Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-1x8c8-minmax-wasmsdot-u2.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c8-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_1x8c8__wasmsdot_u2(
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 <= 1);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (1 * 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, 8 * sizeof(int8_t));
46 float* c0 = c;
47
48 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
49 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
50 XNN_FORCE_REALIZATION(vmin);
51 XNN_FORCE_REALIZATION(vmax);
52
53 do {
54 const v128_t vinput_zero_point = wasm_v128_load32_splat(&quantization_params->zero_point);
55 v128_t vksum0123 = wasm_v128_load(w); w = (const int32_t*) w + 4;
56 v128_t vsum0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
57 v128_t vacc0x01 = wasm_u64x2_extend_low_u32x4(vsum0x0123);
58 v128_t vacc0x23 = wasm_u64x2_extend_high_u32x4(vsum0x0123);
59 v128_t vksum4567 = wasm_v128_load(w); w = (const int32_t*) w + 4;
60 v128_t vsum0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point);
61 v128_t vacc0x45 = wasm_u64x2_extend_low_u32x4(vsum0x4567);
62 v128_t vacc0x67 = wasm_u64x2_extend_high_u32x4(vsum0x4567);
63
64 size_t p = ks;
65 do {
66 const int8_t* restrict a0 = a[0];
67 if XNN_UNPREDICTABLE(a0 != zero) {
68 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
69 } else {
70 a0 = zero_data;
71 }
72 a += 1;
73
74 size_t k = kc;
75 while (k >= 16 * sizeof(int8_t)) {
76 const v128_t va0x01 = wasm_v128_load64_splat(a0);
77 const v128_t va0x23 = wasm_v128_load64_splat((const int8_t*) a0 + 8);
78 a0 += 16;
79
80 const v128_t vb01x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
81 const v128_t vb23x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
82 const v128_t vb45x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
83 const v128_t vb67x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
84 const v128_t vb01x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
85 const v128_t vb23x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
86 const v128_t vb45x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
87 const v128_t vb67x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
88
89 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va0x01, vacc0x01);
90 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va0x01, vacc0x23);
91 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va0x01, vacc0x45);
92 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va0x01, vacc0x67);
93 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va0x23, vacc0x01);
94 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va0x23, vacc0x23);
95 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va0x23, vacc0x45);
96 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va0x23, vacc0x67);
97
98 k -= 16 * sizeof(int8_t);
99 }
100
101 if (k != 0) {
102 const v128_t va0x01 = wasm_v128_load64_splat(a0);
103 a0 += 8;
104
105 const v128_t vb01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
106 const v128_t vb23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
107 const v128_t vb45 = wasm_v128_load(w); w = (const int8_t*) w + 16;
108 const v128_t vb67 = wasm_v128_load(w); w = (const int8_t*) w + 16;
109
110 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va0x01, vacc0x01);
111 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va0x01, vacc0x23);
112 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va0x01, vacc0x45);
113 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va0x01, vacc0x67);
114
115 k -= 8 * sizeof(int8_t);
116 }
117 p -= 1 * sizeof(void*);
118 } while (p != 0);
119
120 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x01, vacc0x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x01, vacc0x23, 1, 3, 5, 7));
121 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x45, vacc0x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x45, vacc0x67, 1, 3, 5, 7));
122
123 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
124 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
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 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale);
130
131 const v128_t vfilter_output_scale0123 = wasm_v128_load(w); w = (const float*) w + 4;
132 const v128_t vfilter_output_scale4567 = wasm_v128_load(w); w = (const float*) w + 4;
133
134 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
135 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567);
136
137 const v128_t vbias0123 = wasm_v128_load(w); w = (const float*) w + 4;
138 const v128_t vbias4567 = wasm_v128_load(w); w = (const float*) w + 4;
139 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
140 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567);
141
142 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
143 vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin);
144
145 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
146 vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax);
147
148 if XNN_LIKELY(nc >= 8) {
149 wasm_v128_store(c0, vacc0x0123);
150 wasm_v128_store(c0 + 4, vacc0x4567);
151
152 c0 = (float*) ((uintptr_t) c0 + cn_stride);
153
154 a = (const int8_t**restrict) ((uintptr_t) a - ks);
155
156 nc -= 8;
157 } else {
158 if (nc & 4) {
159 wasm_v128_store(c0, vacc0x0123);
160 vacc0x0123 = vacc0x4567;
161 c0 += 4;
162 }
163 if (nc & 2) {
164 wasm_v128_store64_lane(c0, vacc0x0123, 0);
165 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
166 c0 += 2;
167 }
168 if (nc & 1) {
169 wasm_v128_store32_lane(c0, vacc0x0123, 0);
170 }
171 nc = 0;
172 }
173 } while (nc != 0);
174 }