commit 91cd2b784615dc3440a0b81bf56b11513cb62d5c
author Marat Dukhan <maratek@google.com> Thu Apr 09 23:57:31 2020 -0700
committer XNNPACK Team <xnnpack-github-robot@google.com> Thu Apr 09 23:58:03 2020 -0700
tree aaa8f9261dfb6e1960a1a83e332b1c8eb41a9880
parent 355ab43fd556d3f6d6e2e2f6573952a33916a5a0

Rename binary operation micro-kernels

PiperOrigin-RevId: 305834937

src/f32-vbinary/gen/vmul-minmax-neon-x8.c

diff --git a/src/f32-vbinary/gen/vmul-minmax-neon-x8.c b/src/f32-vbinary/gen/vmul-minmax-neon-x8.c
new file mode 100644
index 0000000..a3d310b
--- /dev/null
+++ b/src/f32-vbinary/gen/vmul-minmax-neon-x8.c
@@ -0,0 +1,75 @@
+// Auto-generated file. Do not edit!
+//   Template: src/f32-vbinary/vop-neon.c.in
+//   Generator: tools/xngen
+//
+// Copyright 2019 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 <arm_neon.h>
+
+#include <xnnpack/common.h>
+#include <xnnpack/vbinary.h>
+
+
+void xnn_f32_vmul_minmax_ukernel__neon_x8(
+    size_t n,
+    const float* a,
+    const float* b,
+    float* y,
+    const union xnn_f32_minmax_params params[restrict static 1])
+{
+  assert(n != 0);
+  assert(n % sizeof(float) == 0);
+
+  const float32x4_t vy_min = vld1q_dup_f32(&params->scalar.min);
+  const float32x4_t vy_max = vld1q_dup_f32(&params->scalar.max);
+
+  for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
+    const float32x4_t va0123 = vld1q_f32(a); a += 4;
+    const float32x4_t vb0123 = vld1q_f32(b); b += 4;
+    const float32x4_t va4567 = vld1q_f32(a); a += 4;
+    const float32x4_t vb4567 = vld1q_f32(b); b += 4;
+
+    float32x4_t vy0123 = vmulq_f32(va0123, vb0123);
+    float32x4_t vy4567 = vmulq_f32(va4567, vb4567);
+
+    vy0123 = vmaxq_f32(vy0123, vy_min);
+    vy4567 = vmaxq_f32(vy4567, vy_min);
+
+    vy0123 = vminq_f32(vy0123, vy_max);
+    vy4567 = vminq_f32(vy4567, vy_max);
+
+    vst1q_f32(y, vy0123); y += 4;
+    vst1q_f32(y, vy4567); y += 4;
+  }
+  for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
+    const float32x4_t va0123 = vld1q_f32(a); a += 4;
+    const float32x4_t vb0123 = vld1q_f32(b); b += 4;
+
+    float32x4_t vy0123 = vmulq_f32(va0123, vb0123);
+    vy0123 = vmaxq_f32(vy0123, vy_min);
+    vy0123 = vminq_f32(vy0123, vy_max);
+    vst1q_f32(y, vy0123); y += 4;
+  }
+  if XNN_UNLIKELY(n != 0) {
+    const float32x4_t va0123 = vld1q_f32(a);
+    const float32x4_t vb0123 = vld1q_f32(b);
+
+    float32x4_t vy0123 = vmulq_f32(va0123, vb0123);
+    vy0123 = vmaxq_f32(vy0123, vy_min);
+    vy0123 = vminq_f32(vy0123, vy_max);
+
+    float32x2_t vy01 = vget_low_f32(vy0123);
+    if (n & (2 * sizeof(float))) {
+      vst1_f32(y, vy01); y += 2;
+      vy01 = vget_high_f32(vy0123);
+    }
+    if (n & (1 * sizeof(float))) {
+      vst1_lane_f32(y, vy01, 0);
+    }
+  }
+}