Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-2x4c2-minmax-fp32-wasmsimd-dot16x2-ld128.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_qs8_qc8w_igemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld128(
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 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 2 * sizeof(int8_t));
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
47 c1 = c0;
48 }
49
50
51 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
52 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
53 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
54 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
55 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
56 XNN_FORCE_REALIZATION(vmagic_bias);
57 XNN_FORCE_REALIZATION(vmagic_min);
58 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
59 XNN_FORCE_REALIZATION(voutput_max);
60
61
62 do {
63 v128_t vacc0x0123 = wasm_v128_load(w);
64 v128_t vacc1x0123 = vacc0x0123;
65 w = (const void*) ((const int32_t*) w + 4);
66
67 size_t p = ks;
68 do {
69 const int8_t* restrict a0 = a[0];
70 if XNN_UNPREDICTABLE(a0 != zero) {
71 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
72 }
73 const int8_t* restrict a1 = a[1];
74 if XNN_UNPREDICTABLE(a1 != zero) {
75 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
76 }
77 a += 2;
78
79 size_t k = kc;
80 while (k >= 8 * sizeof(int8_t)) {
81 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
82 a0 += 8;
83 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
84 a1 += 8;
85
86 const v128_t vb01 = wasm_v128_load(w);
87 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
88 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
89
90 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
91 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
92 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
93 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
94
95 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
96 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
97 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
98 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
99 const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
100 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
101 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
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
108 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
109 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
110 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
111 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
112
113 w = (const void*) ((const int8_t*) w + 32);
114 k -= 8 * sizeof(int8_t);
115 }
116 if (k != 0) {
117 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
118 a0 = (const int8_t*) ((uintptr_t) a0 + k);
119 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
120 a1 = (const int8_t*) ((uintptr_t) a1 + k);
121
122 const v128_t vxb0 = wasm_i16x8_load8x8(w);
123 w = (const int8_t*) w + 8;
124
125 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
126 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
127 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
128 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
129
130 if (k > 2 * sizeof(int8_t)) {
131 const v128_t vxb1 = wasm_i16x8_load8x8(w);
132 w = (const int8_t*) w + 8;
133
134 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
135 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
136 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
137 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
138
139 if (k > 4 * sizeof(int8_t)) {
140 const v128_t vxb2 = wasm_i16x8_load8x8(w);
141 w = (const int8_t*) w + 8;
142
143 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
144 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
145 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
146 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
147 }
148 }
149 }
150 p -= 2 * sizeof(void*);
151 } while (p != 0);
152
153 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
154 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
155
156 const v128_t vscale0123 = wasm_v128_load(w);
157 w = (const float*) w + 4;
158 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
159 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
160
161 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
162 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
163
164 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
165 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
166
167 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
168 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
169
170 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
171
172 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc01x0123);
173
174 vout = wasm_i8x16_min(vout, voutput_max);
175
176 if (nc >= 4) {
177 wasm_v128_store32_lane(c1, vout, 1);
178 wasm_v128_store32_lane(c0, vout, 0);
179
180 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
181 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
182
183 a = (const int8_t**restrict) ((uintptr_t) a - ks);
184
185 nc -= 4;
186 } else {
187 if (nc & 2) {
188 wasm_v128_store16_lane(c1, vout, 2);
189 c1 += 2;
190 wasm_v128_store16_lane(c0, vout, 0);
191 c0 += 2;
192
193 vout = wasm_u32x4_shr(vout, 16);
194 }
195 if (nc & 1) {
196 wasm_v128_store8_lane(c1, vout, 4);
197 wasm_v128_store8_lane(c0, vout, 0);
198 }
199
200 nc = 0;
201 }
202 } while (nc != 0);
203 }