Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x8c16-minmax-fp32-wasmusdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c16-wasmdot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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_1x8c16__wasmusdot(
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, 16 * 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 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
57 XNN_FORCE_REALIZATION(vsign_mask);
58 do {
59 v128_t vacc0x0 = wasm_v128_load32_zero(w);
60 v128_t vacc0x1 = wasm_v128_load32_zero((const int32_t*) w + 1);
61 v128_t vacc0x2 = wasm_v128_load32_zero((const int32_t*) w + 2);
62 v128_t vacc0x3 = wasm_v128_load32_zero((const int32_t*) w + 3);
63 v128_t vacc0x4 = wasm_v128_load32_zero((const int32_t*) w + 4);
64 v128_t vacc0x5 = wasm_v128_load32_zero((const int32_t*) w + 5);
65 v128_t vacc0x6 = wasm_v128_load32_zero((const int32_t*) w + 6);
66 v128_t vacc0x7 = wasm_v128_load32_zero((const int32_t*) w + 7);
67 w = (const int32_t*) w + 8;
68
69 size_t p = ks;
70 do {
71 const int8_t* restrict a0 = a[0];
72 if XNN_UNPREDICTABLE(a0 != zero) {
73 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
74 }
75 a += 1;
76
77 size_t k = kc;
78 do {
79 const v128_t va0 = wasm_v128_xor(wasm_v128_load(a0), vsign_mask);
80 a0 += 16;
81
82 const v128_t vb0 = wasm_v128_load(w);
83 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
84 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
85 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
86 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
87 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
88 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
89 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
90 const v128_t vb4 = wasm_v128_load((const int8_t*) w + 64);
91 vacc0x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va0, vacc0x4);
92 const v128_t vb5 = wasm_v128_load((const int8_t*) w + 80);
93 vacc0x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va0, vacc0x5);
94 const v128_t vb6 = wasm_v128_load((const int8_t*) w + 96);
95 vacc0x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va0, vacc0x6);
96 const v128_t vb7 = wasm_v128_load((const int8_t*) w + 112);
97 vacc0x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va0, vacc0x7);
98
99 w = (const int8_t*) w + 128;
100 k -= 16 * sizeof(int8_t);
101 } while (k != 0);
102 p -= 1 * sizeof(void*);
103 } while (p != 0);
104
105 const v128_t vacc0x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x0, vacc0x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x0, vacc0x2, 2, 6, 3, 7));
106 const v128_t vacc0x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x1, vacc0x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x1, vacc0x3, 2, 6, 3, 7));
107 const v128_t vacc0x46 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x4, vacc0x6, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x4, vacc0x6, 2, 6, 3, 7));
108 const v128_t vacc0x57 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x5, vacc0x7, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x5, vacc0x7, 2, 6, 3, 7));
109
110 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
111 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x46, vacc0x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x46, vacc0x57, 2, 6, 3, 7));
112
113 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
114 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
115
116 const v128_t vscale0123 = wasm_v128_load(w);
117 w = (const float*) w + 4;
118 const v128_t vscale4567 = wasm_v128_load(w);
119 w = (const float*) w + 4;
120 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
121 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vscale4567);
122
123 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
124 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vmagic_bias);
125
126 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
127 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vmagic_min);
128
129 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
130 vacc0x4567 = wasm_i32x4_sub(vacc0x4567, vmagic_bias_less_output_zero_point);
131
132 v128_t vacc0x01234567 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x4567);
133
134 v128_t vout0x01234567_0x01234567 = wasm_i8x16_narrow_i16x8(vacc0x01234567, vacc0x01234567);
135
136 vout0x01234567_0x01234567 = wasm_i8x16_min(vout0x01234567_0x01234567, voutput_max);
137
138 if (nc >= 8) {
139 wasm_v128_store64_lane(c0, vout0x01234567_0x01234567, 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 -= 8;
146 } else {
147 if (nc & 4) {
148 wasm_v128_store32_lane(c0, vout0x01234567_0x01234567, 0);
149 c0 += 4;
150
151 vout0x01234567_0x01234567 = wasm_u64x2_shr(vout0x01234567_0x01234567, 32);
152 }
153 if (nc & 2) {
154 wasm_v128_store16_lane(c0, vout0x01234567_0x01234567, 0);
155 c0 += 2;
156
157 vout0x01234567_0x01234567 = wasm_u32x4_shr(vout0x01234567_0x01234567, 16);
158 }
159 if (nc & 1) {
160 wasm_v128_store8_lane(c0, vout0x01234567_0x01234567, 0);
161 }
162
163 nc = 0;
164 }
165 } while (nc != 0);
166 }