COMPMID-1096 - Add fast_math flag to CLConvolutionLayer

COMPMID-1103 - CLWinogradConvolutionLayer mismatches

Change-Id: Iceaa9482a1790ec39d2720c220261aaea8043978
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/129398
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

6 files changed, 81 insertions, 68 deletions

diff --git a/arm_compute/core/NEON/kernels/NEReshapeLayerKernel.h b/arm_compute/core/NEON/kernels/NEReshapeLayerKernel.h
index cd70198d6c..0a3fc44881 100644
--- a/arm_compute/core/NEON/kernels/NEReshapeLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEReshapeLayerKernel.h
@@ -40,7 +40,7 @@ public:
     }
     /** Set the input and output of the kernel
      *
-     * @param[in]  input  Source tensor. Data type supported: U8/S8/QS8/U16/S16/QS16/U32/S32/F16/F32
+     * @param[in]  input  Source tensor. Data type supported: U8/S8/QS8/U16/S16/QS16/QASYMM8/U32/S32/F16/F32
      * @param[out] output Destination tensor. Data type supported: Same as @p input
      */
     void configure(const ITensor *input, ITensor *output);
diff --git a/arm_compute/runtime/CL/functions/CLConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLConvolutionLayer.h
index a1cd15515f..5c05334a56 100644
--- a/arm_compute/runtime/CL/functions/CLConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLConvolutionLayer.h
@@ -37,6 +37,7 @@ namespace arm_compute
 /** Basic function to compute the convolution layer. This function calls the following OpenCL kernels/functions:
  *
  * -# @ref CLGEMMConvolutionLayer
+ * -# @ref CLWinogradConvolutionLayer
  * -# @ref CLDirectConvolutionLayer
  */
 class CLConvolutionLayer : public IFunction
@@ -46,57 +47,63 @@ public:
     CLConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
     /** Set the input and output tensors.
      *
-     * @param[in]  input        Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
-     *                          while every optional dimension from 4 and above represent a batch of inputs.
-     *                          Data types supported: QS8/QASYMM8/QS16/F16/F32.
-     * @param[in]  weights      Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input.
-     * @param[in]  biases       Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
-     *                          Data type supported: Should match @p input data type, except for input of QASYMM8 type where biases should be of S32 type.
-     * @param[out] output       Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
-     *                          Data types supported: Same as @p input.
-     * @param[in]  conv_info    Contains padding and stride information described in @ref PadStrideInfo.
-     * @param[in]  weights_info Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. Data type supported: Same as @p input.
-     * @param[in]  dilation     (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     * @param[in]  act_info     (Optional) Activation layer information in case of a fused activation.
+     * @param[in]  input            Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
+     *                              while every optional dimension from 4 and above represent a batch of inputs.
+     *                              Data types supported: QS8/QASYMM8/QS16/F16/F32.
+     * @param[in]  weights          Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input.
+     * @param[in]  biases           Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
+     *                              Data type supported: Should match @p input data type, except for input of QASYMM8 type where biases should be of S32 type.
+     * @param[out] output           Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
+     *                              Data types supported: Same as @p input.
+     * @param[in]  conv_info        Contains padding and stride information described in @ref PadStrideInfo.
+     * @param[in]  weights_info     Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. Data type supported: Same as @p input.
+     * @param[in]  dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in]  act_info         (Optional) Activation layer information in case of a fused activation.
+     * @param[in]  enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation
+     *                                available which may introduce a drop of accuracy as well. Default is false
      */
     void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo(),
-                   const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo());
+                   const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false);
     /** Static function to check if given info will lead to a valid configuration of @ref CLConvolutionLayer
      *
-     * @param[in] input        Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
-     *                         while every optional dimension from 4 and above represent a batch of inputs.
-     *                         Data types supported: QS8/QASYMM8/QS16/F16/F32.
-     * @param[in] weights      Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input.
-     * @param[in] biases       Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported:Same as @p input.
-     * @param[in] output       Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
-     *                         Data types supported: Same as @p input.
-     * @param[in] conv_info    Contains padding and stride information described in @ref PadStrideInfo.
-     * @param[in] weights_info Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. Data type supported: Same as @p input.
-     * @param[in] dilation     (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     * @param[in] act_info     (Optional) Activation layer information in case of a fused activation.
+     * @param[in] input            Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
+     *                             while every optional dimension from 4 and above represent a batch of inputs.
+     *                             Data types supported: QS8/QASYMM8/QS16/F16/F32.
+     * @param[in] weights          Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input.
+     * @param[in] biases           Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported:Same as @p input.
+     * @param[in] output           Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
+     *                             Data types supported: Same as @p input.
+     * @param[in] conv_info        Contains padding and stride information described in @ref PadStrideInfo.
+     * @param[in] weights_info     Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. Data type supported: Same as @p input.
+     * @param[in] dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in] act_info         (Optional) Activation layer information in case of a fused activation.
+     * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation
+     *                             available which may introduce a drop of accuracy as well. Default is false
      *
      * @return a status
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
-                           const WeightsInfo &weights_info = WeightsInfo(), const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo());
+                           const WeightsInfo &weights_info = WeightsInfo(), const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false);
     /** Static function to check if given info will return the convolution called by @ref CLConvolutionLayer
      *
-     * @param[in] input        Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
-     *                         while every optional dimension from 4 and above represent a batch of inputs.
-     *                         Data types supported: QS8/QASYMM8/QS16/F16/F32.
-     * @param[in] weights      Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input.
-     * @param[in] output       Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
-     *                         Data types supported: Same as @p input.
-     * @param[in] conv_info    Contains padding and stride information described in @ref PadStrideInfo.
-     * @param[in] weights_info Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. Data type supported: Same as @p input.
-     * @param[in] act_info     (Optional) Activation layer information in case of a fused activation.
-     * @param[in] gpu_target   Specifies the @p GPUTarget.
-     * @param[in] dilation     (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in] input            Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
+     *                             while every optional dimension from 4 and above represent a batch of inputs.
+     *                             Data types supported: QS8/QASYMM8/QS16/F16/F32.
+     * @param[in] weights          Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input.
+     * @param[in] output           Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
+     *                             Data types supported: Same as @p input.
+     * @param[in] conv_info        Contains padding and stride information described in @ref PadStrideInfo.
+     * @param[in] weights_info     Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. Data type supported: Same as @p input.
+     * @param[in] act_info         (Optional) Activation layer information in case of a fused activation.
+     * @param[in] gpu_target       Specifies the @p GPUTarget.
+     * @param[in] dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation
+     *                             available which may introduce a drop of accuracy as well. Default is false
      *
      * @return a status
      */
     static ConvolutionMethod get_convolution_method(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, const PadStrideInfo &conv_info,
-                                                    const WeightsInfo &weights_info, const ActivationLayerInfo &act_info, const GPUTarget gpu_target, const Size2D &dilation = Size2D(1U, 1U));
+                                                    const WeightsInfo &weights_info, const ActivationLayerInfo &act_info, const GPUTarget gpu_target, const Size2D &dilation = Size2D(1U, 1U), bool enable_fast_math = false);
     // Inherited methods overridden:
     void run() override;
 
