Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-4x4c8-minmax-sse2-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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/igemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_4x4c8__sse2_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const int8_t* zero_data,
31 const union xnn_f32_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(ks != 0);
39 assert(ks % (4 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 float* c0 = c;
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 c3 = c2;
58 }
59
60 const __m128 vmin = _mm_set1_ps(params->scalar.min);
61 const __m128 vmax = _mm_set1_ps(params->scalar.max);
62 XNN_FORCE_REALIZATION(vmin);
63 XNN_FORCE_REALIZATION(vmax);
64 __m128i vinput_zero_point = _mm_cvtsi32_si128(*((const int*) &quantization_params->zero_point));
65 vinput_zero_point = _mm_shuffle_epi32(vinput_zero_point, _MM_SHUFFLE(0, 0, 0, 0));
66 const __m128 vinput_scale = _mm_load1_ps(&quantization_params->inv_scale);
67
68
69 do {
70 const __m128i vksum = _mm_load_si128((const __m128i*) w);
71 const __m128i vzero = _mm_setzero_si128();
72 const __m128i vksum_lo = _mm_srli_epi32(_mm_slli_epi32(vksum, 16), 16);
73 const __m128i vksum_hi = _mm_srli_epi32(vksum, 16);
74
75 __m128i vzpprodksumhi0 = _mm_mulhi_epu16(vinput_zero_point, vksum_lo);
76 const __m128i vzpprodksumlo0 = _mm_mullo_epi16(vinput_zero_point, vksum_lo);
77 __m128i vzpprodksumhi1 = _mm_mulhi_epu16(vinput_zero_point, vksum_lo);
78 const __m128i vzpprodksumlo1 = _mm_mullo_epi16(vinput_zero_point, vksum_lo);
79 __m128i vzpprodksumhi2 = _mm_mulhi_epu16(vinput_zero_point, vksum_lo);
80 const __m128i vzpprodksumlo2 = _mm_mullo_epi16(vinput_zero_point, vksum_lo);
81 __m128i vzpprodksumhi3 = _mm_mulhi_epu16(vinput_zero_point, vksum_lo);
82 const __m128i vzpprodksumlo3 = _mm_mullo_epi16(vinput_zero_point, vksum_lo);
83
84 vzpprodksumhi0 = _mm_add_epi16(vzpprodksumhi0, _mm_mullo_epi16(vinput_zero_point, vksum_hi));
85 vzpprodksumhi0 = _mm_sub_epi16(vzpprodksumhi0, _mm_and_si128(_mm_srai_epi16(vinput_zero_point, 15), vksum_lo));
86 vzpprodksumhi1 = _mm_add_epi16(vzpprodksumhi1, _mm_mullo_epi16(vinput_zero_point, vksum_hi));
87 vzpprodksumhi1 = _mm_sub_epi16(vzpprodksumhi1, _mm_and_si128(_mm_srai_epi16(vinput_zero_point, 15), vksum_lo));
88 vzpprodksumhi2 = _mm_add_epi16(vzpprodksumhi2, _mm_mullo_epi16(vinput_zero_point, vksum_hi));
89 vzpprodksumhi2 = _mm_sub_epi16(vzpprodksumhi2, _mm_and_si128(_mm_srai_epi16(vinput_zero_point, 15), vksum_lo));
90 vzpprodksumhi3 = _mm_add_epi16(vzpprodksumhi3, _mm_mullo_epi16(vinput_zero_point, vksum_hi));
91 vzpprodksumhi3 = _mm_sub_epi16(vzpprodksumhi3, _mm_and_si128(_mm_srai_epi16(vinput_zero_point, 15), vksum_lo));
92
93 vzpprodksumhi0 = _mm_slli_si128(vzpprodksumhi0, 2);
94 vzpprodksumhi1 = _mm_slli_si128(vzpprodksumhi1, 2);
95 vzpprodksumhi2 = _mm_slli_si128(vzpprodksumhi2, 2);
96 vzpprodksumhi3 = _mm_slli_si128(vzpprodksumhi3, 2);
97
98 const __m128i vksumzp0 = _mm_or_si128(vzpprodksumhi0, vzpprodksumlo0);
99 const __m128i vksumzp1 = _mm_or_si128(vzpprodksumhi1, vzpprodksumlo1);
100 const __m128i vksumzp2 = _mm_or_si128(vzpprodksumhi2, vzpprodksumlo2);
101 const __m128i vksumzp3 = _mm_or_si128(vzpprodksumhi3, vzpprodksumlo3);
102
103 const __m128i vksum010 = _mm_unpacklo_epi32(vksumzp0, vzero);
104 const __m128i vksum230 = _mm_unpackhi_epi32(vksumzp0, vzero);
105 const __m128i vksum011 = _mm_unpacklo_epi32(vksumzp1, vzero);
106 const __m128i vksum231 = _mm_unpackhi_epi32(vksumzp1, vzero);
107 const __m128i vksum012 = _mm_unpacklo_epi32(vksumzp2, vzero);
108 const __m128i vksum232 = _mm_unpackhi_epi32(vksumzp2, vzero);
109 const __m128i vksum013 = _mm_unpacklo_epi32(vksumzp3, vzero);
110 const __m128i vksum233 = _mm_unpackhi_epi32(vksumzp3, vzero);
111
