From 9167c9cd1c684218f76a3c0ec97574dd6f381b98 Mon Sep 17 00:00:00 2001 From: Gunes Bayir Date: Wed, 6 Mar 2024 09:58:40 +0000 Subject: Prefer indirect Gemm vs. Direct convolution if supported Indirect GEMM uses optimized assembly path while Direct Conv uses the fallback Acl kernel for convolution. In certain cases, where the input tensor is large and filter size is greater than 7 (e.g. 9x9 filters), heuristics fall back to Direct Conv algorithm where it could still prefer the assembly path if the data layout is NHWC. This is more important when SME2 kernels are present. Resolves: COMPMID-6900 Change-Id: Ia611c975eee0423615113fcaeaa8f9eef0421456 Signed-off-by: Gunes Bayir Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/11254 Benchmark: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Anitha Raj Comments-Addressed: Arm Jenkins --- docs/user_guide/release_version_and_change_log.dox | 1 + 1 file changed, 1 insertion(+) (limited to 'docs/user_guide') diff --git a/docs/user_guide/release_version_and_change_log.dox b/docs/user_guide/release_version_and_change_log.dox index b788957dda..21a5a368ad 100644 --- a/docs/user_guide/release_version_and_change_log.dox +++ b/docs/user_guide/release_version_and_change_log.dox @@ -43,6 +43,7 @@ If there is more than one release in a month then an extra sequential number is v24.04 Public major release - Optimize start-up time of @ref NEConvolutionLayer for some input configurations where GeMM is selected as the convolution algorithm + - Optimize @ref NEConvolutionLayer for input tensor size > 1e7 bytes and weight tensor height > 7 v24.02 Public major release - Replace template writer with compute kernel writer in dynamic fusion. -- cgit v1.2.1