Enable FFT for FP16

Resolves: COMPMID-4051

Change-Id: I0c0bf97212dd281c19d5081e6247e7dc0c23cd6b
Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4687
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

5 files changed, 70 insertions, 62 deletions

diff --git a/arm_compute/runtime/CL/functions/CLFFT1D.h b/arm_compute/runtime/CL/functions/CLFFT1D.h
index e88ee7650d..731bad5c32 100644
--- a/arm_compute/runtime/CL/functions/CLFFT1D.h
+++ b/arm_compute/runtime/CL/functions/CLFFT1D.h
@@ -61,7 +61,7 @@ public:
     ~CLFFT1D();
     /** Initialise the function's source, destinations and border mode.
      *
-     * @param[in]  input  Source tensor. Data types supported: F32.
+     * @param[in]  input  Source tensor. Data types supported: F16/F32.
      * @param[out] output Destination tensor. Data types and data layouts supported: Same as @p input.
      * @param[in]  config FFT related configuration
      */
@@ -69,14 +69,14 @@ public:
     /** Initialise the function's source, destinations and border mode.
      *
      * @param[in]  compile_context The compile context to be used.
-     * @param[in]  input           Source tensor. Data types supported: F32.
+     * @param[in]  input           Source tensor. Data types supported: F16/F32.
      * @param[out] output          Destination tensor. Data types and data layouts supported: Same as @p input.
      * @param[in]  config          FFT related configuration
      */
     void configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const FFT1DInfo &config);
     /** Static function to check if given info will lead to a valid configuration of @ref CLFFT1D.
      *
-     * @param[in] input  Source tensor info. Data types supported: F32.
+     * @param[in] input  Source tensor info. Data types supported: F16/F32.
      * @param[in] output Destination tensor info. Data types and data layouts supported: Same as @p input.
      * @param[in] config FFT related configuration
      *
diff --git a/arm_compute/runtime/CL/functions/CLFFT2D.h b/arm_compute/runtime/CL/functions/CLFFT2D.h
index c54127f209..adc8e46cb2 100644
--- a/arm_compute/runtime/CL/functions/CLFFT2D.h
+++ b/arm_compute/runtime/CL/functions/CLFFT2D.h
@@ -58,7 +58,7 @@ public:
     ~CLFFT2D();
     /** Initialise the function's source, destinations and border mode.
      *
-     * @param[in]  input  Source tensor. Data types supported: F32.
+     * @param[in]  input  Source tensor. Data types supported: F16/F32.
      * @param[out] output Destination tensor. Data types and data layouts supported: Same as @p input.
      * @param[in]  config FFT related configuration
      */
@@ -66,14 +66,14 @@ public:
     /** Initialise the function's source, destinations and border mode.
      *
      * @param[in]  compile_context The compile context to be used.
-     * @param[in]  input           Source tensor. Data types supported: F32.
+     * @param[in]  input           Source tensor. Data types supported: F16/F32.
      * @param[out] output          Destination tensor. Data types and data layouts supported: Same as @p input.
      * @param[in]  config          FFT related configuration
      */
     void configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const FFT2DInfo &config);
     /** Static function to check if given info will lead to a valid configuration of @ref CLFFT2D.
      *
-     * @param[in] input  Source tensor info. Data types supported: F32.
+     * @param[in] input  Source tensor info. Data types supported: F16/F32.
      * @param[in] output Destination tensor info. Data types and data layouts supported: Same as @p input.
      * @param[in] config FFT related configuration
      *
diff --git a/arm_compute/runtime/CL/functions/CLFFTConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLFFTConvolutionLayer.h
index 53ce63333b..5085f5a66c 100644
--- a/arm_compute/runtime/CL/functions/CLFFTConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLFFTConvolutionLayer.h
@@ -73,53 +73,59 @@ public:
      *
      * @note: This function only works with any square kernel size and unit strides for both NCHW and NHWC data layout
      *
-     * @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:  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[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);
     /** Set the input and output tensors.
      *
      * @note: This function only works with any square kernel size and unit strides for both NCHW and NHWC data layout
      *
-     * @param[in]  compile_context The compile context to be used.
-     * @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]  compile_context  The compile context to be used.
+     * @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: 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[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(const CLCompileContext &compile_context, 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 CLFFTConvolutionLayer
      *
      * @note: This function only works with any square kernel size and unit strides for both NCHW and NHWC data layout
      *
