Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-3x4c8-minmax-sse2-ld128.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c8-sse.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 <emmintrin.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18
19
20 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_3x4c8__sse2_ld128(
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 struct xnn_f32_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 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 __m128 vmin = _mm_set1_ps(params->scalar.min);
59 const __m128 vmax = _mm_set1_ps(params->scalar.max);
60 XNN_FORCE_REALIZATION(vmin);
61 XNN_FORCE_REALIZATION(vmax);
62
63 const __m128i vmask = _mm_set1_epi8(0xF0);
64 XNN_FORCE_REALIZATION(vmask);
65
66 do {
67 const __m128i vksum = _mm_load_si128((const __m128i*) w);
68 const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
69 const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
70 const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));
71 __m128i vinput_zero_point2 = _mm_cvtsi32_si128(*((const int*) &quantization_params[2].zero_point));
72 vinput_zero_point2 = _mm_shuffle_epi32(vinput_zero_point2, _MM_SHUFFLE(0, 0, 0, 0));
73 const __m128i vzero = _mm_setzero_si128();
74 const __m128i vksum_lo = _mm_srli_epi32(_mm_slli_epi32(vksum, 16), 16);
75 const __m128i vksum_hi = _mm_srli_epi32(vksum, 16);
76
77 __m128i vzpprodksumhi0 = _mm_mulhi_epu16(vinput_zero_point0, vksum_lo);
78 const __m128i vzpprodksumlo0 = _mm_mullo_epi16(vinput_zero_point0, vksum_lo);
79 __m128i vzpprodksumhi1 = _mm_mulhi_epu16(vinput_zero_point1, vksum_lo);
80 const __m128i vzpprodksumlo1 = _mm_mullo_epi16(vinput_zero_point1, vksum_lo);
81 __m128i vzpprodksumhi2 = _mm_mulhi_epu16(vinput_zero_point2, vksum_lo);
82 const __m128i vzpprodksumlo2 = _mm_mullo_epi16(vinput_zero_point2, vksum_lo);
83
84 vzpprodksumhi0 = _mm_add_epi16(vzpprodksumhi0, _mm_mullo_epi16(vinput_zero_point0, vksum_hi));
85 vzpprodksumhi0 = _mm_sub_epi16(vzpprodksumhi0, _mm_and_si128(_mm_srai_epi16(vinput_zero_point0, 15), vksum_lo));
86 vzpprodksumhi1 = _mm_add_epi16(vzpprodksumhi1, _mm_mullo_epi16(vinput_zero_point1, vksum_hi));
87 vzpprodksumhi1 = _mm_sub_epi16(vzpprodksumhi1, _mm_and_si128(_mm_srai_epi16(vinput_zero_point1, 15), vksum_lo));
88 vzpprodksumhi2 = _mm_add_epi16(vzpprodksumhi2, _mm_mullo_epi16(vinput_zero_point2, vksum_hi));
89 vzpprodksumhi2 = _mm_sub_epi16(vzpprodksumhi2, _mm_and_si128(_mm_srai_epi16(vinput_zero_point2, 15), vksum_lo));
90
91 vzpprodksumhi0 = _mm_slli_si128(vzpprodksumhi0, 2);
92 vzpprodksumhi1 = _mm_slli_si128(vzpprodksumhi1, 2);
93 vzpprodksumhi2 = _mm_slli_si128(vzpprodksumhi2, 2);
94
95 const __m128i vksumzp0 = _mm_or_si128(vzpprodksumhi0, vzpprodksumlo0);
96 const __m128i vksumzp1 = _mm_or_si128(vzpprodksumhi1, vzpprodksumlo1);
97 const __m128i vksumzp2 = _mm_or_si128(vzpprodksumhi2, vzpprodksumlo2);
98
99 const __m128i vksum010 = _mm_unpacklo_epi32(vksumzp0, vzero);
100 const __m128i vksum230 = _mm_unpackhi_epi32(vksumzp0, vzero);
101 const __m128i vksum011 = _mm_unpacklo_epi32(vksumzp1, vzero);
102 const __m128i vksum231 = _mm_unpackhi_epi32(vksumzp1, vzero);
103 const __m128i vksum012 = _mm_unpacklo_epi32(vksumzp2, vzero);
104 const __m128i vksum232 = _mm_unpackhi_epi32(vksumzp2, vzero);
105
106 __m128i vacc0x0 = _mm_unpacklo_epi64(vksum010, vzero);
107 __m128i vacc0x1 = _mm_unpackhi_epi64(vksum010, vzero);
108 __m128i vacc1x0 = _mm_unpacklo_epi64(vksum011, vzero);
109 __m128i vacc1x1 = _mm_unpackhi_epi64(vksum011, vzero);
110 __m128i vacc2x0 = _mm_unpacklo_epi64(vksum012, vzero);
111 __m128i vacc2x1 = _mm_unpackhi_epi64(vksum012, vzero);
112
113 __m128i vacc0x2 = _mm_unpacklo_epi64(vksum230, vzero);
114 __m128i vacc0x3 = _mm_unpackhi_epi64(vksum230, vzero);
115 __m128i vacc1x2 = _mm_unpacklo_epi64(vksum231, vzero);
116 __m128i vacc1x3 = _mm_unpackhi_epi64(vksum231, vzero);
117 __m128i vacc2x2 = _mm_unpacklo_epi64(vksum232, vzero);
118 __m128i vacc2x3 = _mm_unpackhi_epi64(vksum232, vzero);
119 w = (const int32_t*) w + 4;
120
121 size_t k = kc;
122
123 while (k >= 16 * sizeof(int8_t)) {
124 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
125 const __m128i vxa0c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c0, va0c0), 8);
126 a0 += 8;
127 const __m128i va1c0 = _mm_loadl_epi64((const __m128i*) a1);
128 const __m128i vxa1c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va1c0, va1c0), 8);
