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>
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
index 2ebb80b..fe866dd 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
@@ -76,46 +76,51 @@
* |QASYMM8_SIGNED |QASYMM8_SIGNED |S32 |QASYMM8_SIGNED |
* |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32 |QASYMM8_SIGNED |
*
- * @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: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
- * @param[in] weights Weights tensor. 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. 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] 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 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] 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: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
+ * @param[in] weights Weights tensor. 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. 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] 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 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 ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *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 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 of @ref NEGEMMConvolutionLayer
*
- * @param[in] input 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[in] output 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] input 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[in] output 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
*
* @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(), 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() override;