Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-2x4c16-minmax-wasmsdot.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_qd8_f32_qc8w_igemm_minmax_ukernel_2x4c16__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 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, 16 * 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 vmin = wasm_v128_load32_splat(&params->scalar.min);
53 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
54 XNN_FORCE_REALIZATION(vmin);
55 XNN_FORCE_REALIZATION(vmax);
56
57 do {
58 v128_t vksum0 = wasm_v128_load32_zero(w);
59 v128_t vksum1 = wasm_v128_load32_zero((const int32_t*) w + 1);
60 v128_t vksum2 = wasm_v128_load32_zero((const int32_t*) w + 2);
61 v128_t vksum3 = wasm_v128_load32_zero((const int32_t*) w + 3);
62 const v128_t vinput_zero_point = wasm_v128_load32_splat(&quantization_params->zero_point);
63 v128_t vacc0x0 = wasm_i32x4_mul(vksum0, vinput_zero_point);
64 v128_t vacc0x1 = wasm_i32x4_mul(vksum1, vinput_zero_point);
65 v128_t vacc0x2 = wasm_i32x4_mul(vksum2, vinput_zero_point);
66 v128_t vacc0x3 = wasm_i32x4_mul(vksum3, vinput_zero_point);
67 v128_t vacc1x0 = wasm_i32x4_mul(vksum0, vinput_zero_point);
68 v128_t vacc1x1 = wasm_i32x4_mul(vksum1, vinput_zero_point);
69 v128_t vacc1x2 = wasm_i32x4_mul(vksum2, vinput_zero_point);
70 v128_t vacc1x3 = wasm_i32x4_mul(vksum3, vinput_zero_point);
71 w = (const int32_t*) w + 4;
72
73 size_t p = ks;
74 do {
75 const int8_t* restrict a0 = a[0];
76 if XNN_UNPREDICTABLE(a0 != zero) {
77 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
78 } else {
79 a0 = zero_data;
80 }
81 const int8_t* restrict a1 = a[1];
82 if XNN_UNPREDICTABLE(a1 != zero) {
83 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
84 } else {
85 a1 = zero_data;
86 }
87 a += 2;
88
89 size_t k = kc;
90 do {
91 const v128_t va0 = wasm_v128_load(a0);
92 a0 += 16;
93 const v128_t va1 = wasm_v128_load(a1);
94 a1 += 16;
95
96 const v128_t vb0 = wasm_v128_load(w);
97 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
98 vacc1x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va1, vacc1x0);
99 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
100 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
101 vacc1x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va1, vacc1x1);
102 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
103 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
104 vacc1x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va1, vacc1x2);
105 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
106 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
107 vacc1x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va1, vacc1x3);
108
109 w = (const int8_t*) w + 64;
110 k -= 16 * sizeof(int8_t);
111 } while (k != 0);
112 p -= 2 * sizeof(void*);
113 } while (p != 0);
114
115 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));
116 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));
117 const v128_t vacc1x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x0, vacc1x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x0, vacc1x2, 2, 6, 3, 7));
118 const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x1, vacc1x3, 2, 6, 3, 7));
119
120 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
121 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
122
123 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
124 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
125
126 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
127
128 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale);
129 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale);
130
131 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
132 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
133 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
134 w = (const float*) w + 4;
135
136 const v128_t vbias0123 = wasm_v128_load(w);
137 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
138 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
139 w = (const float*) w + 4;
140
141 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
142 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
143
144 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
145 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
146
147 if XNN_LIKELY(nc >= 4) {
148 wasm_v128_store(c1, vacc1x0123);
149 wasm_v128_store(c0, vacc0x0123);
150
151 c0 = (float*) ((uintptr_t) c0 + cn_stride);
152 c1 = (float*) ((uintptr_t) c1 + cn_stride);
153
154 a = (const int8_t**restrict) ((uintptr_t) a - ks);
155
156 nc -= 4;
157 } else {
158 if (nc & 2) {
159 wasm_v128_store64_lane(c1, vacc1x0123, 0);
160 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
161 c1 += 2;
162 wasm_v128_store64_lane(c0, vacc0x0123, 0);
163 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
164 c0 += 2;
165 }
166 if (nc & 1) {
167 wasm_v128_store32_lane(c1, vacc1x0123, 0);
168 wasm_v128_store32_lane(c0, vacc0x0123, 0);
169 }
170 nc = 0;
171 }
172 } while (nc != 0);
173 }