129 a1 += 8;
130 const __m128i va2c0 = _mm_loadl_epi64((const __m128i*) a2);
131 const __m128i vxa2c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va2c0, va2c0), 8);
132 a2 += 8;
133
134 const __m128i vb01c01 = _mm_load_si128((const __m128i*) w);
135 const __m128i vbs01c0 = _mm_slli_epi32(vb01c01, 4);
136 const __m128i vb01c0 = _mm_and_si128(vbs01c0, vmask);
137 const __m128i vsb01c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c0);
138 const __m128i vxb0c0 = _mm_unpacklo_epi8(vb01c0, vsb01c0);
139 const __m128i vxb1c0 = _mm_unpackhi_epi8(vb01c0, vsb01c0);
140
141 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
142 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
143 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
144 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
145 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c0, vxb0c0));
146 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c0, vxb1c0));
147 const __m128i vb23c01 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
148 const __m128i vbs23c0 = _mm_slli_epi32(vb23c01, 4);
149 const __m128i vb23c0 = _mm_and_si128(vbs23c0, vmask);
150 const __m128i vsb23c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c0);
151 const __m128i vxb2c0 = _mm_unpacklo_epi8(vb23c0, vsb23c0);
152 const __m128i vxb3c0 = _mm_unpackhi_epi8(vb23c0, vsb23c0);
153
154 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
155 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
156 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
157 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
158 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c0, vxb2c0));
159 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c0, vxb3c0));
160
161 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
162 const __m128i vxa0c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c1, va0c1), 8);
163 a0 += 8;
164 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
165 const __m128i vxa1c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1c1, va1c1), 8);
166 a1 += 8;
167 const __m128i va2c1 = _mm_loadl_epi64((const __m128i*) a2);
168 const __m128i vxa2c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va2c1, va2c1), 8);
169 a2 += 8;
170
171 const __m128i vb01c1 = _mm_and_si128(vb01c01, vmask);
172 const __m128i vsb01c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c1);
173 const __m128i vxb0c1 = _mm_unpacklo_epi8(vb01c1, vsb01c1);
174 const __m128i vxb1c1 = _mm_unpackhi_epi8(vb01c1, vsb01c1);
175
176 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
177 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
178 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
179 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
180 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c1, vxb0c1));
181 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c1, vxb1c1));
182 const __m128i vb23c1 = _mm_and_si128(vb23c01, vmask);
183 const __m128i vsb23c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c1);
184 const __m128i vxb2c1 = _mm_unpacklo_epi8(vb23c1, vsb23c1);
185 const __m128i vxb3c1 = _mm_unpackhi_epi8(vb23c1, vsb23c1);
186
187 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
188 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
189 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
190 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
191 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c1, vxb2c1));
192 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c1, vxb3c1));
193
194
195 w = (const int8_t*) w + 32;
196 k -= 16 * sizeof(int8_t);
197 }
198
199 while (k >= 8 * sizeof(int8_t)) {
200 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
201 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
202 a0 += 8;
203 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
204 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
205 a1 += 8;
206 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
207 const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
208 a2 += 8;
209
210 __m128i vb01 = _mm_load_si128((const __m128i*) w);
211 vb01 = _mm_slli_epi32(vb01, 4);
212 vb01 = _mm_and_si128(vb01, vmask);
213
214 const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
215 const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
216 const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
217
218 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
