Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-2x4c2s4-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_2x4c2s4__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 <= 2);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (2 * 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
51 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);
52 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
53 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
54 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
55 XNN_FORCE_REALIZATION(voutput_zero_point);
56 XNN_FORCE_REALIZATION(voutput_min);
57
58
59 do {
60 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
61 __m128i vacc1x0123 = vacc0x0123;
62 w = (const void*) ((const int32_t*) w + 4);
63
64 size_t p = ks;
65 do {
66 const int8_t* restrict a0 = a[0];
67 if XNN_UNPREDICTABLE(a0 != zero) {
68 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
69 }
70 const int8_t* restrict a1 = a[1];
71 if XNN_UNPREDICTABLE(a1 != zero) {
72 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
73 }
74 a += 2;
75
76 size_t k = kc;
77 do {
78 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
79 __m128i vxa0 = _mm_cvtepi8_epi16(va0);
80 a0 += 8;
81 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
82 __m128i vxa1 = _mm_cvtepi8_epi16(va1);
83 a1 += 8;
84
85 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
86 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
87 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
88
89 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb0));
90 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
91 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb0));
92 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
93
94 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb1));
95 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
96 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb1));
97 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
98 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
99 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
100 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
101
102 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb2));
103 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
104 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb2));
105 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
106
107 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb3));
108 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb3));
109
110 w = (const void*) ((const int8_t*) w + 32);
111 k -= 8 * sizeof(int8_t);
112 } while (k != 0);
113 p -= 2 * sizeof(void*);
114 } while (p != 0);
115
116 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
117 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
118
119 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
120 w = (const float*) w + 4;
121 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
122 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
123
124 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
125 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
126
127 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
128 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
129
130 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
131
132
133 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
134
135 vout = _mm_max_epi8(vout, voutput_min);
136
137 if (nc >= 4) {
138 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
139 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
140 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
141 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
142
143 a = (const int8_t**restrict) ((uintptr_t) a - ks);
144
145 nc -= 4;
146 } else {
147 if (nc & 2) {
148 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
149 c1 += 2;
150 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
151 c0 += 2;
152 vout = _mm_srli_epi32(vout, 16);
153 }
154 if (nc & 1) {
155 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
156 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
157 }
158
159 nc = 0;
160 }
161 } while (nc != 0);
162 }