Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-2x4c8-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_qb4w_gemm_minmax_ukernel_2x4c8__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_qb4w_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 size_t bl = params->scalar.blocksize;
47 assert(bl <= round_up_po2(kc, 2));
48 assert(bl != 0);
49 assert(bl % 32 == 0);
50 kc = round_up_po2(kc, 8 * sizeof(int8_t));
51 const int8_t* a0 = a;
52 float* c0 = c;
53 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
54 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
55 if XNN_UNPREDICTABLE(mr != 2) {
56 a1 = a0;
57 c1 = c0;
58 }
59
60 const __m128 vmin = _mm_set1_ps(params->scalar.min);
61 const __m128 vmax = _mm_set1_ps(params->scalar.max);
62 XNN_FORCE_REALIZATION(vmin);
63 XNN_FORCE_REALIZATION(vmax);
64
65 const __m128i vmask = _mm_set1_epi8(0xF0);
66 XNN_FORCE_REALIZATION(vmask);
67
68 do {
69 const __m128 vksum = _mm_load_ps((const float*) w);
70 const __m128i vinput_zero_point0 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[0].zero_point));
71 const __m128i vinput_zero_point1 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[1].zero_point));
72
73 __m128 vinput_zero_point0_float = _mm_cvtepi32_ps(vinput_zero_point0);
74 __m128 vout0x0123 = _mm_mul_ps(vksum, vinput_zero_point0_float);
75 __m128 vinput_zero_point1_float = _mm_cvtepi32_ps(vinput_zero_point1);
76 __m128 vout1x0123 = _mm_mul_ps(vksum, vinput_zero_point1_float);
77 w = (const int32_t*) w + 4;
78
79 for (size_t kb=0; kb < kc; kb += bl) {
80 __m128i vacc0x0 = _mm_setzero_si128();
81 __m128i vacc0x1 = _mm_setzero_si128();
82 __m128i vacc0x2 = _mm_setzero_si128();
83 __m128i vacc0x3 = _mm_setzero_si128();
84 __m128i vacc1x0 = _mm_setzero_si128();
85 __m128i vacc1x1 = _mm_setzero_si128();
86 __m128i vacc1x2 = _mm_setzero_si128();
87 __m128i vacc1x3 = _mm_setzero_si128();
88 size_t k = bl;
89
90 while (k >= 16 * sizeof(int8_t)) {
91 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
92 const __m128i vxa0c0 = _mm_cvtepi8_epi16(va0c0);
93 a0 += 8;
94 const __m128i va1c0 = _mm_loadl_epi64((const __m128i*) a1);
95 const __m128i vxa1c0 = _mm_cvtepi8_epi16(va1c0);
96 a1 += 8;
97
98 const __m128i vb01c01 = _mm_loadu_si128((const __m128i*) w);
99 const __m128i vbs01c0 = _mm_slli_epi32(vb01c01, 4);
100 const __m128i vb01c0 = _mm_and_si128(vbs01c0, vmask);
101 const __m128i vsb01c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c0);
102 const __m128i vxb0c0 = _mm_unpacklo_epi8(vb01c0, vsb01c0);
103 const __m128i vxb1c0 = _mm_unpackhi_epi8(vb01c0, vsb01c0);
104
105 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
106 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
107 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
108 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
109 const __m128i vb23c01 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
110 const __m128i vbs23c0 = _mm_slli_epi32(vb23c01, 4);
111 const __m128i vb23c0 = _mm_and_si128(vbs23c0, vmask);
112 const __m128i vsb23c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c0);
113 const __m128i vxb2c0 = _mm_unpacklo_epi8(vb23c0, vsb23c0);
114 const __m128i vxb3c0 = _mm_unpackhi_epi8(vb23c0, vsb23c0);
115
116 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
117 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
118 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
119 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
120
121 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
122 const __m128i vxa0c1 = _mm_cvtepi8_epi16(va0c1);
123 a0 += 8;
124 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
125 const __m128i vxa1c1 = _mm_cvtepi8_epi16(va1c1);
126 a1 += 8;
127
128 const __m128i vb01c1 = _mm_and_si128(vb01c01, vmask);
129 const __m128i vsb01c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c1);
130 const __m128i vxb0c1 = _mm_unpacklo_epi8(vb01c1, vsb01c1);
131 const __m128i vxb1c1 = _mm_unpackhi_epi8(vb01c1, vsb01c1);
132
133 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
134 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
135 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
136 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
137 const __m128i vb23c1 = _mm_and_si128(vb23c01, vmask);
