commit 3267092403cf0984b1f5980ff553130be81e3e2c
author Frank Barchard <fbarchard@google.com> Sat Nov 30 21:58:51 2019 -0800
committer XNNPACK Team <xnnpack-github-robot@google.com> Sat Nov 30 21:59:25 2019 -0800
tree a3007e8df529f2cd11b18f62be58b948274c9510
parent 72d6afbfc871abcaf83c13291d1a0d0e8a2b4e4c

Enable AARCH32 4x8 GEMM kernel

Improves performance on Exynos by using VLDM.
Comparable performance on other CPUs but consistent
performance independent of compiler.

PiperOrigin-RevId: 283189038

src/init.c

diff --git a/src/init.c b/src/init.c
index e7ce5b3..60190bc 100644
--- a/src/init.c
+++ b/src/init.c
@@ -131,14 +131,25 @@
 
   /**************************** F32 micro-kernels ****************************/
   #ifndef XNN_NO_F32_OPERATORS
-    xnn_params.f32.gemm = (struct gemm_parameters) {
-      .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x8__neon_lane_ld128,
-      .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x8__neon_lane_ld128,
-      .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__neon_lane_ld64,
-      .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__neon_lane_ld64,
-      .mr = 4,
-      .nr = 8,
-    };
+    #if XNN_ENABLE_ASSEMBLY
+      xnn_params.f32.gemm = (struct gemm_parameters) {
+        .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x8__aarch32_neon_ld64,
+        .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x8__neon_lane_ld128,
+        .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__neon_lane_ld64,
+        .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__neon_lane_ld64,
+        .mr = 4,
+        .nr = 8,
+      };
+    #else  // XNN_ENABLE_ASSEMBLY
+      xnn_params.f32.gemm = (struct gemm_parameters) {
+        .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x8__neon_lane_ld128,
+        .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x8__neon_lane_ld128,
+        .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__neon_lane_ld64,
+        .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__neon_lane_ld64,
+        .mr = 4,
+        .nr = 8,
+      };
+    #endif  // XNN_ENABLE_ASSEMBLY
     xnn_params.f32.gemm2 = (struct gemm_parameters) {
       .gemm = NULL,
       .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x2__neon_lane_ld64,
@@ -390,7 +401,7 @@
         .mr = 6,
         .nr = 8,
       };
-    #endif
+    #endif  // XNN_ENABLE_ASSEMBLY
 
     xnn_params.f32.gemm2 = (struct gemm_parameters) {
       .gemm = NULL,