Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-4x4c8-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
18
19 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_4x4c8__sse2_ld64(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 float* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
30 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(int8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 kc = round_up_po2(kc, 8 * sizeof(int8_t));
42 const int8_t* a0 = a;
43 float* c0 = c;
44 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
45 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 a1 = a0;
48 c1 = c0;
49 }
50 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 a2 = a1;
54 c2 = c1;
55 }
56 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
57 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
58 if XNN_UNPREDICTABLE(mr != 4) {
59 a3 = a2;
60 c3 = c2;
61 }
62
63 const __m128 vmin = _mm_set1_ps(params->scalar.min);
64 const __m128 vmax = _mm_set1_ps(params->scalar.max);
65 XNN_FORCE_REALIZATION(vmin);
66 XNN_FORCE_REALIZATION(vmax);
67
68
69 do {
70 const __m128i vksum = _mm_load_si128((const __m128i*) w);
71 const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
72 const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
73 const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));
74 const __m128i vinput_zero_point23 = _mm_loadu_si128((const __m128i*) &quantization_params[2]);
75 const __m128i vinput_zero_point2 = _mm_shuffle_epi32(vinput_zero_point23, _MM_SHUFFLE(0, 0, 0, 0));
76 const __m128i vinput_zero_point3 = _mm_shuffle_epi32(vinput_zero_point23, _MM_SHUFFLE(2, 2, 2, 2));
77 const __m128i vzero = _mm_setzero_si128();
78 const __m128i vksum_lo = _mm_srli_epi32(_mm_slli_epi32(vksum, 16), 16);
79 const __m128i vksum_hi = _mm_srli_epi32(vksum, 16);
80
81 __m128i vzpprodksumhi0 = _mm_mulhi_epu16(vinput_zero_point0, vksum_lo);
82 const __m128i vzpprodksumlo0 = _mm_mullo_epi16(vinput_zero_point0, vksum_lo);
83 __m128i vzpprodksumhi1 = _mm_mulhi_epu16(vinput_zero_point1, vksum_lo);
84 const __m128i vzpprodksumlo1 = _mm_mullo_epi16(vinput_zero_point1, vksum_lo);
85 __m128i vzpprodksumhi2 = _mm_mulhi_epu16(vinput_zero_point2, vksum_lo);
86 const __m128i vzpprodksumlo2 = _mm_mullo_epi16(vinput_zero_point2, vksum_lo);
87 __m128i vzpprodksumhi3 = _mm_mulhi_epu16(vinput_zero_point3, vksum_lo);
88 const __m128i vzpprodksumlo3 = _mm_mullo_epi16(vinput_zero_point3, vksum_lo);
89
90 vzpprodksumhi0 = _mm_add_epi16(vzpprodksumhi0, _mm_mullo_epi16(vinput_zero_point0, vksum_hi));
91 vzpprodksumhi0 = _mm_sub_epi16(vzpprodksumhi0, _mm_and_si128(_mm_srai_epi16(vinput_zero_point0, 15), vksum_lo));
92 vzpprodksumhi1 = _mm_add_epi16(vzpprodksumhi1, _mm_mullo_epi16(vinput_zero_point1, vksum_hi));
93 vzpprodksumhi1 = _mm_sub_epi16(vzpprodksumhi1, _mm_and_si128(_mm_srai_epi16(vinput_zero_point1, 15), vksum_lo));
94 vzpprodksumhi2 = _mm_add_epi16(vzpprodksumhi2, _mm_mullo_epi16(vinput_zero_point2, vksum_hi));
95 vzpprodksumhi2 = _mm_sub_epi16(vzpprodksumhi2, _mm_and_si128(_mm_srai_epi16(vinput_zero_point2, 15), vksum_lo));
96 vzpprodksumhi3 = _mm_add_epi16(vzpprodksumhi3, _mm_mullo_epi16(vinput_zero_point3, vksum_hi));
97 vzpprodksumhi3 = _mm_sub_epi16(vzpprodksumhi3, _mm_and_si128(_mm_srai_epi16(vinput_zero_point3, 15), vksum_lo));
98
99 vzpprodksumhi0 = _mm_slli_si128(vzpprodksumhi0, 2);
100 vzpprodksumhi1 = _mm_slli_si128(vzpprodksumhi1, 2);
101 vzpprodksumhi2 = _mm_slli_si128(vzpprodksumhi2, 2);
102 vzpprodksumhi3 = _mm_slli_si128(vzpprodksumhi3, 2);
103
104 const __m128i vksumzp0 = _mm_or_si128(vzpprodksumhi0, vzpprodksumlo0);
105 const __m128i vksumzp1 = _mm_or_si128(vzpprodksumhi1, vzpprodksumlo1);
106 const __m128i vksumzp2 = _mm_or_si128(vzpprodksumhi2, vzpprodksumlo2);
107 const __m128i vksumzp3 = _mm_or_si128(vzpprodksumhi3, vzpprodksumlo3);
108
109 const __m128i vksum010 = _mm_unpacklo_epi32(vksumzp0, vzero);
110 const __m128i vksum230 = _mm_unpackhi_epi32(vksumzp0, vzero);
111 const __m128i vksum011 = _mm_unpacklo_epi32(vksumzp1, vzero);
112 const __m128i vksum231 = _mm_unpackhi_epi32(vksumzp1, vzero);
113 const __m128i vksum012 = _mm_unpacklo_epi32(vksumzp2, vzero);
114 const __m128i vksum232 = _mm_unpackhi_epi32(vksumzp2, vzero);
115 const __m128i vksum013 = _mm_unpacklo_epi32(vksumzp3, vzero);
116 const __m128i vksum233 = _mm_unpackhi_epi32(vksumzp3, vzero);
117
118 __m128i vacc0x0 = _mm_unpacklo_epi64(vksum010, vzero);
119 __m128i vacc0x1 = _mm_unpackhi_epi64(vksum010, vzero);
120 __m128i vacc1x0 = _mm_unpacklo_epi64(vksum011, vzero);
121 __m128i vacc1x1 = _mm_unpackhi_epi64(vksum011, vzero);
122 __m128i vacc2x0 = _mm_unpacklo_epi64(vksum012, vzero);
123 __m128i vacc2x1 = _mm_unpackhi_epi64(vksum012, vzero);
124 __m128i vacc3x0 = _mm_unpacklo_epi64(vksum013, vzero);
125 __m128i vacc3x1 = _mm_unpackhi_epi64(vksum013, vzero);
126
127 __m128i vacc0x2 = _mm_unpacklo_epi64(vksum230, vzero);
128 __m128i vacc0x3 = _mm_unpackhi_epi64(vksum230, vzero);
129 __m128i vacc1x2 = _mm_unpacklo_epi64(vksum231, vzero);
130 __m128i vacc1x3 = _mm_unpackhi_epi64(vksum231, vzero);
131 __m128i vacc2x2 = _mm_unpacklo_epi64(vksum232, vzero);
132 __m128i vacc2x3 = _mm_unpackhi_epi64(vksum232, vzero);
133 __m128i vacc3x2 = _mm_unpacklo_epi64(vksum233, vzero);
134 __m128i vacc3x3 = _mm_unpackhi_epi64(vksum233, vzero);
135 w = (const int32_t*) w + 4;
136
137 size_t k = kc;
138
139
140 while (k >= 8 * sizeof(int8_t)) {
141 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
142 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
143 a0 += 8;
144 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
145 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
146 a1 += 8;
147 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
148 const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
149 a2 += 8;
150 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
151 const __m128i vxa3 = _mm_srai_epi16(_mm_unpacklo_epi8(va3, va3), 8);
152 a3 += 8;
153
154 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
155
156 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
157
158 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
159 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
160 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
161 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3, vxb0));
162 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
163
164 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
165
166 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
167 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
168 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
169 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3, vxb1));
170 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
171
172 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
173
174 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
175 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
176 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
177 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3, vxb2));
178 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
179
180 const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
181
182 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
183 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
184 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
185 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3, vxb3));
186
187 w = (const int8_t*) w + 32;
188 k -= 8 * sizeof(int8_t);
189 }
190
191 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
192 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
193 const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
194 const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
195 const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(vacc2x0, vacc2x2));
196 const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(vacc2x1, vacc2x3));
197 const __m128i vacc3x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc3x0, vacc3x2), _mm_unpackhi_epi32(vacc3x0, vacc3x2));
198 const __m128i vacc3x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc3x1, vacc3x3), _mm_unpackhi_epi32(vacc3x1, vacc3x3));
199
200 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
201 __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
202 __m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc2x02, vacc2x13));
203 __m128i vacc3x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc3x02, vacc3x13), _mm_unpackhi_epi32(vacc3x02, vacc3x13));
204
205 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
206 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
207 __m128 vout2x0123 = _mm_cvtepi32_ps(vacc2x0123);
208 __m128 vout3x0123 = _mm_cvtepi32_ps(vacc3x0123);
209
210 const __m128i vinput_scale01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
211 const __m128 vinput_scale0 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(1, 1, 1, 1)));
212 const __m128 vinput_scale1 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(3, 3, 3, 3)));
213 const __m128i vinput_scale23 = _mm_loadu_si128((const __m128i*) &quantization_params[2]);
214 const __m128 vinput_scale2 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale23, _MM_SHUFFLE(1, 1, 1, 1)));
215 const __m128 vinput_scale3 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale23, _MM_SHUFFLE(3, 3, 3, 3)));
216
217 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
218 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
219 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale2);
220 vout3x0123 = _mm_mul_ps(vout3x0123, vinput_scale3);
221
222 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
223 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
224 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
225 vout2x0123 = _mm_mul_ps(vout2x0123, vfilter_output_scale0123);
226 vout3x0123 = _mm_mul_ps(vout3x0123, vfilter_output_scale0123);
227
228 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
229 w = (const float*) w + 8;
230 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
231 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
232 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
233 vout3x0123 = _mm_add_ps(vout3x0123, vbias0123);
234
235 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
236 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
237 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
238 vout3x0123 = _mm_max_ps(vout3x0123, vmin);
239
240 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
241 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
242 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
243 vout3x0123 = _mm_min_ps(vout3x0123, vmax);
244
245 if XNN_LIKELY(nc >= 4) {
246 _mm_storeu_ps(c0, vout0x0123);
247 _mm_storeu_ps(c1, vout1x0123);
248 _mm_storeu_ps(c2, vout2x0123);
249 _mm_storeu_ps(c3, vout3x0123);
250
251 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
252 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
253 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
254 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
255
256 c0 = (float*) ((uintptr_t) c0 + cn_stride);
257 c1 = (float*) ((uintptr_t) c1 + cn_stride);
258 c2 = (float*) ((uintptr_t) c2 + cn_stride);
259 c3 = (float*) ((uintptr_t) c3 + cn_stride);
260
261 nc -= 4;
262 } else {
263 if (nc & 2) {
264 _mm_storel_pi((__m64*) c0, vout0x0123);
265 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
266 c0 += 2;
267 _mm_storel_pi((__m64*) c1, vout1x0123);
268 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
269 c1 += 2;
270 _mm_storel_pi((__m64*) c2, vout2x0123);
271 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
272 c2 += 2;
273 _mm_storel_pi((__m64*) c3, vout3x0123);
274 vout3x0123 = _mm_unpackhi_ps(vout3x0123, vout3x0123);
275 c3 += 2;
276 }
277 if (nc & 1) {
278 _mm_store_ss(c0, vout0x0123);
279 _mm_store_ss(c1, vout1x0123);
280 _mm_store_ss(c2, vout2x0123);
281 _mm_store_ss(c3, vout3x0123);
282 }
283 nc = 0;
284 }
285 } while (nc != 0);
286 }