Debian Codegraph @

src/qd8-f16-qc8w-gemm/gen/qd8-f16-qc8w-gemm-2x8c8-minmax-avxvnni-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx8c8-avxvnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 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/gemm.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18 #include "xnnpack/prefetch.h"
19
20
21 void xnn_qd8_f16_qc8w_gemm_minmax_ukernel_2x8c8__avxvnni_prfm(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 const int8_t* restrict a,
26 size_t a_stride,
27 const void* restrict w,
28 xnn_float16* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 2);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(kc % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8 * sizeof(int8_t));
44 const int8_t* a0 = a;
45 uint16_t* c0 = (uint16_t*) c;
46 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
47 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr != 2) {
49 a1 = a0;
50 c1 = c0;
51 }
52
53 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
54 XNN_FORCE_REALIZATION(vsign_mask);
55 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
56 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
57 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
58 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
59 // XNN_FORCE_REALIZATION(voutput_min);
60 // XNN_FORCE_REALIZATION(voutput_max);
61 do {
62 const __m256i vksum01234567 = _mm256_load_si256(w);
63 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
64 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
65 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
66 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
67 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
68 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
69 __m256i vacc1x0x0123 = _mm256_setzero_si256();
70 __m256i vacc1x0x4567 = _mm256_setzero_si256();
71 __m256i vacc1x1x0123 = _mm256_setzero_si256();
72 __m256i vacc1x1x4567 = _mm256_setzero_si256();
73 w = (const int32_t*) w + 8;
74
75 size_t k = kc;
76 while (k >= 16 * sizeof(int8_t)) {
77 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
78 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
79 a0 += 16;
80 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
81 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
82 a1 += 16;
83
84 const __m256i vb01234567x0123 = _mm256_load_si256(w);
85 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
86 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
87 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
88 xnn_prefetch_to_l1((const int8_t*) w + 896);
89
90 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
91 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
92 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
93 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
94 xnn_prefetch_to_l1((const int8_t*) w + 960);
95 vacc1x0x0123 = _mm256_dpbusd_avx_epi32(vacc1x0x0123, va0x89ABCDEF, vb01234567x4567);
96 vacc1x0x4567 = _mm256_dpbusd_avx_epi32(vacc1x0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
97 vacc1x1x0123 = _mm256_dpbusd_avx_epi32(vacc1x1x0123, va1x89ABCDEF, vb01234567x4567);
98 vacc1x1x4567 = _mm256_dpbusd_avx_epi32(vacc1x1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
99
100 w = (const int8_t*) w + 128;
101 k -= 16 * sizeof(int8_t);
102 }
103
104 if (k != 0) {
105 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
106 a0 += 8;
107 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
108 a1 += 8;
109
110 const __m256i vb01234567x0123 = _mm256_load_si256(w);
111 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
112
113 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
114 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
115 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
116 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
117 xnn_prefetch_to_l1((const int8_t*) w + 960);
118
119 w = (const int8_t*) w + 64;
120 k -= 8 * sizeof(int8_t);
121 }
122 vacc0x0123 = _mm256_add_epi32(vacc0x0123, vacc1x0x0123);
123 vacc0x4567 = _mm256_add_epi32(vacc0x4567, vacc1x0x4567);
124 vacc1x0123 = _mm256_add_epi32(vacc1x0123, vacc1x1x0123);
125 vacc1x4567 = _mm256_add_epi32(vacc1x4567, vacc1x1x4567);
126
127 // Add adjacent pairs
128 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
129 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
130 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
131 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
132
133 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
134 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
135
136 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
137 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
138
139 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
140 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
141 w = (const float*) w + 16;
142
143 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
144 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
145
146 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
147 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
148
149 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
150 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
151
152 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
153 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
154 if XNN_LIKELY(nc >= 8) {
155 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
156 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
157 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
158 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
159 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
160 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
161 nc -= 8;
162 } else {
163 if (nc & 4) {
164 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
165 c0 += 4;
166 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
167 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
168 c1 += 4;
169 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
170 }
171 if (nc & 2) {
172 _mm_storeu_si32(c0, vfp16out0x01234567);
173 c0 += 2;
174 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
175 _mm_storeu_si32(c1, vfp16out1x01234567);
176 c1 += 2;
177 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
178 }
179 if (nc & 1) {
180 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
181 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
182 }
183 nc = 0;
184 }
185 } while (nc != 0);
186 }