Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-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_qc4w_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_qc4w_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 kc = round_up_po2(kc, 8 * sizeof(int8_t));
43 const int8_t* a0 = a;
44 float* c0 = c;
45 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
46 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 a1 = a0;
49 c1 = c0;
50 }
51 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
52 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 a2 = a1;
55 c2 = c1;
56 }
57
58 const __m128 vmin = _mm_set1_ps(params->scalar.min);
59 const __m128 vmax = _mm_set1_ps(params->scalar.max);
60 XNN_FORCE_REALIZATION(vmin);
61 XNN_FORCE_REALIZATION(vmax);
62
63 const __m128i vmask = _mm_set1_epi8(0xF0);
64 XNN_FORCE_REALIZATION(vmask);
65
66 do {
67 const __m128i vksum = _mm_load_si128((const __m128i*) w);
68 const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
69 const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
70 const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));
71 __m128i vinput_zero_point2 = _mm_cvtsi32_si128(*((const int*) &quantization_params[2].zero_point));
72 vinput_zero_point2 = _mm_shuffle_epi32(vinput_zero_point2, _MM_SHUFFLE(0, 0, 0, 0));
73 const __m128i vzero = _mm_setzero_si128();
74 const __m128i vinit0 = _mm_mullo_epi32(vksum, vinput_zero_point0);
75 const __m128i vinit1 = _mm_mullo_epi32(vksum, vinput_zero_point1);
76 const __m128i vinit2 = _mm_mullo_epi32(vksum, vinput_zero_point2);
77 __m128i vacc0x0 = _mm_blend_epi16(vinit0, vzero, 0xFC);
78 __m128i vacc0x1 = _mm_blend_epi16(vinit0, vzero, 0xF3);
79 __m128i vacc0x2 = _mm_blend_epi16(vinit0, vzero, 0xCF);
80 __m128i vacc0x3 = _mm_blend_epi16(vinit0, vzero, 0x3F);
81 __m128i vacc1x0 = _mm_blend_epi16(vinit1, vzero, 0xFC);
82 __m128i vacc1x1 = _mm_blend_epi16(vinit1, vzero, 0xF3);
83 __m128i vacc1x2 = _mm_blend_epi16(vinit1, vzero, 0xCF);
84 __m128i vacc1x3 = _mm_blend_epi16(vinit1, vzero, 0x3F);
85 __m128i vacc2x0 = _mm_blend_epi16(vinit2, vzero, 0xFC);
86 __m128i vacc2x1 = _mm_blend_epi16(vinit2, vzero, 0xF3);
87 __m128i vacc2x2 = _mm_blend_epi16(vinit2, vzero, 0xCF);
88 __m128i vacc2x3 = _mm_blend_epi16(vinit2, vzero, 0x3F);
89 w = (const int32_t*) w + 4;
90
91 size_t k = kc;
92
93 while (k >= 16 * sizeof(int8_t)) {
94 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
95 const __m128i vxa0c0 = _mm_cvtepi8_epi16(va0c0);
96 a0 += 8;
97 const __m128i va1c0 = _mm_loadl_epi64((const __m128i*) a1);
98 const __m128i vxa1c0 = _mm_cvtepi8_epi16(va1c0);
99 a1 += 8;
100 const __m128i va2c0 = _mm_loadl_epi64((const __m128i*) a2);
101 const __m128i vxa2c0 = _mm_cvtepi8_epi16(va2c0);
102 a2 += 8;
103
104 const __m128i vb01c01 = _mm_load_si128((const __m128i*) w);
105 const __m128i vbs01c0 = _mm_slli_epi32(vb01c01, 4);
106 const __m128i vb01c0 = _mm_and_si128(vbs01c0, vmask);
107 const __m128i vsb01c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c0);
108 const __m128i vxb0c0 = _mm_unpacklo_epi8(vb01c0, vsb01c0);
109 const __m128i vxb1c0 = _mm_unpackhi_epi8(vb01c0, vsb01c0);
110
111 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
112 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
113 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
114 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
115 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c0, vxb0c0));
116 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c0, vxb1c0));
117 const __m128i vb23c01 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
118 const __m128i vbs23c0 = _mm_slli_epi32(vb23c01, 4);
119 const __m128i vb23c0 = _mm_and_si128(vbs23c0, vmask);
120 const __m128i vsb23c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c0);
121 const __m128i vxb2c0 = _mm_unpacklo_epi8(vb23c0, vsb23c0);
122 const __m128i vxb3c0 = _mm_unpackhi_epi8(vb23c0, vsb23c0);
123
124 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
125 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
126 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
127 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
128 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c0, vxb2c0));
129 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c0, vxb3c0));
130
131 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
132 const __m128i vxa0c1 = _mm_cvtepi8_epi16(va0c1);
133 a0 += 8;
134 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
135 const __m128i vxa1c1 = _mm_cvtepi8_epi16(va1c1);
136 a1 += 8;
137 const __m128i va2c1 = _mm_loadl_epi64((const __m128i*) a2);
138 const __m128i vxa2c1 = _mm_cvtepi8_epi16(va2c1);
139 a2 += 8;
140
141 const __m128i vb01c1 = _mm_and_si128(vb01c01, vmask);
142 const __m128i vsb01c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c1);
143 const __m128i vxb0c1 = _mm_unpacklo_epi8(vb01c1, vsb01c1);
144 const __m128i vxb1c1 = _mm_unpackhi_epi8(vb01c1, vsb01c1);
145
146 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
147 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
148 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
149 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
150 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c1, vxb0c1));
151 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c1, vxb1c1));
152 const __m128i vb23c1 = _mm_and_si128(vb23c01, vmask);
153 const __m128i vsb23c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c1);
154 const __m128i vxb2c1 = _mm_unpacklo_epi8(vb23c1, vsb23c1);
155 const __m128i vxb3c1 = _mm_unpackhi_epi8(vb23c1, vsb23c1);
