1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h>
15 #include <xnnpack/math.h>
16
17
18
xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2__avx(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)])19 void xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2__avx(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 2);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 2);
41 const int8_t* a0 = a;
42 int8_t* c0 = c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr != 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49
50 do {
51 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
52 __m128i vacc1x0123 = vacc0x0123;
53 w = (const void*) ((const int32_t*) w + 4);
54
55 size_t k = kc;
56 while (k >= 8 * sizeof(int8_t)) {
57 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
58 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
59 a0 += 8;
60 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
61 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
62 a1 += 8;
63
64 const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
65
66 vacc0x0123 = _mm_add_epi32(vacc0x0123,
67 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
68 vacc1x0123 = _mm_add_epi32(vacc1x0123,
69 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
70 const __m128i vxb1 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 8));
71
72 vacc0x0123 = _mm_add_epi32(vacc0x0123,
73 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
74 vacc1x0123 = _mm_add_epi32(vacc1x0123,
75 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
76 const __m128i vxb2 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 16));
77
78 vacc0x0123 = _mm_add_epi32(vacc0x0123,
79 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
80 vacc1x0123 = _mm_add_epi32(vacc1x0123,
81 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
82 const __m128i vxb3 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 24));
83
84 vacc0x0123 = _mm_add_epi32(vacc0x0123,
85 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
86 vacc1x0123 = _mm_add_epi32(vacc1x0123,
87 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
88
89 w = (const void*) ((const int16_t*) w + 32);
90 k -= 8 * sizeof(int8_t);
91 }
92 if (k != 0) {
93 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
94 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
95 a0 = (const int8_t*) ((uintptr_t) a0 + k);
96 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
97 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
98 a1 = (const int8_t*) ((uintptr_t) a1 + k);
99
100 const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
101 w = (const void*) ((const int16_t*) w + 8);
102
103 vacc0x0123 = _mm_add_epi32(vacc0x0123,
104 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
105 vacc1x0123 = _mm_add_epi32(vacc1x0123,
106 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
107
108 if (k > 2 * sizeof(int8_t)) {
109 const __m128i vxb1 = _mm_load_si128((const __m128i*) w);
110 w = (const void*) ((const int16_t*) w + 8);
111
112 vacc0x0123 = _mm_add_epi32(vacc0x0123,
113 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
114 vacc1x0123 = _mm_add_epi32(vacc1x0123,
115 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
116
117 if (k > 4 * sizeof(int8_t)) {
118 const __m128i vxb2 = _mm_load_si128((const __m128i*) w);
119 w = (const void*) ((const int16_t*) w + 8);
120
121 vacc0x0123 = _mm_add_epi32(vacc0x0123,
122 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
123 vacc1x0123 = _mm_add_epi32(vacc1x0123,
124 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
125 }
126 }
127 }
128
129 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
130 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
131
132 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
133 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
134 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
135
136 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
137 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
138 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
139
140 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
141 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
142
143 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
144 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
145
146
147 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
148
149 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
150
151 if (nc >= 4) {
152 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
153 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
154
155 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
156 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
157
158 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
159 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
160
161 nc -= 4;
162 } else {
163 if (nc & 2) {
164 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
165 c0 += 2;
166 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
167 c1 += 2;
168 vout = _mm_srli_epi32(vout, 16);
169 }
170 if (nc & 1) {
171 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
172 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
173 }
174
175 nc = 0;
176 }
177 } while (nc != 0);
178 }
179