Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-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_f16_qc4w_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 xnn_float16* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const struct xnn_f16_qc4w_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 uint16_t* c0 = (uint16_t*) c;
45 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
46 uint16_t* c1 = (uint16_t*) ((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 uint16_t* c2 = (uint16_t*) ((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_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
64 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
65 // XNN_FORCE_REALIZATION(voutput_min);
66 // XNN_FORCE_REALIZATION(voutput_max);
67 const __m256i vmask = _mm256_set1_epi8(0xF0);
68 XNN_FORCE_REALIZATION(vmask);
69 do {
70 const __m256i vksum01234567 = _mm256_load_si256(w);
71 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
72 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
73 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
74 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
75 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
76 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
77 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
78 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
79 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
80 w = (const int32_t*) w + 8;
81
82 size_t k = kc;
83 while (k >= 16 * sizeof(int8_t)) {
84 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
85 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
86 a0 += 16;
87 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
88 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
89 a1 += 16;
90 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
91 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
92 a2 += 16;
93
94 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
95 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
96 const __m256i vbs01234567x0123 = _mm256_slli_epi32(vbb01234567x01234567, 4);
97 const __m256i vbs89ABCDEFx0123 = _mm256_slli_epi32(vbb89ABCDEFx01234567, 4);
98 const __m256i vb01234567x4567 = _mm256_and_si256(vbb01234567x01234567, vmask);
99 const __m256i vb89ABCDEFx4567 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
100 const __m256i vb01234567x0123 = _mm256_and_si256(vbs01234567x0123, vmask);
101 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbs89ABCDEFx0123, vmask);
102
103 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
104 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
105 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
106 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
107 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
108 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
109 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
110 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
111 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
112 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
113 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
114 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
115
116 w = (const int8_t*) w + 64;
117 k -= 16 * sizeof(int8_t);
118 }
119
120 if (k != 0) {
121 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
122 a0 += 8;
123 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
124 a1 += 8;
125 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
126 a2 += 8;
127
128 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
129 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
130 const __m256i vb01234567x0123 = _mm256_slli_epi32(vbb01234567x01234567, 4);
131 const __m256i vb89ABCDEFx0123 = _mm256_slli_epi32(vbb89ABCDEFx01234567, 4);
132
133 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
134 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
135 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
136 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
137 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
138 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
139
140 w = (const int8_t*) w + 64;
141 k -= 8 * sizeof(int8_t);
142 }
143
144 // Add adjacent pairs
145 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
146 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
147 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
148 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
149 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
150 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
151
152 vacc0x01234567 = _mm256_srai_epi32(vacc0x01234567, 4);
153 vacc1x01234567 = _mm256_srai_epi32(vacc1x01234567, 4);
154 vacc2x01234567 = _mm256_srai_epi32(vacc2x01234567, 4);
155 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
156 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
157 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
158
159 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
160 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
161 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
162
163 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
164 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
165 w = (const float*) w + 16;
166
167 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
168 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
169 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
170
171 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
172 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
173 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
174
175 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
176 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
177 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
178
179 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
180 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
181 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
182 if XNN_LIKELY(nc >= 8) {
183 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
184 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
185 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
186 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
187 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
188 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
189 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
190 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
191 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
192 nc -= 8;
193 } else {
194 if (nc & 4) {
195 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
196 c0 += 4;
197 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
198 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
199 c1 += 4;
200 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
201 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
202 c2 += 4;
203 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
204 }
205 if (nc & 2) {
206 _mm_storeu_si32(c0, vfp16out0x01234567);
207 c0 += 2;
208 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
209 _mm_storeu_si32(c1, vfp16out1x01234567);
210 c1 += 2;
211 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
212 _mm_storeu_si32(c2, vfp16out2x01234567);
213 c2 += 2;
214 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
215 }
216 if (nc & 1) {
217 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
218 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
219 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
220 }
221 nc = 0;
222 }
223 } while (nc != 0);
224 }