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 <smmintrin.h>
13
14 #include <xnnpack/igemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_igemm_minmax_ukernel_2x4c8__sse41_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_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_ukernel_2x4c8__sse41_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8);
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
47 c1 = c0;
48 }
49
50 do {
51 __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
52 __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
53 __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
54 __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
55 __m128i vacc1x0 = vacc0x0;
56 __m128i vacc1x1 = vacc0x1;
57 __m128i vacc1x2 = vacc0x2;
58 __m128i vacc1x3 = vacc0x3;
59 w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
60
61 size_t p = ks;
62 do {
63 const int8_t* restrict a0 = a[0];
64 if XNN_UNPREDICTABLE(a0 != zero) {
65 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
66 }
67 const int8_t* restrict a1 = a[1];
68 if XNN_UNPREDICTABLE(a1 != zero) {
69 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
70 }
71 a += 2;
72
73 size_t k = 0;
74 while (k < kc) {
75 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
76 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
77 a0 += 8;
78 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
79 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
80 a1 += 8;
81
82 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
83 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
84
85 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
86 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
87 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
88 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
89
90 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
91 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
92 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
93 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
94
95 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
96 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
97 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
98 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
99
100 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
101 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
102
103 w = (const void*) ((uintptr_t) w + 32);
104 k += 8 * sizeof(int8_t);
105 }
106 p -= 2 * sizeof(void*);
107 } while (p != 0);
108
109 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
110 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
111 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
112 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
113
114 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
115 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
116
117 const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
118 const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
119
120 const __m128i vacc0x1133 = _mm_shuffle_epi32(vacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
121 const __m128i vacc1x1133 = _mm_shuffle_epi32(vacc1x0123, _MM_SHUFFLE(3, 3, 1, 1));
122
123 const __m128i vprod0x02 = _mm_add_epi64(_mm_mul_epi32(vacc0x0123, vmultiplier), vrounding);
124 const __m128i vprod1x02 = _mm_add_epi64(_mm_mul_epi32(vacc1x0123, vmultiplier), vrounding);
125
126 const __m128i vprod0x13 = _mm_add_epi64(_mm_mul_epi32(vacc0x1133, vmultiplier), vrounding);
127 const __m128i vprod1x13 = _mm_add_epi64(_mm_mul_epi32(vacc1x1133, vmultiplier), vrounding);
128
129 const __m128i vq31prod0x02 = _mm_srli_epi64(vprod0x02, 31);
130 const __m128i vq31prod0x13 = _mm_add_epi64(vprod0x13, vprod0x13);
131 const __m128i vq31prod1x02 = _mm_srli_epi64(vprod1x02, 31);
132 const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13);
133
134 const __m128i vq31prod0x0123 = _mm_blend_epi16(vq31prod0x02, vq31prod0x13, 0xCC);
135 const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC);
136
137 const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
138 const __m128i vrem0x0123 =
139 _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
140 const __m128i vrem1x0123 =
141 _mm_add_epi32(_mm_and_si128(vq31prod1x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod1x0123));
142
143 const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
144 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
145 vacc0x0123 =
146 _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
147 vacc1x0123 =
148 _mm_sub_epi32(_mm_sra_epi32(vq31prod1x0123, vshift), _mm_cmpgt_epi32(vrem1x0123, vremainder_threshold));
149
150 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
151 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
152
153 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
154 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
155 vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
156
157 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
158
159 if (nc >= 4) {
160 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
161 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
162 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
163 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
164
165 a = (const int8_t**restrict) ((uintptr_t) a - ks);
166
167 nc -= 4;
168 } else {
169 if (nc & 2) {
170 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
171 c1 += 2;
172 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
173 c0 += 2;
174 vout = _mm_srli_epi32(vout, 16);
175 }
176 if (nc & 1) {
177 *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
178 *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
179 }
180
181 nc = 0;
182 }
183 } while (nc != 0);
184 }
185