Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-1x4c8-minmax-wasmsimd-dot16x2-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c8-wasmsimd-dot16x2.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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_qc4w_gemm_minmax_ukernel_1x4c8__wasmsimd_dot16x2_ld64(
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 struct xnn_f32_qc4w_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 const v128_t vmask = wasm_i8x16_const_splat(0xF0);
50 do {
51 const v128_t vksum0123 = wasm_v128_load(w);
52 const v128_t vinput_zero_point0 = wasm_v128_load32_splat(&quantization_params[0].zero_point);
53 const v128_t vinit0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
54
55 const v128_t vzero = wasm_i32x4_const_splat(0);
56 v128_t vacc0x0 = wasm_v32x4_shuffle(vinit0x0123, vzero, 0, 5, 6, 7);
57 v128_t vacc0x1 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 1, 6, 7);
58 v128_t vacc0x2 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 2, 7);
59 v128_t vacc0x3 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 6, 3);
60 w = (const int32_t*) w + 4;
61
62 size_t k = kc;
63 while (k >= 16 * sizeof(int8_t)) {
64 v128_t vxa0;
65 vxa0 = wasm_i16x8_load8x8(a0);
66 a0 += 8;
67 const v128_t vb0 = wasm_v128_load(w);
68 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 16);
69 const v128_t vbm0 = wasm_i8x16_shl(vb0, 4);
70 const v128_t vbm2 = wasm_i8x16_shl(vb2, 4);
71 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vbm0);
72 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vbm0);
73 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vbm2);
74 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vbm2);
75 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
76 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
77 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
78 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
79 vxa0 = wasm_i16x8_load8x8(a0);
80 a0 += 8;
81 const v128_t vbm4 = wasm_v128_and(vb0, vmask);
82 const v128_t vbm6 = wasm_v128_and(vb2, vmask);
83 const v128_t vxb4 = wasm_i16x8_extend_low_i8x16(vbm4);
84 const v128_t vxb5 = wasm_i16x8_extend_high_i8x16(vbm4);
85 const v128_t vxb6 = wasm_i16x8_extend_low_i8x16(vbm6);
86 const v128_t vxb7 = wasm_i16x8_extend_high_i8x16(vbm6);
87 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb4));
88 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb5));
89 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb6));
90 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb7));
91 w = (const int8_t*) w + 32;
92 k -= 16 * sizeof(int8_t);
93 }
94
95 while (k >= 8 * sizeof(int8_t)) {
96 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
97 a0 += 8;
98
99 const v128_t vb0 = wasm_v128_load(w);
100 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 16);
101 const v128_t vbm0 = wasm_i8x16_shl(vb0, 4);
102 const v128_t vbm2 = wasm_i8x16_shl(vb2, 4);
103 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vbm0);
104 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vbm0);
105 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vbm2);
106 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vbm2);
107
108 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
109
110 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
111
112 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
113
114 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
115
116 w = (const int8_t*) w + 32; // only low nibble used
117 k -= 8 * sizeof(int8_t);
118 };
119
120 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));
121 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));
122
123 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
124
125 vacc0x0123 = wasm_i32x4_shr(vacc0x0123, 4);
126 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
127
128 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
129
130 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
131
132 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
133 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, 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 w = (const float*) w + 4;
139
140 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
141
142 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
143
144 if XNN_LIKELY(nc >= 4) {
145 wasm_v128_store(c0, vacc0x0123);
146
147 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
148
149 c0 = (float*) ((uintptr_t) c0 + cn_stride);
150
151 nc -= 4;
152 } else {
153 if (nc & 2) {
154 wasm_v128_store64_lane(c0, vacc0x0123, 0);
155 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
156 c0 += 2;
157 }
158 if (nc & 1) {
159 wasm_v128_store32_lane(c0, vacc0x0123, 0);
160 }
161 nc = 0;
162 }
163 } while (nc != 0);
164 }