Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-3x4c2s4-minmax-fp32-wasmsimd-dot16x2-ld128.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2s4-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/common.h"
15 #include "xnnpack/igemm.h"
16 #include "xnnpack/math.h"
17
18
19 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_3x4c2s4__wasmsimd_dot16x2_ld128(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 3);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (3 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
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 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53
54 kc = round_up_po2(kc, 8 * sizeof(int8_t));
55
56 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
57 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
58 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
59 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
60 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
61 XNN_FORCE_REALIZATION(vmagic_bias);
62 XNN_FORCE_REALIZATION(vmagic_min);
63 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
64 XNN_FORCE_REALIZATION(voutput_max);
65
66 do {
67 v128_t vacc0x0123 = wasm_v128_load(w);
68 v128_t vacc1x0123 = vacc0x0123;
69 v128_t vacc2x0123 = vacc0x0123;
70 w = (const void*) ((const int32_t*) w + 4);
71
72 size_t p = ks;
73 do {
74 const int8_t* restrict a0 = a[0];
75 if XNN_UNPREDICTABLE(a0 != zero) {
76 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
77 }
78 const int8_t* restrict a1 = a[1];
79 if XNN_UNPREDICTABLE(a1 != zero) {
80 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
81 }
82 const int8_t* restrict a2 = a[2];
83 if XNN_UNPREDICTABLE(a2 != zero) {
84 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
85 }
86 a += 3;
87
88 size_t k = kc;
89 do {
90 v128_t vxa0 = wasm_i16x8_load8x8(a0);
91 a0 += 8;
92 v128_t vxa1 = wasm_i16x8_load8x8(a1);
93 a1 += 8;
94 v128_t vxa2 = wasm_i16x8_load8x8(a2);
95 a2 += 8;
96
97 const v128_t vb01 = wasm_v128_load(w);
98 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
99 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
100
101 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb0));
102 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
103 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb0));
104 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
105 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb0));
106 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
107
108 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb1));
109 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
110 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb1));
111 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
112 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb1));
113 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
114 const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
115 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
116 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
117
118 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb2));
119 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
120 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb2));
121 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
122 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb2));
123 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
124
125 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb3));
126 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb3));
127 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb3));
128
129 w = (const int8_t*) w + 32;
130 k -= 8 * sizeof(int8_t);
131 } while (k != 0);
132 p -= 3 * sizeof(void*);
133 } while (p != 0);
134
135 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
136 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
137 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
138
139 const v128_t vscale0123 = wasm_v128_load(w);
140 w = (const float*) w + 4;
141 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
142 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
143 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale0123);
144
145 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
146 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
147 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
148
149 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
150 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
151 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
152
153 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
154 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
155 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
156
157 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
158 v128_t vacc22x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x0123);
159
160 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc22x0123);
161
162 vout = wasm_i8x16_min(vout, voutput_max);
163
164 if (nc >= 4) {
165 wasm_v128_store32_lane(c2, vout, 2);
166 wasm_v128_store32_lane(c1, vout, 1);
167 wasm_v128_store32_lane(c0, vout, 0);
168
169 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
170 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
171 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
172
173 a = (const int8_t**restrict) ((uintptr_t) a - ks);
174
175 nc -= 4;
176 } else {
177 if (nc & 2) {
178 wasm_v128_store16_lane(c2, vout, 4);
179 c2 += 2;
180 wasm_v128_store16_lane(c1, vout, 2);
181 c1 += 2;
182 wasm_v128_store16_lane(c0, vout, 0);
183 c0 += 2;
184
185 vout = wasm_u32x4_shr(vout, 16);
186 }
187 if (nc & 1) {
188 wasm_v128_store8_lane(c2, vout, 8);
189 wasm_v128_store8_lane(c1, vout, 4);
190 wasm_v128_store8_lane(c0, vout, 0);
191 }
192
193 nc = 0;
194 }
195 } while (nc != 0);
196 }