COMPMID-421: Added FP16 support in Convolutional layer (Neon)

The test suite for FP16 is conditionally compiled in when the target platform is arch=arm64-8.2-a

Change-Id: I1686157e83809a00a91058bff80dbecf692fb356
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/78740
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

5 files changed, 47 insertions, 13 deletions

diff --git a/arm_compute/core/NEON/kernels/NEIm2ColKernel.h b/arm_compute/core/NEON/kernels/NEIm2ColKernel.h
index ebaafb467f..e219ce2e0e 100644
--- a/arm_compute/core/NEON/kernels/NEIm2ColKernel.h
+++ b/arm_compute/core/NEON/kernels/NEIm2ColKernel.h
@@ -72,7 +72,7 @@ public:
     /** Set the input and output of the kernel.
      *
      * @param[in]  input          The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
-     *                            while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QS8/F32
+     *                            while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QS8/F16/F32
      * @param[out] output         The output tensor. Data types supported: Same as @p input
      * @param[in]  convolved_dims The convolved output dimensions.
      * @param[in]  conv_info      Contains padding and stride information described in @ref PadStrideInfo.
diff --git a/src/core/NEON/kernels/NEIm2ColKernel.cpp b/src/core/NEON/kernels/NEIm2ColKernel.cpp
index c7c23d5d06..875c08ed42 100644
--- a/src/core/NEON/kernels/NEIm2ColKernel.cpp
+++ b/src/core/NEON/kernels/NEIm2ColKernel.cpp
@@ -266,8 +266,8 @@ NEIm2ColKernel::NEIm2ColKernel()
 
 void NEIm2ColKernel::configure(const ITensor *input, ITensor *output, std::pair<unsigned int, unsigned int> convolved_dims, const PadStrideInfo &conv_info, bool has_bias)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::QS8);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32, DataType::QS8);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32, DataType::QS8);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F16, DataType::F32, DataType::QS8);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
 
     _input          = input;
@@ -296,6 +296,11 @@ void NEIm2ColKernel::configure(const ITensor *input, ITensor *output, std::pair<
             case DataType::F32:
                 _func = &NEIm2ColKernel::run_reduced<float>;
                 break;
+#ifdef ARM_COMPUTE_ENABLE_FP16
+            case DataType::F16:
+                _func = &NEIm2ColKernel::run_reduced<float16_t>;
+                break;
+#endif
             case DataType::QS8:
                 _func = &NEIm2ColKernel::run_reduced<qint8_t>;
                 break;
@@ -311,6 +316,11 @@ void NEIm2ColKernel::configure(const ITensor *input, ITensor *output, std::pair<
             case DataType::F32:
                 _func = ((pad_x == 0) && (pad_y == 0)) ? &NEIm2ColKernel::run_generic<float, false> : &NEIm2ColKernel::run_generic<float, true>;
                 break;
+#ifdef ARM_COMPUTE_ENABLE_FP16
+            case DataType::F16:
+                _func = ((pad_x == 0) && (pad_y == 0)) ? &NEIm2ColKernel::run_generic<float16_t, false> : &NEIm2ColKernel::run_generic<float16_t, true>;
+                break;
+#endif
             case DataType::QS8:
                 _func = ((pad_x == 0) && (pad_y == 0)) ? &NEIm2ColKernel::run_generic<qint8_t, false> : &NEIm2ColKernel::run_generic<qint8_t, true>;
                 break;
diff --git a/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp b/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp
index f31cde719a..4f52bf6279 100644
--- a/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp
+++ b/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp
@@ -95,7 +95,7 @@ NEWeightsReshapeKernel::NEWeightsReshapeKernel()
 
 void NEWeightsReshapeKernel::configure(const ITensor *input, const ITensor *bias, ITensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::QS8);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16, DataType::QS8);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
     ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != input->info()->dimension(1));
 
@@ -113,7 +113,7 @@ void NEWeightsReshapeKernel::configure(const ITensor *input, const ITensor *bias
     set_fixed_point_position_if_zero(*output->info(), fixed_point_position);
 