112 __m128i vacc0x0 = _mm_unpacklo_epi64(vksum010, vzero);
113 __m128i vacc0x1 = _mm_unpackhi_epi64(vksum010, vzero);
114 __m128i vacc1x0 = _mm_unpacklo_epi64(vksum011, vzero);
115 __m128i vacc1x1 = _mm_unpackhi_epi64(vksum011, vzero);
116 __m128i vacc2x0 = _mm_unpacklo_epi64(vksum012, vzero);
117 __m128i vacc2x1 = _mm_unpackhi_epi64(vksum012, vzero);
118 __m128i vacc3x0 = _mm_unpacklo_epi64(vksum013, vzero);
119 __m128i vacc3x1 = _mm_unpackhi_epi64(vksum013, vzero);
120
121 __m128i vacc0x2 = _mm_unpacklo_epi64(vksum230, vzero);
122 __m128i vacc0x3 = _mm_unpackhi_epi64(vksum230, vzero);
123 __m128i vacc1x2 = _mm_unpacklo_epi64(vksum231, vzero);
124 __m128i vacc1x3 = _mm_unpackhi_epi64(vksum231, vzero);
125 __m128i vacc2x2 = _mm_unpacklo_epi64(vksum232, vzero);
126 __m128i vacc2x3 = _mm_unpackhi_epi64(vksum232, vzero);
127 __m128i vacc3x2 = _mm_unpacklo_epi64(vksum233, vzero);
128 __m128i vacc3x3 = _mm_unpackhi_epi64(vksum233, vzero);
129 w = (const int32_t*) w + 4;
130
131 size_t p = ks;
132 do {
133 const int8_t* restrict a0 = a[0];
134 if XNN_UNPREDICTABLE(a0 != zero) {
135 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
136 } else {
137 a0 = zero_data;
138 }
139 const int8_t* restrict a1 = a[1];
140 if XNN_UNPREDICTABLE(a1 != zero) {
141 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
142 } else {
143 a1 = zero_data;
144 }
145 const int8_t* restrict a2 = a[2];
146 if XNN_UNPREDICTABLE(a2 != zero) {
147 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
148 } else {
149 a2 = zero_data;
150 }
151 const int8_t* restrict a3 = a[3];
152 if XNN_UNPREDICTABLE(a3 != zero) {
153 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
154 } else {
155 a3 = zero_data;
156 }
157 a += 4;
158
159 size_t k = 0;
160 while (k < kc) {
161 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
162 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
163 a0 += 8;
164 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
165 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
166 a1 += 8;
167 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
168 const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
169 a2 += 8;
170 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
171 const __m128i vxa3 = _mm_srai_epi16(_mm_unpacklo_epi8(va3, va3), 8);
172 a3 += 8;
173
174 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
175 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
176
177 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
178 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
179 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
180 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3, vxb0));
181 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
182 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
183
184 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
185 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
186 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
187 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3, vxb1));
188 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
189 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
190
191 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
192 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
193 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
194 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3, vxb2));
195 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
196 const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
197
198 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
199 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
200 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
201 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3, vxb3));
202
203 w = (const void*) ((const int8_t*) w + 32);
