commit 0e97d6fae0a2c18e775b612f97e69a0e6c6b2c23
author Marat Dukhan <maratek@google.com> Fri Jun 26 19:35:09 2020 -0700
committer XNNPACK Team <xnnpack-github-robot@google.com> Fri Jun 26 19:35:46 2020 -0700
tree 8fbe31695c8014e907055fa37cfd3fc47930a402
parent 6b73c4fab7d12650fe365b1638abfd7ef4396ae2

Vector Elementwise Binary with Scalar microkernels in WAsm SIMD

PiperOrigin-RevId: 318585022

src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c

diff --git a/src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c b/src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c
new file mode 100644
index 0000000..37eb288
--- /dev/null
+++ b/src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c
@@ -0,0 +1,63 @@
+// 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_vdivc_minmax_ukernel__wasmsimd_arm_x4(
+    size_t n,
+    const float* a,
+    const float* b,
+    float* y,
+    const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
+{
+  assert(n != 0);
+  assert(n % sizeof(float) == 0);
+
+  const v128_t vy_min = wasm_v32x4_load_splat(&params->scalar.min);
+  const v128_t vy_max = wasm_v32x4_load_splat(&params->scalar.max);
+
+  const v128_t vb = wasm_v32x4_load_splat(b);
+  for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
+    const v128_t va0123 = wasm_v128_load(a);
+    a += 4;
+
+    v128_t vy0123 = wasm_f32x4_div(va0123, vb);
+
+
+    vy0123 = wasm_f32x4_max(vy0123, vy_min);
+
+    vy0123 = wasm_f32x4_min(vy0123, vy_max);
+
+    wasm_v128_store(y, vy0123);
+    y += 4;
+  }
+  if XNN_UNLIKELY(n != 0) {
+    const v128_t va = wasm_v128_load(a);
+
+    v128_t vy = wasm_f32x4_div(va, vb);
+
+    vy = wasm_f32x4_max(vy, vy_min);
+    vy = wasm_f32x4_min(vy, vy_max);
+
+    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);
+    }
+  }
+}