Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-2x4c8-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 #include "xnnpack/unaligned.h"
18
19
20 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_2x4c8__sse2_ld64(
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 <= 2);
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
52 const __m128 vmin = _mm_set1_ps(params->scalar.min);
53 const __m128 vmax = _mm_set1_ps(params->scalar.max);
54 XNN_FORCE_REALIZATION(vmin);
55 XNN_FORCE_REALIZATION(vmax);
56
57 const __m128i vmask = _mm_set1_epi8(0xF0);
58 XNN_FORCE_REALIZATION(vmask);
59
60 do {
61 const __m128i vksum = _mm_load_si128((const __m128i*) w);
62 const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
63 const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
64 const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));
65 const __m128i vzero = _mm_setzero_si128();
66 const __m128i vksum_lo = _mm_srli_epi32(_mm_slli_epi32(vksum, 16), 16);
67 const __m128i vksum_hi = _mm_srli_epi32(vksum, 16);
68
69 __m128i vzpprodksumhi0 = _mm_mulhi_epu16(vinput_zero_point0, vksum_lo);
70 const __m128i vzpprodksumlo0 = _mm_mullo_epi16(vinput_zero_point0, vksum_lo);
71 __m128i vzpprodksumhi1 = _mm_mulhi_epu16(vinput_zero_point1, vksum_lo);
72 const __m128i vzpprodksumlo1 = _mm_mullo_epi16(vinput_zero_point1, vksum_lo);
73
74 vzpprodksumhi0 = _mm_add_epi16(vzpprodksumhi0, _mm_mullo_epi16(vinput_zero_point0, vksum_hi));
75 vzpprodksumhi0 = _mm_sub_epi16(vzpprodksumhi0, _mm_and_si128(_mm_srai_epi16(vinput_zero_point0, 15), vksum_lo));
76 vzpprodksumhi1 = _mm_add_epi16(vzpprodksumhi1, _mm_mullo_epi16(vinput_zero_point1, vksum_hi));
77 vzpprodksumhi1 = _mm_sub_epi16(vzpprodksumhi1, _mm_and_si128(_mm_srai_epi16(vinput_zero_point1, 15), vksum_lo));
78
79 vzpprodksumhi0 = _mm_slli_si128(vzpprodksumhi0, 2);
80 vzpprodksumhi1 = _mm_slli_si128(vzpprodksumhi1, 2);
81
82 const __m128i vksumzp0 = _mm_or_si128(vzpprodksumhi0, vzpprodksumlo0);
83 const __m128i vksumzp1 = _mm_or_si128(vzpprodksumhi1, vzpprodksumlo1);
84
85 const __m128i vksum010 = _mm_unpacklo_epi32(vksumzp0, vzero);
86 const __m128i vksum230 = _mm_unpackhi_epi32(vksumzp0, vzero);
87 const __m128i vksum011 = _mm_unpacklo_epi32(vksumzp1, vzero);
88 const __m128i vksum231 = _mm_unpackhi_epi32(vksumzp1, vzero);
89
90 __m128i vacc0x0 = _mm_unpacklo_epi64(vksum010, vzero);
91 __m128i vacc0x1 = _mm_unpackhi_epi64(vksum010, vzero);
92 __m128i vacc1x0 = _mm_unpacklo_epi64(vksum011, vzero);
93 __m128i vacc1x1 = _mm_unpackhi_epi64(vksum011, vzero);
94
95 __m128i vacc0x2 = _mm_unpacklo_epi64(vksum230, vzero);
96 __m128i vacc0x3 = _mm_unpackhi_epi64(vksum230, vzero);
97 __m128i vacc1x2 = _mm_unpacklo_epi64(vksum231, vzero);
98 __m128i vacc1x3 = _mm_unpackhi_epi64(vksum231, vzero);
99 w = (const int32_t*) w + 4;
100
101 size_t k = kc;
102
103 while (k >= 16 * sizeof(int8_t)) {
104 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
105 const __m128i vxa0c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c0, va0c0), 8);
106 a0 += 8;
107 const __m128i va1c0 = _mm_loadl_epi64((const __m128i*) a1);
108 const __m128i vxa1c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va1c0, va1c0), 8);
109 a1 += 8;
110
111 const __m128i vb0c01 = _mm_loadl_epi64((const __m128i*) w);
112 const __m128i vbs0c0 = _mm_slli_epi32(vb0c01, 4);
113 const __m128i vb0c0 = _mm_and_si128(vbs0c0, vmask);
114 const __m128i vxb0c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0c0, vb0c0), 8);
115
116 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
117 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
118 const __m128i vb1c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
119 const __m128i vbs1c0 = _mm_slli_epi32(vb1c01, 4);
120 const __m128i vb1c0 = _mm_and_si128(vbs1c0, vmask);
121 const __m128i vxb1c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1c0, vb1c0), 8);
122
123 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
124 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
125 const __m128i vb2c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
126 const __m128i vbs2c0 = _mm_slli_epi32(vb2c01, 4);
127 const __m128i vb2c0 = _mm_and_si128(vbs2c0, vmask);
128 const __m128i vxb2c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2c0, vb2c0), 8);
129
130 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
131 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
132 const __m128i vb3c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
133 const __m128i vbs3c0 = _mm_slli_epi32(vb3c01, 4);
134 const __m128i vb3c0 = _mm_and_si128(vbs3c0, vmask);
135 const __m128i vxb3c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3c0, vb3c0), 8);
136
137 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
138 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
139
140 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
141 const __m128i vxa0c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c1, va0c1), 8);
142 a0 += 8;
143 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