diff --git a/arm_compute/runtime/CL/functions/CLWinogradConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLWinogradConvolutionLayer.h
index 2cf1f77fb4..a27976959c 100644
--- a/arm_compute/runtime/CL/functions/CLWinogradConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLWinogradConvolutionLayer.h
@@ -51,38 +51,44 @@ public:
     CLWinogradConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
     /** Set the input and output tensors.
      *
-     * @note: This function only works with 3x3 kernels and unit strides
+     * @note: This function only works with 3x3 and 5x5 kernels along with unit strides
+     * @note  Some Winograd configurations (i.e. F(4x4, 3x3) and F(4x4, 5x5)) are supported only with enable_fast_math = true
      *
-     * @param[in]  input     Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
-     *                       while every optional dimension from 4 and above represent a batch of inputs.
-     *                       Data types supported: F32.
-     * @param[in]  weights   Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input.
-     * @param[in]  biases    Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input
-     * @param[out] output    Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
-     *                       Data types supported: Same as @p input.
-     * @param[in]  conv_info Contains padding and stride information described in @ref PadStrideInfo.
-     * @param[in]  act_info  (Optional) Activation layer information in case of a fused activation.
+     * @param[in]  input            Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
+     *                              while every optional dimension from 4 and above represent a batch of inputs.
+     *                              Data types supported: F32.
+     * @param[in]  weights          Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input.
+     * @param[in]  biases           Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input
+     * @param[out] output           Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
+     *                              Data types supported: Same as @p input.
+     * @param[in]  conv_info        Contains padding and stride information described in @ref PadStrideInfo.
+     * @param[in]  act_info         (Optional) Activation layer information in case of a fused activation.
+     * @param[in]  enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation
+     *                              available which may introduce a drop of accuracy as well. Default is false
      */
     void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
-                   const ActivationLayerInfo &act_info = ActivationLayerInfo());
+                   const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false);
     /** Static function to check if given info will lead to a valid configuration of @ref CLWinogradConvolutionLayer
      *
-     * @note: This function only works with 3x3 kernels and unit strides
+     * @note: This function only works with 3x3 and 5x5 kernels along with unit strides
+     * @note  Some Winograd configurations (i.e. F(4x4, 3x3) and F(4x4, 5x5)) are supported only with enable_fast_math = true
      *
-     * @param[in]  input     Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
-     *                       while every optional dimension from 4 and above represent a batch of inputs.
-     *                       Data types supported: F32.
-     * @param[in]  weights   Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input.
-     * @param[in]  biases    Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input
-     * @param[out] output    Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
-     *                       Data types supported: Same as @p input.
-     * @param[in]  conv_info Contains padding and stride information described in @ref PadStrideInfo.
-     * @param[in]  act_info  (Optional) Activation layer information in case of a fused activation.
+     * @param[in]  input            Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
+     *                              while every optional dimension from 4 and above represent a batch of inputs.
+     *                              Data types supported: F32.
+     * @param[in]  weights          Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input.
+     * @param[in]  biases           Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input
+     * @param[out] output           Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
+     *                              Data types supported: Same as @p input.
+     * @param[in]  conv_info        Contains padding and stride information described in @ref PadStrideInfo.
+     * @param[in]  act_info         (Optional) Activation layer information in case of a fused activation.
+     * @param[in]  enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation
+     *                              available which may introduce a drop of accuracy as well. Default is false
      *
      * @return a status
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
-                           const ActivationLayerInfo &act_info = ActivationLayerInfo());
+                           const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false);
 
