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_1x4c8__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_1x4c8__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 <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (1 * 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
46 do {
47 __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
48 __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
49 __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
50 __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
51 w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
52
53 size_t p = ks;
54 do {
55 const int8_t* restrict a0 = a[0];
56 if XNN_UNPREDICTABLE(a0 != zero) {
57 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
58 }
59 a += 1;
60
61 size_t k = 0;
62 while (k < kc) {
63 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
64 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
65 a0 += 8;
66
67 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
68 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
69
70 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
71 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
72 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
73
74 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
75 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
76 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
77
78 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
79 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
80 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
81
82 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
83
84 w = (const void*) ((uintptr_t) w + 32);
85 k += 8 * sizeof(int8_t);
86 }
87 p -= 1 * sizeof(void*);
88 } while (p != 0);
89
90 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
91 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
92
93 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
94
95 const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
96 const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
97
98 const __m128i vacc0x1133 = _mm_shuffle_epi32(vacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
99
100 const __m128i vprod0x02 = _mm_add_epi64(_mm_mul_epi32(vacc0x0123, vmultiplier), vrounding);
101
102 const __m128i vprod0x13 = _mm_add_epi64(_mm_mul_epi32(vacc0x1133, vmultiplier), vrounding);
103
104 const __m128i vq31prod0x02 = _mm_srli_epi64(vprod0x02, 31);
105 const __m128i vq31prod0x13 = _mm_add_epi64(vprod0x13, vprod0x13);
106
107 const __m128i vq31prod0x0123 = _mm_blend_epi16(vq31prod0x02, vq31prod0x13, 0xCC);
108
109 const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
110 const __m128i vrem0x0123 =
111 _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
112
113 const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
114 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
115 vacc0x0123 =
116 _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
117
118 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
119 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
120
121 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
122 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
123 vacc00x0123 = _mm_min_epi16(_mm_max_epi16(vacc00x0123, voutput_min), voutput_max);
124
125 __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
126
127 if (nc >= 4) {
128 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
129 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
130
131 a = (const int8_t**restrict) ((uintptr_t) a - ks);
132
133 nc -= 4;
134 } else {
135 if (nc & 2) {
136 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
137 c0 += 2;
138 vout = _mm_srli_epi32(vout, 16);
139 }
140 if (nc & 1) {
141 *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
142 }
143
144 nc = 0;
145 }
146 } while (nc != 0);
147 }
148