1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2-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_fp32_ukernel_1x4c2__sse41_ld128(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_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_fp32_ukernel_1x4c2__sse41_ld128(
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_conv_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, 2);
44 int8_t* c0 = c;
45
46 do {
47 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
48 w = (const void*) ((const int32_t*) w + 4);
49
50 size_t p = ks;
51 do {
52 const int8_t* restrict a0 = a[0];
53 if XNN_UNPREDICTABLE(a0 != zero) {
54 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
55 }
56 a += 1;
57
58 size_t k = kc;
59 while (k >= 8 * sizeof(int8_t)) {
60 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
61 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
62 a0 += 8;
63
64 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
65 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
66 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
67
68 vacc0x0123 = _mm_add_epi32(vacc0x0123,
69 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
70
71 vacc0x0123 = _mm_add_epi32(vacc0x0123,
72 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
73 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
74 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
75 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
76
77 vacc0x0123 = _mm_add_epi32(vacc0x0123,
78 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
79
80 vacc0x0123 = _mm_add_epi32(vacc0x0123,
81 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
82
83 w = (const void*) ((const int8_t*) w + 32);
84 k -= 8 * sizeof(int8_t);
85 }
86 if (k != 0) {
87 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
88 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
89 a0 = (const int8_t*) ((uintptr_t) a0 + k);
90
91 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
92 w = (const void*) ((const int8_t*) w + 8);
93 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
94
95 vacc0x0123 = _mm_add_epi32(vacc0x0123,
96 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
97
98 if (k > 2 * sizeof(int8_t)) {
99 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
100 w = (const void*) ((const int8_t*) w + 8);
101 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
102
103 vacc0x0123 = _mm_add_epi32(vacc0x0123,
104 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
105
106 if (k > 4 * sizeof(int8_t)) {
107 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
108 w = (const void*) ((const int8_t*) w + 8);
109 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
110
111 vacc0x0123 = _mm_add_epi32(vacc0x0123,
112 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
113 }
114 }
115 }
116 p -= 1 * sizeof(void*);
117 } while (p != 0);
118
119 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
120
121 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
122 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
123
124 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
125 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
126
127 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
128
129 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
130 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
131
132
133 __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
134
135 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
136
137 if (nc >= 4) {
138 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
139 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
140
141 a = (const int8_t**restrict) ((uintptr_t) a - ks);
142
143 nc -= 4;
144 } else {
145 if (nc & 2) {
146 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
147 c0 += 2;
148 vout = _mm_srli_epi32(vout, 16);
149 }
150 if (nc & 1) {
151 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
152 }
153
154 nc = 0;
155 }
156 } while (nc != 0);
157 }
158