Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-3x8c8-minmax-avxvnni.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
19
20 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_3x8c8__avxvnni(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const int8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
31 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 3);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(kc % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 8 * sizeof(int8_t));
43 const int8_t* a0 = a;
44 float* c0 = c;
45 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
46 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 a1 = a0;
49 c1 = c0;
50 }
51 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
52 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 a2 = a1;
55 c2 = c1;
56 }
57
58 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
59 XNN_FORCE_REALIZATION(vsign_mask);
60 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
61 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
62 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
63 const __m256 voutput_min = _mm256_set1_ps(params->scalar.min);
64 const __m256 voutput_max = _mm256_set1_ps(params->scalar.max);
65 // XNN_FORCE_REALIZATION(voutput_min);
66 // XNN_FORCE_REALIZATION(voutput_max);
67 do {
68 const __m256i vksum01234567 = _mm256_load_si256(w);
69 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
70 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
71 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
72 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
73 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
74 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
75 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
76 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
77 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
78 w = (const int32_t*) w + 8;
79
80 size_t k = kc;
81 while (k >= 16 * sizeof(int8_t)) {
82 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
83 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
84 a0 += 16;
85 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
86 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
87 a1 += 16;
88 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
89 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
90 a2 += 16;
91
92 const __m256i vb01234567x0123 = _mm256_load_si256(w);
93 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
94 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
95 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
96
97 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
98 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
99 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
100 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
101 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
102 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
103 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
104 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
105 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
106 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
107 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
108 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
109
110 w = (const int8_t*) w + 128;
111 k -= 16 * sizeof(int8_t);
112 }
113
114 if (k != 0) {
115 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
116 a0 += 8;
117 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
118 a1 += 8;
119 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
120 a2 += 8;
121
122 const __m256i vb01234567x0123 = _mm256_load_si256(w);
123 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
124
125 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
126 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
127 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
128 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
129 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
130 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
131
132 w = (const int8_t*) w + 64;
133 k -= 8 * sizeof(int8_t);
134 }
135
136 // Add adjacent pairs
137 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
138 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
139 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
140 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
141 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
142 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
143
144 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
145 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
146 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
147
148 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
149 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
150 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
151
152 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
153 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
154 w = (const float*) w + 16;
155
156 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
157 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
158 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
159
160 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
161 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
162 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
163
164 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
165 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
166 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
167
168 if XNN_LIKELY(nc >= 8) {
169 _mm256_storeu_ps(c0, vout0x01234567);
170 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
171 c0 = (float*) ((uintptr_t) c0 + cn_stride);
172 _mm256_storeu_ps(c1, vout1x01234567);
173 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
174 c1 = (float*) ((uintptr_t) c1 + cn_stride);
175 _mm256_storeu_ps(c2, vout2x01234567);
176 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
177 c2 = (float*) ((uintptr_t) c2 + cn_stride);
178 nc -= 8;
179 } else {
180 __m128 vout0x0123 = _mm256_castps256_ps128(vout0x01234567);
181 __m128 vout1x0123 = _mm256_castps256_ps128(vout1x01234567);
182 __m128 vout2x0123 = _mm256_castps256_ps128(vout2x01234567);
183 if (nc & 4) {
184 _mm_storeu_ps(c0, vout0x0123);
185 c0 += 4;
186 vout0x0123 = _mm256_extractf128_ps(vout0x01234567, 1);
187 _mm_storeu_ps(c1, vout1x0123);
188 c1 += 4;
189 vout1x0123 = _mm256_extractf128_ps(vout1x01234567, 1);
190 _mm_storeu_ps(c2, vout2x0123);
191 c2 += 4;
192 vout2x0123 = _mm256_extractf128_ps(vout2x01234567, 1);
193 }
194 if (nc & 2) {
195 _mm_storel_pi((__m64*) c0, vout0x0123);
196 c0 += 2;
197 vout0x0123 = _mm_movehl_ps(vout0x0123, vout0x0123);
198 _mm_storel_pi((__m64*) c1, vout1x0123);
199 c1 += 2;
200 vout1x0123 = _mm_movehl_ps(vout1x0123, vout1x0123);
201 _mm_storel_pi((__m64*) c2, vout2x0123);
202 c2 += 2;
203 vout2x0123 = _mm_movehl_ps(vout2x0123, vout2x0123);
204 }
205 if (nc & 1) {
206 _mm_store_ss(c0, vout0x0123);
207 _mm_store_ss(c1, vout1x0123);
208 _mm_store_ss(c2, vout2x0123);
209 }
210 nc = 0;
211 }
212 } while (nc != 0);
213 }