COMPMID-616 - Optimizing GEMMLowp on NEON intrinsics
Change-Id: Ibbeff5d37249b6e8fc34ad496035a1511c9da5a3
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/94072
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
diff --git a/src/runtime/NEON/functions/NEGEMMLowp.cpp b/src/runtime/NEON/functions/NEGEMMLowp.cpp
index 12136cb..ab7fa07 100644
--- a/src/runtime/NEON/functions/NEGEMMLowp.cpp
+++ b/src/runtime/NEON/functions/NEGEMMLowp.cpp
@@ -31,64 +31,80 @@
#include "arm_compute/core/Types.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/runtime/NEON/NEScheduler.h"
+#include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h"
#include "arm_compute/runtime/TensorAllocator.h"
#include "support/ToolchainSupport.h"
using namespace arm_compute;
-#define NEGEMMLOWP_VALIDATE_DIMENSIONS(a, b, output) \
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN((a), 1, DataType::U8); \
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN((b), 1, DataType::U8); \
- ARM_COMPUTE_ERROR_ON_MSG((a)->info()->dimension(0) != (b)->info()->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B"); \
- ARM_COMPUTE_ERROR_ON_MSG((a)->info()->dimension(1) != (output)->info()->dimension(1), "The C matrix must have the same number of rows as the matrix A"); \
- ARM_COMPUTE_ERROR_ON_MSG((b)->info()->dimension(0) != (output)->info()->dimension(0), "The C matrix must have the same number of columns as the matrix C");
-
NEGEMMLowp::NEGEMMLowp(std::shared_ptr<IMemoryManager> memory_manager)
- : _memory_group(std::move(memory_manager)), _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _mm_optimised_kernel(nullptr), _interleave_blocked(), _interleave_blocked_transposed(), _tmp_a(),
- _tmp_b()
+ : _memory_group(std::move(memory_manager)), _mm_func(), _mtx_a_reduction_kernel(), _mtx_b_reduction_kernel(), _finalize_kernel(), _vector_sum_col(), _vector_sum_row(), _mm_output(), _a_offset(0),
+ _b_offset(0)
{
}
-void NEGEMMLowp::configure(const ITensor *a, const ITensor *b, ITensor *output)
+void NEGEMMLowp::configure(const ITensor *a, const ITensor *b, ITensor *output, int32_t a_offset, int32_t b_offset, int32_t c_offset, int32_t output_mult_int, int32_t shift)
{
- NEGEMMLOWP_VALIDATE_DIMENSIONS(a, b, output);
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U32);
+ ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN((a), 1, DataType::U8);
+ ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, output);
+ ARM_COMPUTE_ERROR_ON_MSG((a)->info()->dimension(0) != (b)->info()->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B");
+ ARM_COMPUTE_ERROR_ON_MSG((a)->info()->dimension(1) != (output)->info()->dimension(1), "The output matrix must have the same number of rows as the matrix A");
+ ARM_COMPUTE_ERROR_ON_MSG((b)->info()->dimension(0) != (output)->info()->dimension(0), "The output matrix must have the same number of columns as the matrix B");
- const struct CPUInfo ci = NEScheduler::get().cpu_info();
- const int cpu_has_dotprod = static_cast<int>(ci.CPU) & static_cast<int>(CPUTarget::DOT);
- if(cpu_has_dotprod != 0)
+ _a_offset = a_offset;
+ _b_offset = b_offset;
+
+ // Initialize matrix multiply output tensor
+ const TensorShape &shape_mm_output = output->info()->tensor_shape();
+ TensorInfo info_mm_output(shape_mm_output, 1, DataType::S32);
+ _mm_output.allocator()->init(info_mm_output);
+ _memory_group.manage(&_mm_output);
+
+ // Initialize Matrix B reduction kernel only if _a_offset is not equal to 0
+ if(_a_offset != 0)
{
-#ifdef ARM_COMPUTE_AARCH64_V8_2
- // NEGEMMLowpAArch64V8P4Kernel only compiled in AArch64 targets
- _mm_optimised_kernel = support::cpp14::make_unique<NEGEMMLowpAArch64V8P4Kernel>();
- TensorShape shape_a_int = a->info()->tensor_shape();
- shape_a_int.set(0, a->info()->dimension(0) * 8.f);
- shape_a_int.set(1, std::ceil(a->info()->dimension(1) / 8.f));
+ TensorShape shape_vector_sum_col = b->info()->tensor_shape();
+ shape_vector_sum_col.remove_dimension(1);
+ TensorInfo info_vector_sum_col(shape_vector_sum_col, 1, DataType::S32);
+ _vector_sum_col.allocator()->init(info_vector_sum_col);
+ _memory_group.manage(&_vector_sum_col);
- TensorShape shape_b_int = b->info()->tensor_shape();
- shape_b_int.set(0, b->info()->dimension(0) * 12.f);
- shape_b_int.set(1, std::ceil(b->info()->dimension(1) / 12.f));
-
- TensorInfo info_a_int(shape_a_int, 1, a->info()->data_type());
- TensorInfo info_b_int(shape_b_int, 1, b->info()->data_type());
- _tmp_a.allocator()->init(info_a_int);
- _tmp_b.allocator()->init(info_b_int);
-
- _memory_group.manage(&_tmp_a);
- _memory_group.manage(&_tmp_b);
-
- _interleave_blocked.configure(a, &_tmp_a, 8, 4, false);
- _interleave_blocked_transposed.configure(b, &_tmp_b, 12, 4, true);
- _mm_optimised_kernel->configure(&_tmp_a, &_tmp_b, output);
-
- _tmp_a.allocator()->allocate();
- _tmp_b.allocator()->allocate();
-#endif /* ARM_COMPUTE_AARCH64_V8_2 */
+ // Configure Matrix B reduction kernel
+ _mtx_b_reduction_kernel.configure(b, &_vector_sum_col, a->info()->dimension(0), false);
}
- else
+
+ // Initialize Matrix A reduction kernel only if _b_offset is not equal to 0
+ if(_b_offset != 0)
{
- ARM_COMPUTE_ERROR("Not implemented");
- //FIXME: This is in the process of being updated, for more info please refer to COMPMID-624.
