Debian Codegraph @

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

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c8-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_qd8_f32_qc8w_gemm_minmax_ukernel_1x8c8__wasmsdot_u2(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
29 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 1);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 8 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 float* c0 = c;
43
44 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
45 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
46 XNN_FORCE_REALIZATION(vmin);
47 XNN_FORCE_REALIZATION(vmax);
48
49 do {
50 const v128_t vinput_zero_point0 = wasm_v128_load32_splat(&quantization_params[0].zero_point);
51 v128_t vksum0123 = wasm_v128_load(w); w = (const int32_t*) w + 4;
52 v128_t vsum0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
53 v128_t vacc0x01 = wasm_u64x2_extend_low_u32x4(vsum0x0123);
54 v128_t vacc0x23 = wasm_u64x2_extend_high_u32x4(vsum0x0123);
55 v128_t vksum4567 = wasm_v128_load(w); w = (const int32_t*) w + 4;
56 v128_t vsum0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point0);
57 v128_t vacc0x45 = wasm_u64x2_extend_low_u32x4(vsum0x4567);
58 v128_t vacc0x67 = wasm_u64x2_extend_high_u32x4(vsum0x4567);
59
60 size_t k = kc;
61 while (k >= 16 * sizeof(int8_t)) {
62 const v128_t va0x01 = wasm_v128_load64_splat(a0);
63 const v128_t va0x23 = wasm_v128_load64_splat((const int8_t*) a0 + 8);
64 a0 += 16;
65
66 const v128_t vb01x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
67 const v128_t vb23x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
68 const v128_t vb45x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
69 const v128_t vb67x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
70 const v128_t vb01x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
71 const v128_t vb23x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
72 const v128_t vb45x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
73 const v128_t vb67x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
74
75 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va0x01, vacc0x01);
76 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va0x01, vacc0x23);
77 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va0x01, vacc0x45);
78 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va0x01, vacc0x67);
79 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va0x23, vacc0x01);
80 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va0x23, vacc0x23);
81 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va0x23, vacc0x45);
82 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va0x23, vacc0x67);
83
84 k -= 16 * sizeof(int8_t);
85 }
86
87 if (k != 0) {
88 const v128_t va0x01 = wasm_v128_load64_splat(a0);
89 a0 += 8;
90
91 const v128_t vb01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
92 const v128_t vb23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
93 const v128_t vb45 = wasm_v128_load(w); w = (const int8_t*) w + 16;
94 const v128_t vb67 = wasm_v128_load(w); w = (const int8_t*) w + 16;
95
96 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va0x01, vacc0x01);
97 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va0x01, vacc0x23);
98 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va0x01, vacc0x45);
99 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va0x01, vacc0x67);
100
101 k -= 8 * sizeof(int8_t);
102 }
103
104 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x01, vacc0x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x01, vacc0x23, 1, 3, 5, 7));
105 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x45, vacc0x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x45, vacc0x67, 1, 3, 5, 7));
106
107 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
108 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
109
110 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
111
112 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
113 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale0);
114
115 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
116 w = (const float*) w + 4;
117 const v128_t vfilter_output_scale4567 = wasm_v128_load(w);
118 w = (const float*) w + 4;
119 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
120 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567);
121
122 const v128_t vbias0123 = wasm_v128_load(w);
123 w = (const float*) w + 4;
124 const v128_t vbias4567 = wasm_v128_load(w);
125 w = (const float*) w + 4;
126 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
127 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567);
128
129 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
130 vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin);
131
132 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
133 vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax);
134
135 if XNN_LIKELY(nc >= 8) {
136 wasm_v128_store(c0, vacc0x0123);
137 wasm_v128_store(c0 + 4, vacc0x4567);
138
139 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
140
141 c0 = (float*) ((uintptr_t) c0 + cn_stride);
142
143 nc -= 8;
144 } else {
145 if (nc & 4) {
146 wasm_v128_store(c0, vacc0x0123);
147 vacc0x0123 = vacc0x4567;
148 c0 += 4;
149 }
150 if (nc & 2) {
151 wasm_v128_store64_lane(c0, vacc0x0123, 0);
152 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
153 c0 += 2;
154 }
155 if (nc & 1) {
156 wasm_v128_store32_lane(c0, vacc0x0123, 0);
157 }
158 nc = 0;
159 }
160 } while (nc != 0);
161 }