Debian Codegraph @

src/qs8-vmulc/gen/qs8-vmulc-minmax-fp32-wasmsimd-mul32-ld64-u16.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_u16(
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 >= 16 * sizeof(int8_t); batch -= 16 * sizeof(int8_t)) {
49 const v128_t va01234567 = wasm_i16x8_load8x8(input_a);
50 const v128_t va89ABCDEF = wasm_i16x8_load8x8(input_a + 8);
51 input_a += 16;
52
53 const v128_t vxa01234567 = wasm_i16x8_sub(va01234567, va_zero_point);
54 const v128_t vxa89ABCDEF = wasm_i16x8_sub(va89ABCDEF, va_zero_point);
55
56 v128_t vacc0123 = wasm_i32x4_extmul_low_i16x8(vxa01234567, vxb);
57 v128_t vacc4567 = wasm_i32x4_extmul_high_i16x8(vxa01234567, vxb);
58 v128_t vacc89AB = wasm_i32x4_extmul_low_i16x8(vxa89ABCDEF, vxb);
59 v128_t vaccCDEF = wasm_i32x4_extmul_high_i16x8(vxa89ABCDEF, vxb);
60
61 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
62 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
63 vacc89AB = wasm_f32x4_convert_i32x4(vacc89AB);
64 vaccCDEF = wasm_f32x4_convert_i32x4(vaccCDEF);
65
66 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
67 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
68 vacc89AB = wasm_f32x4_mul(vacc89AB, vscale);
69 vaccCDEF = wasm_f32x4_mul(vaccCDEF, vscale);
70
71 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
72 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
73 vacc89AB = wasm_f32x4_add(vacc89AB, vmagic_bias);
74 vaccCDEF = wasm_f32x4_add(vaccCDEF, vmagic_bias);
75
76 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
77 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
78 vacc89AB = wasm_i32x4_max(vacc89AB, vmagic_min);
79 vaccCDEF = wasm_i32x4_max(vaccCDEF, vmagic_min);
80
81 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
82 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
83 vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_output_zero_point);
84 vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_output_zero_point);
85
86 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
87 v128_t vout89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
88
89 v128_t vout0123456789ABCDEF = wasm_i8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
90
91 vout0123456789ABCDEF = wasm_i8x16_min(vout0123456789ABCDEF, voutput_max);
92
93 wasm_v128_store(output, vout0123456789ABCDEF);
94 output += 16;
95 }
96 if XNN_UNLIKELY(batch != 0) {
97 do {
98 const v128_t va01234567 = wasm_i16x8_load8x8(input_a);
99 input_a += 8;
100
101 const v128_t vxa01234567 = wasm_i16x8_sub(va01234567, va_zero_point);
102
103 v128_t vacc0123 = wasm_i32x4_extmul_low_i16x8(vxa01234567, vxb);
104 v128_t vacc4567 = wasm_i32x4_extmul_high_i16x8(vxa01234567, vxb);
105
106 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
107 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
108
109 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
110 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
111
112 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
113 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
114
115 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
116 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
117
118 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
119 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
120
121 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
122 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
123 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
124
125 if XNN_LIKELY(batch >= (8 * sizeof(int8_t))) {
126 wasm_v128_store64_lane(output, vout0123456701234567, 0);
127 output += 8;
128 batch -= 8 * sizeof(int8_t);
129 } else {
130 if (batch & (4 * sizeof(int8_t))) {
131 wasm_v128_store32_lane(output, vout0123456701234567, 0);
132 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
133 output += 4;
134 }
135 if (batch & (2 * sizeof(int8_t))) {
136 wasm_v128_store16_lane(output, vout0123456701234567, 0);
137 vout0123456701234567 = wasm_u32x4_shr(vout0123456701234567, 16);
138 output += 2;
139 }
140 if (batch & (1 * sizeof(int8_t))) {
141 wasm_v128_store8_lane(output, vout0123456701234567, 0);
142 }
143 batch = 0;
144 }
145 } while (batch != 0);
146 }
147 }