xref: /aosp_15_r20/external/XNNPACK/src/qu8-gemm/gen/1x4c2s4-minmax-fp32-wasmsimd-dot16x2-ld64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/MRx4c2s4-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 
18 
xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2s4__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)])19 void xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld64(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     const uint8_t* restrict a,
24     size_t a_stride,
25     const void* restrict w,
26     uint8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31   assert(mr != 0);
32   assert(mr <= 1);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(kc % sizeof(uint8_t) == 0);
36   assert(a != NULL);
37   assert(w != NULL);
38   assert(c != NULL);
39 
40   const uint8_t* a0 = a;
41   uint8_t* c0 = c;
42 
43   kc = round_up_po2(kc, 8 * sizeof(uint8_t));
44   do {
45     v128_t vacc0x0123 = wasm_v128_load(w);
46     w = (const void*) ((const int32_t*) w + 4);
47 
48     const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
49     size_t k = kc;
50     do {
51       v128_t vxa0 = wasm_u16x8_load8x8((const v128_t*) a0);
52       a0 += 8;
53 
54       const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
55 
56       vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb0));
57       vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
58       const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 8), vb_zero_point);
59 
60       vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb1));
61       vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
62       const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 16), vb_zero_point);
63 
64       vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb2));
65       vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
66       const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 24), vb_zero_point);
67 
68       vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb3));
69 
70       w = (const uint8_t*) w + 32;
71       k -= 8 * sizeof(uint8_t);
72     } while (k != 0);
73 
74     vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
75 
76     const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
77     vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
78 
79     const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
80     vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
81 
82     const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
83     vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
84 
85     const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
86     vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
87 
88     v128_t vacc00x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x0123);
89 
90     v128_t vout = wasm_u8x16_narrow_i16x8(vacc00x0123, vacc00x0123);
91 
92     const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
93     vout = wasm_u8x16_min(vout, voutput_max);
94 
95     if (nc >= 4) {
96       *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
97 
98       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
99 
100       a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
101 
102       nc -= 4;
103     } else {
104       uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
105       if (nc & 2) {
106         *((uint16_t*) c0) = (uint16_t) vout0;
107         vout0 >>= 16;
108         c0 += 2;
109       }
110       if (nc & 1) {
111         *c0 = (uint8_t) vout0;
112       }
113 
114       nc = 0;
115     }
116   } while (nc != 0);
117 }
118