Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-3x4c8-minmax-sse41-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 <smmintrin.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__sse41_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_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_cvtepi8_epi16(va0c0);
104 a0 += 8;
105 const __m128i va1c0 = _mm_loadl_epi64((const __m128i*) a1);
106 const __m128i vxa1c0 = _mm_cvtepi8_epi16(va1c0);
107 a1 += 8;
108 const __m128i va2c0 = _mm_loadl_epi64((const __m128i*) a2);
109 const __m128i vxa2c0 = _mm_cvtepi8_epi16(va2c0);
110 a2 += 8;
111
112 const __m128i vb01c01 = _mm_loadu_si128((const __m128i*) w);
113 const __m128i vbs01c0 = _mm_slli_epi32(vb01c01, 4);
114 const __m128i vb01c0 = _mm_and_si128(vbs01c0, vmask);
115 const __m128i vsb01c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c0);
116 const __m128i vxb0c0 = _mm_unpacklo_epi8(vb01c0, vsb01c0);
117 const __m128i vxb1c0 = _mm_unpackhi_epi8(vb01c0, vsb01c0);
118
119 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
120 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
121 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
122 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
123 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c0, vxb0c0));
124 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c0, vxb1c0));
125 const __m128i vb23c01 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
126 const __m128i vbs23c0 = _mm_slli_epi32(vb23c01, 4);
127 const __m128i vb23c0 = _mm_and_si128(vbs23c0, vmask);
128 const __m128i vsb23c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c0);
129 const __m128i vxb2c0 = _mm_unpacklo_epi8(vb23c0, vsb23c0);
130 const __m128i vxb3c0 = _mm_unpackhi_epi8(vb23c0, vsb23c0);
131
132 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
133 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
134 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
135 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
136 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c0, vxb2c0));
137 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c0, vxb3c0));
138
139 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
140 const __m128i vxa0c1 = _mm_cvtepi8_epi16(va0c1);
141 a0 += 8;
142 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
143 const __m128i vxa1c1 = _mm_cvtepi8_epi16(va1c1);
144 a1 += 8;
145 const __m128i va2c1 = _mm_loadl_epi64((const __m128i*) a2);
146 const __m128i vxa2c1 = _mm_cvtepi8_epi16(va2c1);
147 a2 += 8;
148
149 const __m128i vb01c1 = _mm_and_si128(vb01c01, vmask);
150 const __m128i vsb01c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c1);
151 const __m128i vxb0c1 = _mm_unpacklo_epi8(vb01c1, vsb01c1);
152 const __m128i vxb1c1 = _mm_unpackhi_epi8(vb01c1, vsb01c1);
153
154 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
155 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
156 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
157 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
158 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c1, vxb0c1));
159 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c1, vxb1c1));
160 const __m128i vb23c1 = _mm_and_si128(vb23c01, vmask);
161 const __m128i vsb23c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c1);
162 const __m128i vxb2c1 = _mm_unpacklo_epi8(vb23c1, vsb23c1);
163 const __m128i vxb3c1 = _mm_unpackhi_epi8(vb23c1, vsb23c1);
164
165 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
166 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
167 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
168 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
169 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c1, vxb2c1));
170 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c1, vxb3c1));
171
172
173 w = (const int8_t*) w + 32;
174 k -= 16 * sizeof(int8_t);
175 }
176
177 while (k >= 8 * sizeof(int8_t)) {
178 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
179 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
180 a0 += 8;
181 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
182 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
183 a1 += 8;
184 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
185 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
186 a2 += 8;
187
188 __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
189 vb01 = _mm_slli_epi32(vb01, 4);
190 vb01 = _mm_and_si128(vb01, vmask);
191
192 const __m128i vxbm1 = _mm_unpackhi_epi8(vb01, vb01);
193 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
194 const __m128i vxb1 = _mm_srai_epi16(vxbm1, 8);
195
196 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
197 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
198 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
199 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
200 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
201 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
202 __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
203 vb23 = _mm_slli_epi32(vb23, 4);
204 vb23 = _mm_and_si128(vb23, vmask);
205
206 const __m128i vxbm3 = _mm_unpackhi_epi8(vb23, vb23);
207 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
208 const __m128i vxb3 = _mm_srai_epi16(vxbm3, 8);
209
210 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
211 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
212 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
213 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
214 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
215 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
216
217 w = (const int8_t*) w + 32;
218 k -= 8 * sizeof(int8_t);
219 }
220 // accumulate float
221 const __m128 vfilter_output_scale0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_loadl_epi64((const __m128i*) w)), 16));
222 w = (const uint16_t*) w + 4;
223
224 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
225 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
226 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
227 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
228 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
229 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
230
231 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
232 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
233 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
234
235 vout0x0123 = _mm_add_ps(vout0x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc0x0123), vfilter_output_scale0123));
236 vout1x0123 = _mm_add_ps(vout1x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc1x0123), vfilter_output_scale0123));
237 vout2x0123 = _mm_add_ps(vout2x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc2x0123), vfilter_output_scale0123));
238 }
239
240 const __m128i vinput_scale01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
241 const __m128 vinput_scale0 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(1, 1, 1, 1)));
242 const __m128 vinput_scale1 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(3, 3, 3, 3)));
243 const __m128 vinput_scale2 = _mm_load1_ps(&quantization_params[2].inv_scale);
244
245 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
246 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
247 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale2);
248
249
250 const __m128 vbias0123 = _mm_loadu_ps((const float*) w);
251 w = (const float*) w + 4;
252 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
253 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
254 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
255
256 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
257 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
258 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
259
260 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
261 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
262 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
263
264 if XNN_LIKELY(nc >= 4) {
265 _mm_storeu_ps(c0, vout0x0123);
266 _mm_storeu_ps(c1, vout1x0123);
267 _mm_storeu_ps(c2, vout2x0123);
268
269 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
270 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
271 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
272
273 c0 = (float*) ((uintptr_t) c0 + cn_stride);
274 c1 = (float*) ((uintptr_t) c1 + cn_stride);
275 c2 = (float*) ((uintptr_t) c2 + cn_stride);
276
277 nc -= 4;
278 } else {
279 if (nc & 2) {
280 _mm_storel_pi((__m64*) c0, vout0x0123);
281 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
282 c0 += 2;
283 _mm_storel_pi((__m64*) c1, vout1x0123);
284 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
285 c1 += 2;
286 _mm_storel_pi((__m64*) c2, vout2x0123);
287 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
288 c2 += 2;
289 }
290 if (nc & 1) {
291 _mm_store_ss(c0, vout0x0123);
292 _mm_store_ss(c1, vout1x0123);
293 _mm_store_ss(c2, vout2x0123);
294 }
295 nc = 0;
296 }
297 } while (nc != 0);
298 }