commit 4810905317c1138b1daae52f3261d7fee464d99a
author Marat Dukhan <maratek@google.com> Tue Aug 31 17:31:57 2021 -0700
committer XNNPACK Team <xnnpack-github-robot@google.com> Tue Aug 31 17:33:43 2021 -0700
tree e9320e86b3a262db2c919a5f534cf88ec586db37
parent 8dc106e400ab4bf7519c7fe66e7977a83fe0f093

Leverage v128.const WAsm SIMD instruction

Warning: this change makes XNNPACK binaries for WebAssembly SIMD incompatible with Chrome versions earlier than 87
PiperOrigin-RevId: 394121736

src/f32-igemm/wasmsimd-splat.c.in

diff --git a/src/f32-igemm/wasmsimd-splat.c.in b/src/f32-igemm/wasmsimd-splat.c.in
index c36ed05..c5fd4f6 100644
--- a/src/f32-igemm/wasmsimd-splat.c.in
+++ b/src/f32-igemm/wasmsimd-splat.c.in
@@ -140,7 +140,7 @@
           $for M in range(MR):
             vacc${M}x${ABC[N:N+4]} = wasm_f32x4_min(vacc${M}x${ABC[N:N+4]}, vmax);
     $elif ACTIVATION == "RELU":
-      const v128_t vzero = wasm_f64x2_splat(0.0);
+      const v128_t vzero = wasm_i32x4_const_splat(0);
       $for N in range(0, NR, 4):
         $for M in range(MR):
           vacc${M}x${ABC[N:N+4]} = wasm_i32x4_max(vacc${M}x${ABC[N:N+4]}, vzero);