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 #ifdef __GNUC__
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_qs8_igemm_minmax_ukernel_1x4c2__xop_ld128(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_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])23 void xnn_qs8_igemm_minmax_ukernel_1x4c2__xop_ld128(
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_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
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*) ((uintptr_t) w + 4 * sizeof(int32_t));
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 vb01 = _mm_loadu_si128((const __m128i*) w);
70 const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
71 const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
72 const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
73
74 vacc0x0123 = _mm_maddd_epi16(
75 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
76
77 vacc0x0123 = _mm_maddd_epi16(
78 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
79 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
80 const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
81 const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
82 const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
83
84 vacc0x0123 = _mm_maddd_epi16(
85 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
86
87 vacc0x0123 = _mm_maddd_epi16(
88 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
89
90 w = (const void*) ((uintptr_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*) ((uintptr_t) w + 8);
100 const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), 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*) ((uintptr_t) w + 8);
108 const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), 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*) ((uintptr_t) w + 8);
116 const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), 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 const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
127 const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
128
129 const __m128i vacc0x1133 = _mm_shuffle_epi32(vacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
130
131 const __m128i vprod0x02 = _mm_add_epi64(_mm_mul_epi32(vacc0x0123, vmultiplier), vrounding);
132
133 const __m128i vprod0x13 = _mm_add_epi64(_mm_mul_epi32(vacc0x1133, vmultiplier), vrounding);
134
135 const __m128i vq31prod0x02 = _mm_srli_epi64(vprod0x02, 31);
136 const __m128i vq31prod0x13 = _mm_add_epi64(vprod0x13, vprod0x13);
137
138 const __m128i vq31prod0x0123 = _mm_blend_epi16(vq31prod0x02, vq31prod0x13, 0xCC);
139
140 const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
141 const __m128i vrem0x0123 =
142 _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
143
144 const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
145 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
146 vacc0x0123 =
147 _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
148
149 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
150 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
151
152 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
153 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
154 vacc00x0123 = _mm_min_epi16(_mm_max_epi16(vacc00x0123, voutput_min), voutput_max);
155
156 __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
157
158 if (nc >= 4) {
159 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
160 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
161
162 a = (const int8_t**restrict) ((uintptr_t) a - ks);
163
164 nc -= 4;
165 } else {
166 if (nc & 2) {
167 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
168 c0 += 2;
169 vout = _mm_srli_epi32(vout, 16);
170 }
171 if (nc & 1) {
172 *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
173 }
174
175 nc = 0;
176 }
177 } while (nc != 0);
178 }
179