Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-3x4c8-minmax-avx-ld128.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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/igemm.h"
19 #include "xnnpack/math.h"
20
21
22 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_3x4c8__avx_ld128(
23 size_t mr,
24 size_t nc,
25 size_t kc,
26 size_t ks,
27 const int8_t** restrict a,
28 const void* restrict w,
29 float* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 size_t a_offset,
33 const int8_t* zero,
34 const int8_t* zero_data,
35 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
36 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
37 {
38 assert(mr != 0);
39 assert(mr <= 3);
40 assert(nc != 0);
41 assert(kc != 0);
42 assert(ks != 0);
43 assert(ks % (3 * sizeof(void*)) == 0);
44 assert(a_offset % sizeof(int8_t) == 0);
45 assert(a != NULL);
46 assert(w != NULL);
47 assert(c != NULL);
48
49 kc = round_up_po2(kc, 8 * sizeof(int8_t));
50 float* c0 = c;
51 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
52 if XNN_UNPREDICTABLE(mr < 2) {
53 c1 = c0;
54 }
55 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
56 if XNN_UNPREDICTABLE(mr <= 2) {
57 c2 = c1;
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 const __m128i vinput_zero_point = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params->zero_point));
65 const __m128 vinput_scale = _mm_broadcast_ss(&quantization_params->inv_scale);
66
67
68 do {
69 const __m128i vksum = _mm_load_si128((const __m128i*) w);
70 const __m128i vzero = _mm_setzero_si128();
71 const __m128i vinit0 = _mm_mullo_epi32(vksum, vinput_zero_point);
72 const __m128i vinit1 = _mm_mullo_epi32(vksum, vinput_zero_point);
73 const __m128i vinit2 = _mm_mullo_epi32(vksum, vinput_zero_point);
74 __m128i vacc0x0 = _mm_blend_epi16(vinit0, vzero, 0xFC);
75 __m128i vacc0x1 = _mm_blend_epi16(vinit0, vzero, 0xF3);
76 __m128i vacc0x2 = _mm_blend_epi16(vinit0, vzero, 0xCF);
77 __m128i vacc0x3 = _mm_blend_epi16(vinit0, vzero, 0x3F);
78 __m128i vacc1x0 = _mm_blend_epi16(vinit1, vzero, 0xFC);
79 __m128i vacc1x1 = _mm_blend_epi16(vinit1, vzero, 0xF3);
80 __m128i vacc1x2 = _mm_blend_epi16(vinit1, vzero, 0xCF);
81 __m128i vacc1x3 = _mm_blend_epi16(vinit1, vzero, 0x3F);
82 __m128i vacc2x0 = _mm_blend_epi16(vinit2, vzero, 0xFC);
83 __m128i vacc2x1 = _mm_blend_epi16(vinit2, vzero, 0xF3);
84 __m128i vacc2x2 = _mm_blend_epi16(vinit2, vzero, 0xCF);
85 __m128i vacc2x3 = _mm_blend_epi16(vinit2, vzero, 0x3F);
86 w = (const int32_t*) w + 4;
87
88 size_t p = ks;
89 do {
90 const int8_t* restrict a0 = a[0];
91 if XNN_UNPREDICTABLE(a0 != zero) {
92 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
93 } else {
94 a0 = zero_data;
95 }
96 const int8_t* restrict a1 = a[1];
97 if XNN_UNPREDICTABLE(a1 != zero) {
98 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
99 } else {
100 a1 = zero_data;
101 }
102 const int8_t* restrict a2 = a[2];
103 if XNN_UNPREDICTABLE(a2 != zero) {
104 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
105 } else {
106 a2 = zero_data;
107 }
108 a += 3;
109
110 size_t k = 0;
111 while (k < kc) {
112 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
113 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
114 a0 += 8;
115 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
116 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
117 a1 += 8;
118 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
119 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
120 a2 += 8;
121
122 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
123 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
124 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
125
126 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
127 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
128 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
129 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
130 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
131 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
132 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
133 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
134 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
135
136 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
137 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
138 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
139 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
140 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
141 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
142
143 w = (const void*) ((const int8_t*) w + 32);
144 k += 8 * sizeof(int8_t);
145 }
146 p -= 3 * sizeof(void*);
147 } while (p != 0);
148
149 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
150 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
151 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
152 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
153 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
154 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
155
156 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
157 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
158 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
159
160 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
161 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
162 __m128 vout2x0123 = _mm_cvtepi32_ps(vacc2x0123);
163
164 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale);
165 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale);
166 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale);
167
168 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
169 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
170 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
171 vout2x0123 = _mm_mul_ps(vout2x0123, vfilter_output_scale0123);
172
173 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
174 w = (const float*) w + 8;
175 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
176 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
177 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
178
179 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
180 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
181 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
182
183 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
184 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
185 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
186
187 if XNN_LIKELY(nc >= 4) {
188 _mm_storeu_ps(c2, vout2x0123);
189 _mm_storeu_ps(c1, vout1x0123);
190 _mm_storeu_ps(c0, vout0x0123);
191
192 a = (const int8_t**restrict) ((uintptr_t) a - ks);
193
194 c0 = (float*) ((uintptr_t) c0 + cn_stride);
195 c1 = (float*) ((uintptr_t) c1 + cn_stride);
196 c2 = (float*) ((uintptr_t) c2 + cn_stride);
197
198 nc -= 4;
199 } else {
200 if (nc & 2) {
201 _mm_storel_pi((__m64*) c2, vout2x0123);
202 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
203 c2 += 2;
204 _mm_storel_pi((__m64*) c1, vout1x0123);
205 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
206 c1 += 2;
207 _mm_storel_pi((__m64*) c0, vout0x0123);
208 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
209 c0 += 2;
210 }
211 if (nc & 1) {
212 _mm_store_ss(c2, vout2x0123);
213 _mm_store_ss(c1, vout1x0123);
214 _mm_store_ss(c0, vout0x0123);
215 }
216 nc = 0;
217 }
218 } while (nc != 0);
219 }