COMPMID-2766: Add support for QASYMM8_SIGNED in NEDepthwiseConvolutionLayer
This patch also adds support for QASYMM8_SIGNED in the generic functions
that uses NEDepthwiseConvolutionLayerNativeKernel.
Change-Id: I74a99e1476cb1ebd2078e076ab2bea703949527b
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2552
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
diff --git a/arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.h b/arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.h
index 682c3c9..1303cf9 100644
--- a/arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.h
+++ b/arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.h
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2019 ARM Limited.
+ * Copyright (c) 2019-2020 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -25,6 +25,7 @@
#define ARM_COMPUTE_NEDEPTHWISECONVOLUTIONLAYERNATIVEKERNEL_H
#include "arm_compute/core/NEON/INEKernel.h"
+#include "arm_compute/core/utils/misc/Requires.h"
#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
#include <arm_neon.h>
@@ -57,11 +58,11 @@
*
* @note Supported data layouts: NHWC
*
- * @param[in] input Source tensor. DataType supported: QASYMM8/F16/F32.
+ * @param[in] input Source tensor. DataType supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
* @param[in] weights Weights tensor. This is a 3D tensor with dimensions [IFM, W, H].
- * Data type supported: Same as @p input or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
+ * Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8/QASYMM8_SIGNED.
* @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.
+ * 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] 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.
@@ -74,11 +75,11 @@
*
* @note Supported data layouts: NHWC
*
- * @param[in] input Source tensor info. DataType supported: QASYMM8/F16/F32.
+ * @param[in] input Source tensor info. DataType supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
* @param[in] weights Weights tensor info. This is a 3D tensor with dimensions [IFM, W, H].
- * Data type supported: Same as @p input or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
+ * Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8/QASYMM8_SIGNED.
* @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.
+ * 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] 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.
@@ -102,7 +103,7 @@
int >::type = 0 >
void run_depthwise(const Window &window);
- template <typename T, typename TW, int S, bool has_biases, bool is_per_channel, typename std::enable_if<std::is_same<T, uint8_t>::value, int>::type = 0>
+ template < typename T, typename TW, int S, bool has_biases, bool is_per_channel, REQUIRES_TA(std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value) >
void run_depthwise(const Window &window);
/** Common signature for all the specialised depthwise convolution native functions
diff --git a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
index 750dab2..ccab671 100644
--- a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -54,12 +54,12 @@
NEDepthwiseConvolutionLayer &operator=(NEDepthwiseConvolutionLayer &&) = default;
/** Initialize the function's source, destination, weights and convolution information.
*
- * @param[in, out] input Source tensor. Data type supported: QASYMM8/F16/F32
+ * @param[in, out] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32
* @param[out] output Destination tensor. Data type supported: same as @p input.
* @param[in] weights Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM].
- * Data type supported: Same as @p input or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
+ * Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8/QASYMM8_SIGNED.
* @param[in] biases Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
- * Data type supported: Same as @p input, S32 when input is QASYMM8.
+ * Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
* @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] act_info (Optional) Activation layer information in case of a fused activation.
@@ -70,12 +70,12 @@
/** Static function to check if given info will lead to a valid configuration of @ref NEDepthwiseConvolutionLayer
*
- * @param[in] input Source tensor. Data type supported: QASYMM8/F16/F32
+ * @param[in] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32
* @param[in] output Destination tensor. Data type supported: same as @p input.
* @param[in] weights Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM].
- * Data type supported: Same as @p input or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
+ * Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8/QASYMM8_SIGNED.
* @param[in] biases Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
- * Data type supported: Same as @p input, S32 when input is QASYMM8.
+ * Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
* @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] act_info (Optional) Activation layer information in case of a fused activation.
@@ -93,15 +93,15 @@
private:
/** Static function to choose the best depthwise convolution function for @ref NEDepthwiseConvolutionLayer
*
- * @param[in] input Source tensor info. Data type supported: QASYMM8/F16/F32
+ * @param[in] input Source tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32
* @param[in] weights Weights tensor info. These are 3D tensors with shape [kernel_x, kernel_y, IFM].
- * Data type supported: Same as @p input or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
+ * Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8/QASYMM8_SIGNED.
* @param[in] biases Biases tensor info. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
- * Data type supported: Same as @p input, S32 when input is QASYMM8.
+ * Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
* @param[in] output Destination tensor. Data type supported: same as @p input.