     // Inherited methods overridden:
     void run() override;
diff --git a/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h
index ce9a3ed4f2..b82ba89f7c 100644
--- a/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h
@@ -52,7 +52,7 @@ public:
      *
      * @param[in]  input        Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
      *                          while every optional dimension from 4 and above represent a batch of inputs.
-     *                          Data types supported: QS8/QASYMM8/QS16/F32.
+     *                          Data types supported: QS8/QASYMM8/QS16/F16/F32.
      * @param[in]  weights      Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input.
      * @param[in]  biases       Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
      *                          Data type supported: Should match @p input data type, except for input of QASYMM8 type where biases should be of S32 type.
diff --git a/arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h b/arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h
index 7d5e397e80..ba96ae6cfa 100644
--- a/arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h
+++ b/arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h
@@ -37,11 +37,11 @@ class NEPixelWiseMultiplication : public INESimpleFunction
 public:
     /** Initialise the kernel's inputs, output and convertion policy.
      *
-     * @param[in, out] input1          An input tensor. Data types supported: U8/QS8/S16/F32.
+     * @param[in, out] input1          An input tensor. Data types supported: U8/QS8/S16/F16/F32.
      *                                 The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0.
      * @param[in, out] input2          An input tensor. Data types supported: same as @p input1.
      *                                 The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0.
-     * @param[out]     output          Output tensor. Data types supported: U8/QS8/S16/F32.
+     * @param[out]     output          Output tensor. Data types supported: U8/QS8/S16/F16/F32.
      * @param[in]      scale           Scale to apply after multiplication.
      *                                 Scale must be positive and its value must be either 1/255 or 1/2^n where n is between 0 and 15. For QS8 and QS16 scale must be 1.
      * @param[in]      overflow_policy Overflow policy.
@@ -50,9 +50,9 @@ public:
     void configure(ITensor *input1, ITensor *input2, ITensor *output, float scale, ConvertPolicy overflow_policy, RoundingPolicy rounding_policy);
     /** Static function to check if given info will lead to a valid configuration of @ref NEPixelWiseMultiplication
      *
-     * @param[in] input1          First tensor info input. Data types supported: U8/QS8/S16/F32.
-     * @param[in] input2          Second tensor info input. Data types supported: U8/QS8/S16/F32.
-     * @param[in] output          Output tensor info. Data types supported: U8/QS8/S16/F32.
+     * @param[in] input1          First tensor info input. Data types supported: U8/QS8/S16/F16/F32.
+     * @param[in] input2          Second tensor info input. Data types supported: U8/QS8/S16/F16/F32.
+     * @param[in] output          Output tensor info. Data types supported: U8/QS8/S16/F16/F32.
      * @param[in] scale           Scale to apply after multiplication. Must be positive.
      * @param[in] overflow_policy Overflow policy.
      * @param[in] rounding_policy Rounding policy.
diff --git a/arm_compute/runtime/NEON/functions/NEReshapeLayer.h b/arm_compute/runtime/NEON/functions/NEReshapeLayer.h
index 369f50e147..0bab534ebc 100644
--- a/arm_compute/runtime/NEON/functions/NEReshapeLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEReshapeLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -37,7 +37,7 @@ class NEReshapeLayer : public INESimpleFunction
 public:
     /** Initialise the kernel's inputs and outputs
      *
-     * @param[in]  input  First tensor input. Data type supported: U8/S8/QS8/U16/S16/QS16/U32/S32/F16/F32
+     * @param[in]  input  First tensor input. Data type supported: U8/S8/QS8/QASYMM8//U16/S16/QS16/U32/S32/F16/F32
      * @param[out] output Output tensor. Data type supported: Same as @p input
      */
     void configure(const ITensor *input, ITensor *output);