xref: /aosp_15_r20/external/XNNPACK/src/qu8-gemm/gen/3x4c2s4-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_3x4c2s4__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_3x4c2s4__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 <= 3);
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   const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
43   uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
44   if XNN_UNPREDICTABLE(mr < 2) {
45     a1 = a0;
46     c1 = c0;
47   }
48   const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
49   uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     a2 = a1;
52     c2 = c1;
53   }
54 
55   kc = round_up_po2(kc, 8 * sizeof(uint8_t));
56   do {
57     v128_t vacc0x0123 = wasm_v128_load(w);
58     v128_t vacc1x0123 = vacc0x0123;
59     v128_t vacc2x0123 = vacc0x0123;
60     w = (const void*) ((const int32_t*) w + 4);
61 
62     const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
63     size_t k = kc;
64     do {
65       v128_t vxa0 = wasm_u16x8_load8x8((const v128_t*) a0);
66       a0 += 8;
67       v128_t vxa1 = wasm_u16x8_load8x8((const v128_t*) a1);
68       a1 += 8;
69       v128_t vxa2 = wasm_u16x8_load8x8((const v128_t*) a2);
70       a2 += 8;
71 
72       const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
73 
74       vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb0));
75       vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
76       vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb0));
77       vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
78       vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb0));
79       vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
80       const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 8), vb_zero_point);
81 
82       vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb1));
83       vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
84       vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb1));
85       vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
86       vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb1));
87       vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
88       const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 16), vb_zero_point);
89 
90       vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb2));
91       vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
92       vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb2));
93       vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
94       vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb2));
95       vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
96       const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 24), vb_zero_point);
97 
98       vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb3));
99       vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb3));
100       vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb3));
101 
102       w = (const uint8_t*) w + 32;
103       k -= 8 * sizeof(uint8_t);
104     } while (k != 0);
105 
106     vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
107     vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
108     vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
109 
110     const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
111     vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
112     vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
113     vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale);
114 
115     const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
116     vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
117     vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
118     vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
119 
120     const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
121     vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
122     vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
123     vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
124 
125     const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
126     vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
127     vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
128     vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
129 
130     v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
131     v128_t vacc22x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x0123);
132 
133     v128_t vout = wasm_u8x16_narrow_i16x8(vacc01x0123, vacc22x0123);
134 
135     const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
136     vout = wasm_u8x16_min(vout, voutput_max);
137 
138     if (nc >= 4) {
139       *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
140       *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
141       *((float*) c2) = (float) wasm_f32x4_extract_lane(vout, 2);
142 
143       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
144       c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
145       c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
146 
147       a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
148       a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
149       a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
150 
151       nc -= 4;
152     } else {
153       uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
154       uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
155       uint32_t vout2 = wasm_i32x4_extract_lane(vout, 2);
156       if (nc & 2) {
157         *((uint16_t*) c0) = (uint16_t) vout0;
158         vout0 >>= 16;
159         c0 += 2;
160         *((uint16_t*) c1) = (uint16_t) vout1;
161         vout1 >>= 16;
162         c1 += 2;
163         *((uint16_t*) c2) = (uint16_t) vout2;
164         vout2 >>= 16;
165         c2 += 2;
166       }
167       if (nc & 1) {
168         *c0 = (uint8_t) vout0;
169         *c1 = (uint8_t) vout1;
170         *c2 = (uint8_t) vout2;
171       }
172 
173       nc = 0;
174     }
175   } while (nc != 0);
176 }
177