Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-2x4c16-minmax-wasmusdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_2x4c16__wasmusdot(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 float* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
28 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 2);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 16 * sizeof(int8_t));
40 const int8_t* a0 = a;
41 float* c0 = c;
42 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr != 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48
49 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
50 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
51 XNN_FORCE_REALIZATION(vmin);
52 XNN_FORCE_REALIZATION(vmax);
53
54 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
55 do {
56 v128_t vksum0 = wasm_v128_load32_zero(w);
57 v128_t vksum1 = wasm_v128_load32_zero((const int32_t*) w + 1);
58 v128_t vksum2 = wasm_v128_load32_zero((const int32_t*) w + 2);
59 v128_t vksum3 = wasm_v128_load32_zero((const int32_t*) w + 3);
60 const v128_t vinput_zero_point0 = wasm_i32x4_splat((int32_t) quantization_params[0].zero_point + 128);
61 const v128_t vinput_zero_point1 = wasm_i32x4_splat((int32_t) quantization_params[1].zero_point + 128);
62 v128_t vacc0x0 = wasm_i32x4_mul(vksum0, vinput_zero_point0);
63 v128_t vacc0x1 = wasm_i32x4_mul(vksum1, vinput_zero_point0);
64 v128_t vacc0x2 = wasm_i32x4_mul(vksum2, vinput_zero_point0);
65 v128_t vacc0x3 = wasm_i32x4_mul(vksum3, vinput_zero_point0);
66 v128_t vacc1x0 = wasm_i32x4_mul(vksum0, vinput_zero_point1);
67 v128_t vacc1x1 = wasm_i32x4_mul(vksum1, vinput_zero_point1);
68 v128_t vacc1x2 = wasm_i32x4_mul(vksum2, vinput_zero_point1);
69 v128_t vacc1x3 = wasm_i32x4_mul(vksum3, vinput_zero_point1);
70 w = (const int32_t*) w + 4;
71
72 size_t k = kc;
73 do {
74 const v128_t va0 = wasm_v128_xor(wasm_v128_load(a0), vsign_mask);
75 a0 += 16;
76 const v128_t va1 = wasm_v128_xor(wasm_v128_load(a1), vsign_mask);
77 a1 += 16;
78
79 const v128_t vb0 = wasm_v128_load(w);
80
81 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
82 vacc1x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va1, vacc1x0);
83 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
84
85 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
86 vacc1x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va1, vacc1x1);
87 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
88
89 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
90 vacc1x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va1, vacc1x2);
91 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
92
93 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
94 vacc1x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va1, vacc1x3);
95
96 w = (const int8_t*) w + 64;
97 k -= 16 * sizeof(int8_t);
98 } while (k != 0);
99
100 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));
101 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));
102 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));
103 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));
104
105 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
106 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
107
108 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
109 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
110
111 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
112 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
113
114 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
115 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
116
117 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
118 w = (const float*) w + 4;
119 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
120 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
121
122 const v128_t vbias0123 = wasm_v128_load(w);
123 w = (const float*) w + 4;
124 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
125 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
126
127 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
128 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
129
130 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
131 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
132
133 if XNN_LIKELY(nc >= 4) {
134 wasm_v128_store(c0, vacc0x0123);
135 wasm_v128_store(c1, vacc1x0123);
136
137 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
138 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
139
140 c0 = (float*) ((uintptr_t) c0 + cn_stride);
141 c1 = (float*) ((uintptr_t) c1 + cn_stride);
142
143 nc -= 4;
144 } else {
145 if (nc & 2) {
146 wasm_v128_store64_lane(c0, vacc0x0123, 0);
147 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
148 c0 += 2;
149 wasm_v128_store64_lane(c1, vacc1x0123, 0);
150 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
151 c1 += 2;
152 }
153 if (nc & 1) {
154 wasm_v128_store32_lane(c0, vacc0x0123, 0);
155 wasm_v128_store32_lane(c1, vacc1x0123, 0);
156 }
157 nc = 0;
158 }
159 } while (nc != 0);
160 }