COMPMID-1026 - Add support for 4x4 output tile in CLWinogradConvolutionLayer
The performance achieved can be found at the following confluence page:
https://confluence.arm.com/display/MLENG/GEMM-based+convolution+vs+Winograd-based+convolution+on+OpenCL
Change-Id: I4b690cfdd4eb4ff0cd17b14fdd49ccaa1d1dc85c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/127729
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
diff --git a/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h b/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h
index 7115710..828e2e5 100644
--- a/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h
+++ b/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h
@@ -49,8 +49,7 @@
/** Set the input and output tensor.
*
* @note Winograd filter transform supports the following configurations:
- * Output tile size: 2x2, 4x4
- * Kernel size: 3x3
+ * F(output tile, kernel size):F(2x2, 3x3), F(4x4, 3x3), F(4x4, 5x5)
* Strides: only unit strides
*
* @param[in] input Source tensor. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout). Data types supported: F32.
@@ -61,8 +60,7 @@
/** Static function to check if given info will lead to a valid configuration of @ref CLWinogradFilterTransformKernel
*
* @note Winograd filter transform supports the following configurations:
- * Output tile size: 2x2, 4x4
- * Kernel size: 3x3
+ * F(output tile, kernel size):F(2x2, 3x3), F(4x4, 3x3), F(4x4, 5x5)
* Strides: only unit strides
*
* @param[in] input Source tensor. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout). Data types supported: F32.
diff --git a/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h b/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h
index 2d1eadf..b92ff2f 100644
--- a/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h
+++ b/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h
@@ -47,8 +47,7 @@
/** Set the input and output of the kernel.
*
* @note Winograd input transform supports the following configurations:
- * Output tile size: 2x2
- * Kernel size: 3x3
+ * F(output tile, kernel size):F(2x2, 3x3), F(4x4, 3x3), F(4x4, 5x5)
* Strides: only unit strides
*
* @param[in] input The input tensor to transform. Data types supported: F32
@@ -59,8 +58,7 @@
/** Static function to check if given info will lead to a valid configuration of @ref CLWinogradInputTransformKernel
*
* @note Winograd input transform supports the following configurations:
- * Output tile size: 2x2
- * Kernel size: 3x3
+ * F(output tile, kernel size):F(2x2, 3x3), F(4x4, 3x3), F(4x4, 5x5)
* Strides: only unit strides
*
* @param[in] input The input tensor to transform. Data types supported: F32
diff --git a/arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h b/arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h
index b0d0bbe..5e64a82 100644
--- a/arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h
+++ b/arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h
@@ -49,11 +49,10 @@
/** Set the input and output tensor.
*
* @note Winograd output transform supports the following configurations:
- * Output tile size: 2x2
- * Kernel size: 3x3
+ * F(output tile, kernel size):F(2x2, 3x3), F(4x4, 3x3), F(4x4, 5x5)
* Strides: only unit strides
*
- * @param[in] input Source tensor with shape [C, N, 16, batches]. Data types supported: F32.
+ * @param[in] input Source tensor with shape [C, N, K, batches]. Data types supported: F32.
* @param[in] bias Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. It can be a nullptr. Data type supported: as @p input
* @param[out] output The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_output_transform_shape. Data types supported: Same as @p input
* @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo
@@ -62,11 +61,10 @@
/** Static function to check if given info will lead to a valid configuration of @ref CLWinogradOutputTransformKernel
*
* @note Winograd output transform supports the following configurations:
- * Output tile size: 2x2
- * Kernel size: 3x3
+ * F(output tile, kernel size):F(2x2, 3x3), F(4x4, 3x3), F(4x4, 5x5)
* Strides: only unit strides
*
- * @param[in] input Source tensor with shape [C, N, 16, batches]. Data types supported: F32.
+ * @param[in] input Source tensor with shape [C, N, K, batches]. Data types supported: F32.
* @param[in] bias Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. It can be a nullptr. Data type supported: as @p input
* @param[out] output The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_output_transform_shape. Data types supported: Same as @p input
* @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo
diff --git a/arm_compute/runtime/CL/functions/CLConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLConvolutionLayer.h
index 7df7659..a1cd155 100644
--- a/arm_compute/runtime/CL/functions/CLConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLConvolutionLayer.h
@@ -26,6 +26,7 @@
#include "arm_compute/runtime/CL/functions/CLDirectConvolutionLayer.h"
#include "arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h"
+#include "arm_compute/runtime/CL/functions/CLWinogradConvolutionLayer.h"
#include "arm_compute/runtime/IFunction.h"
#include "arm_compute/runtime/IMemoryManager.h"
@@ -84,7 +85,6 @@
* 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.
@@ -95,7 +95,7 @@
*
* @return a status
*/
- static ConvolutionMethod get_convolution_method(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+ 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));
// Inherited methods overridden:
void run() override;
diff --git a/arm_compute/runtime/CL/functions/CLWinogradInputTransform.h b/arm_compute/runtime/CL/functions/CLWinogradInputTransform.h
index 0e0d6bf..64d3e80 100644
--- a/arm_compute/runtime/CL/functions/CLWinogradInputTransform.h
+++ b/arm_compute/runtime/CL/functions/CLWinogradInputTransform.h
@@ -40,8 +40,7 @@
/** Set the input and output tensors.
*
* @note Winograd input transform supports the following configurations:
- * Output tile size: 2x2
- * Kernel size: 3x3
+ * F(output tile, kernel size):F(2x2, 3x3), F(4x4, 3x3), F(4x4, 5x5)
* Strides: only unit strides
*
* @param[in] input The input tensor to transform. Data types supported: F32
@@ -52,8 +51,7 @@
/** Static function to check if given info will lead to a valid configuration of @ref CLWinogradInputTransform.
*
* @note Winograd input transform supports the following configurations:
- * Output tile size: 2x2
- * Kernel size: 3x3
+ * F(output tile, kernel size):F(2x2, 3x3), F(4x4, 3x3), F(4x4, 5x5)
* Strides: only unit strides
*
* @param[in] input The input tensor to transform. Data types supported: F32