commit fb0ab0b87874452edebd5cbaf9491a879b9ecb0b
author Frank Barchard <fbarchard@google.com> Wed Mar 03 07:54:34 2021 -0800
committer XNNPACK Team <xnnpack-github-robot@google.com> Wed Mar 03 07:55:20 2021 -0800
tree c7930d7c9dba89b73e0eb31013fcd597f38704b5
parent da78da1388909044ff76dc2b81ab14884e121c5e

QS8 enable 4x8c4__neondot for ARM32

PiperOrigin-RevId: 360669439

src/init.c

diff --git a/src/init.c b/src/init.c
index 3939a10..8f5d21d 100644
--- a/src/init.c
+++ b/src/init.c
@@ -104,13 +104,23 @@
     #ifndef XNN_NO_QS8_OPERATORS
       init_flags |= XNN_INIT_FLAG_QS8;
 
-      xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup);
-      xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup);
-      xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x8c2__neon_mlal_padal_dup);
-      xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x8c2__neon_mlal_padal_dup);
-      xnn_params.qs8.gemm.mr = 2;
-      xnn_params.qs8.gemm.nr = 8;
-      xnn_params.qs8.gemm.log2_kr = 1;
+      if (cpuinfo_has_arm_neon_dot()) {
+        xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot);
+        xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x8c4__neondot);
+        xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot);
+        xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x8c4__neondot);
+        xnn_params.qs8.gemm.mr = 4;
+        xnn_params.qs8.gemm.nr = 8;
+        xnn_params.qs8.gemm.log2_kr = 2;
+      } else {
+        xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup);
+        xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup);
+        xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x8c2__neon_mlal_padal_dup);
+        xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x8c2__neon_mlal_padal_dup);
+        xnn_params.qs8.gemm.mr = 2;
+        xnn_params.qs8.gemm.nr = 8;
+        xnn_params.qs8.gemm.log2_kr = 1;
+      }
 
       xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_ukernel_up8x9__neon_mul16;
       xnn_params.qs8.dwconv[0].channel_tile = 8;