Add enable_fast_math for NEDeconvolutionLayer

Resolves: [ONCPUML-1128]

Signed-off-by: Annop Wongwathanarat <annop.wongwathanarat@arm.com>
Change-Id: I287a71222d3f0289d8cccfcb15383b0a930a55e6
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8952
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 29 insertions, 20 deletions

diff --git a/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h
index 4ee2d10a22..15124d6041 100644
--- a/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2021 Arm Limited.
+ * Copyright (c) 2017-2021, 2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -103,25 +103,34 @@ public:
      * |QASYMM8_SIGNED |QASYMM8_SIGNED     |S32    |QASYMM8_SIGNED |
      * |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32    |QASYMM8_SIGNED |
      *
-     * @param[in,out] input   Input tensor. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs. Data types supported: F32/F16/QASYMM8/QASYMM8_SIGNED.
-     * @param[in]     weights The 4d weights with dimensions [width, height, IFM, OFM]. Data type supported: Same as @p input, also could be QSYMM8_PER_CHANNEL if input is QASYMM8/QASYMM8_SIGNED.
-     * @param[in]     bias    Optional, ignored if NULL. The biases have one dimension. Data type supported: Data types supported: S32 for QASYMM8/QASYMM8_SIGNED input, F32 for F32 input, F16 for F16 input.
-     * @param[out]    output  Output tensor. The output has the same number of dimensions as the @p input.
-     * @param[in]     info    Contains padding and policies to be used in the deconvolution, this is decribed in @ref PadStrideInfo.
+     * @param[in,out] input            Input tensor. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs.
+     *                                 Data types supported: F32/F16/QASYMM8/QASYMM8_SIGNED.
+     * @param[in]     weights          The 4d weights with dimensions [width, height, IFM, OFM].
+     *                                 Data type supported: Same as @p input, also could be QSYMM8_PER_CHANNEL if input is QASYMM8/QASYMM8_SIGNED.
+     * @param[in]     bias             Optional, ignored if NULL. The biases have one dimension.
+     *                                 Data type supported: Data types supported: S32 for QASYMM8/QASYMM8_SIGNED input, F32 for F32 input, F16 for F16 input.
+     * @param[out]    output           Output tensor. The output has the same number of dimensions as the @p input.
+     * @param[in]     info             Contains padding and policies to be used in the deconvolution, this is decribed in @ref PadStrideInfo.
+     * @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(ITensor *input, const ITensor *weights, const ITensor *bias, ITensor *output, const PadStrideInfo &info);
+    void configure(ITensor *input, const ITensor *weights, const ITensor *bias, ITensor *output, const PadStrideInfo &info, bool enable_fast_math = false);
     /** Static function to check if given info will lead to a valid configuration of @ref NEDeconvolutionLayer
      *
-     * @param[in] input   Input tensor info. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs. Data types supported: F32/F16/QASYMM8/QASYMM8_SIGNED.
-     * @param[in] weights The 4d weights info with dimensions [width, height, IFM, OFM]. Data type supported: Same as @p input, also could be QSYMM8_PER_CHANNEL if input is QASYMM8/QASYMM8_SIGNED.
-     * @param[in] bias    (Optional) The biases have one dimension. Data type supported: Data types supported: S32 for QASYMM8/QASYMM8_SIGNED input, F32 for F32 input, F16 for F16 input.
-     * @param[in] output  Output tensor info. The output has the same number of dimensions as the @p input.
-     * @param[in] info    Contains padding and policies to be used in the deconvolution, this is decribed in @ref PadStrideInfo.
+     * @param[in] input            Input tensor info. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs.
+     *                             Data types supported: F32/F16/QASYMM8/QASYMM8_SIGNED.
+     * @param[in] weights          The 4d weights info with dimensions [width, height, IFM, OFM].
+     *                             Data type supported: Same as @p input, also could be QSYMM8_PER_CHANNEL if input is QASYMM8/QASYMM8_SIGNED.
+     * @param[in] bias             (Optional) The biases have one dimension. Data type supported: Data types supported: S32 for QASYMM8/QASYMM8_SIGNED input, F32 for F32 input, F16 for F16 input.
+     * @param[in] output           Output tensor info. The output has the same number of dimensions as the @p input.
+     * @param[in] info             Contains padding and policies to be used in the deconvolution, this is decribed in @ref PadStrideInfo.
+     * @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 *bias, const ITensorInfo *output, const PadStrideInfo &info);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, const ITensorInfo *output, const PadStrideInfo &info, bool enable_fast_math = false);
 
     // Inherited methods overridden:
     void run() override;
diff --git a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
index e04c6467eb..1a75c14896 100644
--- a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2021 Arm Limited.
+ * Copyright (c) 2017-2021, 2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -81,7 +81,7 @@ NEDeconvolutionLayer::NEDeconvolutionLayer(std::shared_ptr<IMemoryManager> memor
 {
 }
 
-Status NEDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, const ITensorInfo *output, const PadStrideInfo &info)
+Status NEDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, const ITensorInfo *output, const PadStrideInfo &info, bool enable_fast_math)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
@@ -148,17 +148,17 @@ Status NEDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInf
     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(batches_idx) != scale_out_info.dimension(batches_idx));
     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(channel_idx) != scale_out_info.dimension(channel_idx));
 
-    ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayer::validate(&scale_out_info, weights, bias, output, conv_info, WeightsInfo()));
+    ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayer::validate(&scale_out_info, weights, bias, output, conv_info, WeightsInfo(), Size2D(1U, 1U), ActivationLayerInfo(), enable_fast_math));
 
     return Status{};
 }
 
-void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, const ITensor *bias, ITensor *output, const PadStrideInfo &info)
+void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, const ITensor *bias, ITensor *output, const PadStrideInfo &info, bool enable_fast_math)
 {
     // Perform validation step
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
-    ARM_COMPUTE_ERROR_THROW_ON(NEDeconvolutionLayer::validate(input->info(), weights->info(), (bias == nullptr) ? nullptr : bias->info(), output->info(), info));
-    ARM_COMPUTE_LOG_PARAMS(input, weights, bias, output, info);
+    ARM_COMPUTE_ERROR_THROW_ON(NEDeconvolutionLayer::validate(input->info(), weights->info(), (bias == nullptr) ? nullptr : bias->info(), output->info(), info, enable_fast_math));
+    ARM_COMPUTE_LOG_PARAMS(input, weights, bias, output, info, enable_fast_math);
 
     const DataLayout   data_layout = input->info()->data_layout();
     const unsigned int width_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
@@ -202,7 +202,7 @@ void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, con
 
     _upsample_f.configure(input, &_scaled_output, upsample_info);
 
-    _conv_f.configure(&_scaled_output, &_weights_flipped, bias, output, conv_info);
+    _conv_f.configure(&_scaled_output, &_weights_flipped, bias, output, conv_info, WeightsInfo(), Size2D(1U, 1U), ActivationLayerInfo(), enable_fast_math);
 
     // Setup flip axis data
     _flip_axis.allocator()->allocate();