Marat Dukhan | e9c4b96 | 2021-04-02 16:56:55 -0700 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/qs8-vaddc/sse-mul16-ld64.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 <smmintrin.h> |
| 13 | |
| 14 | #include <xnnpack/vadd.h> |
| 15 | |
| 16 | |
| 17 | void xnn_qs8_vaddc_minmax_ukernel__avx_mul16_ld64_x16( |
| 18 | size_t n, |
| 19 | const int8_t* input_x, |
| 20 | const int8_t* input_y, |
| 21 | int8_t* output, |
Marat Dukhan | 6e0fc39 | 2021-07-19 18:38:24 -0700 | [diff] [blame^] | 22 | const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN |
Marat Dukhan | e9c4b96 | 2021-04-02 16:56:55 -0700 | [diff] [blame] | 23 | { |
| 24 | const __m128i vx_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.x_multiplier_lo); |
| 25 | const __m128i vx_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.x_multiplier_hi); |
| 26 | const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask); |
| 27 | const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold); |
| 28 | const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift); |
| 29 | const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point); |
| 30 | const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min); |
| 31 | const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max); |
| 32 | |
| 33 | __m128i vzero_point_product = _mm_cvtsi32_si128(params->sse2.y_multiplier[0] * (int32_t) *input_y); |
| 34 | vzero_point_product = _mm_shuffle_epi32(vzero_point_product, _MM_SHUFFLE(0, 0, 0, 0)); |
| 35 | vzero_point_product = _mm_add_epi32(vzero_point_product, _mm_load_si128((const __m128i*) params->sse2.zero_point_product)); |
| 36 | for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) { |
| 37 | const __m128i vx01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_x)); |
| 38 | const __m128i vx89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_x + 8))); |
| 39 | input_x += 16; |
| 40 | |
| 41 | |
| 42 | __m128i vxprod01234567hi = _mm_mulhi_epu16(vx01234567, vx_multiplier_lo); |
| 43 | const __m128i vxprod01234567lo = _mm_mullo_epi16(vx01234567, vx_multiplier_lo); |
| 44 | __m128i vxprod89ABCDEFhi = _mm_mulhi_epu16(vx89ABCDEF, vx_multiplier_lo); |
| 45 | const __m128i vxprod89ABCDEFlo = _mm_mullo_epi16(vx89ABCDEF, vx_multiplier_lo); |
| 46 | |
| 47 | vxprod01234567hi = _mm_add_epi16(vxprod01234567hi, _mm_mullo_epi16(vx01234567, vx_multiplier_hi)); |
| 48 | vxprod89ABCDEFhi = _mm_add_epi16(vxprod89ABCDEFhi, _mm_mullo_epi16(vx89ABCDEF, vx_multiplier_hi)); |
| 49 | |
| 50 | vxprod01234567hi = _mm_sub_epi16(vxprod01234567hi, _mm_and_si128(_mm_srai_epi16(vx01234567, 15), vx_multiplier_lo)); |
| 51 | vxprod89ABCDEFhi = _mm_sub_epi16(vxprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(vx89ABCDEF, 15), vx_multiplier_lo)); |
| 52 | |
| 53 | __m128i vacc0123 = _mm_add_epi32(vzero_point_product, _mm_unpacklo_epi16(vxprod01234567lo, vxprod01234567hi)); |
| 54 | __m128i vacc4567 = _mm_add_epi32(vzero_point_product, _mm_unpackhi_epi16(vxprod01234567lo, vxprod01234567hi)); |
| 55 | __m128i vacc89AB = _mm_add_epi32(vzero_point_product, _mm_unpacklo_epi16(vxprod89ABCDEFlo, vxprod89ABCDEFhi)); |
| 56 | __m128i vaccCDEF = _mm_add_epi32(vzero_point_product, _mm_unpackhi_epi16(vxprod89ABCDEFlo, vxprod89ABCDEFhi)); |
| 57 | |
| 58 | const __m128i vrem0123 = _mm_add_epi32(_mm_and_si128(vacc0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0123)); |
| 59 | const __m128i vrem4567 = _mm_add_epi32(_mm_and_si128(vacc4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567)); |
| 60 | const __m128i vrem89AB = _mm_add_epi32(_mm_and_si128(vacc89AB, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc89AB)); |
