Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-2x4c16-minmax-fp32-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_qs8_qc8w_igemm_minmax_fp32_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 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 16 * sizeof(int8_t));
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
47 c1 = c0;
48 }
49
50 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
51 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
52 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
53 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
54 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
55 XNN_FORCE_REALIZATION(vmagic_bias);
56 XNN_FORCE_REALIZATION(vmagic_min);
57 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
58 XNN_FORCE_REALIZATION(voutput_max);
59
60 do {
61 v128_t vacc0x0 = wasm_v128_load32_zero(w);
62 v128_t vacc0x1 = wasm_v128_load32_zero((const int32_t*) w + 1);
63 v128_t vacc0x2 = wasm_v128_load32_zero((const int32_t*) w + 2);
64 v128_t vacc0x3 = wasm_v128_load32_zero((const int32_t*) w + 3);
65 v128_t vacc1x0 = vacc0x0;
66 v128_t vacc1x1 = vacc0x1;
67 v128_t vacc1x2 = vacc0x2;
68 v128_t vacc1x3 = vacc0x3;
69 w = (const int32_t*) w + 4;
70
71 size_t p = ks;
72 do {
73 const int8_t* restrict a0 = a[0];
74 if XNN_UNPREDICTABLE(a0 != zero) {
75 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
76 }
77 const int8_t* restrict a1 = a[1];
78 if XNN_UNPREDICTABLE(a1 != zero) {
79 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
80 }
81 a += 2;
82
83 size_t k = kc;
84 do {
85 const v128_t va0 = wasm_v128_load(a0);
86 a0 += 16;
87 const v128_t va1 = wasm_v128_load(a1);
88 a1 += 16;
89
90 const v128_t vb0 = wasm_v128_load(w);
91 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
92 vacc1x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va1, vacc1x0);
93 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
94 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
95 vacc1x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va1, vacc1x1);
96 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
97 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
98 vacc1x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va1, vacc1x2);
99 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
100 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
101 vacc1x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va1, vacc1x3);
102
103 w = (const int8_t*) w + 64;
104 k -= 16 * sizeof(int8_t);
105 } while (k != 0);
106 p -= 2 * sizeof(void*);
107 } while (p != 0);
108
109 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));
110 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));
111 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));
112 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));
113
114 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
115 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
116
117 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
118 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
119
120 const v128_t vscale0123 = wasm_v128_load(w);
121 w = (const float*) w + 4;
122 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
123 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
124
125 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
126 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
127
128 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
129 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
130
131 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
132 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
133
134 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
135
136 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc01x0123);
137
138 vout = wasm_i8x16_min(vout, voutput_max);
139
140 if (nc >= 4) {
141 wasm_v128_store32_lane(c1, vout, 1);
142 wasm_v128_store32_lane(c0, vout, 0);
143
144 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
145 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
146
147 a = (const int8_t**restrict) ((uintptr_t) a - ks);
148
149 nc -= 4;
150 } else {
151 if (nc & 2) {
152 wasm_v128_store16_lane(c1, vout, 2);
153 c1 += 2;
154 wasm_v128_store16_lane(c0, vout, 0);
155 c0 += 2;
156
157 vout = wasm_u32x4_shr(vout, 16);
158 }
159 if (nc & 1) {
160 wasm_v128_store8_lane(c1, vout, 4);
161 wasm_v128_store8_lane(c0, vout, 0);
162 }
163
164 nc = 0;
165 }
166 } while (nc != 0);
167 }