xref: /aosp_15_r20/external/XNNPACK/src/qu8-vadd/gen/minmax-wasmsimd-x8.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker //   Template: src/qs8-vadd/wasmsimd.c.in
3*4bdc9457SAndroid Build Coastguard Worker //   Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2020 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker 
10*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker 
12*4bdc9457SAndroid Build Coastguard Worker #include <wasm_simd128.h>
13*4bdc9457SAndroid Build Coastguard Worker 
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/vadd.h>
15*4bdc9457SAndroid Build Coastguard Worker 
16*4bdc9457SAndroid Build Coastguard Worker 
xnn_qu8_vadd_minmax_ukernel__wasmsimd_x8(size_t n,const uint8_t * input_a,const uint8_t * input_b,uint8_t * output,const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17*4bdc9457SAndroid Build Coastguard Worker void xnn_qu8_vadd_minmax_ukernel__wasmsimd_x8(
18*4bdc9457SAndroid Build Coastguard Worker     size_t n,
19*4bdc9457SAndroid Build Coastguard Worker     const uint8_t* input_a,
20*4bdc9457SAndroid Build Coastguard Worker     const uint8_t* input_b,
21*4bdc9457SAndroid Build Coastguard Worker     uint8_t* output,
22*4bdc9457SAndroid Build Coastguard Worker     const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23*4bdc9457SAndroid Build Coastguard Worker {
24*4bdc9457SAndroid Build Coastguard Worker   const v128_t vbias = wasm_v128_load64_splat(params->wasmsimd.bias);
25*4bdc9457SAndroid Build Coastguard Worker   const v128_t va_multiplier = wasm_v128_load64_splat(params->wasmsimd.a_multiplier);
26*4bdc9457SAndroid Build Coastguard Worker   const v128_t vb_multiplier = wasm_v128_load64_splat(params->wasmsimd.b_multiplier);
27*4bdc9457SAndroid Build Coastguard Worker   const uint32_t vshift = params->wasmsimd.shift;
28*4bdc9457SAndroid Build Coastguard Worker   const v128_t voutput_zero_point = wasm_v128_load64_splat(params->wasmsimd.output_zero_point);
29*4bdc9457SAndroid Build Coastguard Worker   const v128_t voutput_min = wasm_v128_load64_splat(params->wasmsimd.output_min);
30*4bdc9457SAndroid Build Coastguard Worker   const v128_t voutput_max = wasm_v128_load64_splat(params->wasmsimd.output_max);
31*4bdc9457SAndroid Build Coastguard Worker 
32*4bdc9457SAndroid Build Coastguard Worker   for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
33*4bdc9457SAndroid Build Coastguard Worker     const v128_t va01234567 = wasm_u16x8_load8x8(input_a);
34*4bdc9457SAndroid Build Coastguard Worker     const v128_t vb01234567 = wasm_u16x8_load8x8(input_b);
35*4bdc9457SAndroid Build Coastguard Worker     input_a += 8;
36*4bdc9457SAndroid Build Coastguard Worker     input_b += 8;
37*4bdc9457SAndroid Build Coastguard Worker 
38*4bdc9457SAndroid Build Coastguard Worker     v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(va01234567), va_multiplier));
39*4bdc9457SAndroid Build Coastguard Worker     v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(va01234567), va_multiplier));
40*4bdc9457SAndroid Build Coastguard Worker 
41*4bdc9457SAndroid Build Coastguard Worker     vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vb01234567), vb_multiplier));
42*4bdc9457SAndroid Build Coastguard Worker     vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vb01234567), vb_multiplier));
43*4bdc9457SAndroid Build Coastguard Worker 
44*4bdc9457SAndroid Build Coastguard Worker     vacc0123 = wasm_i32x4_shr(vacc0123, vshift);
45*4bdc9457SAndroid Build Coastguard Worker     vacc4567 = wasm_i32x4_shr(vacc4567, vshift);
46*4bdc9457SAndroid Build Coastguard Worker 
47*4bdc9457SAndroid Build Coastguard Worker     v128_t vout01234567 = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc0123, vacc4567), voutput_zero_point);
48*4bdc9457SAndroid Build Coastguard Worker 
