Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-2x4c2-minmax-wasmsimd-dot16x2-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2-wasmsimd-dot16x2.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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 <wasm_simd128.h>
13
14 #include "xnnpack/igemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_2x4c2__wasmsimd_dot16x2_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const int8_t* zero_data,
31 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 2);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (2 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 2 * sizeof(int8_t));
46 float* c0 = c;
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr != 2) {
49 c1 = c0;
50 }
51
52 const v128_t vinput_zero_point = wasm_v128_load32_splat(&quantization_params->zero_point);
53 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
54 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
55 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
56 XNN_FORCE_REALIZATION(vmin);
57 XNN_FORCE_REALIZATION(vmax);
58
59
60 do {
61 const v128_t vksum0123 = wasm_v128_load(w);
62 v128_t vacc0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
63 v128_t vacc1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
64 w = (const void*) ((const int32_t*) w + 4);
65
66 size_t p = ks;
67 do {
68 const int8_t* restrict a0 = a[0];
69 if XNN_UNPREDICTABLE(a0 != zero) {
70 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
71 } else {
72 a0 = zero_data;
73 }
74 const int8_t* restrict a1 = a[1];
75 if XNN_UNPREDICTABLE(a1 != zero) {
76 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
77 } else {
78 a1 = zero_data;
79 }
80 a += 2;
81
82 size_t k = kc;
83 while (k >= 8 * sizeof(int8_t)) {
84 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
85 a0 += 8;
86 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
87 a1 += 8;
88
89 const v128_t vxb0 = wasm_i16x8_load8x8(w);
90
91 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
92 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
93 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
94 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
95 const v128_t vxb1 = wasm_i16x8_load8x8((const int8_t*) w + 8);
96
97 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
98 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
99 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
100 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
101 const v128_t vxb2 = wasm_i16x8_load8x8((const int8_t*) w + 16);
102
103 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
104 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
105 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
106 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
107 const v128_t vxb3 = wasm_i16x8_load8x8((const int8_t*) w + 24);
108
109 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
110 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
111 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
112 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
113
114 w = (const void*) ((const int8_t*) w + 32);
115 k -= 8 * sizeof(int8_t);
116 }
117 if (k != 0) {
118 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
119 a0 = (const int8_t*) ((uintptr_t) a0 + k);
120 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
121 a1 = (const int8_t*) ((uintptr_t) a1 + k);
122
123 const v128_t vxb0 = wasm_i16x8_load8x8(w);
124 w = (const int8_t*) w + 8;
125
126 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
127 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
128 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
129 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
130
131 if (k > 2 * sizeof(int8_t)) {
132 const v128_t vxb1 = wasm_i16x8_load8x8(w);
133 w = (const int8_t*) w + 8;
134
135 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
136 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
137 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
138 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
139
140 if (k > 4 * sizeof(int8_t)) {
141 const v128_t vxb2 = wasm_i16x8_load8x8(w);
142 w = (const int8_t*) w + 8;
143
144 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
145 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
146 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
147 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
148 }
149 }
150 }
151 p -= 2 * sizeof(void*);
152 } while (p != 0);
153
154 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
155 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
156
157 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale);
158 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale);
159
160 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
161 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
162 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
163 w = (const float*) w + 4;
164
165 const v128_t vbias0123 = wasm_v128_load(w);
166 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
167 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
168 w = (const float*) w + 4;
169
170 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
171 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
172
173 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
174 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
175
176 if XNN_LIKELY(nc >= 4) {
177 wasm_v128_store(c1, vacc1x0123);
178 wasm_v128_store(c0, vacc0x0123);
179
180 a = (const int8_t**restrict) ((uintptr_t) a - ks);
181
182 c0 = (float*) ((uintptr_t) c0 + cn_stride);
183 c1 = (float*) ((uintptr_t) c1 + cn_stride);
184
185 nc -= 4;
186 } else {
187 if (nc & 2) {
188 wasm_v128_store64_lane(c1, vacc1x0123, 0);
189 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
190 c1 += 2;
191 wasm_v128_store64_lane(c0, vacc0x0123, 0);
192 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
193 c0 += 2;
194 }
195 if (nc & 1) {
196 wasm_v128_store32_lane(c1, vacc1x0123, 0);
197 wasm_v128_store32_lane(c0, vacc0x0123, 0);
198 }
199 nc = 0;
200 }
201 } while (nc != 0);
202 }