Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-vbinary/vopc-avx512f.c.in |
| 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2019 Google LLC |
| 6 | // |
| 7 | // This source code is licensed under the BSD-style license found in the |
| 8 | // LICENSE file in the root directory of this source tree. |
| 9 | |
| 10 | #include <assert.h> |
| 11 | |
| 12 | #include <immintrin.h> |
| 13 | |
| 14 | #include <xnnpack/common.h> |
Marat Dukhan | 9e0e8ee | 2019-12-13 15:34:13 -0800 | [diff] [blame^] | 15 | #include <xnnpack/intrinsics-polyfill.h> |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 16 | #include <xnnpack/vbinary.h> |
| 17 | |
| 18 | |
| 19 | void xnn_f32_vmulc_ukernel__avx512f_x16( |
| 20 | size_t n, |
| 21 | const float* a, |
| 22 | const float* b, |
| 23 | float* y, |
| 24 | const union xnn_f32_output_params params[restrict static 1]) |
| 25 | { |
| 26 | assert(n != 0); |
| 27 | assert(n % sizeof(float) == 0); |
| 28 | |
| 29 | const __m512 vy_min = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.min)); |
| 30 | const __m512 vy_max = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.max)); |
| 31 | |
| 32 | const __m512 vb = _mm512_set1_ps(*b); |
| 33 | for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) { |
| 34 | const __m512 va0123456789ABCDEF = _mm512_loadu_ps(a); |
| 35 | a += 16; |
| 36 | |
| 37 | __m512 vy0123456789ABCDEF = _mm512_mul_ps(va0123456789ABCDEF, vb); |
| 38 | |
| 39 | vy0123456789ABCDEF = _mm512_max_ps(vy0123456789ABCDEF, vy_min); |
| 40 | |
| 41 | vy0123456789ABCDEF = _mm512_min_ps(vy0123456789ABCDEF, vy_max); |
| 42 | |
| 43 | _mm512_storeu_ps(y, vy0123456789ABCDEF); |
| 44 | y += 16; |
| 45 | } |
| 46 | for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) { |
| 47 | const __m512 va = _mm512_loadu_ps(a); |
| 48 | a += 16; |
| 49 | |
| 50 | __m512 vy = _mm512_mul_ps(va, vb); |
| 51 | vy = _mm512_max_ps(vy, vy_min); |
| 52 | vy = _mm512_min_ps(vy, vy_max); |
| 53 | _mm512_storeu_ps(y, vy); |
| 54 | y += 16; |
| 55 | } |
| 56 | if XNN_UNLIKELY(n != 0) { |
| 57 | assert(n >= 1 * sizeof(float)); |
| 58 | assert(n <= 15 * sizeof(float)); |
| 59 | // Prepare mask for valid 32-bit elements (depends on n). |
| 60 | n >>= 2 /* log2(sizeof(float)) */; |
| 61 | const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1))); |
| 62 | |
| 63 | const __m512 va = _mm512_maskz_loadu_ps(vmask, a); |
| 64 | |
| 65 | __m512 vy = _mm512_mul_ps(va, vb); |
| 66 | vy = _mm512_max_ps(vy, vy_min); |
| 67 | vy = _mm512_min_ps(vy, vy_max); |
| 68 | _mm512_mask_storeu_ps(y, vmask, vy); |
| 69 | } |
| 70 | } |