| 61 | const __m128i vremCDEF = _mm_add_epi32(_mm_and_si128(vaccCDEF, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vaccCDEF)); |
| 62 | |
| 63 | vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_threshold)); |
| 64 | vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)); |
| 65 | vacc89AB = _mm_sub_epi32(_mm_sra_epi32(vacc89AB, vshift), _mm_cmpgt_epi32(vrem89AB, vremainder_threshold)); |
| 66 | vaccCDEF = _mm_sub_epi32(_mm_sra_epi32(vaccCDEF, vshift), _mm_cmpgt_epi32(vremCDEF, vremainder_threshold)); |
| 67 | |
| 68 | __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); |
| 69 | __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point); |
| 70 | |
| 71 | vout01234567 = _mm_max_epi16(vout01234567, voutput_min); |
| 72 | vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min); |
| 73 | |
| 74 | vout01234567 = _mm_min_epi16(vout01234567, voutput_max); |
| 75 | vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max); |
| 76 | |
| 77 | const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF); |
| 78 | |
| 79 | _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF); |
| 80 | output += 16; |
| 81 | } |
| 82 | if XNN_UNLIKELY(n != 0) { |
| 83 | do { |
| 84 | const __m128i vx01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_x)); |
| 85 | input_x += 8; |
| 86 | |
| 87 | |
| 88 | __m128i vxprod01234567hi = _mm_mulhi_epu16(vx01234567, vx_multiplier_lo); |
| 89 | const __m128i vxprod01234567lo = _mm_mullo_epi16(vx01234567, vx_multiplier_lo); |
| 90 | |
| 91 | vxprod01234567hi = _mm_add_epi16(vxprod01234567hi, _mm_mullo_epi16(vx01234567, vx_multiplier_hi)); |
| 92 | |
| 93 | vxprod01234567hi = _mm_sub_epi16(vxprod01234567hi, _mm_and_si128(_mm_srai_epi16(vx01234567, 15), vx_multiplier_lo)); |
| 94 | |
| 95 | __m128i vacc0123 = _mm_add_epi32(vzero_point_product, _mm_unpacklo_epi16(vxprod01234567lo, vxprod01234567hi)); |
| 96 | __m128i vacc4567 = _mm_add_epi32(vzero_point_product, _mm_unpackhi_epi16(vxprod01234567lo, vxprod01234567hi)); |
| 97 | |
| 98 | const __m128i vrem0123 = _mm_add_epi32(_mm_and_si128(vacc0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0123)); |
| 99 | const __m128i vrem4567 = _mm_add_epi32(_mm_and_si128(vacc4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567)); |
| 100 | |
| 101 | vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_threshold)); |
| 102 | vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)); |
| 103 | |
| 104 | __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); |
| 105 | vout01234567 = _mm_max_epi16(vout01234567, voutput_min); |
| 106 | vout01234567 = _mm_min_epi16(vout01234567, voutput_max); |
| 107 | |
| 108 | __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); |
| 109 | |
| 110 | if XNN_LIKELY(n >= (8 * sizeof(int8_t))) { |
| 111 | _mm_storel_epi64((__m128i*) output, vout0123456701234567); |
| 112 | output += 8; |
| 113 | n -= 8 * sizeof(int8_t); |
| 114 | } else { |
| 115 | if (n & (4 * sizeof(int8_t))) { |
| 116 | *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567); |
| 117 | vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); |
| 118 | output += 4; |
| 119 | } |
| 120 | if (n & (2 * sizeof(int8_t))) { |
| 121 | *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0); |
| 122 | vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16); |
| 123 | output += 2; |
| 124 | } |
| 125 | if (n & (1 * sizeof(int8_t))) { |
| 126 | *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0); |
| 127 | } |
| 128 | n = 0; |
| 129 | } |
| 130 | } while (n != 0); |
| 131 | } |
| 132 | } |