vbinary microkernels unrolled to x8 for scalar and web assembly and x16 web assembly simd
PiperOrigin-RevId: 329407843
diff --git a/src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c b/src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c
new file mode 100644
index 0000000..d086deb
--- /dev/null
+++ b/src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c
@@ -0,0 +1,77 @@
+// Auto-generated file. Do not edit!
+// Template: src/f32-vbinary/vopc-wasmsimd.c.in
+// Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <wasm_simd128.h>
+
+#include <xnnpack/common.h>
+#include <xnnpack/vbinary.h>
+
+
+void xnn_f32_vmaxc_ukernel__wasmsimd_arm_x16(
+ size_t n,
+ const float* a,
+ const float* b,
+ float* y,
+ const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
+{
+ assert(n != 0);
+ assert(n % sizeof(float) == 0);
+ assert(a != NULL);
+ assert(b != NULL);
+ assert(y != NULL);
+
+ const v128_t vb = wasm_v32x4_load_splat(b);
+ for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
+ const v128_t va0123 = wasm_v128_load(a);
+ const v128_t va4567 = wasm_v128_load(a + 4);
+ const v128_t va89AB = wasm_v128_load(a + 8);
+ const v128_t vaCDEF = wasm_v128_load(a + 12);
+ a += 16;
+
+ v128_t vy0123 = wasm_f32x4_max(va0123, vb);
+ v128_t vy4567 = wasm_f32x4_max(va4567, vb);
+ v128_t vy89AB = wasm_f32x4_max(va89AB, vb);
+ v128_t vyCDEF = wasm_f32x4_max(vaCDEF, vb);
+
+
+
+ wasm_v128_store(y, vy0123);
+ wasm_v128_store(y + 4, vy4567);
+ wasm_v128_store(y + 8, vy89AB);
+ wasm_v128_store(y + 12, vyCDEF);
+ y += 16;
+ }
+ for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
+ const v128_t va = wasm_v128_load(a);
+ a += 4;
+
+ v128_t vy = wasm_f32x4_max(va, vb);
+
+
+ wasm_v128_store(y, vy);
+ y += 4;
+ }
+ if XNN_UNLIKELY(n != 0) {
+ const v128_t va = wasm_v128_load(a);
+
+ v128_t vy = wasm_f32x4_max(va, vb);
+
+
+ if (n & (2 * sizeof(float))) {
+ *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
+ vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
+ y += 2;
+ }
+ if (n & (1 * sizeof(float))) {
+ *y = wasm_f32x4_extract_lane(vy, 0);
+ }
+ }
+}