Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-1x4c8-minmax-sse41-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 <smmintrin.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__sse41_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 vinit0 = _mm_mullo_epi32(vksum, vinput_zero_point0);
60 __m128i vacc0x0 = _mm_blend_epi16(vinit0, vzero, 0xFC);
61 __m128i vacc0x1 = _mm_blend_epi16(vinit0, vzero, 0xF3);
62 __m128i vacc0x2 = _mm_blend_epi16(vinit0, vzero, 0xCF);
63 __m128i vacc0x3 = _mm_blend_epi16(vinit0, vzero, 0x3F);
64 w = (const int32_t*) w + 4;
65
66 size_t k = kc;
67
68 while (k >= 16 * sizeof(int8_t)) {
69 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
70 const __m128i vxa0c0 = _mm_cvtepi8_epi16(va0c0);
71 a0 += 8;
72
73 const __m128i vb0c01 = _mm_loadl_epi64((const __m128i*) w);
74 const __m128i vbs0c0 = _mm_slli_epi32(vb0c01, 4);
75 const __m128i vb0c0 = _mm_and_si128(vbs0c0, vmask);
76 const __m128i vxb0c0 = _mm_cvtepi8_epi16(vb0c0);
77
78 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
79 const __m128i vb1c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
80 const __m128i vbs1c0 = _mm_slli_epi32(vb1c01, 4);
81 const __m128i vb1c0 = _mm_and_si128(vbs1c0, vmask);
82 const __m128i vxb1c0 = _mm_cvtepi8_epi16(vb1c0);
83
84 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
85 const __m128i vb2c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
86 const __m128i vbs2c0 = _mm_slli_epi32(vb2c01, 4);
87 const __m128i vb2c0 = _mm_and_si128(vbs2c0, vmask);
88 const __m128i vxb2c0 = _mm_cvtepi8_epi16(vb2c0);
89
90 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
91 const __m128i vb3c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
92 const __m128i vbs3c0 = _mm_slli_epi32(vb3c01, 4);
93 const __m128i vb3c0 = _mm_and_si128(vbs3c0, vmask);
94 const __m128i vxb3c0 = _mm_cvtepi8_epi16(vb3c0);
95
96 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
97
98 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
99 const __m128i vxa0c1 = _mm_cvtepi8_epi16(va0c1);
100 a0 += 8;
101
102 const __m128i vb0c1 = _mm_and_si128(vb0c01, vmask);
103 const __m128i vxb0c1 = _mm_cvtepi8_epi16(vb0c1);
104
105 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
106 const __m128i vb1c1 = _mm_and_si128(vb1c01, vmask);
107 const __m128i vxb1c1 = _mm_cvtepi8_epi16(vb1c1);
108
109 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
110 const __m128i vb2c1 = _mm_and_si128(vb2c01, vmask);
111 const __m128i vxb2c1 = _mm_cvtepi8_epi16(vb2c1);
112
113 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
114 const __m128i vb3c1 = _mm_and_si128(vb3c01, vmask);
115 const __m128i vxb3c1 = _mm_cvtepi8_epi16(vb3c1);
116
117 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
118
119
120 w = (const int8_t*) w + 32;
121 k -= 16 * sizeof(int8_t);
122 }
123
124 while (k >= 8 * sizeof(int8_t)) {
125 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
126 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
127 a0 += 8;
128
129 __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
130 vb0 = _mm_slli_epi32(vb0, 4);
131 vb0 = _mm_and_si128(vb0, vmask);
132
133 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
134
135 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
136 __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
137 vb1 = _mm_slli_epi32(vb1, 4);
138 vb1 = _mm_and_si128(vb1, vmask);
139
140 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
141
142 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
143 __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
144 vb2 = _mm_slli_epi32(vb2, 4);
145 vb2 = _mm_and_si128(vb2, vmask);
146
147 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
148
149 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
150 __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
151 vb3 = _mm_slli_epi32(vb3, 4);
152 vb3 = _mm_and_si128(vb3, vmask);
153
154 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
155
156 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
157
158 w = (const int8_t*) w + 32;
159 k -= 8 * sizeof(int8_t);
160 }
161
162 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
163 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
164
165 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
166
167 vacc0x0123 = _mm_srai_epi32(vacc0x0123, 4);
168 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
169
170 const __m128 vinput_scale0 = _mm_load1_ps(&quantization_params[0].inv_scale);
171
172 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
173
174 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
175 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
176
177 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
178 w = (const float*) w + 8;
179 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
180
181 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
182
183 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
184
185 if XNN_LIKELY(nc >= 4) {
186 _mm_storeu_ps(c0, vout0x0123);
187
188 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
189
190 c0 = (float*) ((uintptr_t) c0 + cn_stride);
191
192 nc -= 4;
193 } else {
194 if (nc & 2) {
195 _mm_storel_pi((__m64*) c0, vout0x0123);
196 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
197 c0 += 2;
198 }
199 if (nc & 1) {
200 _mm_store_ss(c0, vout0x0123);
201 }
202 nc = 0;
203 }
204 } while (nc != 0);
205 }