Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-2x4c8-minmax-avx-ld64.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_2x4c8__avx_ld64(
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 <= 2);
40 assert(nc != 0);
41 assert(kc != 0);
42 assert(ks != 0);
43 assert(ks % (2 * 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
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 const __m128i vinput_zero_point = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params->zero_point));
61 const __m128 vinput_scale = _mm_broadcast_ss(&quantization_params->inv_scale);
62
63
64 do {
65 const __m128i vksum = _mm_load_si128((const __m128i*) w);
66 const __m128i vzero = _mm_setzero_si128();
67 const __m128i vinit0 = _mm_mullo_epi32(vksum, vinput_zero_point);
68 const __m128i vinit1 = _mm_mullo_epi32(vksum, vinput_zero_point);
69 __m128i vacc0x0 = _mm_blend_epi16(vinit0, vzero, 0xFC);
70 __m128i vacc0x1 = _mm_blend_epi16(vinit0, vzero, 0xF3);
71 __m128i vacc0x2 = _mm_blend_epi16(vinit0, vzero, 0xCF);
72 __m128i vacc0x3 = _mm_blend_epi16(vinit0, vzero, 0x3F);
73 __m128i vacc1x0 = _mm_blend_epi16(vinit1, vzero, 0xFC);
74 __m128i vacc1x1 = _mm_blend_epi16(vinit1, vzero, 0xF3);
75 __m128i vacc1x2 = _mm_blend_epi16(vinit1, vzero, 0xCF);
76 __m128i vacc1x3 = _mm_blend_epi16(vinit1, vzero, 0x3F);
77 w = (const int32_t*) w + 4;
78
79 size_t p = ks;
80 do {
81 const int8_t* restrict a0 = a[0];
82 if XNN_UNPREDICTABLE(a0 != zero) {
83 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
84 } else {
85 a0 = zero_data;
86 }
87 const int8_t* restrict a1 = a[1];
88 if XNN_UNPREDICTABLE(a1 != zero) {
89 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
90 } else {
91 a1 = zero_data;
92 }
93 a += 2;
94
95 size_t k = 0;
96 while (k < kc) {
97 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
98 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
99 a0 += 8;
100 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
101 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
102 a1 += 8;
103
104 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
105 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
106
107 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
108 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
109 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
110 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
111
112 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
113 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
114 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
115 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
116
117 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
118 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
119 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
120 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
121
122 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
123 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
124
125 w = (const void*) ((const int8_t*) w + 32);
126 k += 8 * sizeof(int8_t);
127 }
128 p -= 2 * sizeof(void*);
129 } while (p != 0);
130
131 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
132 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
133 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
134 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
135
136 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
137 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
138
139 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
140 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
141
142 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale);
143 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale);
144
145 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
146 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
147 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
148
149 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
150 w = (const float*) w + 8;
151 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
152 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
153
154 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
155 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
156
157 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
158 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
159
160 if XNN_LIKELY(nc >= 4) {
161 _mm_storeu_ps(c1, vout1x0123);
162 _mm_storeu_ps(c0, vout0x0123);
163
164 a = (const int8_t**restrict) ((uintptr_t) a - ks);
165
166 c0 = (float*) ((uintptr_t) c0 + cn_stride);
167 c1 = (float*) ((uintptr_t) c1 + cn_stride);
168
169 nc -= 4;
170 } else {
171 if (nc & 2) {
172 _mm_storel_pi((__m64*) c1, vout1x0123);
173 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
174 c1 += 2;
175 _mm_storel_pi((__m64*) c0, vout0x0123);
176 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
177 c0 += 2;
178 }
179 if (nc & 1) {
180 _mm_store_ss(c1, vout1x0123);
181 _mm_store_ss(c0, vout0x0123);
182 }
183 nc = 0;
184 }
185 } while (nc != 0);
186 }