xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/2x4c2-minmax-fp32-wasmsimd-dot16x2-ld128.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_qs8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld128(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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld128(
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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31   assert(mr != 0);
32   assert(mr <= 2);
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   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr != 2) {
46     a1 = a0;
47     c1 = c0;
48   }
49 
50   do {
51     v128_t vacc0x0123 = wasm_v128_load(w);
52     v128_t vacc1x0123 = vacc0x0123;
53     w = (const void*) ((const int32_t*) w + 4);
54 
55     size_t k = kc;
56     while (k >= 8 * sizeof(int8_t)) {
57       const v128_t vxa0 = wasm_i16x8_load8x8((const v128_t*) a0);
58       a0 += 8;
59       const v128_t vxa1 = wasm_i16x8_load8x8((const v128_t*) a1);
60       a1 += 8;
61 
62       const v128_t vb01 = wasm_v128_load(w);
63       const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
64       const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
65 
66       vacc0x0123 = wasm_i32x4_add(vacc0x0123,
67         wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
68       vacc1x0123 = wasm_i32x4_add(vacc1x0123,
69         wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
70 
71       vacc0x0123 = wasm_i32x4_add(vacc0x0123,
72         wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
73       vacc1x0123 = wasm_i32x4_add(vacc1x0123,
74         wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
75       const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
76       const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
77       const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
78 
79       vacc0x0123 = wasm_i32x4_add(vacc0x0123,
80         wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
81       vacc1x0123 = wasm_i32x4_add(vacc1x0123,
82         wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
83 
84       vacc0x0123 = wasm_i32x4_add(vacc0x0123,
85         wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
86       vacc1x0123 = wasm_i32x4_add(vacc1x0123,
87         wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
88 
89       w = (const void*) ((const int8_t*) w + 32);
90       k -= 8 * sizeof(int8_t);
91     }
92     if (k != 0) {
93       const v128_t vxa0 = wasm_i16x8_load8x8(a0);
94       a0 = (const int8_t*) ((uintptr_t) a0 + k);
95       const v128_t vxa1 = wasm_i16x8_load8x8(a1);
96       a1 = (const int8_t*) ((uintptr_t) a1 + k);
97 
98       const v128_t vxb0 = wasm_i16x8_load8x8(w);
99       w = (const void*) ((const int8_t*) w + 8);
100 
101       vacc0x0123 = wasm_i32x4_add(vacc0x0123,
102         wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
103       vacc1x0123 = wasm_i32x4_add(vacc1x0123,
104         wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
105 
106       if (k > 2 * sizeof(int8_t)) {
107         const v128_t vxb1 = wasm_i16x8_load8x8(w);
108         w = (const void*) ((const int8_t*) w + 8);
109 
110         vacc0x0123 = wasm_i32x4_add(vacc0x0123,
111           wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
112         vacc1x0123 = wasm_i32x4_add(vacc1x0123,
113           wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
114 
115         if (k > 4 * sizeof(int8_t)) {
116           const v128_t vxb2 = wasm_i16x8_load8x8(w);
117           w = (const void*) ((const int8_t*) w + 8);
118 
119           vacc0x0123 = wasm_i32x4_add(vacc0x0123,
120             wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
121           vacc1x0123 = wasm_i32x4_add(vacc1x0123,
122             wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
123         }
124       }
125     }
126 
127     vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
128     vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
129 
130     const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
131     vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
132     vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
133 
134     const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
135     vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
136     vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
137 
138     const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
139     vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
140     vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
141 
142     const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
143     vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
144     vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
145 
146     v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
147 
148     v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc01x0123);
149 
150     const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
151     vout = wasm_i8x16_min(vout, voutput_max);
152 
153     if (nc >= 4) {
154       *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
155       *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
156 
157       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
158       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
159 
160       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
161       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
162 
163       nc -= 4;
164     } else {
165       uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
166       uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
167       if (nc & 2) {
168         *((uint16_t*) c0) = (uint16_t) vout0;
169         vout0 >>= 16;
170         c0 += 2;
171         *((uint16_t*) c1) = (uint16_t) vout1;
172         vout1 >>= 16;
173         c1 += 2;
174       }
175       if (nc & 1) {
176         *c0 = (int8_t) vout0;
177         *c1 = (int8_t) vout1;
178       }
179 
180       nc = 0;
181     }
182   } while (nc != 0);
183 }
184