* @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] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for 3x3 QASYMM8 supported.
+ * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for 3x3 quantized are supported.
* @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
*
* @return a Depthwise Convolution Function
@@ -136,10 +136,10 @@
NEDepthwiseConvolutionLayerOptimizedInternal &operator=(NEDepthwiseConvolutionLayerOptimizedInternal &&) = default;
/** Initialize the function's source, destination, kernels and border_size.
*
- * @param[in, out] input Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
+ * @param[in, out] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32. (Written to only for border filling).
* @param[in] weights Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
* @param[in] biases Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
- * Data type supported: Same as @p input, S32 when input is QASYMM8.
+ * 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] 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.
@@ -151,10 +151,10 @@
/** Static function to check if given info will lead to a valid configuration of @ref NEDepthwiseConvolutionLayer3x3
*
- * @param[in] input Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
+ * @param[in] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32. (Written to only for border filling).
* @param[in] weights Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
* @param[in] biases Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
- * Data type supported: Same as @p input, S32 when input is QASYMM8.
+ * Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
* @param[in] output Destination tensor. Data type supported: same as @p input.
* @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.
@@ -173,10 +173,10 @@
private:
/** Configure the kernels/functions for the generic pipeline.
*
- * @param[in, out] input Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
+ * @param[in, out] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32. (Written to only for border filling).
* @param[in] weights Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
* @param[in] biases Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
- * Data type supported: Same as @p input, S32 when input is QASYMM8.
+ * 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] conv_info Padding and stride information to use for the convolution.
* @param[in] depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
@@ -188,10 +188,10 @@
unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation = Size2D(1U, 1U));
/** Configure the kernels/functions for the optimized pipeline.
*
- * @param[in] input Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
+ * @param[in] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32. (Written to only for border filling).
* @param[in] weights Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
* @param[in] biases Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
- * Data type supported: Same as @p input, S32 when input is QASYMM8.
+ * 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] conv_info Padding and stride information to use for the convolution.
* @param[in] depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
@@ -247,12 +247,12 @@
NEDepthwiseConvolutionLayerGeneric &operator=(NEDepthwiseConvolutionLayerGeneric &&) = default;
/** Initialize the function's source, destination, weights and convolution information.
*
- * @param[in, out] input Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
+ * @param[in, out] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32. (Written to only for border filling).
* @param[out] output Destination tensor. Data type supported: same as @p input.
* @param[in] weights Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM].
- * Data type supported: Same as @p input or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
+ * Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8/QASYMM8_SIGNED.
* @param[in] biases Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
- * Data type supported: Same as @p input, S32 when input is QASYMM8.
+ * Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
* @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] act_info (Optional) Activation layer information in case of a fused activation.
@@ -263,12 +263,12 @@
/** Static function to check if given info will lead to a valid configuration of @ref NEDepthwiseConvolutionLayerGeneric
*
- * @param[in] input Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
+ * @param[in] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32. (Written to only for border filling).
* @param[in] output Destination tensor. Data type supported: same as @p input.
* @param[in] weights Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM].
- * Data type supported: Same as @p input or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
+ * Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8/QASYMM8_SIGNED.
* @param[in] biases Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
- * Data type supported: Same as @p input, S32 when input is QASYMM8.
+ * Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
* @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] act_info (Optional) Activation layer information in case of a fused activation.
@@ -330,10 +330,10 @@
NEDepthwiseConvolutionLayerOptimized &operator=(NEDepthwiseConvolutionLayerOptimized &&) = default;
/** Initialize the function's source, destination, kernels and border_size.
*
- * @param[in, out] input Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
+ * @param[in, out] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32. (Written to only for border filling).
* @param[in] weights Weights tensor. These are 3D tensors with shape [W, H, IFM]. Data type supported: Same as @p input.
* @param[in] biases Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
- * Data type supported: Same as @p input, S32 when input is QASYMM8.
+ * 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] 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.
@@ -346,10 +346,10 @@
/** Static function to check if given info will lead to a valid configuration of @ref NEDepthwiseConvolutionLayerOptimized
*
- * @param[in] input Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
+ * @param[in] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32. (Written to only for border filling).
* @param[in] weights Weights tensor. These are 3D tensors with shape [W, H, IFM]. Data type supported: Same as @p input.
* @param[in] biases Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
- * Data type supported: Same as @p input, S32 when input is QASYMM8.
+ * Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
* @param[in] output Destination tensor. Data type supported: same as @p input.
* @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.