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 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/gemm.h>
20 #include <xnnpack/math.h>
21
22
xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld128(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_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])23 void xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld128(
24 size_t mr,
25 size_t nc,
26 size_t kc,
27 const int8_t* restrict a,
28 size_t a_stride,
29 const void* restrict w,
30 int8_t* restrict c,
31 size_t cm_stride,
32 size_t cn_stride,
33 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 2);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(kc % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 8);
45 const int8_t* a0 = a;
46 int8_t* c0 = c;
47 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
48 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr != 2) {
50 a1 = a0;
51 c1 = c0;
52 }
53
54 do {
55 __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
56 __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
57 __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
58 __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
59 __m128i vacc1x0 = vacc0x0;
60 __m128i vacc1x1 = vacc0x1;
61 __m128i vacc1x2 = vacc0x2;
62 __m128i vacc1x3 = vacc0x3;
63 w = (const void*) ((const int32_t*) w + 4);
64
65 size_t k = 0;
66 while (k < kc) {
67 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
68 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
69 a0 += 8;
70 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
71 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
72 a1 += 8;
73
74 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
75 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
76 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
77
78 vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
79 vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
80 vacc1x0 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0);
81 vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
82 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
83 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
84 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
85
86 vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
87 vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
88 vacc1x2 = _mm_maddd_epi16(vxa1, vxb2, vacc1x2);
89 vacc1x3 = _mm_maddd_epi16(vxa1, vxb3, vacc1x3);
90
91 w = (const void*) ((const int8_t*) w + 32);
92 k += 8 * sizeof(int8_t);
93 }
94
95 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
96 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
97 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
98 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
99
100 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
101 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
102
103 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
104 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
105
106 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
107 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
108 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
109
110 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_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->fp32_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->fp32_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