Marat Dukhan | 2b9efd8 | 2020-06-08 01:09:31 -0700 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-vunary/avx512f.c.in |
| 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2020 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> |
| 15 | #include <xnnpack/intrinsics-polyfill.h> |
| 16 | #include <xnnpack/vunary.h> |
| 17 | |
| 18 | |
| 19 | void xnn_f32_vneg_ukernel__avx512f_x16( |
| 20 | size_t n, |
| 21 | const float* x, |
| 22 | float* y, |
| 23 | const union xnn_f32_neg_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 24 | { |
| 25 | assert(n != 0); |
| 26 | assert(n % sizeof(float) == 0); |
Frank Barchard | 0822dde | 2020-07-04 12:47:24 -0700 | [diff] [blame] | 27 | assert(x != NULL); |
| 28 | assert(y != NULL); |
Marat Dukhan | 2b9efd8 | 2020-06-08 01:09:31 -0700 | [diff] [blame] | 29 | |
| 30 | const __m512i vsign_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse.sign_mask)); |
| 31 | for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) { |
Marat Dukhan | 34ccfba | 2020-06-09 08:46:51 -0700 | [diff] [blame] | 32 | const __m512i vx0123456789ABCDEF = _mm512_loadu_si512(x); |
Marat Dukhan | 2b9efd8 | 2020-06-08 01:09:31 -0700 | [diff] [blame] | 33 | x += 16; |
| 34 | |
| 35 | const __m512i vy0123456789ABCDEF = _mm512_xor_epi32(vx0123456789ABCDEF, vsign_mask); |
| 36 | |
Marat Dukhan | 34ccfba | 2020-06-09 08:46:51 -0700 | [diff] [blame] | 37 | _mm512_storeu_si512(y, vy0123456789ABCDEF); |
Marat Dukhan | 2b9efd8 | 2020-06-08 01:09:31 -0700 | [diff] [blame] | 38 | y += 16; |
| 39 | } |
| 40 | if XNN_UNLIKELY(n != 0) { |
| 41 | assert(n >= 1 * sizeof(float)); |
| 42 | assert(n <= 15 * sizeof(float)); |
| 43 | // Prepare mask for valid 32-bit elements (depends on n). |
| 44 | n >>= 2 /* log2(sizeof(float)) */; |
| 45 | const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1))); |
| 46 | |
| 47 | const __m512i vx = _mm512_maskz_loadu_epi32(vmask, x); |
| 48 | const __m512i vy = _mm512_xor_epi32(vx, vsign_mask); |
| 49 | _mm512_mask_storeu_epi32(y, vmask, vy); |
| 50 | } |
| 51 | } |