xref: /aosp_15_r20/external/XNNPACK/src/qu8-gemm/gen/1x4c8-minmax-fp32-wasmsimd-dot16x2-ld64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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 
xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld64(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld64(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const uint8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     uint8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(mr != 0);
31   assert(mr <= 1);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(uint8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   kc = round_up_po2(kc, 8);
40   const uint8_t* a0 = a;
41   uint8_t* c0 = c;
42 
43   const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
44   do {
45     v128_t vacc0x0 = wasm_v128_load32_zero(w);
46     v128_t vacc0x1 = wasm_v128_load32_zero((const int32_t*) w + 1);
47     v128_t vacc0x2 = wasm_v128_load32_zero((const int32_t*) w + 2);
48     v128_t vacc0x3 = wasm_v128_load32_zero((const int32_t*) w + 3);
49     w = (const void*) ((const int32_t*) w + 4);
50 
51     size_t k = 0;
52     while (k < kc) {
53       const v128_t vxa0 = wasm_u16x8_load8x8(a0);
54       a0 += 8;
55 
56       const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
57 
58       vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
59       const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8((const int8_t*) w + 8), vb_zero_point);
60 
61       vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
62       const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8((const int8_t*) w + 16), vb_zero_point);
63 
64       vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
65       const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_load8x8((const int8_t*) w + 24), vb_zero_point);
66 
67       vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
68 
69       w = (const void*) ((const uint8_t*) w + 32);
70       k += 8 * sizeof(uint8_t);
71     }
72 
73     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));
74     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));
75 
76     v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
77 
78     vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
79 
80     const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
81     vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
82 
83     const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
84     vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
85 
86     const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
87     vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
88 
89     const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
90     vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
91 
92     v128_t vacc00x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x0123);
93 
94     v128_t vout = wasm_u8x16_narrow_i16x8(vacc00x0123, vacc00x0123);
95 
96     const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
97     vout = wasm_u8x16_min(vout, voutput_max);
98 
99     if (nc >= 4) {
100       *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
101 
102       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
103 
104       a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
105 
106       nc -= 4;
107     } else {
108       uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
109       if (nc & 2) {
110         *((uint16_t*) c0) = (uint16_t) vout0;
111         vout0 >>= 16;
112         c0 += 2;
113       }
114       if (nc & 1) {
115         *c0 = (uint8_t) vout0;
116       }
117 
118       nc = 0;
119     }
120   } while (nc != 0);
121 }
122