156
157 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
158 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
159 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
160 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
161 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c1, vxb2c1));
162 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c1, vxb3c1));
163
164
165 w = (const int8_t*) w + 32;
166 k -= 16 * sizeof(int8_t);
167 }
168
169 while (k >= 8 * sizeof(int8_t)) {
170 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
171 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
172 a0 += 8;
173 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
174 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
175 a1 += 8;
176 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
177 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
178 a2 += 8;
179
180 __m128i vb01 = _mm_load_si128((const __m128i*) w);
181 vb01 = _mm_slli_epi32(vb01, 4);
182 vb01 = _mm_and_si128(vb01, vmask);
183
184 const __m128i vxbm1 = _mm_unpackhi_epi8(vb01, vb01);
185 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
186 const __m128i vxb1 = _mm_srai_epi16(vxbm1, 8);
187
188 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
189 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
190 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
191 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
192 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
193 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
194 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
195 vb23 = _mm_slli_epi32(vb23, 4);
196 vb23 = _mm_and_si128(vb23, vmask);
197
198 const __m128i vxbm3 = _mm_unpackhi_epi8(vb23, vb23);
199 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
200 const __m128i vxb3 = _mm_srai_epi16(vxbm3, 8);
201
202 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
203 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
204 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
205 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
206 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
207 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
208
209 w = (const int8_t*) w + 32;
210 k -= 8 * sizeof(int8_t);
211 }
212
213 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
214 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
215 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
216 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
217 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
218 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
219
220 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
221 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
222 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
223
224 vacc0x0123 = _mm_srai_epi32(vacc0x0123, 4);
225 vacc1x0123 = _mm_srai_epi32(vacc1x0123, 4);
226 vacc2x0123 = _mm_srai_epi32(vacc2x0123, 4);
227 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
228 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
229 __m128 vout2x0123 = _mm_cvtepi32_ps(vacc2x0123);
230
231 const __m128i vinput_scale01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
232 const __m128 vinput_scale0 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(1, 1, 1, 1)));
233 const __m128 vinput_scale1 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(3, 3, 3, 3)));
234 const __m128 vinput_scale2 = _mm_load1_ps(&quantization_params[2].inv_scale);
235
236 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
237 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
238 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale2);
239
240 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
241 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
242 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
243 vout2x0123 = _mm_mul_ps(vout2x0123, vfilter_output_scale0123);
244
245 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
246 w = (const float*) w + 8;
247 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
248 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
249 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
250
251 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
252 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
253 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
254
255 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
256 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
257 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
258
259 if XNN_LIKELY(nc >= 4) {
260 _mm_storeu_ps(c0, vout0x0123);
261 _mm_storeu_ps(c1, vout1x0123);
262 _mm_storeu_ps(c2, vout2x0123);
263
264 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
265 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
266 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
267
268 c0 = (float*) ((uintptr_t) c0 + cn_stride);
269 c1 = (float*) ((uintptr_t) c1 + cn_stride);
270 c2 = (float*) ((uintptr_t) c2 + cn_stride);
271
272 nc -= 4;
273 } else {
274 if (nc & 2) {
275 _mm_storel_pi((__m64*) c0, vout0x0123);
276 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
277 c0 += 2;
278 _mm_storel_pi((__m64*) c1, vout1x0123);
279 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
280 c1 += 2;
281 _mm_storel_pi((__m64*) c2, vout2x0123);
282 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
283 c2 += 2;
284 }
285 if (nc & 1) {
286 _mm_store_ss(c0, vout0x0123);
287 _mm_store_ss(c1, vout1x0123);
288 _mm_store_ss(c2, vout2x0123);
289 }
290 nc = 0;
291 }
292 } while (nc != 0);
293 }