Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-1x4c8-minmax-avx-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 #ifdef _MSC_VER
13 #include <intrin.h>
14 #else
15 #include <x86intrin.h>
16 #endif
17
18 #include "xnnpack/common.h"
19 #include "xnnpack/gemm.h"
20 #include "xnnpack/math.h"
21 #include "xnnpack/unaligned.h"
22
23
24 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_1x4c8__avx_ld64(
25 size_t mr,
26 size_t nc,
27 size_t kc,
28 const int8_t* restrict a,
29 size_t a_stride,
30 const void* restrict w,
31 float* restrict c,
32 size_t cm_stride,
33 size_t cn_stride,
34 const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
35 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 1);
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 float* c0 = c;
49
50 const __m128 vmin = _mm_set1_ps(params->scalar.min);
51 const __m128 vmax = _mm_set1_ps(params->scalar.max);
52 XNN_FORCE_REALIZATION(vmin);
53 XNN_FORCE_REALIZATION(vmax);
54
55 const __m128i vmask = _mm_set1_epi8(0xF0);
56 XNN_FORCE_REALIZATION(vmask);
57
58 do {
59 const __m128i vksum = _mm_load_si128((const __m128i*) w);
60 const __m128i vinput_zero_point0 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[0].zero_point));
61 const __m128i vzero = _mm_setzero_si128();
62 const __m128i vinit0 = _mm_mullo_epi32(vksum, vinput_zero_point0);
63 __m128i vacc0x0 = _mm_blend_epi16(vinit0, vzero, 0xFC);
64 __m128i vacc0x1 = _mm_blend_epi16(vinit0, vzero, 0xF3);
65 __m128i vacc0x2 = _mm_blend_epi16(vinit0, vzero, 0xCF);
66 __m128i vacc0x3 = _mm_blend_epi16(vinit0, vzero, 0x3F);
67 w = (const int32_t*) w + 4;
68
69 size_t k = kc;
70
71 while (k >= 16 * sizeof(int8_t)) {
72 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
73 const __m128i vxa0c0 = _mm_cvtepi8_epi16(va0c0);
74 a0 += 8;
75
76 const __m128i vb0c01 = _mm_loadl_epi64((const __m128i*) w);
77 const __m128i vbs0c0 = _mm_slli_epi32(vb0c01, 4);
78 const __m128i vb0c0 = _mm_and_si128(vbs0c0, vmask);
79 const __m128i vxb0c0 = _mm_cvtepi8_epi16(vb0c0);
80
81 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
82 const __m128i vb1c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
83 const __m128i vbs1c0 = _mm_slli_epi32(vb1c01, 4);
84 const __m128i vb1c0 = _mm_and_si128(vbs1c0, vmask);
85 const __m128i vxb1c0 = _mm_cvtepi8_epi16(vb1c0);
86
87 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
88 const __m128i vb2c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
89 const __m128i vbs2c0 = _mm_slli_epi32(vb2c01, 4);
90 const __m128i vb2c0 = _mm_and_si128(vbs2c0, vmask);
91 const __m128i vxb2c0 = _mm_cvtepi8_epi16(vb2c0);
92
93 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
94 const __m128i vb3c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
95 const __m128i vbs3c0 = _mm_slli_epi32(vb3c01, 4);
96 const __m128i vb3c0 = _mm_and_si128(vbs3c0, vmask);
97 const __m128i vxb3c0 = _mm_cvtepi8_epi16(vb3c0);
98
99 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
100
101 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
102 const __m128i vxa0c1 = _mm_cvtepi8_epi16(va0c1);
103 a0 += 8;
104
105 const __m128i vb0c1 = _mm_and_si128(vb0c01, vmask);
106 const __m128i vxb0c1 = _mm_cvtepi8_epi16(vb0c1);
107
108 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
109 const __m128i vb1c1 = _mm_and_si128(vb1c01, vmask);
110 const __m128i vxb1c1 = _mm_cvtepi8_epi16(vb1c1);
111
112 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
113 const __m128i vb2c1 = _mm_and_si128(vb2c01, vmask);
114 const __m128i vxb2c1 = _mm_cvtepi8_epi16(vb2c1);
115
116 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
117 const __m128i vb3c1 = _mm_and_si128(vb3c01, vmask);
118 const __m128i vxb3c1 = _mm_cvtepi8_epi16(vb3c1);
119
120 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
121
122
123 w = (const int8_t*) w + 32;
124 k -= 16 * sizeof(int8_t);
125 }
126
127 while (k >= 8 * sizeof(int8_t)) {
128 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
129 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
130 a0 += 8;
131
132 __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
133 vb0 = _mm_slli_epi32(vb0, 4);
134 vb0 = _mm_and_si128(vb0, vmask);
135
136 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
137
138 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
139 __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
140 vb1 = _mm_slli_epi32(vb1, 4);
141 vb1 = _mm_and_si128(vb1, vmask);
142
143 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
144
145 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
146 __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
147 vb2 = _mm_slli_epi32(vb2, 4);
148 vb2 = _mm_and_si128(vb2, vmask);
149
150 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
151
152 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
153 __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
154 vb3 = _mm_slli_epi32(vb3, 4);
155 vb3 = _mm_and_si128(vb3, vmask);
156
157 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
158
159 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
160
161 w = (const int8_t*) w + 32;
162 k -= 8 * sizeof(int8_t);
163 }
164
165 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
166 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
167
168 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
169
170 vacc0x0123 = _mm_srai_epi32(vacc0x0123, 4);
171 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
172
173 const __m128 vinput_scale0 = _mm_broadcast_ss(&quantization_params[0].inv_scale);
174
175 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
176
177 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
178 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
179
180 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
181 w = (const float*) w + 8;
182 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
183
184 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
185
186 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
187
188 if XNN_LIKELY(nc >= 4) {
189 _mm_storeu_ps(c0, vout0x0123);
190
191 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
192
193 c0 = (float*) ((uintptr_t) c0 + cn_stride);
194
195 nc -= 4;
196 } else {
197 if (nc & 2) {
198 _mm_storel_pi((__m64*) c0, vout0x0123);
199 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
200 c0 += 2;
201 }
202 if (nc & 1) {
203 _mm_store_ss(c0, vout0x0123);
204 }
205 nc = 0;
206 }
207 } while (nc != 0);
208 }