+ TensorShape shape_vector_sum_row = a->info()->tensor_shape();
+ shape_vector_sum_row.set(Window::DimX, a->info()->dimension(1));
+ shape_vector_sum_row.remove_dimension(1);
+ TensorInfo info_vector_sum_row(shape_vector_sum_row, 1, DataType::S32);
+ _vector_sum_row.allocator()->init(info_vector_sum_row);
+ _memory_group.manage(&_vector_sum_row);
+
+ // Configure Matrix A reduction kernel
+ _mtx_a_reduction_kernel.configure(a, &_vector_sum_row, a->info()->dimension(0), false);
+ }
+
+ // Configure matrix multiply function
+ _mm_func.configure(a, b, &_mm_output);
+
+ // Configure finalize kernel
+ _finalize_kernel.configure(_a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, &_mm_output, output, a->info()->dimension(0), a_offset, b_offset, c_offset,
+ output_mult_int, shift);
+
+ // Allocate tensors
+ _mm_output.allocator()->allocate();
+
+ if(_a_offset != 0)
+ {
+ _vector_sum_col.allocator()->allocate();
+ }
+
+ if(_b_offset != 0)
+ {
+ _vector_sum_row.allocator()->allocate();
}
}
@@ -96,55 +112,23 @@
{
_memory_group.acquire();
- if(_mm_optimised_kernel != nullptr)
+ // Run matrix A reduction kernel only if _b_offset is not equal to 0
+ if(_b_offset != 0)
{
- NEScheduler::get().schedule(&_interleave_blocked, Window::DimY);
- NEScheduler::get().schedule(&_interleave_blocked_transposed, Window::DimY);
- NEScheduler::get().schedule(_mm_optimised_kernel.get(), Window::DimY);
+ NEScheduler::get().schedule(&_mtx_a_reduction_kernel, Window::DimX);
}
- else
+
+ // Run matrix B reduction kernel only if _a_offset is not equal to 0
+ if(_a_offset != 0)
{
- /* Run interleave kernel */
- NEScheduler::get().schedule(&_interleave_kernel, Window::DimY);
- /* Run transpose kernel */
- NEScheduler::get().schedule(&_transpose_kernel, Window::DimY);
- /* Run matrix multiply kernel */
- NEScheduler::get().schedule(&_mm_kernel, Window::DimY);
+ NEScheduler::get().schedule(&_mtx_b_reduction_kernel, Window::DimX);
}
+ // Run matrix multiply core function
+ _mm_func.run();
+
+ // Run finalise kernel
+ NEScheduler::get().schedule(&_finalize_kernel, Window::DimY);
+
_memory_group.release();
-}
-
-void NEGEMMLowp::configure(const ITensor *a, const ITensor *b, ITensor *output, int32_t a_offset, int32_t b_offset, int32_t output_offset, int32_t output_mult_int, int32_t shift)
-{
- NEGEMMLOWP_VALIDATE_DIMENSIONS(a, b, output);
- ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, output);
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8);
-
- /* The interleaved output matrix will have the following shape: [ a_height * 4, ceil(a_width / 4.0f) ] */
- TensorShape shape_tmp_a = a->info()->tensor_shape();
- shape_tmp_a.set(0, a->info()->dimension(0) * 4);
- shape_tmp_a.set(1, std::ceil(a->info()->dimension(1) / 4.f));
-
- TensorShape shape_tmp_b = b->info()->tensor_shape();
- shape_tmp_b.set(0, b->info()->dimension(1) * 16);
- shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / 16.f));
-
- TensorInfo info_a(shape_tmp_a, 1, a->info()->data_type());
- TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type());
- _tmp_a.allocator()->init(info_a);
- _tmp_b.allocator()->init(info_b);
-
- // Manage intermediate buffers
- _memory_group.manage(&_tmp_a);
- _memory_group.manage(&_tmp_b);
-
- _interleave_kernel.configure(a, &_tmp_a);
- _transpose_kernel.configure(b, &_tmp_b);
- _mm_kernel.configure(&_tmp_a, &_tmp_b, output, a_offset, b_offset, output_offset, output_mult_int, shift);
-
- _tmp_a.allocator()->allocate();
- _tmp_b.allocator()->allocate();
-}
-
-#undef NEGEMMLOWP_VALIDATE_DIMENSIONS
+}
\ No newline at end of file