Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-4x8c8-minmax-avx2-madd-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_qc4w_gemm_minmax_ukernel_4x8c8__avx2_madd_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 struct xnn_f16_qc4w_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 <= 4);
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 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
53 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 a2 = a1;
56 c2 = c1;
57 }
58 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
59 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
60 if XNN_UNPREDICTABLE(mr != 4) {
61 a3 = a2;
62 c3 = c2;
63 }
64
65 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
66 XNN_FORCE_REALIZATION(vsign_mask);
67 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
68 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
69 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
70 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
71 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
72 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
73 // XNN_FORCE_REALIZATION(voutput_min);
74 // XNN_FORCE_REALIZATION(voutput_max);
75 const __m256i vmask = _mm256_set1_epi8(0x0F);
76 XNN_FORCE_REALIZATION(vmask);
77 do {
78 const __m256i vksum01234567 = _mm256_load_si256(w);
79 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
80 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
81 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
82 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
83 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
84 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
85 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
86 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
87 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
88 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
89 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
90 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
91 w = (const int32_t*) w + 8;
92
93 size_t k = kc;
94 while (k >= 16 * sizeof(int8_t)) {
95 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
96 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
97 a0 += 16;
98 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
99 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
100 a1 += 16;
101 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
102 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
103 a2 += 16;
104 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
105 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
106 a3 += 16;
107
108 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
109 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
110 const __m256i vbs01234567x0123 = _mm256_srli_epi32(vbb01234567x01234567, 4);
111 const __m256i vbs89ABCDEFx0123 = _mm256_srli_epi32(vbb89ABCDEFx01234567, 4);
112 const __m256i vb01234567x0123 = _mm256_and_si256(vbb01234567x01234567, vmask);
113 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
114 const __m256i vb01234567x4567 = _mm256_and_si256(vbs01234567x0123, vmask);
115 const __m256i vb89ABCDEFx4567 = _mm256_and_si256(vbs89ABCDEFx0123, vmask);
116
117 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x01234567, vb01234567x0123);
118 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
119 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x01234567, vb01234567x0123);
120 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
121 vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x01234567, vb01234567x0123);
122 vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
123 vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x01234567, vb01234567x0123);
124 vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
125 xnn_prefetch_to_l1((const int8_t*) w + 960);
126 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
127 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
128 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
129 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
130 vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
131 vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
132 vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
133 vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
134
135 w = (const int8_t*) w + 64;
136 k -= 16 * sizeof(int8_t);
137 }
138
139 if (k != 0) {
140 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
141 a0 += 8;
142 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
143 a1 += 8;
144 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
145 a2 += 8;
146 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
147 a3 += 8;
148
149 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
150 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
151 const __m256i vb01234567x0123 = _mm256_and_si256(vbb01234567x01234567, vmask);
152 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
153
154 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x01234567, vb01234567x0123);
155 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
156 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x01234567, vb01234567x0123);
157 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
158 vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x01234567, vb01234567x0123);
159 vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
160 vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x01234567, vb01234567x0123);
161 vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
162 xnn_prefetch_to_l1((const int8_t*) w + 960);
163
164 w = (const int8_t*) w + 64;
165 k -= 8 * sizeof(int8_t);
166 }
167
168 // Add adjacent pairs
169 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
170 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
171 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
172 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
173 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
174 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
175 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
176 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
177
178 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
179 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
180 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
181 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
182
183 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
184 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
185 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
186 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
187
188 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
189 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
190 w = (const float*) w + 16;
191
192 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
193 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
194 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
195 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
196
197 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
198 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
199 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
200 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
201
202 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
203 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
204 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
205 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
206
207 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
208 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
209 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
210 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
211 if XNN_LIKELY(nc >= 8) {
212 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
213 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
214 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
215 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
216 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
217 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
218 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
219 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
220 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
221 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
222 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
223 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
224 nc -= 8;
225 } else {
226 if (nc & 4) {
227 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
228 c0 += 4;
229 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
230 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
231 c1 += 4;
232 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
233 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
234 c2 += 4;
235 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
236 _mm_storel_epi64((__m128i*) c3, vfp16out3x01234567);
237 c3 += 4;
238 vfp16out3x01234567 = _mm_unpackhi_epi64(vfp16out3x01234567, vfp16out3x01234567);
239 }
240 if (nc & 2) {
241 _mm_storeu_si32(c0, vfp16out0x01234567);
242 c0 += 2;
243 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
244 _mm_storeu_si32(c1, vfp16out1x01234567);
245 c1 += 2;
246 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
247 _mm_storeu_si32(c2, vfp16out2x01234567);
248 c2 += 2;
249 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
250 _mm_storeu_si32(c3, vfp16out3x01234567);
251 c3 += 2;
252 vfp16out3x01234567 = _mm_srli_epi64(vfp16out3x01234567, 32);
253 }
254 if (nc & 1) {
255 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
256 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
257 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
258 *c3 = (uint16_t) _mm_extract_epi16(vfp16out3x01234567, 0);
259 }
260 nc = 0;
261 }
262 } while (nc != 0);
263 }