diff options
Diffstat (limited to 'src/core/NEON')
-rw-r--r-- | src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp | 41 | ||||
-rw-r--r-- | src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp | 19 |
2 files changed, 32 insertions, 28 deletions
diff --git a/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp b/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp index 49c67d19bb..8cdf175d8a 100644 --- a/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp +++ b/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp @@ -52,13 +52,14 @@ class convolver_3x3 { public: static void convolve(const Window &window, unsigned int num_elems_written_per_iteration, - const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info) + const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier) { const int input_offset = -input->info()->quantization_info().offset; const int weights_offset = -weights->info()->quantization_info().offset; const int input_stride_x = input->info()->strides_in_bytes().x(); const int input_stride_y = input->info()->strides_in_bytes().y(); + const int input_stride_z = input->info()->strides_in_bytes().z(); const int output_stride_y = output->info()->strides_in_bytes().y(); const int kernel_stride_y = weights->info()->strides_in_bytes().y(); const int kernel_stride_z = weights->info()->strides_in_bytes().z(); @@ -93,7 +94,7 @@ public: int ih = 0; int oh = 0; - const uint8_t *input_ptr = in.ptr() - conv_pad_x * input_stride_x - conv_pad_y * input_stride_y; + const uint8_t *input_ptr = in.ptr() - conv_pad_x * input_stride_x - conv_pad_y * input_stride_y - (id.z() - id.z() / depth_multiplier) * input_stride_z; const uint8_t *ptr_weights_base = weights_ptr + id.z() * kernel_stride_z; const auto ptr_weights_r0 = reinterpret_cast<const T1 *>(ptr_weights_base); @@ -125,19 +126,19 @@ public: template <typename T1, typename T2> inline void convolve_3x3(const Window &window, unsigned int num_elems_written_per_iteration, - const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info) + const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier) { const unsigned int conv_stride_x = std::get<0>(conv_info.stride()); switch(conv_stride_x) { case 1: - convolver_3x3<T1, T2, 1>::convolve(window, num_elems_written_per_iteration, input, weights, output, conv_info); + convolver_3x3<T1, T2, 1>::convolve(window, num_elems_written_per_iteration, input, weights, output, conv_info, depth_multiplier); break; case 2: - convolver_3x3<T1, T2, 2>::convolve(window, num_elems_written_per_iteration, input, weights, output, conv_info); + convolver_3x3<T1, T2, 2>::convolve(window, num_elems_written_per_iteration, input, weights, output, conv_info, depth_multiplier); break; case 3: - convolver_3x3<T1, T2, 3>::convolve(window, num_elems_written_per_iteration, input, weights, output, conv_info); + convolver_3x3<T1, T2, 3>::convolve(window, num_elems_written_per_iteration, input, weights, output, conv_info, depth_multiplier); break; default: ARM_COMPUTE_ERROR("Not implemented"); @@ -146,7 +147,7 @@ inline void convolve_3x3(const Window &window, unsigned int num_elems_written_pe } // namespace NEDepthwiseConvolutionLayer3x3Kernel::NEDepthwiseConvolutionLayer3x3Kernel() - : _border_size(0), _input(), _output(), _weights(), _conv_info(), _convolver(nullptr), _num_elems_written_per_iteration(0), _run_optimized(false) + : _border_size(0), _input(), _output(), _weights(), _conv_info(), _convolver(nullptr), _num_elems_written_per_iteration(0), _run_optimized(false), _depth_multiplier(1) { } @@ -155,20 +156,22 @@ BorderSize NEDepthwiseConvolutionLayer3x3Kernel::border_size() const return _border_size; } -void NEDepthwiseConvolutionLayer3x3Kernel::configure(const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info, DataLayout data_layout) +void NEDepthwiseConvolutionLayer3x3Kernel::configure(const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier, + DataLayout data_layout) { ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F32); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); - _input = input; - _output = output; - _weights = weights; - _conv_info = conv_info; - _convolver = nullptr; + _input = input; + _output = output; + _weights = weights; + _conv_info = conv_info; + _depth_multiplier = depth_multiplier; + _convolver = nullptr; _run_optimized = NEDepthwiseConvolutionLayer3x3Kernel::is_optimized_execution_possible(input->info()->tensor_shape(), conv_info, - input->info()->data_type(), + input->info()->data_type(), depth_multiplier, data_layout); (_run_optimized) ? configure_optimized() : configure_generic(); @@ -182,7 +185,7 @@ void NEDepthwiseConvolutionLayer3x3Kernel::run(const Window &window, const Threa (_run_optimized) ? run_optimized(window, info) : run_generic(window, info); } -bool NEDepthwiseConvolutionLayer3x3Kernel::is_optimized_execution_possible(TensorShape input_shape, PadStrideInfo conv_info, DataType dt, DataLayout data_layout) +bool NEDepthwiseConvolutionLayer3x3Kernel::is_optimized_execution_possible(TensorShape input_shape, PadStrideInfo conv_info, DataType dt, unsigned int depth_multiplier, DataLayout data_layout) { // Reshape input shape if in NHWC format TensorShape in_shape{ input_shape }; @@ -210,7 +213,7 @@ bool NEDepthwiseConvolutionLayer3x3Kernel::is_optimized_execution_possible(Tenso bool is_valid_padding = (pad_top == 0) && (pad_right == 0) && (pad_bottom == 0) && (pad_left == 0); bool supported_padding = is_same_padding || is_valid_padding; - return supported_datatype && supported_strides && supported_padding; + return supported_datatype && supported_strides && supported_padding && (depth_multiplier == 1); } void NEDepthwiseConvolutionLayer3x3Kernel::generate_convolver() @@ -227,7 +230,7 @@ void NEDepthwiseConvolutionLayer3x3Kernel::configure_generic() ARM_COMPUTE_ERROR_ON(_weights->info()->dimension(0) != 3 || _weights->info()->dimension(1) != 3); // Get convolved dimensions - const TensorShape output_shape = compute_depthwise_convolution_shape(*_input->info(), *_weights->info(), _conv_info); + const TensorShape output_shape = compute_depthwise_convolution_shape(*_input->info(), *_weights->info(), _conv_info, _depth_multiplier); const DataType output_dt = (_input->info()->data_type() == DataType::QASYMM8) ? DataType::S32 : _input->info()->data_type(); // Output auto inizialitation if not yet initialized @@ -317,10 +320,10 @@ void NEDepthwiseConvolutionLayer3x3Kernel::run_generic(const Window &window, con switch(_input->info()->data_type()) { case DataType::F32: - convolve_3x3<float, float>(window, _num_elems_written_per_iteration, _input, _weights, _output, _conv_info); + convolve_3x3<float, float>(window, _num_elems_written_per_iteration, _input, _weights, _output, _conv_info, _depth_multiplier); break; case DataType::QASYMM8: - convolve_3x3<uint8_t, int32_t>(window, _num_elems_written_per_iteration, _input, _weights, _output, _conv_info); + convolve_3x3<uint8_t, int32_t>(window, _num_elems_written_per_iteration, _input, _weights, _output, _conv_info, _depth_multiplier); break; default: ARM_COMPUTE_ERROR("Not implemented"); diff --git a/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp b/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp index b924d9f8bd..cfd8eacfdd 100644 --- a/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp +++ b/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp @@ -85,7 +85,7 @@ void NEDepthwiseIm2ColKernel::run_generic(const Window &window) const int src_y = -pad_top + src_pixel_linear / max_initial_x * stride_y; // Get pointers - const uint8_t *const input_ptr = in.ptr() + id.z() * input_stride_z; + const uint8_t *const input_ptr = in.ptr() + id.z() / _depth_multiplier * input_stride_z; auto output_ptr = reinterpret_cast<T *>(out.ptr()); const int height = src_y + _kernel_dims.height; const int width = src_x + _kernel_dims.width; @@ -114,24 +114,25 @@ void NEDepthwiseIm2ColKernel::run_generic(const Window &window) } NEDepthwiseIm2ColKernel::NEDepthwiseIm2ColKernel() - : _func(nullptr), _input(nullptr), _output(nullptr), _kernel_dims(), _conv_info(), _has_bias() + : _func(nullptr), _input(nullptr), _output(nullptr), _kernel_dims(), _conv_info(), _has_bias(), _depth_multiplier(1) { } -void NEDepthwiseIm2ColKernel::configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias) +void NEDepthwiseIm2ColKernel::configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier) { ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); ARM_COMPUTE_ERROR_ON(is_data_type_quantized_asymmetric(input->info()->data_type()) && has_bias); - ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != output->info()->dimension(2)); + ARM_COMPUTE_ERROR_ON((input->info()->dimension(2) * depth_multiplier) != output->info()->dimension(2)); ARM_COMPUTE_ERROR_ON(output->info()->dimension(0) != (kernel_dims.width * kernel_dims.height + ((has_bias) ? 1 : 0))); - _input = input; - _output = output; - _kernel_dims = kernel_dims; - _conv_info = conv_info; - _has_bias = has_bias; + _input = input; + _output = output; + _kernel_dims = kernel_dims; + _conv_info = conv_info; + _has_bias = has_bias; + _depth_multiplier = depth_multiplier; // Configure kernel window Window win = calculate_max_window(*input->info(), Steps()); |