138 const __m128i vsb23c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c1);
139 const __m128i vxb2c1 = _mm_unpacklo_epi8(vb23c1, vsb23c1);
140 const __m128i vxb3c1 = _mm_unpackhi_epi8(vb23c1, vsb23c1);
141
142 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
143 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
144 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
145 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
146
147
148 w = (const int8_t*) w + 32;
149 k -= 16 * sizeof(int8_t);
150 }
151
152 while (k >= 8 * sizeof(int8_t)) {
153 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
154 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
155 a0 += 8;
156 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
157 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
158 a1 += 8;
159
160 __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
161 vb01 = _mm_slli_epi32(vb01, 4);
162 vb01 = _mm_and_si128(vb01, vmask);
163
164 const __m128i vxbm1 = _mm_unpackhi_epi8(vb01, vb01);
165 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
166 const __m128i vxb1 = _mm_srai_epi16(vxbm1, 8);
167
168 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
169 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
170 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
171 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
172 __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
173 vb23 = _mm_slli_epi32(vb23, 4);
174 vb23 = _mm_and_si128(vb23, vmask);
175
176 const __m128i vxbm3 = _mm_unpackhi_epi8(vb23, vb23);
177 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
178 const __m128i vxb3 = _mm_srai_epi16(vxbm3, 8);
179
180 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
181 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
182 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
183 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
184
185 w = (const int8_t*) w + 32;
186 k -= 8 * sizeof(int8_t);
187 }
188 // accumulate float
189 const __m128 vfilter_output_scale0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_loadl_epi64((const __m128i*) w)), 16));
190 w = (const uint16_t*) w + 4;
191
192 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
193 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
194 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
195 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
196
197 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
198 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
199
200 vout0x0123 = _mm_add_ps(vout0x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc0x0123), vfilter_output_scale0123));
201 vout1x0123 = _mm_add_ps(vout1x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc1x0123), vfilter_output_scale0123));
202 }
203
204 const __m128 vinput_scale0 = _mm_broadcast_ss(&quantization_params[0].inv_scale);
205 const __m128 vinput_scale1 = _mm_broadcast_ss(&quantization_params[1].inv_scale);
206
207 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
208 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
209
210
211 const __m128 vbias0123 = _mm_loadu_ps((const float*) w);
212 w = (const float*) w + 4;
213 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
214 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
215
216 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
217 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
218
219 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
220 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
221
222 if XNN_LIKELY(nc >= 4) {
223 _mm_storeu_ps(c0, vout0x0123);
224 _mm_storeu_ps(c1, vout1x0123);
225
226 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
227 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
228
229 c0 = (float*) ((uintptr_t) c0 + cn_stride);
230 c1 = (float*) ((uintptr_t) c1 + cn_stride);
231
232 nc -= 4;
233 } else {
234 if (nc & 2) {
235 _mm_storel_pi((__m64*) c0, vout0x0123);
236 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
237 c0 += 2;
238 _mm_storel_pi((__m64*) c1, vout1x0123);
239 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
240 c1 += 2;
241 }
242 if (nc & 1) {
243 _mm_store_ss(c0, vout0x0123);
244 _mm_store_ss(c1, vout1x0123);
245 }
246 nc = 0;
247 }
248 } while (nc != 0);
249 }