49*4bdc9457SAndroid Build Coastguard Worker     v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567);
50*4bdc9457SAndroid Build Coastguard Worker 
51*4bdc9457SAndroid Build Coastguard Worker     vout0123456701234567 = wasm_u8x16_max(vout0123456701234567, voutput_min);
52*4bdc9457SAndroid Build Coastguard Worker 
53*4bdc9457SAndroid Build Coastguard Worker     vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max);
54*4bdc9457SAndroid Build Coastguard Worker 
55*4bdc9457SAndroid Build Coastguard Worker     *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
56*4bdc9457SAndroid Build Coastguard Worker     output += 8;
57*4bdc9457SAndroid Build Coastguard Worker   }
58*4bdc9457SAndroid Build Coastguard Worker   if XNN_UNLIKELY(n != 0) {
59*4bdc9457SAndroid Build Coastguard Worker     {
60*4bdc9457SAndroid Build Coastguard Worker       const v128_t va01234567 = wasm_u16x8_load8x8(input_a);
61*4bdc9457SAndroid Build Coastguard Worker       const v128_t vb01234567 = wasm_u16x8_load8x8(input_b);
62*4bdc9457SAndroid Build Coastguard Worker 
63*4bdc9457SAndroid Build Coastguard Worker       v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(va01234567), va_multiplier));
64*4bdc9457SAndroid Build Coastguard Worker       v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(va01234567), va_multiplier));
65*4bdc9457SAndroid Build Coastguard Worker 
66*4bdc9457SAndroid Build Coastguard Worker       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vb01234567), vb_multiplier));
67*4bdc9457SAndroid Build Coastguard Worker       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vb01234567), vb_multiplier));
68*4bdc9457SAndroid Build Coastguard Worker 
69*4bdc9457SAndroid Build Coastguard Worker       vacc0123 = wasm_i32x4_shr(vacc0123, vshift);
70*4bdc9457SAndroid Build Coastguard Worker       vacc4567 = wasm_i32x4_shr(vacc4567, vshift);
71*4bdc9457SAndroid Build Coastguard Worker 
72*4bdc9457SAndroid Build Coastguard Worker       v128_t vout01234567 = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc0123, vacc4567), voutput_zero_point);
73*4bdc9457SAndroid Build Coastguard Worker 
74*4bdc9457SAndroid Build Coastguard Worker       v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567);
75*4bdc9457SAndroid Build Coastguard Worker       vout0123456701234567 = wasm_u8x16_max(vout0123456701234567, voutput_min);
76*4bdc9457SAndroid Build Coastguard Worker       vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max);
77*4bdc9457SAndroid Build Coastguard Worker 
78*4bdc9457SAndroid Build Coastguard Worker       if (n & (4 * sizeof(uint8_t))) {
79*4bdc9457SAndroid Build Coastguard Worker         *((float*) output) = (float) wasm_f32x4_extract_lane(vout0123456701234567, 0);
80*4bdc9457SAndroid Build Coastguard Worker         vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
81*4bdc9457SAndroid Build Coastguard Worker         output += 4;
82*4bdc9457SAndroid Build Coastguard Worker       }
83*4bdc9457SAndroid Build Coastguard Worker       uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
84*4bdc9457SAndroid Build Coastguard Worker       if (n & (2 * sizeof(uint8_t))) {
85*4bdc9457SAndroid Build Coastguard Worker         *((uint16_t*) output) = (uint16_t) vout0123;
86*4bdc9457SAndroid Build Coastguard Worker         vout0123 >>= 16;
87*4bdc9457SAndroid Build Coastguard Worker         output += 2;
88*4bdc9457SAndroid Build Coastguard Worker       }
89*4bdc9457SAndroid Build Coastguard Worker       if (n & (1 * sizeof(uint8_t))) {
90*4bdc9457SAndroid Build Coastguard Worker         *output = (uint8_t) vout0123;
91*4bdc9457SAndroid Build Coastguard Worker       }
92*4bdc9457SAndroid Build Coastguard Worker     }
93*4bdc9457SAndroid Build Coastguard Worker   }
94*4bdc9457SAndroid Build Coastguard Worker }
95