Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-2x4c8-minmax-avx-ld128.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_2x4c8__avx_ld128(
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 <= 2);
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 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
50 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
51 if XNN_UNPREDICTABLE(mr != 2) {
52 a1 = a0;
53 c1 = c0;
54 }
55
56 const __m128 vmin = _mm_set1_ps(params->scalar.min);
57 const __m128 vmax = _mm_set1_ps(params->scalar.max);
58 XNN_FORCE_REALIZATION(vmin);
59 XNN_FORCE_REALIZATION(vmax);
60
61 const __m128i vmask = _mm_set1_epi8(0xF0);
62 XNN_FORCE_REALIZATION(vmask);
63
64 do {
65 const __m128i vksum = _mm_load_si128((const __m128i*) w);
66 const __m128i vinput_zero_point0 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[0].zero_point));
67 const __m128i vinput_zero_point1 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[1].zero_point));
68 const __m128i vzero = _mm_setzero_si128();
69 const __m128i vinit0 = _mm_mullo_epi32(vksum, vinput_zero_point0);
70 const __m128i vinit1 = _mm_mullo_epi32(vksum, vinput_zero_point1);
71 __m128i vacc0x0 = _mm_blend_epi16(vinit0, vzero, 0xFC);
72 __m128i vacc0x1 = _mm_blend_epi16(vinit0, vzero, 0xF3);
73 __m128i vacc0x2 = _mm_blend_epi16(vinit0, vzero, 0xCF);
74 __m128i vacc0x3 = _mm_blend_epi16(vinit0, vzero, 0x3F);
75 __m128i vacc1x0 = _mm_blend_epi16(vinit1, vzero, 0xFC);
76 __m128i vacc1x1 = _mm_blend_epi16(vinit1, vzero, 0xF3);
77 __m128i vacc1x2 = _mm_blend_epi16(vinit1, vzero, 0xCF);
78 __m128i vacc1x3 = _mm_blend_epi16(vinit1, vzero, 0x3F);
79 w = (const int32_t*) w + 4;
80
81 size_t k = kc;
82
83 while (k >= 16 * sizeof(int8_t)) {
84 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
85 const __m128i vxa0c0 = _mm_cvtepi8_epi16(va0c0);
86 a0 += 8;
87 const __m128i va1c0 = _mm_loadl_epi64((const __m128i*) a1);
88 const __m128i vxa1c0 = _mm_cvtepi8_epi16(va1c0);
89 a1 += 8;
90
91 const __m128i vb01c01 = _mm_load_si128((const __m128i*) w);
92 const __m128i vbs01c0 = _mm_slli_epi32(vb01c01, 4);
93 const __m128i vb01c0 = _mm_and_si128(vbs01c0, vmask);
94 const __m128i vsb01c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c0);
95 const __m128i vxb0c0 = _mm_unpacklo_epi8(vb01c0, vsb01c0);
96 const __m128i vxb1c0 = _mm_unpackhi_epi8(vb01c0, vsb01c0);
97
98 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
99 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
100 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
101 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
102 const __m128i vb23c01 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
103 const __m128i vbs23c0 = _mm_slli_epi32(vb23c01, 4);
104 const __m128i vb23c0 = _mm_and_si128(vbs23c0, vmask);
105 const __m128i vsb23c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c0);
106 const __m128i vxb2c0 = _mm_unpacklo_epi8(vb23c0, vsb23c0);
107 const __m128i vxb3c0 = _mm_unpackhi_epi8(vb23c0, vsb23c0);
108
109 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
110 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
111 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
112 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
113
114 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
115 const __m128i vxa0c1 = _mm_cvtepi8_epi16(va0c1);
116 a0 += 8;
117 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
118 const __m128i vxa1c1 = _mm_cvtepi8_epi16(va1c1);
119 a1 += 8;
120
121 const __m128i vb01c1 = _mm_and_si128(vb01c01, vmask);
122 const __m128i vsb01c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c1);
123 const __m128i vxb0c1 = _mm_unpacklo_epi8(vb01c1, vsb01c1);
124 const __m128i vxb1c1 = _mm_unpackhi_epi8(vb01c1, vsb01c1);
125
126 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
127 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
128 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
129 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
130 const __m128i vb23c1 = _mm_and_si128(vb23c01, vmask);
131 const __m128i vsb23c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c1);
132 const __m128i vxb2c1 = _mm_unpacklo_epi8(vb23c1, vsb23c1);
133 const __m128i vxb3c1 = _mm_unpackhi_epi8(vb23c1, vsb23c1);
134
135 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
136 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
137 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
138 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
139
140
141 w = (const int8_t*) w + 32;
142 k -= 16 * sizeof(int8_t);
143 }
144
145 while (k >= 8 * sizeof(int8_t)) {
146 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
147 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
148 a0 += 8;
149 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
150 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
151 a1 += 8;
152
153 __m128i vb01 = _mm_load_si128((const __m128i*) w);
154 vb01 = _mm_slli_epi32(vb01, 4);
155 vb01 = _mm_and_si128(vb01, vmask);
156
157 const __m128i vxbm1 = _mm_unpackhi_epi8(vb01, vb01);
158 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
159 const __m128i vxb1 = _mm_srai_epi16(vxbm1, 8);
160
161 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
162 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
163 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
164 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
165 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
166 vb23 = _mm_slli_epi32(vb23, 4);
167 vb23 = _mm_and_si128(vb23, vmask);
168
169 const __m128i vxbm3 = _mm_unpackhi_epi8(vb23, vb23);
170 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
171 const __m128i vxb3 = _mm_srai_epi16(vxbm3, 8);
172
173 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
174 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
175 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
176 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
177
178 w = (const int8_t*) w + 32;
179 k -= 8 * sizeof(int8_t);
180 }
181
182 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
183 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
184 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
185 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
186
187 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
188 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
189
190 vacc0x0123 = _mm_srai_epi32(vacc0x0123, 4);
191 vacc1x0123 = _mm_srai_epi32(vacc1x0123, 4);
192 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
193 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
194
195 const __m128 vinput_scale0 = _mm_broadcast_ss(&quantization_params[0].inv_scale);
196 const __m128 vinput_scale1 = _mm_broadcast_ss(&quantization_params[1].inv_scale);
197
198 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
199 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
200
201 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
202 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
203 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
204
205 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
206 w = (const float*) w + 8;
207 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
208 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
209
210 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
211 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
212
213 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
214 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
215
216 if XNN_LIKELY(nc >= 4) {
217 _mm_storeu_ps(c0, vout0x0123);
218 _mm_storeu_ps(c1, vout1x0123);
219
220 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
221 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
222
223 c0 = (float*) ((uintptr_t) c0 + cn_stride);
224 c1 = (float*) ((uintptr_t) c1 + cn_stride);
225
226 nc -= 4;
227 } else {
228 if (nc & 2) {
229 _mm_storel_pi((__m64*) c0, vout0x0123);
230 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
231 c0 += 2;
232 _mm_storel_pi((__m64*) c1, vout1x0123);
233 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
234 c1 += 2;
235 }
236 if (nc & 1) {
237 _mm_store_ss(c0, vout0x0123);
238 _mm_store_ss(c1, vout1x0123);
239 }
240 nc = 0;
241 }
242 } while (nc != 0);
243 }