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32, DataType::QS8);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F16, DataType::F32, DataType::QS8);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
 
@@ -127,7 +127,7 @@ void NEWeightsReshapeKernel::configure(const ITensor *input, const ITensor *bias
         set_shape_if_empty(*bias->info(), bias_shape);
 
         ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(bias->info()->tensor_shape(), bias_shape);
-        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::F32, DataType::QS8);
+        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::F16, DataType::F32, DataType::QS8);
         ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
         ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
     }
@@ -143,6 +143,11 @@ void NEWeightsReshapeKernel::configure(const ITensor *input, const ITensor *bias
             _func = &weights_reshape<uint32_t>;
             break;
         }
+        case DataType::F16:
+        {
+            _func = &weights_reshape<uint16_t>;
+            break;
+        }
         case DataType::QS8:
         {
             _func = &weights_reshape<uint8_t>;
diff --git a/src/runtime/NEON/functions/NEConvolutionLayer.cpp b/src/runtime/NEON/functions/NEConvolutionLayer.cpp
index bd688cffb6..82c33d54bb 100644
--- a/src/runtime/NEON/functions/NEConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEConvolutionLayer.cpp
@@ -40,15 +40,15 @@ NEConvolutionLayerReshapeWeights::NEConvolutionLayerReshapeWeights()
 
 void NEConvolutionLayerReshapeWeights::configure(const ITensor *weights, const ITensor *biases, ITensor *output, bool transpose1xW)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QS8, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QS8, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(weights, output);
     ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4);
 
     if(biases != nullptr)
     {
-        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::QS8, DataType::F32);
+        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::QS8, DataType::F16, DataType::F32);
         ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
         ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(weights, biases);
         ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(3));
@@ -96,9 +96,9 @@ NEConvolutionLayer::NEConvolutionLayer()
 
 void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QS8, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QS8, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights, output);
     ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && weights->info()->dimension(2) != input->info()->dimension(2));
@@ -106,7 +106,7 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights,
 
     if(biases != nullptr)
     {
-        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::QS8, DataType::F32);
+        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::QS8, DataType::F16, DataType::F32);
         ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
         ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases);
         ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && biases->info()->dimension(0) != weights->info()->dimension(3));
diff --git a/tests/validation/NEON/ConvolutionLayer.cpp b/tests/validation/NEON/ConvolutionLayer.cpp
index ee2b24db96..40c1e2e969 100644
--- a/tests/validation/NEON/ConvolutionLayer.cpp
+++ b/tests/validation/NEON/ConvolutionLayer.cpp
@@ -130,6 +130,25 @@ BOOST_DATA_TEST_CASE(Configuration,
     validate(dst.info()->valid_region(), dst_valid_region);
 }
 
+#ifdef ARM_COMPUTE_ENABLE_FP16
+BOOST_AUTO_TEST_SUITE(Float16)
+BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit"))
+BOOST_DATA_TEST_CASE(SmallConvolutionLayer,
+                     SmallConvolutionLayerDataset() * boost::unit_test::data::make(DataType::F16),
+                     conv_set, dt)
+{
+    // Compute function
+    Tensor dst = compute_convolution_layer(conv_set.src_shape, conv_set.weights_shape, conv_set.bias_shape, conv_set.dst_shape, dt, conv_set.info, 0);
+
+    // Compute reference
+    RawTensor ref_dst = Reference::compute_reference_convolution_layer(conv_set.src_shape, conv_set.weights_shape, conv_set.bias_shape, conv_set.dst_shape, dt, conv_set.info, 0);
+
+    // Validate output
+    validate(NEAccessor(dst), ref_dst, tolerance_f32);
+}
+BOOST_AUTO_TEST_SUITE_END()
+#endif
+
 BOOST_AUTO_TEST_SUITE(Float)
 BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit"))
 BOOST_DATA_TEST_CASE(SmallConvolutionLayer,