diff options
4 files changed, 28 insertions, 21 deletions
diff --git a/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp b/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp index 2086b1de03..dc47bb0adc 100644 --- a/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp +++ b/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp @@ -42,6 +42,17 @@ CLDepthwiseVectorToTensorKernel::CLDepthwiseVectorToTensorKernel() void CLDepthwiseVectorToTensorKernel::configure(const ICLTensor *input, ICLTensor *output, size_t conv_w, size_t conv_h) { ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); + ARM_COMPUTE_ERROR_ON_NULLPTR(output); + + TensorShape output_shape = input->info()->tensor_shape(); + output_shape.set(0, conv_w); + output_shape.set(1, conv_h); + output_shape.set(2, input->info()->tensor_shape()[0] / (conv_w * conv_h)); + + // Output auto inizialitation if not yet initialized + auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position()); + + ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); diff --git a/src/runtime/CL/functions/CLDepthwiseConvolution.cpp b/src/runtime/CL/functions/CLDepthwiseConvolution.cpp index ffd9f40f4f..bfc792c942 100644 --- a/src/runtime/CL/functions/CLDepthwiseConvolution.cpp +++ b/src/runtime/CL/functions/CLDepthwiseConvolution.cpp @@ -60,7 +60,6 @@ CLDepthwiseConvolution::CLDepthwiseConvolution() void CLDepthwiseConvolution::configure(ICLTensor *input, ICLTensor *output, const ICLTensor *weights, const ICLTensor *biases, const PadStrideInfo &conv_info) { ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != weights->info()->dimension(2)); @@ -78,28 +77,28 @@ void CLDepthwiseConvolution::configure(ICLTensor *input, ICLTensor *output, cons const size_t patch_size = weights_w * weights_h + ((has_bias) ? 1 : 0); const size_t conv_size = conv_w * conv_h; + // Im2Col configuration TensorShape shape_im2col = input->info()->tensor_shape(); shape_im2col.set(0, patch_size); shape_im2col.set(1, conv_size); shape_im2col.set(2, weights_z); + const TensorInfo info_im2col(shape_im2col, 1, input->info()->data_type(), input->info()->fixed_point_position()); + _input_reshaped.allocator()->init(info_im2col); + _im2col_kernel.configure(input, &_input_reshaped, Size2D(weights_w, weights_h), conv_info, has_bias); + // Weights reshape configuration const TensorShape shape_weights_reshape(patch_size, weights_z); - TensorShape shape_v2mm_out = output->info()->tensor_shape(); + const TensorInfo info_weights_reshape(shape_weights_reshape, 1, weights->info()->data_type(), weights->info()->fixed_point_position()); + _weights_reshaped.allocator()->init(info_weights_reshape); + _weights_reshape_kernel.configure(weights, &_weights_reshaped, biases); + + // GEMV configuration + TensorShape shape_v2mm_out = input->info()->tensor_shape(); shape_v2mm_out.set(0, conv_size * weights_z); shape_v2mm_out.set(1, 1); shape_v2mm_out.set(2, 1); - - const TensorInfo info_im2col(shape_im2col, 1, input->info()->data_type(), input->info()->fixed_point_position()); - const TensorInfo info_weights_reshape(shape_weights_reshape, 1, weights->info()->data_type(), weights->info()->fixed_point_position()); const TensorInfo info_v2mm_out(shape_v2mm_out, 1, input->info()->data_type(), input->info()->fixed_point_position()); - - _input_reshaped.allocator()->init(info_im2col); - _weights_reshaped.allocator()->init(info_weights_reshape); _v2mm_output.allocator()->init(info_v2mm_out); - - // Configure kernels - _im2col_kernel.configure(input, &_input_reshaped, Size2D(weights_w, weights_h), conv_info, has_bias); - _weights_reshape_kernel.configure(weights, &_weights_reshaped, biases); _v2mm_kernel.configure(&_input_reshaped, &_weights_reshaped, &_v2mm_output); _vector_to_tensor_kernel.configure(&_v2mm_output, output, conv_w, conv_h); diff --git a/tests/datasets/MobileNetDepthwiseConvolutionDataset.h b/tests/datasets/MobileNetDepthwiseConvolutionDataset.h index c2690d9ad6..918815f41e 100644 --- a/tests/datasets/MobileNetDepthwiseConvolutionDataset.h +++ b/tests/datasets/MobileNetDepthwiseConvolutionDataset.h @@ -42,13 +42,13 @@ class MobileNetDepthwiseConvolutionDataset final : public DepthwiseConvolutionDa public: MobileNetDepthwiseConvolutionDataset() { - add_config(TensorShape(7U, 7U, 1024U), TensorShape(3U, 3U, 1024U), TensorShape(1024U), TensorShape(3U, 3U, 1024U), PadStrideInfo(2, 2, 1, 1)); - add_config(TensorShape(14U, 14U, 512U), TensorShape(3U, 3U, 512U), TensorShape(512U), TensorShape(7U, 7U, 512U), PadStrideInfo(2, 2, 1, 1)); - add_config(TensorShape(28U, 28U, 256U), TensorShape(3U, 3U, 256U), TensorShape(256U), TensorShape(14U, 14U, 256U), PadStrideInfo(2, 2, 1, 1)); + add_config(TensorShape(7U, 7U, 1024U), TensorShape(3U, 3U, 1024U), TensorShape(1024U), TensorShape(3U, 3U, 1024U), PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR)); + add_config(TensorShape(14U, 14U, 512U), TensorShape(3U, 3U, 512U), TensorShape(512U), TensorShape(7U, 7U, 512U), PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR)); + add_config(TensorShape(28U, 28U, 256U), TensorShape(3U, 3U, 256U), TensorShape(256U), TensorShape(14U, 14U, 256U), PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR)); add_config(TensorShape(28U, 28U, 256U), TensorShape(3U, 3U, 256U), TensorShape(256U), TensorShape(28U, 28U, 256U), PadStrideInfo(1, 1, 1, 1)); - add_config(TensorShape(56U, 56U, 128U), TensorShape(3U, 3U, 128U), TensorShape(128U), TensorShape(28U, 28U, 128U), PadStrideInfo(2, 2, 1, 1)); + add_config(TensorShape(56U, 56U, 128U), TensorShape(3U, 3U, 128U), TensorShape(128U), TensorShape(28U, 28U, 128U), PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR)); add_config(TensorShape(56U, 56U, 128U), TensorShape(3U, 3U, 128U), TensorShape(128U), TensorShape(56U, 56U, 128U), PadStrideInfo(1, 1, 1, 1)); - add_config(TensorShape(112U, 112U, 64U), TensorShape(3U, 3U, 64U), TensorShape(64U), TensorShape(56U, 56U, 64U), PadStrideInfo(2, 2, 1, 1)); + add_config(TensorShape(112U, 112U, 64U), TensorShape(3U, 3U, 64U), TensorShape(64U), TensorShape(56U, 56U, 64U), PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR)); add_config(TensorShape(112U, 112U, 32U), TensorShape(3U, 3U, 32U), TensorShape(32U), TensorShape(112U, 112U, 32U), PadStrideInfo(1, 1, 1, 1)); } }; diff --git a/tests/datasets/MobileNetDepthwiseSeparableConvolutionLayerDataset.h b/tests/datasets/MobileNetDepthwiseSeparableConvolutionLayerDataset.h index d0f602daf0..eef0000925 100644 --- a/tests/datasets/MobileNetDepthwiseSeparableConvolutionLayerDataset.h +++ b/tests/datasets/MobileNetDepthwiseSeparableConvolutionLayerDataset.h @@ -43,10 +43,7 @@ public: MobileNetDepthwiseSeparableConvolutionLayerDataset() { add_config(TensorShape(14U, 14U, 512U), TensorShape(3U, 3U, 512U), TensorShape(512U), TensorShape(14U, 14U, 512U), TensorShape(1U, 1U, 512U, 512U), TensorShape(512U), TensorShape(14U, 14U, 512U), - PadStrideInfo(1, 1, 1, - 1, - DimensionRoundingType::FLOOR), - PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::FLOOR)); + PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::FLOOR), PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::FLOOR)); } }; } // namespace datasets |