1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2-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 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/igemm.h>
20 #include <xnnpack/math.h>
21
22
xnn_qc8_igemm_minmax_fp32_ukernel_1x4c2__xop_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])23 void xnn_qc8_igemm_minmax_fp32_ukernel_1x4c2__xop_ld64(
24 size_t mr,
25 size_t nc,
26 size_t kc,
27 size_t ks,
28 const int8_t** restrict a,
29 const void* restrict w,
30 int8_t* restrict c,
31 size_t cm_stride,
32 size_t cn_stride,
33 size_t a_offset,
34 const int8_t* zero,
35 const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 1);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(ks != 0);
42 assert(ks % (1 * sizeof(void*)) == 0);
43 assert(a_offset % sizeof(int8_t) == 0);
44 assert(a != NULL);
45 assert(w != NULL);
46 assert(c != NULL);
47
48 kc = round_up_po2(kc, 2);
49 int8_t* c0 = c;
50
51 do {
52 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
53 w = (const void*) ((const int32_t*) w + 4);
54
55 size_t p = ks;
56 do {
57 const int8_t* restrict a0 = a[0];
58 if XNN_UNPREDICTABLE(a0 != zero) {
59 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
60 }
61 a += 1;
62
63 size_t k = kc;
64 while (k >= 8 * sizeof(int8_t)) {
65 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
66 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
67 a0 += 8;
68
69 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
70 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
71
72 vacc0x0123 = _mm_maddd_epi16(
73 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
74 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
75 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
76
77 vacc0x0123 = _mm_maddd_epi16(
78 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
79 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
80 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
81
82 vacc0x0123 = _mm_maddd_epi16(
83 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
84 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
85 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
86
87 vacc0x0123 = _mm_maddd_epi16(
88 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
89
90 w = (const void*) ((const int8_t*) w + 32);
91 k -= 8 * sizeof(int8_t);
92 }
93 if (k != 0) {
94 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
95 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
96 a0 = (const int8_t*) ((uintptr_t) a0 + k);
97
98 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
99 w = (const void*) ((const int8_t*) w + 8);
100 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
101
102 vacc0x0123 = _mm_maddd_epi16(
103 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
104
105 if (k > 2 * sizeof(int8_t)) {
106 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
107 w = (const void*) ((const int8_t*) w + 8);
108 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
109
110 vacc0x0123 = _mm_maddd_epi16(
111 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
112
113 if (k > 4 * sizeof(int8_t)) {
114 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
115 w = (const void*) ((const int8_t*) w + 8);
116 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
117
118 vacc0x0123 = _mm_maddd_epi16(
119 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
120 }
121 }
122 }
123 p -= 1 * sizeof(void*);
124 } while (p != 0);
125
126 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
127
128 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
129 w = (const void*) ((const float*) w + 4);
130 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
131
132 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse4.output_max_less_zero_point);
133 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
134
135 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
136
137 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
138 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
139
140
141 __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
142
143 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
144
145 if (nc >= 4) {
146 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
147 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
148
149 a = (const int8_t**restrict) ((uintptr_t) a - ks);
150
151 nc -= 4;
152 } else {
153 if (nc & 2) {
154 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
155 c0 += 2;
156 vout = _mm_srli_epi32(vout, 16);
157 }
158 if (nc & 1) {
159 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
160 }
161
162 nc = 0;
163 }
164 } while (nc != 0);
165 }
166