1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2s4-sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 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_4x4c2s4__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_4x4c2s4__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 <= 4);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (4 * 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, 8 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr != 4) {
56 c3 = c2;
57 }
58
59 do {
60 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
61 __m128i vacc1x0123 = vacc0x0123;
62 __m128i vacc2x0123 = vacc0x0123;
63 __m128i vacc3x0123 = vacc0x0123;
64 w = (const void*) ((const int32_t*) w + 4);
65
66 size_t p = ks;
67 do {
68 const int8_t* restrict a0 = a[0];
69 if XNN_UNPREDICTABLE(a0 != zero) {
70 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
71 }
72 const int8_t* restrict a1 = a[1];
73 if XNN_UNPREDICTABLE(a1 != zero) {
74 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
75 }
76 const int8_t* restrict a2 = a[2];
77 if XNN_UNPREDICTABLE(a2 != zero) {
78 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
79 }
80 const int8_t* restrict a3 = a[3];
81 if XNN_UNPREDICTABLE(a3 != zero) {
82 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
83 }
84 a += 4;
85
86 size_t k = kc;
87 do {
88 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
89 __m128i vxa0 = _mm_cvtepi8_epi16(va0);
90 a0 += 8;
91 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
92 __m128i vxa1 = _mm_cvtepi8_epi16(va1);
93 a1 += 8;
94 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
95 __m128i vxa2 = _mm_cvtepi8_epi16(va2);
96 a2 += 8;
97 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
98 __m128i vxa3 = _mm_cvtepi8_epi16(va3);
99 a3 += 8;
100
101 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
102 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
103 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
104
105 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb0));
106 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
107 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb0));
108 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
109 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb0));
110 vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
111 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(vxa3, vxb0));
112 vxa3 = _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 3, 2, 1));
113
114 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb1));
115 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
116 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb1));
117 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
118 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb1));
119 vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
120 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(vxa3, vxb1));
121 vxa3 = _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 3, 2, 1));
122 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
123 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
124 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
125
126 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb2));
127 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
128 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb2));
129 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
130 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb2));
131 vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
132 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(vxa3, vxb2));
133 vxa3 = _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 3, 2, 1));
134
135 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb3));
136 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb3));
137 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb3));
138 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(vxa3, vxb3));
139
140 w = (const void*) ((const int8_t*) w + 32);
141 k -= 8 * sizeof(int8_t);
142 } while (k != 0);
143 p -= 4 * sizeof(void*);
144 } while (p != 0);
145
146 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
147 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
148 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
149 __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
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 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
156 vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale0123);
157
158 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
159 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
160 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
161 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
162 vscaled3x0123 = _mm_min_ps(vscaled3x0123, voutput_max_less_zero_point);
163
164 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
165 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
166 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
167 vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
168
169 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
170 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
171 __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
172
173
174 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
175
176 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
177
178 if (nc >= 4) {
179 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout, 3));
180 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
181 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
182 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
183 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
184 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
185 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
186 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
187
188 a = (const int8_t**restrict) ((uintptr_t) a - ks);
189
190 nc -= 4;
191 } else {
192 if (nc & 2) {
193 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(vout, 6));
194 c3 += 2;
195 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
196 c2 += 2;
197 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
198 c1 += 2;
199 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
200 c0 += 2;
201 vout = _mm_srli_epi32(vout, 16);
202 }
203 if (nc & 1) {
204 *c3 = (int8_t) _mm_extract_epi8(vout, 12);
205 *c2 = (int8_t) _mm_extract_epi8(vout, 8);
206 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
207 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
208 }
209
210 nc = 0;
211 }
212 } while (nc != 0);
213 }
214