Debian Codegraph @

src/qs8-vmulc/gen/qs8-vmulc-minmax-fp32-wasmsimd-mul32-ld64-u8.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vmulc/wasmsimd-mul32-ld64.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/common.h"
15 #include "xnnpack/math.h"
16 #include "xnnpack/vbinary.h"
17
18
19 void xnn_qs8_vmulc_minmax_fp32_ukernel__wasmsimd_mul32_ld64_u8(
20 size_t batch,
21 const int8_t* input_a,
22 const int8_t* input_b,
23 int8_t* output,
24 const union xnn_qs8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 assert(batch != 0);
27 assert(batch % sizeof(int8_t) == 0);
28 assert(input_a != NULL);
29 assert(input_b != NULL);
30 assert(output != NULL);
31
32 const float output_min_less_zero_point = (float) ((int32_t) params->scalar.output_min - (int32_t) params->scalar.output_zero_point);
33 const v128_t va_zero_point = wasm_i16x8_splat(params->scalar.a_zero_point);
34 const v128_t vscale = wasm_v128_load32_splat(&params->scalar.scale);
35 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
36 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
37 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->scalar.output_zero_point);
38 const v128_t voutput_max = wasm_v128_load8_splat(&params->scalar.output_max);
39 const v128_t vxb = wasm_i16x8_splat((int16_t) *input_b - params->scalar.b_zero_point);
40
41 XNN_FORCE_REALIZATION(va_zero_point);
42 XNN_FORCE_REALIZATION(vscale);
43 XNN_FORCE_REALIZATION(vmagic_bias);
44 XNN_FORCE_REALIZATION(vmagic_min);
45 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
46 XNN_FORCE_REALIZATION(voutput_max);
47
48 for (; batch >= 8 * sizeof(int8_t); batch -= 8 * sizeof(int8_t)) {
49 const v128_t va01234567 = wasm_i16x8_load8x8(input_a);
50 input_a += 8;
51
52 const v128_t vxa01234567 = wasm_i16x8_sub(va01234567, va_zero_point);
53
54 v128_t vacc0123 = wasm_i32x4_extmul_low_i16x8(vxa01234567, vxb);
55 v128_t vacc4567 = wasm_i32x4_extmul_high_i16x8(vxa01234567, vxb);
56
57 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
58 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
59
60 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
61 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
62
63 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
64 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
65
66 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
67 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
68
69 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
70 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
71
72 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
73
74 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
75
76 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
77
78 wasm_v128_store64_lane(output, vout0123456701234567, 0);
79 output += 8;
80 }
81 if XNN_UNLIKELY(batch != 0) {
82 {
83 const v128_t va01234567 = wasm_i16x8_load8x8(input_a);
84
85 const v128_t vxa01234567 = wasm_i16x8_sub(va01234567, va_zero_point);
86
87 v128_t vacc0123 = wasm_i32x4_extmul_low_i16x8(vxa01234567, vxb);
88 v128_t vacc4567 = wasm_i32x4_extmul_high_i16x8(vxa01234567, vxb);
89
90 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
91 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
92
93 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
94 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
95
96 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
97 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
98
99 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
100 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
101
102 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
103 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
104
105 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
106 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
107 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
108
109 if (batch & (4 * sizeof(int8_t))) {
110 wasm_v128_store32_lane(output, vout0123456701234567, 0);
111 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
112 output += 4;
113 }
114 if (batch & (2 * sizeof(int8_t))) {
115 wasm_v128_store16_lane(output, vout0123456701234567, 0);
116 vout0123456701234567 = wasm_u32x4_shr(vout0123456701234567, 16);
117 output += 2;
118 }
119 if (batch & (1 * sizeof(int8_t))) {
120 wasm_v128_store8_lane(output, vout0123456701234567, 0);
121 }
122 }
123 }
124 }