1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld64(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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 2);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 8);
40 const int8_t* a0 = a;
41 int8_t* c0 = c;
42 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr != 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48
49 do {
50 __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
51 __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
52 __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
53 __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
54 __m128i vacc1x0 = vacc0x0;
55 __m128i vacc1x1 = vacc0x1;
56 __m128i vacc1x2 = vacc0x2;
57 __m128i vacc1x3 = vacc0x3;
58 w = (const void*) ((const int32_t*) w + 4);
59
60 size_t k = 0;
61 while (k < kc) {
62 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
63 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
64 a0 += 8;
65 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
66 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
67 a1 += 8;
68
69 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
70 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
71
72 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
73 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
74 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
75 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
76
77 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
78 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
79 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
80 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
81
82 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
83 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
84 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
85 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
86
87 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
88 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
89
90 w = (const void*) ((const int8_t*) w + 32);
91 k += 8 * sizeof(int8_t);
92 }
93
94 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
95 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
96 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
97 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
98
99 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
100 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
101
102 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
103 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
104
105 const __m128 vscale0123 = _mm_load_ps((const float*) w);
106 w = (const void*) ((const float*) w + 4);
107 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
108 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
109
110 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse4.output_max_less_zero_point);
111 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
112 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
113
114 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
115 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
116
117 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
118 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
119
120
121 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
122
123 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
124
125 if (nc >= 4) {
126 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
127 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
128
129 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
130 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
131
132 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
133 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
134
135 nc -= 4;
136 } else {
137 if (nc & 2) {
138 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
139 c0 += 2;
140 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
141 c1 += 2;
142 vout = _mm_srli_epi32(vout, 16);
143 }
144 if (nc & 1) {
145 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
146 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
147 }
148
149 nc = 0;
150 }
151 } while (nc != 0);
152 }
153