Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-3x4c2s4-minmax-fp32-avx-ld128.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2s4-sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 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 <smmintrin.h>
13
14 #include "xnnpack/igemm.h"
15 #include "xnnpack/math.h"
16 #include "xnnpack/unaligned.h"
17
18
19 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_3x4c2s4__avx_ld128(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 3);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (3 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 8 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54
55 const __m128 voutput_max_less_zero_point = _mm_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
56 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
57 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
58 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
59 XNN_FORCE_REALIZATION(voutput_zero_point);
60 XNN_FORCE_REALIZATION(voutput_min);
61
62
63 do {
64 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
65 __m128i vacc1x0123 = vacc0x0123;
66 __m128i vacc2x0123 = vacc0x0123;
67 w = (const void*) ((const int32_t*) w + 4);
68
69 size_t p = ks;
70 do {
71 const int8_t* restrict a0 = a[0];
72 if XNN_UNPREDICTABLE(a0 != zero) {
73 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
74 }
75 const int8_t* restrict a1 = a[1];
76 if XNN_UNPREDICTABLE(a1 != zero) {
77 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
78 }
79 const int8_t* restrict a2 = a[2];
80 if XNN_UNPREDICTABLE(a2 != zero) {
81 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
82 }
83 a += 3;
84
85 size_t k = kc;
86 do {
87 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
88 __m128i vxa0 = _mm_cvtepi8_epi16(va0);
89 a0 += 8;
90 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
91 __m128i vxa1 = _mm_cvtepi8_epi16(va1);
92 a1 += 8;
93 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
94 __m128i vxa2 = _mm_cvtepi8_epi16(va2);
95 a2 += 8;
96
97 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
98 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
99 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
100
101 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb0));
102 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
103 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb0));
104 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
105 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb0));
106 vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
107
108 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb1));
109 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
110 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb1));
111 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
112 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb1));
113 vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
114 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
115 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
116 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
117
118 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb2));
119 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
120 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb2));
121 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
122 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb2));
123 vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
124
125 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb3));
126 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb3));
127 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb3));
128
129 w = (const void*) ((const int8_t*) w + 32);
130 k -= 8 * sizeof(int8_t);
131 } while (k != 0);
132 p -= 3 * sizeof(void*);
133 } while (p != 0);
134
135 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
136 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
137 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
138
139 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
140 w = (const float*) w + 4;
141 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
142 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
143 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
144
145 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
146 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
147 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
148
149 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
150 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
151 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
152
153 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
154 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
155
156
157 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
158
159 vout = _mm_max_epi8(vout, voutput_min);
160
161 if (nc >= 4) {
162 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
163 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
164 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
165 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
166 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
167 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
168
169 a = (const int8_t**restrict) ((uintptr_t) a - ks);
170
171 nc -= 4;
172 } else {
173 if (nc & 2) {
174 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
175 c2 += 2;
176 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
177 c1 += 2;
178 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
179 c0 += 2;
180 vout = _mm_srli_epi32(vout, 16);
181 }
182 if (nc & 1) {
183 *c2 = (int8_t) _mm_extract_epi8(vout, 8);
184 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
185 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
186 }
187
188 nc = 0;
189 }
190 } while (nc != 0);
191 }