Expose fast_math mode for GEMM through BFloat16

Resolves: COMPMID-4641

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I7ccc663b2692d40c370794caa906b5be8fd25a32
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5977
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 43 insertions, 36 deletions

diff --git a/src/runtime/cpu/operators/CpuGemmConvolution.h b/src/runtime/cpu/operators/CpuGemmConvolution.h
index 8b41cb4a91..578586e7d1 100644
--- a/src/runtime/cpu/operators/CpuGemmConvolution.h
+++ b/src/runtime/cpu/operators/CpuGemmConvolution.h
@@ -87,24 +87,26 @@ public:
      * |QASYMM8_SIGNED |QASYMM8_SIGNED     |S32      |QASYMM8_SIGNED |
      * |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32      |QASYMM8_SIGNED |
      *
-     * @param[in]  src          Source tensor info. 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: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
-     * @param[in]  weights      Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
-     *                          Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
-     * @param[in]  biases       Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
-     *                          Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
-     * @param[out] dst          Destination tensor info. 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 NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights
-     *                          tensor has also been transposed with cpu::kernels::CpuGemmTranspose1xWKernel. 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. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported.
-     * @param[in]  num_groups   (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is not supported
+     * @param[in]  src              Source tensor info. 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: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
+     * @param[in]  weights          Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
+     *                              Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
+     * @param[in]  biases           Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
+     *                              Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
+     * @param[out] dst              Destination tensor info. 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 NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights
+     *                              tensor has also been transposed with cpu::kernels::CpuGemmTranspose1xWKernel. 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. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported.
+     * @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
+     * @param[in]  num_groups       (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is not supported
      */
     void configure(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo(),
-                   const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(), unsigned int num_groups = 1);
+                   const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false, unsigned int num_groups = 1);
     /** Static function to check if given info will lead to a valid configuration
      *
      * Similar to CpuGemmConvolution::configure()
@@ -112,7 +114,8 @@ public:
      * @return a status
      */
     static Status validate(const ITensorInfo *src, 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(), unsigned int num_groups = 1);
+                           const WeightsInfo &weights_info = WeightsInfo(), const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(),
+                           bool enable_fast_math = false, unsigned int num_groups = 1);
 
     // Inherited methods overridden:
     void run(ITensorPack &tensors) override;
@@ -122,33 +125,37 @@ public:
 private:
     /** Configures the appropriate matrix multiply routine
      *
-     * @param[in]  src           Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
-     * @param[in]  weights       Weights tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
-     * @param[in]  biases        Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
-     *                           Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
-     * @param[out] dst           Output tensor info. Data types supported: Same as @p input,
-     *                           except for input of QASYMM8/QASYMM8_SIGNED type where output should be of S32 type.
-     * @param[in]  act_info      (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported.
-     * @param[in]  gemm_3d_depth (Optional) Depth of GEMM 3D (Defaults to 1)
+     * @param[in]  src              Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
+     * @param[in]  weights          Weights tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
+     * @param[in]  biases           Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
+     *                              Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
+     * @param[out] dst              Output tensor info. Data types supported: Same as @p input,
+     *                              except for input of QASYMM8/QASYMM8_SIGNED type where output should be of S32 type.
+     * @param[in]  act_info         (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported.
+     * @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
+     * @param[in]  gemm_3d_depth    (Optional) Depth of GEMM 3D (Defaults to 1)
      */
     void configure_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo(),
-                      int gemm_3d_depth = 1);
+                      bool enable_fast_math = false, int gemm_3d_depth = 1);
     /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMConvolutionLayer matrix multiply routines
      *
-     * @param[in] src           Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
-     * @param[in] weights       Weights tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
-     * @param[in] biases        Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
-     *                          Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
-     * @param[in] dst           Output tensor info. Data types supported: Same as @p input,
-     *                          except for input of QASYMM8/QASYMM8_SIGNED type where output should be of S32 type.
-     * @param[in] act_info      (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported.
-     * @param[in] gemm_3d_depth (Optional) Depth of GEMM 3D (Defaults to 1)
-     * @param[in] skip_im2col   (Optional) Flag which specifies if im2col has to be skipped. i.e. 1x1 convolution with NHWC data layout. (Default to false)
+     * @param[in] src              Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
+     * @param[in] weights          Weights tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
+     * @param[in] biases           Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
+     *                             Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
+     * @param[in] dst              Output tensor info. Data types supported: Same as @p input,
+     *                             except for input of QASYMM8/QASYMM8_SIGNED type where output should be of S32 type.
+     * @param[in] act_info         (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported.
+     * @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
+     * @param[in] gemm_3d_depth    (Optional) Depth of GEMM 3D (Defaults to 1)
+     * @param[in] skip_im2col      (Optional) Flag which specifies if im2col has to be skipped. i.e. 1x1 convolution with NHWC data layout. (Default to false)
      *
      * @return a status
      */
     static Status validate_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const ActivationLayerInfo &act_info = ActivationLayerInfo(),
-                              int gemm_3d_depth = 1, bool skip_im2col = false);
+                              bool enable_fast_math = false, int gemm_3d_depth = 1, bool skip_im2col = false);
     /** Static function to check if GEMM3D is supported in @ref NEGEMM or in @ref CpuGemmMLowpMatrixMultiplyCore
      *
      * @param[in] src           Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.