COMPMID-1277 - Optimizing CLIm2ColKernel for NHWC.
This patch includes:
- Im2Col optimizations for NHWC using a new data layout
- Refactoring of CLIm2ColKernel adding validation method and auto-init
- Removed im2col_reduced from CLIm2ColKernel and created a new kernel CLFlattenLayerKernel
Change-Id: I1620640b6796baa268324b33ae92cdd8de53e27c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/141241
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
diff --git a/arm_compute/runtime/CL/functions/CLFlattenLayer.h b/arm_compute/runtime/CL/functions/CLFlattenLayer.h
index 88df4a7..ebc0e5e 100644
--- a/arm_compute/runtime/CL/functions/CLFlattenLayer.h
+++ b/arm_compute/runtime/CL/functions/CLFlattenLayer.h
@@ -33,7 +33,7 @@
/** Basic function to execute flatten. This function calls the following OpenCL kernel:
*
-* -# @ref CLIm2ColKernel
+* -# @ref CLFlattenLayerKernel
*
*/
class CLFlattenLayer : public ICLSimpleFunction
@@ -41,11 +41,22 @@
public:
/** Initialise the kernel's input and output.
*
- * @param[in] input First input tensor to flatten with at least 3 dimensions. The dimensions over the third will be interpreted as batches. Data types supported: F16/F32
+ * @param[in] input First input tensor to flatten with at least 3 dimensions.
+ * The dimensions above the third will be interpreted as batches. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
* @param[out] output Output tensor with shape [w*h*d, input_batches] where:
- * w = width input tensor, h = height input tensor and d = depth input tensor. Data type supported: same as @p input
+ * w = width input tensor, h = height input tensor and d = depth input tensor. Data type supported: same as @p input
*/
void configure(const ICLTensor *input, ICLTensor *output);
+ /** Static function to check if given info will lead to a valid configuration of @ref CLTranspose
+ *
+ * @param[in] input First input tensor to flatten with at least 3 dimensions.
+ * The dimensions above the third will be interpreted as batches. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[out] output Output tensor with shape [w*h*d, input_batches] where:
+ * w = width input tensor, h = height input tensor and d = depth input tensor. Data type supported: same as @p input
+ *
+ * @return a status
+ */
+ static Status validate(const ITensorInfo *input, const ITensorInfo *output);
};
} // namespace arm_compute
diff --git a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
index e8fe8e4..450cd83 100644
--- a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
+++ b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
@@ -27,11 +27,11 @@
#include "arm_compute/runtime/CL/ICLSimpleFunction.h"
#include "arm_compute/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.h"
-#include "arm_compute/core/CL/kernels/CLIm2ColKernel.h"
#include "arm_compute/core/CL/kernels/CLTransposeKernel.h"
#include "arm_compute/runtime/CL/CLMemoryGroup.h"
#include "arm_compute/runtime/CL/CLTensor.h"
#include "arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h"
+#include "arm_compute/runtime/CL/functions/CLFlattenLayer.h"
#include "arm_compute/runtime/CL/functions/CLGEMM.h"
#include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h"
#include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h"
@@ -130,14 +130,14 @@
void configure_mm(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output);
CLMemoryGroup _memory_group;
- CLIm2ColKernel _im2col_kernel;
CLConvertFullyConnectedWeights _convert_weights;
+ CLFlattenLayer _flatten_layer;
CLFullyConnectedLayerReshapeWeights _reshape_weights_kernel;
CLGEMM _mm_gemm;
CLGEMMLowpMatrixMultiplyCore _mm_gemmlowp;
CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage;
CLGEMMMatrixAccumulateBiasesKernel _accumulate_biases_kernel;
- CLTensor _im2col_output;
+ CLTensor _flatten_output;
CLTensor _gemmlowp_output;
CLTensor _converted_weights_output;
CLTensor _reshape_weights_output;