204 k += 8 * sizeof(int8_t);
205 }
206 p -= 4 * sizeof(void*);
207 } while (p != 0);
208
209 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
210 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
211 const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
212 const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
213 const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(vacc2x0, vacc2x2));
214 const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(vacc2x1, vacc2x3));
215 const __m128i vacc3x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc3x0, vacc3x2), _mm_unpackhi_epi32(vacc3x0, vacc3x2));
216 const __m128i vacc3x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc3x1, vacc3x3), _mm_unpackhi_epi32(vacc3x1, vacc3x3));
217
218 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
219 __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
220 __m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc2x02, vacc2x13));
221 __m128i vacc3x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc3x02, vacc3x13), _mm_unpackhi_epi32(vacc3x02, vacc3x13));
222
223 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
224 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
225 __m128 vout2x0123 = _mm_cvtepi32_ps(vacc2x0123);
226 __m128 vout3x0123 = _mm_cvtepi32_ps(vacc3x0123);
227
228 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale);
229 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale);
230 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale);
231 vout3x0123 = _mm_mul_ps(vout3x0123, vinput_scale);
232
233 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
234 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
235 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
236 vout2x0123 = _mm_mul_ps(vout2x0123, vfilter_output_scale0123);
237 vout3x0123 = _mm_mul_ps(vout3x0123, vfilter_output_scale0123);
238
239 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
240 w = (const float*) w + 8;
241 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
242 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
243 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
244 vout3x0123 = _mm_add_ps(vout3x0123, vbias0123);
245
246 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
247 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
248 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
249 vout3x0123 = _mm_max_ps(vout3x0123, vmin);
250
251 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
252 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
253 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
254 vout3x0123 = _mm_min_ps(vout3x0123, vmax);
255
256 if XNN_LIKELY(nc >= 4) {
257 _mm_storeu_ps(c3, vout3x0123);
258 _mm_storeu_ps(c2, vout2x0123);
259 _mm_storeu_ps(c1, vout1x0123);
260 _mm_storeu_ps(c0, vout0x0123);
261
262 a = (const int8_t**restrict) ((uintptr_t) a - ks);
263
264 c0 = (float*) ((uintptr_t) c0 + cn_stride);
265 c1 = (float*) ((uintptr_t) c1 + cn_stride);
266 c2 = (float*) ((uintptr_t) c2 + cn_stride);
267 c3 = (float*) ((uintptr_t) c3 + cn_stride);
268
269 nc -= 4;
270 } else {
271 if (nc & 2) {
272 _mm_storel_pi((__m64*) c3, vout3x0123);
273 vout3x0123 = _mm_unpackhi_ps(vout3x0123, vout3x0123);
274 c3 += 2;
275 _mm_storel_pi((__m64*) c2, vout2x0123);
276 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
277 c2 += 2;
278 _mm_storel_pi((__m64*) c1, vout1x0123);
279 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
280 c1 += 2;
281 _mm_storel_pi((__m64*) c0, vout0x0123);
282 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
283 c0 += 2;
284 }
285 if (nc & 1) {
286 _mm_store_ss(c3, vout3x0123);
287 _mm_store_ss(c2, vout2x0123);
288 _mm_store_ss(c1, vout1x0123);
289 _mm_store_ss(c0, vout0x0123);
290 }
291 nc = 0;
292 }
293 } while (nc != 0);
294 }