1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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 #ifdef __GNUC__
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/gemm.h>
20 #include <xnnpack/math.h>
21
22
xnn_qs8_gemm_xw_minmax_ukernel_1x4c2__xop(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_gemm_xw_params params[restrict XNN_MIN_ELEMENTS (1)])23 void xnn_qs8_gemm_xw_minmax_ukernel_1x4c2__xop(
24 size_t mr,
25 size_t nc,
26 size_t kc,
27 const int8_t* restrict a,
28 size_t a_stride,
29 const void* restrict w,
30 int8_t* restrict c,
31 size_t cm_stride,
32 size_t cn_stride,
33 const union xnn_qs8_gemm_xw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
34 {
35 assert(mr != 0);
36 assert(mr <= 1);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(kc % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 2);
45 const int8_t* a0 = a;
46 int8_t* c0 = c;
47
48 do {
49 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
50 w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
51
52 size_t k = kc;
53 while (k >= 8 * sizeof(int8_t)) {
54 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
55 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
56 a0 += 8;
57
58 const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
59
60 vacc0x0123 = _mm_maddd_epi16(
61 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
62 const __m128i vxb1 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 8 * sizeof(int16_t)));
63
64 vacc0x0123 = _mm_maddd_epi16(
65 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
66 const __m128i vxb2 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int16_t)));
67
68 vacc0x0123 = _mm_maddd_epi16(
69 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
70 const __m128i vxb3 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 24 * sizeof(int16_t)));
71
72 vacc0x0123 = _mm_maddd_epi16(
73 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
74
75 w = (const void*) ((uintptr_t) w + 32 * sizeof(int16_t));
76 k -= 8 * sizeof(int8_t);
77 }
78 if (k != 0) {
79 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
80 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
81 a0 = (const int8_t*) ((uintptr_t) a0 + k);
82
83 const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
84 w = (const void*) ((uintptr_t) w + 8 * sizeof(int16_t));
85
86 vacc0x0123 = _mm_maddd_epi16(
87 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
88
89 if (k > 2 * sizeof(int8_t)) {
90 const __m128i vxb1 = _mm_load_si128((const __m128i*) w);
91 w = (const void*) ((uintptr_t) w + 8 * sizeof(int16_t));
92
93 vacc0x0123 = _mm_maddd_epi16(
94 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
95
96 if (k > 4 * sizeof(int8_t)) {
97 const __m128i vxb2 = _mm_load_si128((const __m128i*) w);
98 w = (const void*) ((uintptr_t) w + 8 * sizeof(int16_t));
99
100 vacc0x0123 = _mm_maddd_epi16(
101 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
102 }
103 }
104 }
105
106 const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
107 const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
108
109 const __m128i vacc0x1133 = _mm_shuffle_epi32(vacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
110
111 const __m128i vprod0x02 = _mm_add_epi64(_mm_mul_epi32(vacc0x0123, vmultiplier), vrounding);
112
113 const __m128i vprod0x13 = _mm_add_epi64(_mm_mul_epi32(vacc0x1133, vmultiplier), vrounding);
114
115 const __m128i vq31prod0x02 = _mm_srli_epi64(vprod0x02, 31);
116 const __m128i vq31prod0x13 = _mm_add_epi64(vprod0x13, vprod0x13);
117
118 const __m128i vq31prod0x0123 = _mm_blend_epi16(vq31prod0x02, vq31prod0x13, 0xCC);
119
120 const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
121 const __m128i vrem0x0123 =
122 _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
123
124 const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
125 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
126 vacc0x0123 =
127 _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
128
129 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
130 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
131
132 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
133 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
134 vacc00x0123 = _mm_min_epi16(_mm_max_epi16(vacc00x0123, voutput_min), voutput_max);
135
136 __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
137
138 if (nc >= 4) {
139 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
140
141 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
142
143 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
144
145 nc -= 4;
146 } else {
147 if (nc & 2) {
148 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
149 c0 += 2;
150 vout = _mm_srli_epi32(vout, 16);
151 }
152 if (nc & 1) {
153 *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
154 }
155
156 nc = 0;
157 }
158 } while (nc != 0);
159 }
160