Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x4c8-minmax-fp32-sse2-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 #include <emmintrin.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_1x4c8__sse2_ld64(
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 <= 1);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (1 * 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
47
48 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);
49 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
50 const __m128i voutput_min = _mm_set1_epi16(params->fp32_scalar.output_min);
51 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
52 XNN_FORCE_REALIZATION(voutput_zero_point);
53 XNN_FORCE_REALIZATION(voutput_min);
54
55
56 do {
57 __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
58 __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
59 __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
60 __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
61 w = (const int32_t*) w + 4;
62
63 size_t p = ks;
64 do {
65 const int8_t* restrict a0 = a[0];
66 if XNN_UNPREDICTABLE(a0 != zero) {
67 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
68 }
69 a += 1;
70
71 size_t k = 0;
72 while (k < kc) {
73 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
74 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
75 a0 += 8;
76
77 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
78 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
79
80 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
81 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
82 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
83
84 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
85 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
86 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
87
88 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
89 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
90 const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
91
92 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
93
94 w = (const void*) ((const int8_t*) w + 32);
95 k += 8 * sizeof(int8_t);
96 }
97 p -= 1 * sizeof(void*);
98 } while (p != 0);
99
100 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
101 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
102
103 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
104
105 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
106
107 const __m128 vscale0123 = _mm_load_ps((const float*) w);
108 w = (const float*) w + 4;
109 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
110
111 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
112
113 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
114
115 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
116
117 vacc00x0123 = _mm_max_epi16(vacc00x0123, voutput_min);
118
119 __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
120
121
122 if (nc >= 4) {
123 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
124 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
125
126 a = (const int8_t**restrict) ((uintptr_t) a - ks);
127
128 nc -= 4;
129 } else {
130 if (nc & 2) {
131 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
132 c0 += 2;
133 vout = _mm_srli_epi32(vout, 16);
134 }
135 if (nc & 1) {
136 *c0 = (int8_t) _mm_cvtsi128_si32(vout);
137 }
138
139 nc = 0;
140 }
141 } while (nc != 0);
142 }