1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/gemm.h>
20 #include <xnnpack/math.h>
21 #include <xnnpack/unaligned.h>
22
23
24
xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2s4__xop(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)])25 void xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2s4__xop(
26 size_t mr,
27 size_t nc,
28 size_t kc,
29 const int8_t* restrict a,
30 size_t a_stride,
31 const void* restrict w,
32 int8_t* restrict c,
33 size_t cm_stride,
34 size_t cn_stride,
35 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 4);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(kc % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 8 * sizeof(int8_t));
47 const int8_t* a0 = a;
48 int8_t* c0 = c;
49 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
50 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
51 if XNN_UNPREDICTABLE(mr < 2) {
52 a1 = a0;
53 c1 = c0;
54 }
55 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
56 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
57 if XNN_UNPREDICTABLE(mr <= 2) {
58 a2 = a1;
59 c2 = c1;
60 }
61 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
62 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
63 if XNN_UNPREDICTABLE(mr != 4) {
64 a3 = a2;
65 c3 = c2;
66 }
67
68 do {
69 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
70 __m128i vacc1x0123 = vacc0x0123;
71 __m128i vacc2x0123 = vacc0x0123;
72 __m128i vacc3x0123 = vacc0x0123;
73 w = (const void*) ((const int32_t*) w + 4);
74
75 size_t k = kc;
76 do {
77 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
78 __m128i vxa0 = _mm_cvtepi8_epi16(va0);
79 a0 += 8;
80 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
81 __m128i vxa1 = _mm_cvtepi8_epi16(va1);
82 a1 += 8;
83 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
84 __m128i vxa2 = _mm_cvtepi8_epi16(va2);
85 a2 += 8;
86 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
87 __m128i vxa3 = _mm_cvtepi8_epi16(va3);
88 a3 += 8;
89
90 const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
91
92 vacc0x0123 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0123);
93 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
94 vacc1x0123 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0123);
95 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
96 vacc2x0123 = _mm_maddd_epi16(vxa2, vxb0, vacc2x0123);
97 vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
98 vacc3x0123 = _mm_maddd_epi16(vxa3, vxb0, vacc3x0123);
99 vxa3 = _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 3, 2, 1));
100 const __m128i vxb1 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 8));
101
102 vacc0x0123 = _mm_maddd_epi16(vxa0, vxb1, vacc0x0123);
103 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
104 vacc1x0123 = _mm_maddd_epi16(vxa1, vxb1, vacc1x0123);
105 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
106 vacc2x0123 = _mm_maddd_epi16(vxa2, vxb1, vacc2x0123);
107 vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
108 vacc3x0123 = _mm_maddd_epi16(vxa3, vxb1, vacc3x0123);
109 vxa3 = _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 3, 2, 1));
110 const __m128i vxb2 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 16));
111
112 vacc0x0123 = _mm_maddd_epi16(vxa0, vxb2, vacc0x0123);
113 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
114 vacc1x0123 = _mm_maddd_epi16(vxa1, vxb2, vacc1x0123);
115 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
116 vacc2x0123 = _mm_maddd_epi16(vxa2, vxb2, vacc2x0123);
117 vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
118 vacc3x0123 = _mm_maddd_epi16(vxa3, vxb2, vacc3x0123);
119 vxa3 = _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 3, 2, 1));
120 const __m128i vxb3 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 24));
121
122 vacc0x0123 = _mm_maddd_epi16(vxa0, vxb3, vacc0x0123);
123 vacc1x0123 = _mm_maddd_epi16(vxa1, vxb3, vacc1x0123);
124 vacc2x0123 = _mm_maddd_epi16(vxa2, vxb3, vacc2x0123);
125 vacc3x0123 = _mm_maddd_epi16(vxa3, vxb3, vacc3x0123);
126
127 w = (const void*) ((const int16_t*) w + 32);
128 k -= 8 * sizeof(int8_t);
129 } while (k != 0);
130
131 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
132 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
133 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
134 __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
135
136 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
137 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
138 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
139 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
140 vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale);
141
142 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
143 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
144 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
145 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
146 vscaled3x0123 = _mm_min_ps(vscaled3x0123, voutput_max_less_zero_point);
147
148 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
149 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
150 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
151 vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
152
153 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
154 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
155 __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
156
157
158 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
159
160 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
161
162 if (nc >= 4) {
163 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
164 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
165 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
166 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout, 3));
167
168 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
169 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
170 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
171 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
172
173 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
174 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
175 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
176 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
177
178 nc -= 4;
179 } else {
180 if (nc & 2) {
181 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
182 c0 += 2;
183 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
184 c1 += 2;
185 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
186 c2 += 2;
187 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(vout, 6));
188 c3 += 2;
189 vout = _mm_srli_epi32(vout, 16);
190 }
191 if (nc & 1) {
192 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
193 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
194 *c2 = (int8_t) _mm_extract_epi8(vout, 8);
195 *c3 = (int8_t) _mm_extract_epi8(vout, 12);
196 }
197
198 nc = 0;
199 }
200 } while (nc != 0);
201 }
202