Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-2x4c8-minmax-avx-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 #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_2x4c8__avx_ld64(
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 <= 2);
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
56 const __m128 vmin = _mm_set1_ps(params->scalar.min);
57 const __m128 vmax = _mm_set1_ps(params->scalar.max);
58 XNN_FORCE_REALIZATION(vmin);
59 XNN_FORCE_REALIZATION(vmax);
60
61 const __m128i vmask = _mm_set1_epi8(0xF0);
62 XNN_FORCE_REALIZATION(vmask);
63
64 do {
65 const __m128i vksum = _mm_load_si128((const __m128i*) w);
66 const __m128i vinput_zero_point0 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[0].zero_point));
67 const __m128i vinput_zero_point1 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[1].zero_point));
68 const __m128i vzero = _mm_setzero_si128();
69 const __m128i vinit0 = _mm_mullo_epi32(vksum, vinput_zero_point0);
70 const __m128i vinit1 = _mm_mullo_epi32(vksum, vinput_zero_point1);
71 __m128i vacc0x0 = _mm_blend_epi16(vinit0, vzero, 0xFC);
72 __m128i vacc0x1 = _mm_blend_epi16(vinit0, vzero, 0xF3);
73 __m128i vacc0x2 = _mm_blend_epi16(vinit0, vzero, 0xCF);
74 __m128i vacc0x3 = _mm_blend_epi16(vinit0, vzero, 0x3F);
75 __m128i vacc1x0 = _mm_blend_epi16(vinit1, vzero, 0xFC);
76 __m128i vacc1x1 = _mm_blend_epi16(vinit1, vzero, 0xF3);
77 __m128i vacc1x2 = _mm_blend_epi16(vinit1, vzero, 0xCF);
78 __m128i vacc1x3 = _mm_blend_epi16(vinit1, vzero, 0x3F);
79 w = (const int32_t*) w + 4;
80
81 size_t k = kc;
82
83 while (k >= 16 * sizeof(int8_t)) {
84 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
85 const __m128i vxa0c0 = _mm_cvtepi8_epi16(va0c0);
86 a0 += 8;
87 const __m128i va1c0 = _mm_loadl_epi64((const __m128i*) a1);
88 const __m128i vxa1c0 = _mm_cvtepi8_epi16(va1c0);
89 a1 += 8;
90
91 const __m128i vb0c01 = _mm_loadl_epi64((const __m128i*) w);
92 const __m128i vbs0c0 = _mm_slli_epi32(vb0c01, 4);
93 const __m128i vb0c0 = _mm_and_si128(vbs0c0, vmask);
94 const __m128i vxb0c0 = _mm_cvtepi8_epi16(vb0c0);
95
96 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
97 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
98 const __m128i vb1c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
99 const __m128i vbs1c0 = _mm_slli_epi32(vb1c01, 4);
100 const __m128i vb1c0 = _mm_and_si128(vbs1c0, vmask);
101 const __m128i vxb1c0 = _mm_cvtepi8_epi16(vb1c0);
102
103 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
104 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
105 const __m128i vb2c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
106 const __m128i vbs2c0 = _mm_slli_epi32(vb2c01, 4);
107 const __m128i vb2c0 = _mm_and_si128(vbs2c0, vmask);
108 const __m128i vxb2c0 = _mm_cvtepi8_epi16(vb2c0);
109
110 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
111 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
112 const __m128i vb3c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
113 const __m128i vbs3c0 = _mm_slli_epi32(vb3c01, 4);
114 const __m128i vb3c0 = _mm_and_si128(vbs3c0, vmask);
115 const __m128i vxb3c0 = _mm_cvtepi8_epi16(vb3c0);
116
117 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
118 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
119
120 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
121 const __m128i vxa0c1 = _mm_cvtepi8_epi16(va0c1);
122 a0 += 8;
123 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
124 const __m128i vxa1c1 = _mm_cvtepi8_epi16(va1c1);
125 a1 += 8;
126
127 const __m128i vb0c1 = _mm_and_si128(vb0c01, vmask);
128 const __m128i vxb0c1 = _mm_cvtepi8_epi16(vb0c1);
129
130 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
131 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
132 const __m128i vb1c1 = _mm_and_si128(vb1c01, vmask);
133 const __m128i vxb1c1 = _mm_cvtepi8_epi16(vb1c1);
134
135 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
136 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
137 const __m128i vb2c1 = _mm_and_si128(vb2c01, vmask);
138 const __m128i vxb2c1 = _mm_cvtepi8_epi16(vb2c1);
139
140 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
141 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
142 const __m128i vb3c1 = _mm_and_si128(vb3c01, vmask);
143 const __m128i vxb3c1 = _mm_cvtepi8_epi16(vb3c1);
144
145 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
146 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
147
148
149 w = (const int8_t*) w + 32;
150 k -= 16 * sizeof(int8_t);
151 }
152
153 while (k >= 8 * sizeof(int8_t)) {
154 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
155 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
156 a0 += 8;
157 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
158 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
159 a1 += 8;
160
161 __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
162 vb0 = _mm_slli_epi32(vb0, 4);
163 vb0 = _mm_and_si128(vb0, vmask);
164
165 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
166
167 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
168 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
169 __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
170 vb1 = _mm_slli_epi32(vb1, 4);
171 vb1 = _mm_and_si128(vb1, vmask);
172
173 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
174
175 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
176 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
177 __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
178 vb2 = _mm_slli_epi32(vb2, 4);
179 vb2 = _mm_and_si128(vb2, vmask);
180
181 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
182
183 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
184 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
185 __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
186 vb3 = _mm_slli_epi32(vb3, 4);
187 vb3 = _mm_and_si128(vb3, vmask);
188
189 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
190
191 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
192 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
193
194 w = (const int8_t*) w + 32;
195 k -= 8 * sizeof(int8_t);
196 }
197
198 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
199 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
200 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
201 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
202
203 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
204 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
205
206 vacc0x0123 = _mm_srai_epi32(vacc0x0123, 4);
207 vacc1x0123 = _mm_srai_epi32(vacc1x0123, 4);
208 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
209 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
210
211 const __m128 vinput_scale0 = _mm_broadcast_ss(&quantization_params[0].inv_scale);
212 const __m128 vinput_scale1 = _mm_broadcast_ss(&quantization_params[1].inv_scale);
213
214 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
215 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
216
217 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
218 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
219 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
220
221 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
222 w = (const float*) w + 8;
223 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
224 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
225
226 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
227 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
228
229 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
230 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
231
232 if XNN_LIKELY(nc >= 4) {
233 _mm_storeu_ps(c0, vout0x0123);
234 _mm_storeu_ps(c1, vout1x0123);
235
236 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
237 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
238
239 c0 = (float*) ((uintptr_t) c0 + cn_stride);
240 c1 = (float*) ((uintptr_t) c1 + cn_stride);
241
242 nc -= 4;
243 } else {
244 if (nc & 2) {
245 _mm_storel_pi((__m64*) c0, vout0x0123);
246 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
247 c0 += 2;
248 _mm_storel_pi((__m64*) c1, vout1x0123);
249 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
250 c1 += 2;
251 }
252 if (nc & 1) {
253 _mm_store_ss(c0, vout0x0123);
254 _mm_store_ss(c1, vout1x0123);
255 }
256 nc = 0;
257 }
258 } while (nc != 0);
259 }