Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-3x4c8-minmax-sse2-ld64.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_qb4w_gemm_minmax_ukernel_3x4c8__sse2_ld64(
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_qb4w_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 size_t bl = params->scalar.blocksize;
43 assert(bl <= round_up_po2(kc, 2));
44 assert(bl != 0);
45 assert(bl % 32 == 0);
46 kc = round_up_po2(kc, 8 * sizeof(int8_t));
47 const int8_t* a0 = a;
48 float* c0 = c;
49 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
50 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
51 if XNN_UNPREDICTABLE(mr < 2) {
52 a1 = a0;
53 c1 = c0;
54 }
55 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
56 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
57 if XNN_UNPREDICTABLE(mr <= 2) {
58 a2 = a1;
59 c2 = c1;
60 }
61
62 const __m128 vmin = _mm_set1_ps(params->scalar.min);
63 const __m128 vmax = _mm_set1_ps(params->scalar.max);
64 XNN_FORCE_REALIZATION(vmin);
65 XNN_FORCE_REALIZATION(vmax);
66
67 const __m128i vmask = _mm_set1_epi8(0xF0);
68 XNN_FORCE_REALIZATION(vmask);
69
70 do {
71 const __m128 vksum = _mm_load_ps((const float*) w);
72 const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
73 const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
74 const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));
75 __m128i vinput_zero_point2 = _mm_cvtsi32_si128(*((const int*) &quantization_params[2].zero_point));
76 vinput_zero_point2 = _mm_shuffle_epi32(vinput_zero_point2, _MM_SHUFFLE(0, 0, 0, 0));
77
78 __m128 vinput_zero_point0_float = _mm_cvtepi32_ps(vinput_zero_point0);
79 __m128 vout0x0123 = _mm_mul_ps(vksum, vinput_zero_point0_float);
80 __m128 vinput_zero_point1_float = _mm_cvtepi32_ps(vinput_zero_point1);
81 __m128 vout1x0123 = _mm_mul_ps(vksum, vinput_zero_point1_float);
82 __m128 vinput_zero_point2_float = _mm_cvtepi32_ps(vinput_zero_point2);
83 __m128 vout2x0123 = _mm_mul_ps(vksum, vinput_zero_point2_float);
84 w = (const int32_t*) w + 4;
85
86 for (size_t kb=0; kb < kc; kb += bl) {
87 __m128i vacc0x0 = _mm_setzero_si128();
88 __m128i vacc0x1 = _mm_setzero_si128();
89 __m128i vacc0x2 = _mm_setzero_si128();
90 __m128i vacc0x3 = _mm_setzero_si128();
91 __m128i vacc1x0 = _mm_setzero_si128();
92 __m128i vacc1x1 = _mm_setzero_si128();
93 __m128i vacc1x2 = _mm_setzero_si128();
94 __m128i vacc1x3 = _mm_setzero_si128();
95 __m128i vacc2x0 = _mm_setzero_si128();
96 __m128i vacc2x1 = _mm_setzero_si128();
97 __m128i vacc2x2 = _mm_setzero_si128();
98 __m128i vacc2x3 = _mm_setzero_si128();
99 size_t k = bl;
100
101 while (k >= 16 * sizeof(int8_t)) {
102 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
103 const __m128i vxa0c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c0, va0c0), 8);
104 a0 += 8;
105 const __m128i va1c0 = _mm_loadl_epi64((const __m128i*) a1);
106 const __m128i vxa1c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va1c0, va1c0), 8);
107 a1 += 8;
108 const __m128i va2c0 = _mm_loadl_epi64((const __m128i*) a2);
109 const __m128i vxa2c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va2c0, va2c0), 8);
110 a2 += 8;
111
112 const __m128i vb0c01 = _mm_loadl_epi64((const __m128i*) w);
113 const __m128i vbs0c0 = _mm_slli_epi32(vb0c01, 4);
114 const __m128i vb0c0 = _mm_and_si128(vbs0c0, vmask);
115 const __m128i vxb0c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0c0, vb0c0), 8);
116
117 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
118 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
119 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c0, vxb0c0));
120 const __m128i vb1c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
121 const __m128i vbs1c0 = _mm_slli_epi32(vb1c01, 4);
122 const __m128i vb1c0 = _mm_and_si128(vbs1c0, vmask);
123 const __m128i vxb1c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1c0, vb1c0), 8);
124
125 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
126 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
127 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c0, vxb1c0));
128 const __m128i vb2c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
129 const __m128i vbs2c0 = _mm_slli_epi32(vb2c01, 4);
130 const __m128i vb2c0 = _mm_and_si128(vbs2c0, vmask);
131 const __m128i vxb2c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2c0, vb2c0), 8);
132
133 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
134 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
135 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c0, vxb2c0));
136 const __m128i vb3c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
137 const __m128i vbs3c0 = _mm_slli_epi32(vb3c01, 4);
138 const __m128i vb3c0 = _mm_and_si128(vbs3c0, vmask);
139 const __m128i vxb3c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3c0, vb3c0), 8);
140
141 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
142 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
143 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c0, vxb3c0));
144
145 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
146 const __m128i vxa0c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c1, va0c1), 8);
147 a0 += 8;
148 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
149 const __m128i vxa1c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1c1, va1c1), 8);
150 a1 += 8;
151 const __m128i va2c1 = _mm_loadl_epi64((const __m128i*) a2);
152 const __m128i vxa2c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va2c1, va2c1), 8);
153 a2 += 8;
154
155 const __m128i vb0c1 = _mm_and_si128(vb0c01, vmask);
156 const __m128i vxb0c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0c1, vb0c1), 8);
157
158 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
159 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
160 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c1, vxb0c1));
161 const __m128i vb1c1 = _mm_and_si128(vb1c01, vmask);
162 const __m128i vxb1c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1c1, vb1c1), 8);
163
164 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
165 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
166 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c1, vxb1c1));
