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 #include <xnnpack/unaligned.h>
17
18
xnn_qc8_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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_igemm_minmax_fp32_ukernel_1x4c2__sse41_ld128(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 1);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (1 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 2 * sizeof(int8_t));
45 int8_t* c0 = c;
46
47 do {
48 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
49 w = (const void*) ((const int32_t*) w + 4);
50
51 size_t p = ks;
52 do {
53 const int8_t* restrict a0 = a[0];
54 if XNN_UNPREDICTABLE(a0 != zero) {
55 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
56 }
57 a += 1;
58
59 size_t k = kc;
60 while (k >= 8 * sizeof(int8_t)) {
61 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
62 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
63 a0 += 8;
64
65 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
66 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
67 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
68
69 vacc0x0123 = _mm_add_epi32(vacc0x0123,
70 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
71
72 vacc0x0123 = _mm_add_epi32(vacc0x0123,
73 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
74 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
75 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
76 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
77
78 vacc0x0123 = _mm_add_epi32(vacc0x0123,
79 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
80
81 vacc0x0123 = _mm_add_epi32(vacc0x0123,
82 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
83
84 w = (const void*) ((const int8_t*) w + 32);
85 k -= 8 * sizeof(int8_t);
86 }
87 if (k != 0) {
88 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
89 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
90 a0 = (const int8_t*) ((uintptr_t) a0 + k);
91
92 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
93 w = (const void*) ((const int8_t*) w + 8);
94 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
95
96 vacc0x0123 = _mm_add_epi32(vacc0x0123,
97 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
98
99 if (k > 2 * sizeof(int8_t)) {
100 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
101 w = (const void*) ((const int8_t*) w + 8);
102 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
103
104 vacc0x0123 = _mm_add_epi32(vacc0x0123,
105 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
106
107 if (k > 4 * sizeof(int8_t)) {
108 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
109 w = (const void*) ((const int8_t*) w + 8);
110 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
111
112 vacc0x0123 = _mm_add_epi32(vacc0x0123,
113 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
114 }
115 }
116 }
117 p -= 1 * sizeof(void*);
118 } while (p != 0);
119
120 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
121
122 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
123 w = (const void*) ((const float*) w + 4);
124 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
125
126 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
127 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
128
129 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
130
131 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
132 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
133
134
135 __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
136
137 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
138
139 if (nc >= 4) {
140 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
141 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
142
143 a = (const int8_t**restrict) ((uintptr_t) a - ks);
144
145 nc -= 4;
146 } else {
147 if (nc & 2) {
148 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
149 c0 += 2;
150 vout = _mm_srli_epi32(vout, 16);
151 }
152 if (nc & 1) {
153 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
154 }
155
156 nc = 0;
157 }
158 } while (nc != 0);
159 }
160