commit c9c62c2fa1c80ba7f11b0d0732740460dfa00e74
author Gian Marco Iodice <gianmarco.iodice@arm.com> Fri Apr 06 10:00:10 2018 +0100
committer Anthony Barbier <anthony.barbier@arm.com> Fri Nov 02 16:49:37 2018 +0000
tree 260052aa5c7172e2afc8517ae13adb75504ee62e
parent 3ab6804a0e30be4d8591c8c84ae6a73940d0f2e2

COMPMID-1056 - Optimizing CLGEMMMatrixMultiplyKernel refactoring the inner loop

Results reported at:
https://confluence.arm.com/display/MLENG/GEMM+FP32+performance%3A+ACL+18.05

Change-Id: I3246c4f19c4d21a7d6a44e4593bc5caffc016f81
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/127838
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

src/runtime/CL/functions/CLFullyConnectedLayer.cpp

diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
index 5dd1f00..9b3bf48 100644
--- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
+++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
@@ -37,10 +37,8 @@
 
 namespace
 {
-Status validate_mm(const ITensorInfo &input, const ITensorInfo &weights, const ITensorInfo &output, bool is_interleaved_transposed)
+Status validate_mm(const ITensorInfo &input, const ITensorInfo &weights, const ITensorInfo &output)
 {
-    const GPUTarget gpu_target = CLScheduler::get().target();
-
     if(is_data_type_quantized_asymmetric(input.data_type()))
     {
         // Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
@@ -55,7 +53,7 @@
     }
     else
     {
-        ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMMatrixMultiplyKernel::validate(&input, &weights, &output, 1.f, is_interleaved_transposed, GEMMReshapeInfo(), gpu_target));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(&input, &weights, nullptr, &output, 1.f, 0.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run */)));
     }
 
     return Status{};
@@ -75,12 +73,12 @@
 }
 
 CLFullyConnectedLayer::CLFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(memory_manager), _im2col_kernel(), _reshape_weights_kernel(), _mm_kernel(), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(), _accumulate_biases_kernel(), _im2col_output(),
-      _gemmlowp_output(), _reshape_weights_output(), _are_weights_reshaped(true), _is_fc_after_conv(true), _accumulate_biases(false), _is_quantized(false), _original_weights(nullptr)
+    : _memory_group(memory_manager), _im2col_kernel(), _reshape_weights_kernel(), _mm_gemm(memory_manager), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(), _accumulate_biases_kernel(),
+      _im2col_output(), _gemmlowp_output(), _reshape_weights_output(), _are_weights_reshaped(true), _is_fc_after_conv(true), _accumulate_biases(false), _is_quantized(false), _original_weights(nullptr)
 {
 }
 
-void CLFullyConnectedLayer::configure_mm(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output, bool is_interleaved_transposed)
+void CLFullyConnectedLayer::configure_mm(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output)
 {
     if(_is_quantized)
     {
@@ -102,8 +100,7 @@
     else
     {
         // Configure matrix multiply kernel
-        _mm_kernel.set_target(CLScheduler::get().target());
-        _mm_kernel.configure(input, weights, output, 1.f, is_interleaved_transposed);
+        _mm_gemm.configure(input, weights, nullptr, output, 1.f, 0.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run */));
     }
 }
 
@@ -122,7 +119,7 @@
     _im2col_kernel.configure(input, &_im2col_output, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false);
 
     // Configure matrix multiply kernel
-    configure_mm(&_im2col_output, weights, output, false);
+    configure_mm(&_im2col_output, weights, output);
 
     // Allocate the output tensor for im2col once all the configure methods have been called
     _im2col_output.allocator()->allocate();
@@ -133,7 +130,7 @@
     ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != weights->info()->dimension(1));
 
     // Configure matrix multiply kernel
-    configure_mm(input, weights, output, false);
+    configure_mm(input, weights, output);
 }
 
 void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose_weights, bool are_weights_reshaped)
@@ -301,7 +298,7 @@
         ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != weights_to_use->dimension(1));
     }
     // Validate matrix multiply kernel
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(*input_to_use, *weights_to_use, *tmp_output, false));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(*input_to_use, *weights_to_use, *tmp_output));
 
     // Validate output stage for asymmetric quantized types
     if(is_quantized)
@@ -341,7 +338,7 @@
     }
     else
     {
-        CLScheduler::get().enqueue(_mm_kernel, !_accumulate_biases);
+        _mm_gemm.run();
     }
 
     // Accumulate biases if provided