144 const __m128i vxa1c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1c1, va1c1), 8);
145 a1 += 8;
146
147 const __m128i vb0c1 = _mm_and_si128(vb0c01, vmask);
148 const __m128i vxb0c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0c1, vb0c1), 8);
149
150 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
151 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
152 const __m128i vb1c1 = _mm_and_si128(vb1c01, vmask);
153 const __m128i vxb1c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1c1, vb1c1), 8);
154
155 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
156 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
157 const __m128i vb2c1 = _mm_and_si128(vb2c01, vmask);
158 const __m128i vxb2c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2c1, vb2c1), 8);
159
160 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
161 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
162 const __m128i vb3c1 = _mm_and_si128(vb3c01, vmask);
163 const __m128i vxb3c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3c1, vb3c1), 8);
164
165 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
166 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
167
168
169 w = (const int8_t*) w + 32;
170 k -= 16 * sizeof(int8_t);
171 }
172
173 while (k >= 8 * sizeof(int8_t)) {
174 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
175 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
176 a0 += 8;
177 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
178 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
179 a1 += 8;
180
181 __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
182 vb0 = _mm_slli_epi32(vb0, 4);
183 vb0 = _mm_and_si128(vb0, vmask);
184
185 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
186
187 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
188 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
189 __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
190 vb1 = _mm_slli_epi32(vb1, 4);
191 vb1 = _mm_and_si128(vb1, vmask);
192
193 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
194
195 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
196 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
197 __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
198 vb2 = _mm_slli_epi32(vb2, 4);
199 vb2 = _mm_and_si128(vb2, vmask);
200
201 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
202
203 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
204 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
205 __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
206 vb3 = _mm_slli_epi32(vb3, 4);
207 vb3 = _mm_and_si128(vb3, vmask);
208
209 const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
210
211 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
212 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
213
214 w = (const int8_t*) w + 32;
215 k -= 8 * sizeof(int8_t);
216 }
217
218 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
219 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
220 const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
221 const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
222
223 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
224 __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
225
226 vacc0x0123 = _mm_srai_epi32(vacc0x0123, 4);
227 vacc1x0123 = _mm_srai_epi32(vacc1x0123, 4);
228 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
229 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
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
235 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
236 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
237
238 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
239 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
240 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
241
242 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
243 w = (const float*) w + 8;
244 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
245 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
246
247 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
248 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
249
250 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
251 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
252
253 if XNN_LIKELY(nc >= 4) {
254 _mm_storeu_ps(c0, vout0x0123);
255 _mm_storeu_ps(c1, vout1x0123);
256
257 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
258 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
259
260 c0 = (float*) ((uintptr_t) c0 + cn_stride);
261 c1 = (float*) ((uintptr_t) c1 + cn_stride);
262
263 nc -= 4;
264 } else {
265 if (nc & 2) {
266 _mm_storel_pi((__m64*) c0, vout0x0123);
267 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
268 c0 += 2;
269 _mm_storel_pi((__m64*) c1, vout1x0123);
270 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
271 c1 += 2;
272 }
273 if (nc & 1) {
274 _mm_store_ss(c0, vout0x0123);
275 _mm_store_ss(c1, vout1x0123);
276 }
277 nc = 0;
278 }
279 } while (nc != 0);
280 }