Marat Dukhan | d454545 | 2022-01-10 16:13:11 -0800 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f16-vbinary/vopc-f16c.c.in |
| 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2022 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/vbinary.h> |
| 16 | |
| 17 | |
| 18 | void xnn_f16_vaddc_minmax_ukernel__f16c_x8( |
| 19 | size_t n, |
| 20 | const void* restrict a_ptr, |
| 21 | const void* restrict b_ptr, |
| 22 | void* restrict y_ptr, |
| 23 | const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 24 | { |
| 25 | assert(n != 0); |
| 26 | assert(n % sizeof(uint16_t) == 0); |
| 27 | assert(a_ptr != NULL); |
| 28 | assert(b_ptr != NULL); |
| 29 | assert(y_ptr != NULL); |
| 30 | |
| 31 | const uint16_t* a = (const uint16_t*) a_ptr; |
| 32 | const uint16_t* b = (const uint16_t*) b_ptr; |
| 33 | uint16_t* y = (uint16_t*) y_ptr; |
| 34 | |
| 35 | const __m256 vy_min = _mm256_load_ps(params->avx.min); |
| 36 | const __m256 vy_max = _mm256_load_ps(params->avx.max); |
| 37 | |
| 38 | const __m256 vb = _mm256_cvtph_ps(_mm_set1_epi16((short) *b)); |
| 39 | for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) { |
| 40 | const __m256 va = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a)); |
| 41 | a += 8; |
| 42 | |
| 43 | __m256 vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(va, vb), _MM_FROUND_NO_EXC)); |
| 44 | |
| 45 | vy = _mm256_max_ps(vy, vy_min); |
| 46 | vy = _mm256_min_ps(vy, vy_max); |
| 47 | |
| 48 | _mm_storeu_si128((__m128i*) y, _mm256_cvtps_ph(vy, _MM_FROUND_NO_EXC)); |
| 49 | y += 8; |
| 50 | } |
| 51 | if XNN_UNLIKELY(n != 0) { |
| 52 | const __m256 va = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a)); |
| 53 | |
| 54 | __m256 vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(va, vb), _MM_FROUND_NO_EXC)); |
| 55 | |
| 56 | vy = _mm256_max_ps(vy, vy_min); |
| 57 | vy = _mm256_min_ps(vy, vy_max); |
| 58 | |
| 59 | __m128i vh = _mm256_cvtps_ph(vy, _MM_FROUND_NO_EXC); |
| 60 | if (n & (4 * sizeof(uint16_t))) { |
| 61 | _mm_storel_epi64((__m128i*) y, vh); |
| 62 | vh = _mm_unpackhi_epi64(vh, vh); |
| 63 | y += 4; |
| 64 | } |
| 65 | if (n & (2 * sizeof(uint16_t))) { |
| 66 | *((uint32_t*) y) = (uint32_t) _mm_cvtsi128_si32(vh); |
| 67 | vh = _mm_srli_epi64(vh, 32); |
| 68 | y += 2; |
| 69 | } |
| 70 | if (n & (1 * sizeof(uint16_t))) { |
| 71 | *y = (uint16_t) _mm_extract_epi16(vh, 0); |
| 72 | } |
| 73 | } |
| 74 | } |