Implement FP GPU depthwise convolution 1x1 kernel for in-place computation

* Implement in-place graph node mutator for 1x1 depthwise convolution

* Add in-place to validation fixture except for
  DepthwiseConvolutionLayerNativeValidationFixture as it would be a
  duplicate test otherwise (DepthwiseConvolutionLayerNative test tests
  the underlying kernel)

Resolves: COMPMID-4432

Change-Id: Id7f10f5ebdce7d49f550c0b62dbaaab7f5b59d29
Signed-off-by: SiCongLi <sicong.li@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5874
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

2 files changed, 67 insertions, 61 deletions

diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
index 4cc0e462c4..eb1cf146af 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
@@ -46,7 +46,13 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights,
                           const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
 {
     ARM_COMPUTE_UNUSED(dwc_info);
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
+    bool in_place = false;
+    if(output == nullptr || output == input)
+    {
+        in_place = true;
+        output   = input;
+    }
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights);
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32);
@@ -58,6 +64,18 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights,
     ARM_COMPUTE_UNUSED(idx_c);
     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_c) != (input->dimension(idx_c) * depth_multiplier));
 
+    // In place restrictions
+    if(in_place)
+    {
+        const int weights_width_idx  = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::WIDTH);
+        const int weights_height_idx = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::HEIGHT);
+        ARM_COMPUTE_RETURN_ERROR_ON(weights->tensor_shape()[weights_width_idx] != 1U || weights->tensor_shape()[weights_height_idx] != 1U);
+        ARM_COMPUTE_RETURN_ERROR_ON(depth_multiplier != 1U);
+        ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride() != std::make_pair(1U, 1U));
+        ARM_COMPUTE_RETURN_ERROR_ON(dilation != Size2D(1U, 1U));
+        ARM_COMPUTE_RETURN_ERROR_ON(conv_info.has_padding()); // Note that in princple padding can be supported with in_place but we choose not to support it
+    }
+
     const ConvolutionInfo info{ conv_info, depth_multiplier, ActivationLayerInfo(), dilation };
     const TensorShape     output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, info);
 
@@ -139,19 +157,24 @@ CLDepthwiseConvolutionLayerNativeKernel::CLDepthwiseConvolutionLayerNativeKernel
     _type = CLKernelType::DEPTHWISE;
 }
 
-void CLDepthwiseConvolutionLayerNativeKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const DWCWeightsKernelInfo &dwc_weights_info,
-                                                        const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation,
+void CLDepthwiseConvolutionLayerNativeKernel::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
+                                                        const DWCWeightsKernelInfo &dwc_weights_info, const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation,
                                                         const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
 {
     configure(CLKernelLibrary::get().get_compile_context(), input, weights, biases, output, dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation, output_multipliers, output_shifts);
 }
 
-void CLDepthwiseConvolutionLayerNativeKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
+void CLDepthwiseConvolutionLayerNativeKernel::configure(const CLCompileContext &compile_context, ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
                                                         const DWCWeightsKernelInfo &dwc_weights_info,
                                                         const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation,
                                                         const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights);
+    if(output == nullptr)
+    {
+        // In-place
+        output = input;
+    }
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(),
                                                   dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation,
                                                   (output_multipliers != nullptr) ? output_multipliers->info() : nullptr, (output_shifts != nullptr) ? output_shifts->info() : nullptr));
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.h b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.h
index 325f4e7067..068131f434 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.h
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -49,64 +49,47 @@ public:
     CLDepthwiseConvolutionLayerNativeKernel &operator=(CLDepthwiseConvolutionLayerNativeKernel &&) = default;
     /** Initialize the function's source, destination and parameters
      *
-     * @param[in]  input              Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/FP32/FP16. Data layout supported: NHWC
-     * @param[in]  weights            Weights tensor. A 3D tensor with dimensions [IFM, N, M].
-     *                                Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
-     * @param[in]  biases             Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
-     *                                Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
-     * @param[out] output             Destination tensor. Data type supported: Same as @p input.
-     * @param[in]  dwc_weights_info   Depthwise convolution layer weights info to retrieve the number of output elements processed by each thread
-     * @param[in]  dwc_info           Depthwise convolution layer info
-     * @param[in]  conv_info          Padding and stride information to use for the convolution.
-     * @param[in]  depth_multiplier   (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in]  dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     * @param[in]  output_multipliers (Optional) Output multipliers tensor for quantized computations. In case of per-channel quantization,
-     *                                the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
-     * @param[in]  output_shifts      (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization,
+     * @param[in]     compile_context    The compile context to be used.
+     * @param[in,out] input              Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/FP32/FP16. Data layout supported: NHWC
+     * @param[in]     weights            Weights tensor. A 3D tensor with dimensions [IFM, N, M].
+     *                                   Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
+     * @param[in]     biases             Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
+     *                                   Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
+     * @param[out]    output             Destination tensor. Pass in nullptr or @p input for in-place operation. Data type supported: Same as @p input.
+     * @param[in]     dwc_weights_info   Depthwise convolution layer weights info to retrieve the number of output elements processed by each thread
+     * @param[in]     dwc_info           Depthwise convolution layer info
+     * @param[in]     conv_info          Padding and stride information to use for the convolution.
+     * @param[in]     depth_multiplier   (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+     * @param[in]     dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in]     output_multipliers (Optional) Output multipliers tensor for quantized computations. In case of per-channel quantization,
+     *                                   the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
+     * @param[in]     output_shifts      (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization,
      *                                the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
+     *
+     * @note: In-place is only supported when
+     *          * data layout: NHWC
+     *          * filter: 1x1
+     *          * @p depth_multiplier: 1
+     *          * strides: 1
+     *          * dilation: 1
+     *          * no padding
+     *          * no change of data layout after configure
      */
-    void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const DWCWeightsKernelInfo &dwc_weights_info,
+    void configure(const CLCompileContext &compile_context, ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const DWCWeightsKernelInfo &dwc_weights_info,
                    const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, const Size2D &dilation = Size2D(1U, 1U),
                    const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr);
+
     /** Initialize the function's source, destination and parameters
      *
-     * @param[in]  compile_context    The compile context to be used.
-     * @param[in]  input              Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/FP32/FP16. Data layout supported: NHWC
-     * @param[in]  weights            Weights tensor. A 3D tensor with dimensions [IFM, N, M].
-     *                                Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
-     * @param[in]  biases             Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
-     *                                Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
-     * @param[out] output             Destination tensor. Data type supported: Same as @p input.
-     * @param[in]  dwc_weights_info   Depthwise convolution layer weights info to retrieve the number of output elements processed by each thread
-     * @param[in]  dwc_info           Depthwise convolution layer info
-     * @param[in]  conv_info          Padding and stride information to use for the convolution.
-     * @param[in]  depth_multiplier   (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in]  dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     * @param[in]  output_multipliers (Optional) Output multipliers tensor for quantized computations. In case of per-channel quantization,
-     *                                the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
-     * @param[in]  output_shifts      (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization,
-     *                                the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
+     * Similar to @ref CLDepthwiseConvolutionLayerNativeKernel::configure()
      */
-    void configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const DWCWeightsKernelInfo &dwc_weights_info,
+    void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const DWCWeightsKernelInfo &dwc_weights_info,
                    const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, const Size2D &dilation = Size2D(1U, 1U),
                    const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr);
+
     /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayerNativeKernel
      *
-     * @param[in] input              Source tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/FP32/FP16. Data layout supported: NHWC
-     * @param[in] weights            Weights tensor info. A 3D tensor with dimensions [IFM, N, M].
-     *                               Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
-     * @param[in] biases             Biases tensor info. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
-     *                               Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
-     * @param[in] output             Destination tensor info. Data type supported: Same as @p input.
-     * @param[in] dwc_weights_info   Depthwise convolution layer weights info to retrieve the number of output elements processed by each thread
-     * @param[in] dwc_info           Depthwise convolution layer info
-     * @param[in] conv_info          Padding and stride information to use for the convolution.
-     * @param[in] depth_multiplier   (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in] dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     * @param[in] output_multipliers (Optional) Output multipliers tensor for quantized computations. In case of per-channel quantization,
-     *                               the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
-     * @param[in] output_shifts      (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization,
-     *                               the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
+     * Similar to @ref CLDepthwiseConvolutionLayerNativeKernel::configure()
      *
      * @return a status
      */
@@ -118,14 +101,14 @@ public:
     void run(const Window &window, cl::CommandQueue &queue) override;
 
 private:
-    const ICLTensor *_input;
-    const ICLTensor *_weights;
-    const ICLTensor *_biases;
-    ICLTensor       *_output;
-    unsigned int     _depth_multiplier;
-    const ICLTensor *_output_multipliers;
-    const ICLTensor *_output_shifts;
-    bool             _is_quantized;
+    const ICLTensor *_input {};
+    const ICLTensor *_weights{};
+    const ICLTensor *_biases{};
+    ICLTensor       *_output{};
+    unsigned int     _depth_multiplier{ 0 };
+    const ICLTensor *_output_multipliers{};
+    const ICLTensor *_output_shifts{};
+    bool             _is_quantized{ false };
 };
 } // namespace arm_compute
 #endif /*ARM_COMPUTE_CLDEPTHWISECONVOLUTIONLAYERNATIVEKERNEL_H */