219 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
220 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
221 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
222 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
223 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
224 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
225 vb23 = _mm_slli_epi32(vb23, 4);
226 vb23 = _mm_and_si128(vb23, vmask);
227
228 const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
229 const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
230 const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
231
232 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
233 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
234 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
235 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
236 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
237 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
238
239 w = (const int8_t*) w + 32;
240 k -= 8 * sizeof(int8_t);
241 }
242
243 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
244 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
245 const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
246 const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
247 const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(vacc2x0, vacc2x2));
248 const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(vacc2x1, vacc2x3));
249
250 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
251 __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
252 __m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc2x02, vacc2x13));
253
254 vacc0x0123 = _mm_srai_epi32(vacc0x0123, 4);
255 vacc1x0123 = _mm_srai_epi32(vacc1x0123, 4);
256 vacc2x0123 = _mm_srai_epi32(vacc2x0123, 4);
257 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
258 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
259 __m128 vout2x0123 = _mm_cvtepi32_ps(vacc2x0123);
260
261 const __m128i vinput_scale01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
262 const __m128 vinput_scale0 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(1, 1, 1, 1)));
263 const __m128 vinput_scale1 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(3, 3, 3, 3)));
264 const __m128 vinput_scale2 = _mm_load1_ps(&quantization_params[2].inv_scale);
265
266 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
267 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
268 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale2);
269
270 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
271 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
272 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
273 vout2x0123 = _mm_mul_ps(vout2x0123, vfilter_output_scale0123);
274
275 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
276 w = (const float*) w + 8;
277 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
278 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
279 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
280
281 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
282 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
283 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
284
285 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
286 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
287 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
288
289 if XNN_LIKELY(nc >= 4) {
290 _mm_storeu_ps(c0, vout0x0123);
291 _mm_storeu_ps(c1, vout1x0123);
292 _mm_storeu_ps(c2, vout2x0123);
293
294 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
295 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
296 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
297
298 c0 = (float*) ((uintptr_t) c0 + cn_stride);
299 c1 = (float*) ((uintptr_t) c1 + cn_stride);
300 c2 = (float*) ((uintptr_t) c2 + cn_stride);
301
302 nc -= 4;
303 } else {
304 if (nc & 2) {
305 _mm_storel_pi((__m64*) c0, vout0x0123);
306 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
307 c0 += 2;
308 _mm_storel_pi((__m64*) c1, vout1x0123);
309 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
310 c1 += 2;
311 _mm_storel_pi((__m64*) c2, vout2x0123);
312 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
313 c2 += 2;
314 }
315 if (nc & 1) {
316 _mm_store_ss(c0, vout0x0123);
317 _mm_store_ss(c1, vout1x0123);
318 _mm_store_ss(c2, vout2x0123);
319 }
320 nc = 0;
321 }
322 } while (nc != 0);
323 }