-     * @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: 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[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/CL/functions/CLPixelWiseMultiplication.h b/arm_compute/runtime/CL/functions/CLPixelWiseMultiplication.h
index 6432cd040d..b48f6eb6cc 100644
--- a/arm_compute/runtime/CL/functions/CLPixelWiseMultiplication.h
+++ b/arm_compute/runtime/CL/functions/CLPixelWiseMultiplication.h
@@ -120,7 +120,7 @@ public:
     /** Initialise the kernel's inputs, output.
      *
      * @param[in]      compile_context The compile context to be used.
-     * @param[in, out] input1          An input tensor. Data types supported: F32. Number of channels supported: 2.
+     * @param[in, out] input1          An input tensor. Data types supported: F16/F32. Number of channels supported: 2.
      *                                 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. Number of channels 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.
@@ -130,7 +130,7 @@ public:
     void configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo());
     /** Static function to check if given info will lead to a valid configuration of @ref CLComplexPixelWiseMultiplication
      *
-     * @param[in] input1   An input tensor info. Data types supported: F32. Number of channels supported: 2.
+     * @param[in] input1   An input tensor info. Data types supported: F16/F32. Number of channels supported: 2.
      * @param[in] input2   An input tensor info. Data types supported: same as @p input1. Number of channels supported: same as @p input1.
      * @param[in] output   The output tensor info, Data types supported: same as @p input1. Number of channels supported: same as @p input1.
      * @param[in] act_info (Optional) Activation layer information in case of a fused activation.
@@ -277,7 +277,7 @@ public:
     CLComplexPixelWiseMultiplication &operator=(CLComplexPixelWiseMultiplication &&);
     /** Initialise the kernel's inputs, output.
      *
-     * @param[in, out] input1   An input tensor. Data types supported: F32. Number of channels supported: 2.
+     * @param[in, out] input1   An input tensor. Data types supported: F16/F32. Number of channels supported: 2.
      *                          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. Number of channels 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.
@@ -288,7 +288,7 @@ public:
     /** Initialise the kernel's inputs, output.
      *
      * @param[in]      compile_context The compile context to be used.
-     * @param[in, out] input1          An input tensor. Data types supported: F32. Number of channels supported: 2.
+     * @param[in, out] input1          An input tensor. Data types supported: F16/F32. Number of channels supported: 2.
      *                                 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. Number of channels 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.
@@ -298,7 +298,7 @@ public:
     void configure(const CLCompileContext &compile_context, ICLTensor *input1, ICLTensor *input2, ICLTensor *output, const ActivationLayerInfo &act_info = ActivationLayerInfo());
     /** Static function to check if given info will lead to a valid configuration of @ref CLComplexPixelWiseMultiplication
      *
-     * @param[in] input1   An input tensor info. Data types supported: F32. Number of channels supported: 2.
+     * @param[in] input1   An input tensor info. Data types supported: F16/F32. Number of channels supported: 2.
      * @param[in] input2   An input tensor info. Data types supported: same as @p input1. Number of channels supported: same as @p input1.
      * @param[in] output   The output tensor info, Data types supported: same as @p input1. Number of channels supported: same as @p input1.
      * @param[in] act_info (Optional) Activation layer information in case of a fused activation.
diff --git a/arm_compute/runtime/NEON/functions/NEFFTConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEFFTConvolutionLayer.h
index 37750e243b..b181e05c1a 100644
--- a/arm_compute/runtime/NEON/functions/NEFFTConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEFFTConvolutionLayer.h
@@ -75,36 +75,38 @@ public:
      *
      * @note: This function only works with any square kernel size and unit strides for both NCHW and NHWC data layout
      *
-     * @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. Unused for NEON backend.
      */
     void configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *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 NEFFTConvolutionLayer
      *
      * @note: This function only works with any square kernel size and unit strides for both NCHW and NHWC data layout
      *
-     * @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[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] 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[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] act_info         (Optional) Activation layer information in case of a fused activation.
+     * @param[in] enable_fast_math (Optional) Enable fast math computation. Unused for NEON backend.
      *
      * @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;