xref: /aosp_15_r20/external/XNNPACK/src/qc8-gemm/gen/1x4c2-minmax-fp32-wasmsimd-dot16x2-ld64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/MRx4c2-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_qc8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld64(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld64(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     const int8_t* restrict a,
24     size_t a_stride,
25     const void* restrict w,
26     int8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     const union xnn_qc8_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(int8_t) == 0);
36   assert(a != NULL);
37   assert(w != NULL);
38   assert(c != NULL);
39 
40   kc = round_up_po2(kc, 2);
41   const int8_t* a0 = a;
42   int8_t* c0 = c;
43 
44   do {
45     v128_t vacc0x0123 = wasm_v128_load(w);
46     w = (const void*) ((const int32_t*) w + 4);
47 
48     size_t k = kc;
49     while (k >= 8 * sizeof(int8_t)) {
50       const v128_t vxa0 = wasm_i16x8_load8x8((const v128_t*) a0);
51       a0 += 8;
52 
53       const v128_t vxb0 = wasm_i16x8_load8x8(w);
54 
55       vacc0x0123 = wasm_i32x4_add(vacc0x0123,
56         wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
57       const v128_t vxb1 = wasm_i16x8_load8x8((const int8_t*) w + 8);
58 
59       vacc0x0123 = wasm_i32x4_add(vacc0x0123,
60         wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
61       const v128_t vxb2 = wasm_i16x8_load8x8((const int8_t*) w + 16);
62 
63       vacc0x0123 = wasm_i32x4_add(vacc0x0123,
64         wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
65       const v128_t vxb3 = wasm_i16x8_load8x8((const int8_t*) w + 24);
66 
67       vacc0x0123 = wasm_i32x4_add(vacc0x0123,
68         wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
69 
70       w = (const void*) ((const int8_t*) w + 32);
71       k -= 8 * sizeof(int8_t);
72     }
73     if (k != 0) {
74       const v128_t vxa0 = wasm_i16x8_load8x8(a0);
75       a0 = (const int8_t*) ((uintptr_t) a0 + k);
76 
77       const v128_t vxb0 = wasm_i16x8_load8x8(w);
78       w = (const void*) ((const int8_t*) w + 8);
79 
80       vacc0x0123 = wasm_i32x4_add(vacc0x0123,
81         wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
82 
83       if (k > 2 * sizeof(int8_t)) {
84         const v128_t vxb1 = wasm_i16x8_load8x8(w);
85         w = (const void*) ((const int8_t*) w + 8);
86 
87         vacc0x0123 = wasm_i32x4_add(vacc0x0123,
88           wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
89 
90         if (k > 4 * sizeof(int8_t)) {
91           const v128_t vxb2 = wasm_i16x8_load8x8(w);
92           w = (const void*) ((const int8_t*) w + 8);
93 
94           vacc0x0123 = wasm_i32x4_add(vacc0x0123,
95             wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
96         }
97       }
98     }
99 
100     vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
101 
102     const v128_t vscale0123 = wasm_v128_load(w);
103     w = (const void*) ((const float*) w + 4);
104     vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
105 
106     const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
107     vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
108 
109     const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
110     vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
111 
112     const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
113     vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
114 
115     v128_t vacc00x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x0123);
116 
117     v128_t vout = wasm_i8x16_narrow_i16x8(vacc00x0123, vacc00x0123);
118 
119     const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
120     vout = wasm_i8x16_min(vout, voutput_max);
121 
122     if (nc >= 4) {
123       *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
124 
125       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
126 
127       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
128 
129       nc -= 4;
130     } else {
131       uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
132       if (nc & 2) {
133         *((uint16_t*) c0) = (uint16_t) vout0;
134         vout0 >>= 16;
135         c0 += 2;
136       }
137       if (nc & 1) {
138         *c0 = (int8_t) vout0;
139       }
140 
141       nc = 0;
142     }
143   } while (nc != 0);
144 }
145