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_qc8_igemm_minmax_fp32_ukernel_2x4c2__avx_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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_igemm_minmax_fp32_ukernel_2x4c2__avx_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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * 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 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
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 p = ks;
56 do {
57 const int8_t* restrict a0 = a[0];
58 if XNN_UNPREDICTABLE(a0 != zero) {
59 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
60 }
61 const int8_t* restrict a1 = a[1];
62 if XNN_UNPREDICTABLE(a1 != zero) {
63 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
64 }
65 a += 2;
66
67 size_t k = kc;
68 while (k >= 8 * sizeof(int8_t)) {
69 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
70 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
71 a0 += 8;
72 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
73 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
74 a1 += 8;
75
76 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
77 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
78 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
79
80 vacc0x0123 = _mm_add_epi32(vacc0x0123,
81 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
82 vacc1x0123 = _mm_add_epi32(vacc1x0123,
83 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
84
85 vacc0x0123 = _mm_add_epi32(vacc0x0123,
86 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
87 vacc1x0123 = _mm_add_epi32(vacc1x0123,
88 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
89 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
90 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
91 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
92
93 vacc0x0123 = _mm_add_epi32(vacc0x0123,
94 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
95 vacc1x0123 = _mm_add_epi32(vacc1x0123,
96 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
97
98 vacc0x0123 = _mm_add_epi32(vacc0x0123,
99 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
100 vacc1x0123 = _mm_add_epi32(vacc1x0123,
101 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
102
103 w = (const void*) ((const int8_t*) w + 32);
104 k -= 8 * sizeof(int8_t);
105 }
106 if (k != 0) {
107 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
108 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
109 a0 = (const int8_t*) ((uintptr_t) a0 + k);
110 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
111 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
112 a1 = (const int8_t*) ((uintptr_t) a1 + k);
113
114 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
115 w = (const void*) ((const int8_t*) w + 8);
116 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
117
118 vacc0x0123 = _mm_add_epi32(vacc0x0123,
119 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
120 vacc1x0123 = _mm_add_epi32(vacc1x0123,
121 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
122
123 if (k > 2 * sizeof(int8_t)) {
124 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
125 w = (const void*) ((const int8_t*) w + 8);
126 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
127
128 vacc0x0123 = _mm_add_epi32(vacc0x0123,
129 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
130 vacc1x0123 = _mm_add_epi32(vacc1x0123,
131 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
132
133 if (k > 4 * sizeof(int8_t)) {
134 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
135 w = (const void*) ((const int8_t*) w + 8);
136 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
137
138 vacc0x0123 = _mm_add_epi32(vacc0x0123,
139 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
140 vacc1x0123 = _mm_add_epi32(vacc1x0123,
141 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
142 }
143 }
144 }
145 p -= 2 * sizeof(void*);
146 } while (p != 0);
147
148 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
149 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
150
151 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
152 w = (const void*) ((const float*) w + 4);
153 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
154 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
155
156 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse4.output_max_less_zero_point);
157 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
158 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
159
160 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
161 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
162
163 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
164 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
165
166
167 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
168
169 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
170
171 if (nc >= 4) {
172 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
173 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
174 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
175 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
176
177 a = (const int8_t**restrict) ((uintptr_t) a - ks);
178
179 nc -= 4;
180 } else {
181 if (nc & 2) {
182 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
183 c1 += 2;
184 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
185 c0 += 2;
186 vout = _mm_srli_epi32(vout, 16);
187 }
188 if (nc & 1) {
189 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
190 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
191 }
192
193 nc = 0;
194 }
195 } while (nc != 0);
196 }
197