diff --git a/doc/src/Getting_Started_FusionAPI.md b/doc/src/Getting_Started_FusionAPI.md index d7419db01f..e67728c9ce 100644 --- a/doc/src/Getting_Started_FusionAPI.md +++ b/doc/src/Getting_Started_FusionAPI.md @@ -244,7 +244,11 @@ The table below outlines the supported fusions as well as any applicable constra ## Performance Comparison to Non-Fused Kernels -The following graphs depicts the speedup gained for a fused Convolution+Bias+Activation over a non-fused version, all configurations have a batch size of 64: +The following graph depicts the speedup gained for a fused Convolution+Bias+Activation over a non-fused version, all configurations have a batch size of 64: ![CBA Graph](cba.png) +Speedup obtained by fusing Batchnorm (spatial mode) with Activation are presented in the graph below: + +![Batchnorm activation fusion](na.png) + diff --git a/doc/src/na.png b/doc/src/na.png new file mode 100644 index 0000000000..b36fe04f16 Binary files /dev/null and b/doc/src/na.png differ