Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x16c4-minmax-fp32-wasmsdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx16c4-wasmdot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 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/gemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x16c4__wasmsdot(
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 <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (1 * 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, 4 * sizeof(int8_t));
44 int8_t* c0 = c;
45
46 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
47 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
48 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
49 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
50 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
51 //XNN_FORCE_REALIZATION(vmagic_bias);
52 //XNN_FORCE_REALIZATION(vmagic_min);
53 //XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
54 //XNN_FORCE_REALIZATION(voutput_max);
55
56 do {
57 v128_t vacc0x0123 = wasm_v128_load(w);
58 v128_t vacc0x4567 = wasm_v128_load((const int32_t*) w + 4);
59 v128_t vacc0x89AB = wasm_v128_load((const int32_t*) w + 8);
60 v128_t vacc0xCDEF = wasm_v128_load((const int32_t*) w + 12);
61 w = (const int32_t*) w + 16;
62
63 size_t p = ks;
64 do {
65 const int8_t* restrict a0 = a[0];
66 if XNN_UNPREDICTABLE(a0 != zero) {
67 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
68 }
69 a += 1;
70
71 size_t k = kc;
72
73 while (k != 0) {
74 v128_t va0x0123 = wasm_v128_load32_splat(a0);
75 a0 += 4;
76
77
78 const v128_t vb0123 = wasm_v128_load((const int8_t*) w);
79 const v128_t vb4567 = wasm_v128_load((const int8_t*) w + 16);
80 const v128_t vb89AB = wasm_v128_load((const int8_t*) w + 32);
81 const v128_t vbCDEF = wasm_v128_load((const int8_t*) w + 48);
82
83 vacc0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va0x0123, vacc0x0123);
84 vacc0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va0x0123, vacc0x4567);
85 vacc0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va0x0123, vacc0x89AB);
86 vacc0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va0x0123, vacc0xCDEF);
87
88 w = (const int8_t*) w + 64;
89 k -= 4 * sizeof(int8_t);
90 }
91
92 p -= 1 * sizeof(void*);
93 } while (p != 0);
94
95 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
96 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
97 vacc0x89AB = wasm_f32x4_convert_i32x4(vacc0x89AB);
98 vacc0xCDEF = wasm_f32x4_convert_i32x4(vacc0xCDEF);
99
100 const v128_t vscale0123 = wasm_v128_load(w);
101 w = (const float*) w + 4;
102 const v128_t vscale4567 = wasm_v128_load(w);
103 w = (const float*) w + 4;
104 const v128_t vscale89AB = wasm_v128_load(w);
105 w = (const float*) w + 4;
106 const v128_t vscaleCDEF = wasm_v128_load(w);
107 w = (const float*) w + 4;
108 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
109 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vscale4567);
110 vacc0x89AB = wasm_f32x4_mul(vacc0x89AB, vscale89AB);
111 vacc0xCDEF = wasm_f32x4_mul(vacc0xCDEF, vscaleCDEF);
112
113 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
114 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vmagic_bias);
115 vacc0x89AB = wasm_f32x4_add(vacc0x89AB, vmagic_bias);
116 vacc0xCDEF = wasm_f32x4_add(vacc0xCDEF, vmagic_bias);
117
118 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
119 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vmagic_min);
120 vacc0x89AB = wasm_i32x4_max(vacc0x89AB, vmagic_min);
121 vacc0xCDEF = wasm_i32x4_max(vacc0xCDEF, vmagic_min);
122
123 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
124 vacc0x4567 = wasm_i32x4_sub(vacc0x4567, vmagic_bias_less_output_zero_point);
125 vacc0x89AB = wasm_i32x4_sub(vacc0x89AB, vmagic_bias_less_output_zero_point);
126 vacc0xCDEF = wasm_i32x4_sub(vacc0xCDEF, vmagic_bias_less_output_zero_point);
127
128 v128_t vacc0x01234567 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x4567);
129 v128_t vacc0x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc0x89AB, vacc0xCDEF);
130
131 vacc0x01234567 = wasm_i8x16_narrow_i16x8(vacc0x01234567, vacc0x01234567);
132 vacc0x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc0x89ABCDEF, vacc0x89ABCDEF);
133
134 vacc0x01234567 = wasm_i8x16_min(vacc0x01234567, voutput_max);
135 vacc0x89ABCDEF = wasm_i8x16_min(vacc0x89ABCDEF, voutput_max);
136
137 if (nc >= 16) {
138 wasm_v128_store64_lane(c0, vacc0x01234567, 0);
139 wasm_v128_store64_lane(c0 + 8, vacc0x89ABCDEF, 0);
140
141 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
142
143 a = (const int8_t**restrict) ((uintptr_t) a - ks);
144
145 nc -= 16;
146 } else {
147 if (nc & 8) {
148 wasm_v128_store64_lane(c0, vacc0x01234567, 0);
149 c0 += 8;
150
151 vacc0x01234567 = vacc0x89ABCDEF;
152 }
153 if (nc & 4) {
154 wasm_v128_store32_lane(c0, vacc0x01234567, 0);
155 c0 += 4;
156
157 vacc0x01234567 = wasm_u64x2_shr(vacc0x01234567, 32);
158 }
159 if (nc & 2) {
160 wasm_v128_store16_lane(c0, vacc0x01234567, 0);
161 c0 += 2;
162
163 vacc0x01234567 = wasm_u32x4_shr(vacc0x01234567, 16);
164 }
165 if (nc & 1) {
166 wasm_v128_store8_lane(c0, vacc0x01234567, 0);
167 }
168
169 nc = 0;
170 }
171 } while (nc != 0);
172 }