs_kernel, false);
}
}
@@ -409,14 +410,7 @@
{
// Run transpose kernel and mark original weights tensor as unused
_tmp_b.allocator()->allocate();
- if(_is_G76_path)
- {
- CLScheduler::get().enqueue(_reshape_rhs_kernel, false);
- }
- else
- {
- CLScheduler::get().enqueue(_transpose_kernel, false);
- }
+ CLScheduler::get().enqueue(_reshape_rhs_kernel, false);
_original_b->mark_as_unused();
}
CLScheduler::get().queue().finish();
diff --git a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
index 2d4d231..cf20bc6 100644
--- a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
@@ -108,6 +108,7 @@
const ICLTensor *matrix_a = a;
const ICLTensor *matrix_b = b;
+ GEMMRHSMatrixInfo rhs_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
@@ -120,6 +121,11 @@
const int depth_output_gemm3d = gemm_info.depth_output_gemm3d();
constexpr int mult_transpose1xW_width = 1;
constexpr int mult_interleave4x4_height = 1;
+ rhs_info.n0 = 16 / b->info()->element_size();
+ rhs_info.k0 = 1;
+ rhs_info.h0 = mult_transpose1xW_width;
+ rhs_info.interleave = false;
+ rhs_info.transpose = false;
// 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);
@@ -142,7 +148,7 @@
_mtx_a_reshape_kernel.configure(a, &_tmp_a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d(), unroll_block);
// Configure transpose kernel
- _mtx_b_reshape_kernel.configure(b, &_tmp_b, mult_transpose1xW_width);
+ _mtx_b_reshape_kernel.configure(b, &_tmp_b, rhs_info);
}
// Initialize matrix B reduction kernel only if _a_offset is not equal to 0
@@ -233,8 +239,9 @@
const ITensorInfo *matrix_a_info = a;
const ITensorInfo *matrix_b_info = b;
- TensorInfo tmp_a_info{};
- TensorInfo tmp_b_info{};
+ TensorInfo tmp_a_info{};
+ TensorInfo tmp_b_info{};
+ GEMMRHSMatrixInfo rhs_info;
bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
const int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1);
@@ -243,6 +250,11 @@
constexpr int mult_transpose1xW_width = 1;
constexpr int mult_interleave4x4_height = 1;
const int depth_output_gemm3d = gemm_info.depth_output_gemm3d();
+ rhs_info.n0 = 16 / b->element_size();
+ rhs_info.k0 = 1;
+ rhs_info.h0 = mult_transpose1xW_width;
+ rhs_info.interleave = false;
+ rhs_info.transpose = false;
bool reshape_matrices = is_interleaved_transposed(m, n, k, gemm_info.reshape_b_only_on_first_run(), CLScheduler::get().target());
@@ -264,8 +276,9 @@
ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMInterleave4x4Kernel::validate(a, &tmp_a_info, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d()));
// Validate transpose kernel
- auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_transpose1xW_with_element_size_shape(*b, mult_transpose1xW_width)));
- ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMTranspose1xWKernel::validate(b, &tmp_b_info, mult_transpose1xW_width));
+
+ 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));
}
TensorInfo info_vector_sum_col, info_vector_sum_row;