VSCALE microkernel and SoftMax Three-Pass algorithm with Reloading
- VSCALE micro-kernel in AVX and AVX512F implementations
- Unit tests
- Micro-benchmark for SoftArgMax using Three-Pass algorithm with Reloading
PiperOrigin-RevId: 275939577
diff --git a/src/f32-vscale/avx-unroll32.c b/src/f32-vscale/avx-unroll32.c
new file mode 100644
index 0000000..50c6b99
--- /dev/null
+++ b/src/f32-vscale/avx-unroll32.c
@@ -0,0 +1,64 @@
+// 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 <immintrin.h>
+
+#include <xnnpack/common.h>
+#include <xnnpack/vscale.h>
+
+
+void xnn_f32_vscale_ukernel__avx_unroll32(
+ size_t n,
+ const float* x,
+ float* y,
+ float c)
+{
+ assert(n != 0);
+ assert(n % sizeof(float) == 0);
+
+ __m256 vc = _mm256_set1_ps(c);
+ for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
+ const __m256 vx0 = _mm256_loadu_ps(x);
+ const __m256 vx1 = _mm256_loadu_ps(x + 8);
+ const __m256 vx2 = _mm256_loadu_ps(x + 16);
+ const __m256 vx3 = _mm256_loadu_ps(x + 24);
+ x += 32;
+
+ const __m256 vy0 = _mm256_mul_ps(vx0, vc);
+ const __m256 vy1 = _mm256_mul_ps(vx1, vc);
+ const __m256 vy2 = _mm256_mul_ps(vx2, vc);
+ const __m256 vy3 = _mm256_mul_ps(vx3, vc);
+
+ _mm256_storeu_ps(y, vy0);
+ _mm256_storeu_ps(y + 8, vy1);
+ _mm256_storeu_ps(y + 16, vy2);
+ _mm256_storeu_ps(y + 24, vy3);
+ y += 32;
+ }
+ for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
+ const __m256 vx = _mm256_loadu_ps(x);
+ x += 8;
+
+ const __m256 vy = _mm256_mul_ps(vx, vc);
+
+ _mm256_storeu_ps(y, vy);
+ y += 8;
+ }
+ if XNN_UNLIKELY(n != 0) {
+ do {
+ const __m128 vx = _mm_load_ss(x);
+ x += 1;
+
+ const __m128 vy = _mm_mul_ss(vx, _mm256_castps256_ps128(vc));
+
+ _mm_store_ss(y, vy);
+ y += 1;
+
+ n -= sizeof(float);
+ } while (n != 0);
+ }
+}