167 const __m128i vb2c1 = _mm_and_si128(vb2c01, vmask);
168 const __m128i vxb2c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2c1, vb2c1), 8);
169
170 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
171 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
172 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c1, vxb2c1));
173 const __m128i vb3c1 = _mm_and_si128(vb3c01, vmask);
174 const __m128i vxb3c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3c1, vb3c1), 8);
175
176 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
177 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
178 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c1, vxb3c1));
179
180
181 w = (const int8_t*) w + 32;
182 k -= 16 * sizeof(int8_t);
183 }
184
185 while (k >= 8 * sizeof(int8_t)) {
186 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
187 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
188 a0 += 8;
189 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
190 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
191 a1 += 8;
192 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
193 const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
194 a2 += 8;
195
196 __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
197 vb0 = _mm_slli_epi32(vb0, 4);
198 vb0 = _mm_and_si128(vb0, vmask);
199
200 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
201
202 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
203 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
204 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
205 __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
206 vb1 = _mm_slli_epi32(vb1, 4);
207 vb1 = _mm_and_si128(vb1, vmask);
208
209 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
210
211 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
212 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
213 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
214 __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
215 vb2 = _mm_slli_epi32(vb2, 4);
216 vb2 = _mm_and_si128(vb2, vmask);
217
218 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
219
220 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
221 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
222 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
223 __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
224 vb3 = _mm_slli_epi32(vb3, 4);
225 vb3 = _mm_and_si128(vb3, vmask);
226
227 const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
228
229 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
230 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
231 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
232
233 w = (const int8_t*) w + 32;
234 k -= 8 * sizeof(int8_t);
235 }
236 // accumulate float
237 const __m128 vfilter_output_scale0123 = _mm_castsi128_ps(_mm_unpacklo_epi16(_mm_setzero_si128(), _mm_loadl_epi64((const __m128i*) w)));
238 w = (const uint16_t*) w + 4;
239
240 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
241 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
242 const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
243 const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
244 const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(vacc2x0, vacc2x2));
245 const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(vacc2x1, vacc2x3));
246
247 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
248 __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
249 __m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc2x02, vacc2x13));
250
251 vout0x0123 = _mm_add_ps(vout0x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc0x0123), vfilter_output_scale0123));
252 vout1x0123 = _mm_add_ps(vout1x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc1x0123), vfilter_output_scale0123));
253 vout2x0123 = _mm_add_ps(vout2x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc2x0123), vfilter_output_scale0123));
254 }
255
256 const __m128i vinput_scale01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
257 const __m128 vinput_scale0 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(1, 1, 1, 1)));
258 const __m128 vinput_scale1 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(3, 3, 3, 3)));
259 const __m128 vinput_scale2 = _mm_load1_ps(&quantization_params[2].inv_scale);
260
261 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
262 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
263 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale2);
264
265
266 const __m128 vbias0123 = _mm_loadu_ps((const float*) w);
267 w = (const float*) w + 4;
268 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
269 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
270 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
271
272 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
273 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
274 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
275
276 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
277 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
278 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
279
280 if XNN_LIKELY(nc >= 4) {
281 _mm_storeu_ps(c0, vout0x0123);
282 _mm_storeu_ps(c1, vout1x0123);
283 _mm_storeu_ps(c2, vout2x0123);
284
285 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
286 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
287 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
288
289 c0 = (float*) ((uintptr_t) c0 + cn_stride);
290 c1 = (float*) ((uintptr_t) c1 + cn_stride);
291 c2 = (float*) ((uintptr_t) c2 + cn_stride);
292
293 nc -= 4;
294 } else {
295 if (nc & 2) {
296 _mm_storel_pi((__m64*) c0, vout0x0123);
297 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
298 c0 += 2;
299 _mm_storel_pi((__m64*) c1, vout1x0123);
300 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
301 c1 += 2;
302 _mm_storel_pi((__m64*) c2, vout2x0123);
303 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
304 c2 += 2;
305 }
306 if (nc & 1) {
307 _mm_store_ss(c0, vout0x0123);
308 _mm_store_ss(c1, vout1x0123);
309 _mm_store_ss(c2, vout2x0123);
310 }
311 nc = 0;
312 }
313 } while (nc != 0);
314 }