COMPMID-2466: Improved ConvLayer documentation.
Added two tables in NEConvolutionLayer and CLConvolutionLayer
showing ACL's criteria to choose the different algorithms.
Change-Id: Ie6e44e72381ab8232c050688c86d746eb8d2fb36
Signed-off-by: Pablo Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1577
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
diff --git a/arm_compute/runtime/CL/functions/CLConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLConvolutionLayer.h
index 96f69f2..04ce1cf 100644
--- a/arm_compute/runtime/CL/functions/CLConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLConvolutionLayer.h
@@ -41,6 +41,32 @@
* -# @ref CLWinogradConvolutionLayer
* -# @ref CLDirectConvolutionLayer
* -# @ref CLFFTConvolutionLayer
+ *
+ * The function selects one of the algorithms mentioned above based on:
+ * - The size of the kernel
+ * - Number of input/output feature maps
+ * - Amount of memory needed
+ *
+ * Generally GEMM-based convolution is executed when neither Winograd nor FFT nor Direct convolution can be performed.
+ *
+ * FP32 Algorithm| Filter Size | Input/Output feature maps |
+ * --------------|-------------------------------------------------------------|-------------------------------------------|
+ * Winograd | 3x3 1x3 3x1 5x1 1x5 5x5(fast maths) 7x1 1x7 | Input channels is greater than 3 |
+ * FFT | Squared kernels and greater than 9x9 | Input feature maps > Output feature maps |
+ * DirectConv | 9x9 | |
+ * GEMM | Any size | |
+ *
+ * Winograd 5x5 requires fast maths enabled.
+ *
+ * FP16 Algorithm| Filter Size | Input/Output feature maps |
+ * --------------|----------------------------|-------------------------------------------|
+ * Winograd | 3x3 1x3 3x1 5x1 1x5 5x5 | Input channels is greater than 3 |
+ * FFT | Not supported | |
+ * DirectConv | 9x9 | |
+ * GEMM | Any size | |
+ *
+ * Winograd FP16 requires fast maths enabled.
+ *
*/
class CLConvolutionLayer : public IFunction
{
diff --git a/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h
index 3c1d20a..4310ab4 100644
--- a/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h
@@ -43,6 +43,32 @@
* -# @ref NEWinogradConvolutionLayer (executed only in case Winograd is required for the operation)
* -# @ref NEDirectConvolutionLayer (executed only in case Direct Convolution is required for the operation)
* -# @ref NEFFTConvolutionLayer (executed only in case FFT is required for the operation)
+ *
+ *
+ * The function selects one of the algorithms mentioned above based on:
+ * - The size of the kernel
+ * - Number of input/output feature maps
+ * - Amount of memory needed
+ *
+ * Generally GEMM-based convolution is executed when neither Winograd nor FFT nor Direct convolution can be performed.
+ *
+ * FP32 Algorithm| Filter Size | Input/Output feature maps |
+ * --------------|----------------------------------------------------|-------------------------------------------|
+ * Winograd | 3x3 1x3 3x1 5x1 1x5 5x5(fast maths) 7x1 1x7 | Input channels is greater than 3 |
+ * FFT | Squared kernels and greater than 9x9 | Input feature maps > Output feature maps |
+ * DirectConv | 9x9 | |
+ * GEMM | Any size | |
+ *
+ * Winograd 5x5 requires fast maths enabled.
+ *
+ * FP16 Algorithm| Filter Size |
+ * --------------|------------------|
+ * Winograd | Not supported |
+ * FFT | Not supported |
+ * DirectConv | 9x9 |
+ * GEMM | Any size |
+ *
+ *
*/
class NEConvolutionLayer : public IFunction
{