Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-1x4c8-minmax-sse2-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c8-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 <emmintrin.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18
19
20 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_1x4c8__sse2_ld64(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const int8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
31 const struct xnn_qd8_quantization_params quantization_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(kc % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 8 * sizeof(int8_t));
43 const int8_t* a0 = a;
44 float* c0 = c;
45
46 const __m128 vmin = _mm_set1_ps(params->scalar.min);
47 const __m128 vmax = _mm_set1_ps(params->scalar.max);
48 XNN_FORCE_REALIZATION(vmin);
49 XNN_FORCE_REALIZATION(vmax);
50
51 const __m128i vmask = _mm_set1_epi8(0xF0);
52 XNN_FORCE_REALIZATION(vmask);
53
54 do {
55 const __m128i vksum = _mm_load_si128((const __m128i*) w);
56 __m128i vinput_zero_point0 = _mm_cvtsi32_si128(*((const int*) &quantization_params[0].zero_point));
57 vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point0, _MM_SHUFFLE(0, 0, 0, 0));
58 const __m128i vzero = _mm_setzero_si128();
59 const __m128i vksum_lo = _mm_srli_epi32(_mm_slli_epi32(vksum, 16), 16);
60 const __m128i vksum_hi = _mm_srli_epi32(vksum, 16);
61
62 __m128i vzpprodksumhi0 = _mm_mulhi_epu16(vinput_zero_point0, vksum_lo);
63 const __m128i vzpprodksumlo0 = _mm_mullo_epi16(vinput_zero_point0, vksum_lo);
64
65 vzpprodksumhi0 = _mm_add_epi16(vzpprodksumhi0, _mm_mullo_epi16(vinput_zero_point0, vksum_hi));
66 vzpprodksumhi0 = _mm_sub_epi16(vzpprodksumhi0, _mm_and_si128(_mm_srai_epi16(vinput_zero_point0, 15), vksum_lo));
67
68 vzpprodksumhi0 = _mm_slli_si128(vzpprodksumhi0, 2);
69
70 const __m128i vksumzp0 = _mm_or_si128(vzpprodksumhi0, vzpprodksumlo0);
71
72 const __m128i vksum010 = _mm_unpacklo_epi32(vksumzp0, vzero);
73 const __m128i vksum230 = _mm_unpackhi_epi32(vksumzp0, vzero);
74
75 __m128i vacc0x0 = _mm_unpacklo_epi64(vksum010, vzero);
76 __m128i vacc0x1 = _mm_unpackhi_epi64(vksum010, vzero);
77
78 __m128i vacc0x2 = _mm_unpacklo_epi64(vksum230, vzero);
79 __m128i vacc0x3 = _mm_unpackhi_epi64(vksum230, vzero);
80 w = (const int32_t*) w + 4;
81
82 size_t k = kc;
83
84 while (k >= 16 * sizeof(int8_t)) {
85 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
86 const __m128i vxa0c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c0, va0c0), 8);
87 a0 += 8;
88
89 const __m128i vb0c01 = _mm_loadl_epi64((const __m128i*) w);
90 const __m128i vbs0c0 = _mm_slli_epi32(vb0c01, 4);
91 const __m128i vb0c0 = _mm_and_si128(vbs0c0, vmask);
92 const __m128i vxb0c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0c0, vb0c0), 8);
93
94 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
95 const __m128i vb1c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
96 const __m128i vbs1c0 = _mm_slli_epi32(vb1c01, 4);
97 const __m128i vb1c0 = _mm_and_si128(vbs1c0, vmask);
98 const __m128i vxb1c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1c0, vb1c0), 8);
99
100 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
101 const __m128i vb2c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
102 const __m128i vbs2c0 = _mm_slli_epi32(vb2c01, 4);
103 const __m128i vb2c0 = _mm_and_si128(vbs2c0, vmask);
104 const __m128i vxb2c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2c0, vb2c0), 8);
105
106 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
107 const __m128i vb3c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
108 const __m128i vbs3c0 = _mm_slli_epi32(vb3c01, 4);
109 const __m128i vb3c0 = _mm_and_si128(vbs3c0, vmask);
110 const __m128i vxb3c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3c0, vb3c0), 8);
111
112 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
113
114 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
115 const __m128i vxa0c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c1, va0c1), 8);
116 a0 += 8;
117
118 const __m128i vb0c1 = _mm_and_si128(vb0c01, vmask);
119 const __m128i vxb0c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0c1, vb0c1), 8);
120
121 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
122 const __m128i vb1c1 = _mm_and_si128(vb1c01, vmask);
123 const __m128i vxb1c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1c1, vb1c1), 8);
124
125 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
126 const __m128i vb2c1 = _mm_and_si128(vb2c01, vmask);
127 const __m128i vxb2c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2c1, vb2c1), 8);
128
129 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
130 const __m128i vb3c1 = _mm_and_si128(vb3c01, vmask);
131 const __m128i vxb3c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3c1, vb3c1), 8);
132
133 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
134
135
136 w = (const int8_t*) w + 32;
137 k -= 16 * sizeof(int8_t);
138 }
139
140 while (k >= 8 * sizeof(int8_t)) {
141 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
142 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
143 a0 += 8;
144
145 __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
146 vb0 = _mm_slli_epi32(vb0, 4);
147 vb0 = _mm_and_si128(vb0, vmask);
148
149 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
150
151 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
152 __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
153 vb1 = _mm_slli_epi32(vb1, 4);
154 vb1 = _mm_and_si128(vb1, vmask);
155
156 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
157
158 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
159 __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
160 vb2 = _mm_slli_epi32(vb2, 4);
161 vb2 = _mm_and_si128(vb2, vmask);
162
163 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
164
165 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
166 __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
167 vb3 = _mm_slli_epi32(vb3, 4);
168 vb3 = _mm_and_si128(vb3, vmask);
169
170 const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
171
172 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
173
174 w = (const int8_t*) w + 32;
175 k -= 8 * sizeof(int8_t);
176 }
177
178 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
179 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
180
181 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
182
183 vacc0x0123 = _mm_srai_epi32(vacc0x0123, 4);
184 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
185
186 const __m128 vinput_scale0 = _mm_load1_ps(&quantization_params[0].inv_scale);
187
188 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
189
190 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
191 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
192
193 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
194 w = (const float*) w + 8;
195 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
196
197 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
198
199 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
200
201 if XNN_LIKELY(nc >= 4) {
202 _mm_storeu_ps(c0, vout0x0123);
203
204 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
205
206 c0 = (float*) ((uintptr_t) c0 + cn_stride);
207
208 nc -= 4;
209 } else {
210 if (nc & 2) {
211 _mm_storel_pi((__m64*) c0, vout0x0123);
212 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
213 c0 += 2;
214 }
215 if (nc & 1) {
216 _mm_store_ss(c0, vout0x0123);
217 }
218 nc = 0;
219 }
220 } while (nc != 0);
221 }