COMPMID-1096 - Add fast_math flag to CLConvolutionLayer

COMPMID-1103 - CLWinogradConvolutionLayer mismatches

Change-Id: Iceaa9482a1790ec39d2720c220261aaea8043978
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/129398
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

1 files changed, 121 insertions, 1 deletions

diff --git a/tests/validation/fixtures/WinogradConvolutionLayerFixture.h b/tests/validation/fixtures/WinogradConvolutionLayerFixture.h
index 249f9d5649..e15931eafb 100644
--- a/tests/validation/fixtures/WinogradConvolutionLayerFixture.h
+++ b/tests/validation/fixtures/WinogradConvolutionLayerFixture.h
@@ -35,6 +35,7 @@
 #include "tests/validation/Helpers.h"
 #include "tests/validation/reference/ActivationLayer.h"
 #include "tests/validation/reference/ConvolutionLayer.h"
+#include "tests/validation/reference/GEMM.h"
 #include "tests/validation/reference/Utils.h"
 #include "tests/validation/reference/Winograd.h"
 
@@ -153,6 +154,123 @@ protected:
 };
 
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class WinogradConvolutionLayerFastMathValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, DataType data_type, ActivationLayerInfo act_info)
+    {
+        ARM_COMPUTE_UNUSED(dilation);
+
+        _target    = compute_target(input_shape, weights_shape, bias_shape, output_shape, info, data_type, act_info);
+        _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, data_type, act_info);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i, float min, float max)
+    {
+        switch(tensor.data_type())
+        {
+            case DataType::F32:
+            {
+                std::uniform_real_distribution<> distribution(min, max);
+                library->fill(tensor, distribution, i);
+                break;
+            }
+            default:
+            {
+                ARM_COMPUTE_ERROR("Not supported");
+                library->fill_tensor_uniform(tensor, i);
+                break;
+            }
+        }
+    }
+
+    TensorType compute_target(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, const PadStrideInfo &info,
+                              DataType data_type, ActivationLayerInfo act_info)
+    {
+        // Create tensors
+        TensorType src     = create_tensor<TensorType>(input_shape, data_type, 1);
+        TensorType weights = create_tensor<TensorType>(weights_shape, data_type, 1);
+        TensorType bias    = create_tensor<TensorType>(bias_shape, data_type, 1);
+        TensorType dst     = create_tensor<TensorType>(output_shape, data_type, 1);
+
+        // Create and configure function
+        FunctionType conv;
+        ARM_COMPUTE_EXPECT(static_cast<bool>(conv.validate(src.info(), weights.info(), bias.info(), dst.info(), info, act_info, true /* Enable fast math */)), framework::LogLevel::ERRORS);
+        conv.configure(&src, &weights, &bias, &dst, info, act_info, true /* Enable fast math */);
+
+        ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(weights.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        src.allocator()->allocate();
+        weights.allocator()->allocate();
+        dst.allocator()->allocate();
+        bias.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!weights.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensors
+        fill(AccessorType(src), 0, -1.f, 1.f);
+        fill(AccessorType(weights), 1, -1.f, 1.f);
+        fill(AccessorType(bias), 2, -1.f, 1.f);
+
+        // Compute Winograd Convolution function
+        conv.run();
+
+        return dst;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, const PadStrideInfo &info,
+                                      DataType data_type, ActivationLayerInfo act_info)
+    {
+        // Create reference
+        SimpleTensor<T> src{ input_shape, data_type, 1 };
+        SimpleTensor<T> weights{ weights_shape, data_type, 1 };
+        SimpleTensor<T> bias{ bias_shape, data_type, 1 };
+
+        // Fill reference
+        fill(src, 0, -1.f, 1.f);
+        fill(weights, 1, -1.f, 1.f);
+        fill(bias, 2, -1.f, 1.f);
+
+        WinogradInfo winograd_info(Size2D(4U, 4U),
+                                   Size2D(weights_shape[0], weights_shape[1]),
+                                   Size2D(input_shape[0], input_shape[1]),
+                                   info,
+                                   src.data_layout());
+
+        // Compute tensor shapes for input, filter and output transforms
+        TensorShape input_transform_shape  = compute_winograd_input_transform_shape(TensorInfo(input_shape, 1, data_type), winograd_info);
+        TensorShape filter_transform_shape = compute_winograd_filter_transform_shape(TensorInfo(weights_shape, 1, data_type), winograd_info);
+        TensorShape batched_gemm_shape     = input_transform_shape;
+        batched_gemm_shape[0]              = filter_transform_shape[0];
+        TensorShape output_transform_shape = compute_winograd_output_transform_shape(TensorInfo(batched_gemm_shape, 1, data_type), winograd_info);
+
+        // Dummy matrix C to perform matrix multiplication
+        SimpleTensor<T> dummy_c{ batched_gemm_shape, data_type, 1 };
+
+        // Compute Winograd-based convolution
+        SimpleTensor<T> input_transform_out  = reference::winograd_input_transform<T>(src, input_transform_shape, winograd_info);
+        SimpleTensor<T> filter_transform_out = reference::winograd_filter_transform<T>(weights, filter_transform_shape, winograd_info);
+        SimpleTensor<T> batched_gemm         = reference::gemm<T>(input_transform_out, filter_transform_out, dummy_c, 1.0f, 0.0f);
+        SimpleTensor<T> conv_out             = reference::winograd_output_transform<T>(batched_gemm, bias, output_transform_shape, winograd_info);
+
+        return (act_info.enabled()) ? reference::activation_layer<T>(conv_out, act_info) : conv_out;
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
 class WinogradInputTransformValidationFixture : public framework::Fixture
 {
 public:
@@ -373,11 +491,13 @@ protected:
     {
         // Create reference
         SimpleTensor<T> src{ input_shape, data_type };
+        SimpleTensor<T> bias{ TensorShape(input_shape[0]), data_type };
 
         // Fill reference
         fill(src, 0, -1.f, 1.f);
+        fill(bias, 1, 0.0f, 0.0f); // Fill with zeros as we validate just the output transform without bias contribution
 
-        return reference::winograd_output_transform<T>(src, output_shape, winograd_info);
+        return reference::winograd_output_transform<T>(src, bias, output_shape, winograd_info);
     }
 
     TensorType      _target{};