COMPMID-3108: Add Winograd 3x3,4x4 FP16 support for NEON
Change-Id: I20680dc74a3d709297539e2132417308a7aecc9d
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3159
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/arm_compute/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h b/arm_compute/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h
index e2e8331..1740df0 100644
--- a/arm_compute/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -31,10 +31,10 @@
namespace arm_compute
{
+// Forward declarations
class ITensor;
/** Interface for the NEON kernel to perform Winograd input transform. */
-template <typename T>
class INEWinogradLayerTransformInputKernel : public INEKernel
{
public:
@@ -97,7 +97,7 @@
/** NEON kernel to perform Winograd input transform. */
template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
-class NEWinogradLayerTransformInputKernel : public INEWinogradLayerTransformInputKernel<T>
+class NEWinogradLayerTransformInputKernel : public INEWinogradLayerTransformInputKernel
{
public:
/** Prevent instances of this class from being copied (As this class contains pointers) */
@@ -168,7 +168,7 @@
/** Configure the output transform kernel.
*
- * @param[in] input_nhwc Input tensor. Data types supported: F32. Layout supported NHWC.
+ * @param[in] input_nhwc Input tensor. Data types supported: F16/F32. Layout supported NHWC.
* @param[in] num_batches Number of batches in input tensor.
* @param[in] num_rows Number of rows in input tensor.
* @param[in] num_cols Number of columns in input tensor.
@@ -199,7 +199,7 @@
/** Static function to check if given info will lead to a valid configuration of @ref NEWinogradLayerTransformInputKernel
*
- * @param[in] input First tensor input info. Data types supported: F32.
+ * @param[in] input First tensor input info. Data types supported: F16/F32.
* @param[in] output Output tensor info. Data types supported: same as @p input.
* @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo
*
@@ -227,7 +227,6 @@
};
/** Interface for the NEON kernel to perform Winograd output transform. */
-template <typename T>
class INEWinogradLayerTransformOutputKernel : public INEKernel
{
public:
@@ -312,7 +311,7 @@
/** NEON kernel to perform Winograd output transform. */
template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
-class NEWinogradLayerTransformOutputKernel : public INEWinogradLayerTransformOutputKernel<T>
+class NEWinogradLayerTransformOutputKernel : public INEWinogradLayerTransformOutputKernel
{
public:
const char *name() const override
@@ -410,7 +409,7 @@
/** Static function to check if given info will lead to a valid configuration of @ref NEWinogradLayerTransformOutputKernel
*
- * @param[in] input Source tensor info with shape [C, N, 16, batches] or [C, N, 36, batches]. Data types supported: F32.
+ * @param[in] input Source tensor info with shape [C, N, 16, batches] or [C, N, 36, batches]. Data types supported: F16/F32.
* @param[in] bias Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. It can be a nullptr. Data type supported: as @p input
* @param[in] output Destination tensor info with shape [output_convolved_dims.width, output_convolved_dims.height, C, batches]. Data type supported: same as @p input
* @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo
@@ -438,7 +437,6 @@
};
/** Interface for the NEON kernel to perform Winograd weights transform. */
-template <typename T>
class INEWinogradLayerTransformWeightsKernel : public INEKernel
{
public:
@@ -488,7 +486,7 @@
/** Static function to check if given info will lead to a valid configuration of @ref NEWinogradLayerTransformWeightsKernel
*
- * @param[in] input First tensor input info. Data types supported: F32.
+ * @param[in] input First tensor input info. Data types supported: F16/F32.
* @param[in] weights Weights tensor info. Data types supported: same as @p input.
*
* @return a status
@@ -498,7 +496,7 @@
/** NEON kernel to perform Winograd weights transform. */
template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
-class NEWinogradLayerTransformWeightsKernel final : public INEWinogradLayerTransformWeightsKernel<T>
+class NEWinogradLayerTransformWeightsKernel final : public INEWinogradLayerTransformWeightsKernel
{
public:
/** Prevent instances of this class from being copied (As this class contains pointers) */
@@ -522,7 +520,7 @@
/** Static function to check if given info will lead to a valid configuration of @ref NEWinogradLayerTransformWeightsKernel
*
* @param[in] input Source tensor info. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout).
- * kernel_x must be 3 and equal to kernel_y. Data types supported: F32.
+ * kernel_x must be 3 and equal to kernel_y. Data types supported: F16/F32.
* @param[in] output Destination tensor info. The output is a 3D tensor with dimensions [OFM, IFM, 16] or [OFM, IFM, 36]. Data type supported: same as @p input
* @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo
*
diff --git a/arm_compute/core/NEON/kernels/convolution/common/utils.hpp b/arm_compute/core/NEON/kernels/convolution/common/utils.hpp
index 25bfa33..99b2282 100644
--- a/arm_compute/core/NEON/kernels/convolution/common/utils.hpp
+++ b/arm_compute/core/NEON/kernels/convolution/common/utils.hpp
@@ -24,6 +24,8 @@
#pragma once
+#include <limits>
+
void PrintMatrix(const float *const m, const int M, const int N, const int row_stride);
constexpr inline int iceildiv(const int a, const int b)
@@ -36,3 +38,23 @@
{
return b * iceildiv(a, b);
}
+
+template<typename T>
+struct TypeBounds
+{
+ static constexpr T lower() noexcept { return std::numeric_limits<T>::has_infinity
+ ? -std::numeric_limits<T>::infinity()
+ : std::numeric_limits<T>::lowest(); };
+ static constexpr T upper() noexcept { return std::numeric_limits<T>::has_infinity
+ ? std::numeric_limits<T>::infinity()
+ : std::numeric_limits<T>::max(); };
+};
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+template<>
+struct TypeBounds<__fp16>
+{
+ static constexpr __fp16 lower() noexcept { return -std::numeric_limits<float>::infinity(); };
+ static constexpr __fp16 upper() noexcept { return std::numeric_limits<float>::infinity(); }
+};
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */