Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-3x4c8-minmax-avx-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 #ifdef _MSC_VER
13 #include <intrin.h>
14 #else
15 #include <x86intrin.h>
16 #endif
17
18 #include "xnnpack/common.h"
19 #include "xnnpack/gemm.h"
20 #include "xnnpack/math.h"
21 #include "xnnpack/unaligned.h"
22
23
24 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_3x4c8__avx_ld128(
25 size_t mr,
26 size_t nc,
27 size_t kc,
28 const int8_t* restrict a,
29 size_t a_stride,
30 const void* restrict w,
31 float* restrict c,
32 size_t cm_stride,
33 size_t cn_stride,
34 const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
35 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 3);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(kc % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
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 __m128i vksum = _mm_load_si128((const __m128i*) w);
72 const __m128i vinput_zero_point0 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[0].zero_point));
73 const __m128i vinput_zero_point1 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[1].zero_point));
74 const __m128i vinput_zero_point2 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[2].zero_point));
75 const __m128i vzero = _mm_setzero_si128();
76 const __m128i vinit0 = _mm_mullo_epi32(vksum, vinput_zero_point0);
77 const __m128i vinit1 = _mm_mullo_epi32(vksum, vinput_zero_point1);
78 const __m128i vinit2 = _mm_mullo_epi32(vksum, vinput_zero_point2);
79 __m128i vacc0x0 = _mm_blend_epi16(vinit0, vzero, 0xFC);
80 __m128i vacc0x1 = _mm_blend_epi16(vinit0, vzero, 0xF3);
81 __m128i vacc0x2 = _mm_blend_epi16(vinit0, vzero, 0xCF);
82 __m128i vacc0x3 = _mm_blend_epi16(vinit0, vzero, 0x3F);
83 __m128i vacc1x0 = _mm_blend_epi16(vinit1, vzero, 0xFC);
84 __m128i vacc1x1 = _mm_blend_epi16(vinit1, vzero, 0xF3);
85 __m128i vacc1x2 = _mm_blend_epi16(vinit1, vzero, 0xCF);
86 __m128i vacc1x3 = _mm_blend_epi16(vinit1, vzero, 0x3F);
87 __m128i vacc2x0 = _mm_blend_epi16(vinit2, vzero, 0xFC);
88 __m128i vacc2x1 = _mm_blend_epi16(vinit2, vzero, 0xF3);
89 __m128i vacc2x2 = _mm_blend_epi16(vinit2, vzero, 0xCF);
90 __m128i vacc2x3 = _mm_blend_epi16(vinit2, vzero, 0x3F);
91 w = (const int32_t*) w + 4;
92
93 size_t k = kc;
94
95 while (k >= 16 * sizeof(int8_t)) {
96 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
97 const __m128i vxa0c0 = _mm_cvtepi8_epi16(va0c0);
98 a0 += 8;
99 const __m128i va1c0 = _mm_loadl_epi64((const __m128i*) a1);
100 const __m128i vxa1c0 = _mm_cvtepi8_epi16(va1c0);
101 a1 += 8;
102 const __m128i va2c0 = _mm_loadl_epi64((const __m128i*) a2);
103 const __m128i vxa2c0 = _mm_cvtepi8_epi16(va2c0);
104 a2 += 8;
105
106 const __m128i vb01c01 = _mm_load_si128((const __m128i*) w);
107 const __m128i vbs01c0 = _mm_slli_epi32(vb01c01, 4);
108 const __m128i vb01c0 = _mm_and_si128(vbs01c0, vmask);
109 const __m128i vsb01c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c0);
110 const __m128i vxb0c0 = _mm_unpacklo_epi8(vb01c0, vsb01c0);
111 const __m128i vxb1c0 = _mm_unpackhi_epi8(vb01c0, vsb01c0);
112
113 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
114 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
115 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
116 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
117 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c0, vxb0c0));
118 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c0, vxb1c0));
119 const __m128i vb23c01 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
120 const __m128i vbs23c0 = _mm_slli_epi32(vb23c01, 4);
121 const __m128i vb23c0 = _mm_and_si128(vbs23c0, vmask);
122 const __m128i vsb23c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c0);
123 const __m128i vxb2c0 = _mm_unpacklo_epi8(vb23c0, vsb23c0);
124 const __m128i vxb3c0 = _mm_unpackhi_epi8(vb23c0, vsb23c0);
125
126 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
127 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
128 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
129 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
130 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c0, vxb2c0));
131 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c0, vxb3c0));
132
133 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
134 const __m128i vxa0c1 = _mm_cvtepi8_epi16(va0c1);
135 a0 += 8;
136 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
137 const __m128i vxa1c1 = _mm_cvtepi8_epi16(va1c1);
138 a1 += 8;
139 const __m128i va2c1 = _mm_loadl_epi64((const __m128i*) a2);
140 const __m128i vxa2c1 = _mm_cvtepi8_epi16(va2c1);
141 a2 += 8;
142
143 const __m128i vb01c1 = _mm_and_si128(vb01c01, vmask);
144 const __m128i vsb01c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c1);
145 const __m128i vxb0c1 = _mm_unpacklo_epi8(vb01c1, vsb01c1);
146 const __m128i vxb1c1 = _mm_unpackhi_epi8(vb01c1, vsb01c1);
147
148 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
149 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
150 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
151 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
152 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c1, vxb0c1));
153 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c1, vxb1c1));
154 const __m128i vb23c1 = _mm_and_si128(vb23c01, vmask);
155 const __m128i vsb23c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c1);
156 const __m128i vxb2c1 = _mm_unpacklo_epi8(vb23c1, vsb23c1);
157 const __m128i vxb3c1 = _mm_unpackhi_epi8(vb23c1, vsb23c1);
158
159 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
160 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
161 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
162 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
163 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c1, vxb2c1));
164 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c1, vxb3c1));
165
166
167 w = (const int8_t*) w + 32;
168 k -= 16 * sizeof(int8_t);
169 }
170
171 while (k >= 8 * sizeof(int8_t)) {
172 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
173 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
174 a0 += 8;
175 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
176 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
177 a1 += 8;
178 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
179 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
180 a2 += 8;
181
182 __m128i vb01 = _mm_load_si128((const __m128i*) w);
183 vb01 = _mm_slli_epi32(vb01, 4);
184 vb01 = _mm_and_si128(vb01, vmask);
185
186 const __m128i vxbm1 = _mm_unpackhi_epi8(vb01, vb01);
187 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
188 const __m128i vxb1 = _mm_srai_epi16(vxbm1, 8);
189
190 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
191 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
192 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
193 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
194 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
195 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
196 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
197 vb23 = _mm_slli_epi32(vb23, 4);
198 vb23 = _mm_and_si128(vb23, vmask);
199
200 const __m128i vxbm3 = _mm_unpackhi_epi8(vb23, vb23);
201 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
202 const __m128i vxb3 = _mm_srai_epi16(vxbm3, 8);
203
204 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
205 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
206 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
207 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
208 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
209 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
210
211 w = (const int8_t*) w + 32;
212 k -= 8 * sizeof(int8_t);
213 }
214
215 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
216 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
217 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
218 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
219 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
220 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
221
222 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
223 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
224 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
225
226 vacc0x0123 = _mm_srai_epi32(vacc0x0123, 4);
227 vacc1x0123 = _mm_srai_epi32(vacc1x0123, 4);
228 vacc2x0123 = _mm_srai_epi32(vacc2x0123, 4);
229 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
230 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
231 __m128 vout2x0123 = _mm_cvtepi32_ps(vacc2x0123);
232
233 const __m128 vinput_scale0 = _mm_broadcast_ss(&quantization_params[0].inv_scale);
234 const __m128 vinput_scale1 = _mm_broadcast_ss(&quantization_params[1].inv_scale);
235 const __m128 vinput_scale2 = _mm_broadcast_ss(&quantization_params[2].inv_scale);
236
237 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
238 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
239 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale2);
240
241 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
242 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
243 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
244 vout2x0123 = _mm_mul_ps(vout2x0123, vfilter_output_scale0123);
245
246 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
247 w = (const float*) w + 8;
248 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
249 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
250 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
251
252 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
253 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
254 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
255
256 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
257 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
258 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
259
260 if XNN_LIKELY(nc >= 4) {
261 _mm_storeu_ps(c0, vout0x0123);
262 _mm_storeu_ps(c1, vout1x0123);
263 _mm_storeu_ps(c2, vout2x0123);
264
265 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
266 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
267 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
268
269 c0 = (float*) ((uintptr_t) c0 + cn_stride);
270 c1 = (float*) ((uintptr_t) c1 + cn_stride);
271 c2 = (float*) ((uintptr_t) c2 + cn_stride);
272
273 nc -= 4;
274 } else {
275 if (nc & 2) {
276 _mm_storel_pi((__m64*) c0, vout0x0123);
277 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
278 c0 += 2;
279 _mm_storel_pi((__m64*) c1, vout1x0123);
280 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
281 c1 += 2;
282 _mm_storel_pi((__m64*) c2, vout2x0123);
283 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
284 c2 += 2;
285 }
286 if (nc & 1) {
287 _mm_store_ss(c0, vout0x0123);
288 _mm_store_ss(c1, vout1x0123);
289 _mm_store_ss(c2, vout2x0123);
290 }
291 nc = 0;
292 }
293 } while (nc != 0);
294 }