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_qc8_igemm_minmax_fp32_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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_igemm_minmax_fp32_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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
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*) ((const int32_t*) w + 4);
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*) ((const int8_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*) ((const int8_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*) ((const int8_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*) ((const int8_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 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
96
97 const __m128 vscale0123 = _mm_load_ps((const float*) w);
98 w = (const void*) ((const float*) w + 4);
99 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
100
101 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse4.output_max_less_zero_point);
102 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
103
104 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
105
106 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
107 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
108
109
110 __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
111
112 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
113
114 if (nc >= 4) {
115 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
116 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
117
118 a = (const int8_t**restrict) ((uintptr_t) a - ks);
119
120 nc -= 4;
121 } else {
122 if (nc & 2) {
123 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
124 c0 += 2;
125 vout = _mm_srli_epi32(vout, 16);
126 }
127 if (nc & 1) {
128 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
129 }
130
131 nc = 0;
132 }
133 } while (nc != 0);
134 }
135