commit 3522c0a6133873558e1509ad514fbb6826fd2039
author Frank Barchard <fbarchard@google.com> Wed Mar 10 15:51:30 2021 -0800
committer XNNPACK Team <xnnpack-github-robot@google.com> Wed Mar 10 15:52:13 2021 -0800
tree 0e27816ffcb103f9bbac2fad0d9eaf3ab256fbe4
parent 671d1b00378384bec26e2a73bd269322478a2ff2

Enable QS8 4x16c4-aarch64-neondot-ld64 IGEMM microkernel

IGEMM use assembly for 4x16

PiperOrigin-RevId: 362160266

src/init.c

diff --git a/src/init.c b/src/init.c
index 47e60f1..9fe2e45 100644
--- a/src/init.c
+++ b/src/init.c
@@ -810,7 +810,7 @@
         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_4x16c4__aarch64_neondot_ld64);
           xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x16c4__aarch64_neondot_ld64);
-          xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot);
+          xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_4x16c4__aarch64_neondot_ld64);
           xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot);
           xnn_params.qs8.gemm.mr = 4;
           xnn_params.qs8.gemm.nr = 16;
@@ -853,7 +853,7 @@
               break;
             default:
               xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_4x16c4__aarch64_neondot_ld64);
-              xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot);
+              xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_4x16c4__aarch64_neondot_ld64);
               break;
           }
           xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x16c4__aarch64_neondot_ld64);