xref: /aosp_15_r20/external/XNNPACK/src/qu8-gemm/gen/3x2-minmax-fp32-wasm-fmagic.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/scalar.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 <xnnpack/math.h>
13 #include <xnnpack/gemm.h>
14 #include <xnnpack/unaligned.h>
15 
16 
xnn_qu8_gemm_minmax_fp32_ukernel_3x2__wasm_fmagic(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)])17 void xnn_qu8_gemm_minmax_fp32_ukernel_3x2__wasm_fmagic(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     const uint8_t* restrict a,
22     size_t a_stride,
23     const void* restrict w,
24     uint8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29   assert(mr != 0);
30   assert(mr <= 3);
31   assert(nc != 0);
32   assert(kc != 0);
33 
34   const uint8_t* a0 = a;
35   uint8_t* c0 = c;
36   const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
37   uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
38   if XNN_UNPREDICTABLE(mr < 2) {
39     a1 = a0;
40     c1 = c0;
41   }
42   const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
43   uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
44   if XNN_UNPREDICTABLE(mr <= 2) {
45     a2 = a1;
46     c2 = c1;
47   }
48 
49   const int32_t vb_zero_point = params->fp32_scalar_fmagic.kernel_zero_point;
50   do {
51     int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
52     int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
53     int32_t vacc1x0 = vacc0x0;
54     int32_t vacc1x1 = vacc0x1;
55     int32_t vacc2x0 = vacc0x0;
56     int32_t vacc2x1 = vacc0x1;
57     w = (const void*) ((const int32_t*) w + 2);
58 
59     size_t k = kc;
60     do {
61       const int32_t va0 = (int32_t) (uint32_t) *a0++;
62       const int32_t va1 = (int32_t) (uint32_t) *a1++;
63       const int32_t va2 = (int32_t) (uint32_t) *a2++;
64 
65       const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
66       const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
67       w = (const void*) ((const uint8_t*) w + 2);
68 
69       vacc0x0 += va0 * vb0;
70       vacc0x1 += va0 * vb1;
71       vacc1x0 += va1 * vb0;
72       vacc1x1 += va1 * vb1;
73       vacc2x0 += va2 * vb0;
74       vacc2x1 += va2 * vb1;
75 
76       k -= sizeof(uint8_t);
77     } while (k != 0);
78 
79     float vfpacc0x0 = (float) vacc0x0;
80     float vfpacc0x1 = (float) vacc0x1;
81     float vfpacc1x0 = (float) vacc1x0;
82     float vfpacc1x1 = (float) vacc1x1;
83     float vfpacc2x0 = (float) vacc2x0;
84     float vfpacc2x1 = (float) vacc2x1;
85 
86     const float vscale = params->fp32_scalar_fmagic.scale;
87     vfpacc0x0 *= vscale;
88     vfpacc0x1 *= vscale;
89     vfpacc1x0 *= vscale;
90     vfpacc1x1 *= vscale;
91     vfpacc2x0 *= vscale;
92     vfpacc2x1 *= vscale;
93 
94     const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
95     vfpacc0x0 = __builtin_wasm_max_f32(vfpacc0x0, voutput_min_less_zero_point);
96     vfpacc0x1 = __builtin_wasm_max_f32(vfpacc0x1, voutput_min_less_zero_point);
97     vfpacc1x0 = __builtin_wasm_max_f32(vfpacc1x0, voutput_min_less_zero_point);
98     vfpacc1x1 = __builtin_wasm_max_f32(vfpacc1x1, voutput_min_less_zero_point);
99     vfpacc2x0 = __builtin_wasm_max_f32(vfpacc2x0, voutput_min_less_zero_point);
100     vfpacc2x1 = __builtin_wasm_max_f32(vfpacc2x1, voutput_min_less_zero_point);
101 
102     const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
103     vfpacc0x0 = __builtin_wasm_min_f32(vfpacc0x0, voutput_max_less_zero_point);
104     vfpacc0x1 = __builtin_wasm_min_f32(vfpacc0x1, voutput_max_less_zero_point);
105     vfpacc1x0 = __builtin_wasm_min_f32(vfpacc1x0, voutput_max_less_zero_point);
106     vfpacc1x1 = __builtin_wasm_min_f32(vfpacc1x1, voutput_max_less_zero_point);
107     vfpacc2x0 = __builtin_wasm_min_f32(vfpacc2x0, voutput_max_less_zero_point);
108     vfpacc2x1 = __builtin_wasm_min_f32(vfpacc2x1, voutput_max_less_zero_point);
109 
110     const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
111     vfpacc0x0 += vmagic_bias;
112     vfpacc0x1 += vmagic_bias;
113     vfpacc1x0 += vmagic_bias;
114     vfpacc1x1 += vmagic_bias;
115     vfpacc2x0 += vmagic_bias;
116     vfpacc2x1 += vmagic_bias;
117 
118     const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
119     int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
120     int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
121     int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0) - vmagic_bias_less_output_zero_point;
122     int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1) - vmagic_bias_less_output_zero_point;
123     int32_t vout2x0 = (int32_t) float_as_uint32(vfpacc2x0) - vmagic_bias_less_output_zero_point;
124     int32_t vout2x1 = (int32_t) float_as_uint32(vfpacc2x1) - vmagic_bias_less_output_zero_point;
125 
126     if XNN_LIKELY(nc >= 2) {
127       c0[0] = (uint8_t) vout0x0;
128       c0[1] = (uint8_t) vout0x1;
129       c1[0] = (uint8_t) vout1x0;
130       c1[1] = (uint8_t) vout1x1;
131       c2[0] = (uint8_t) vout2x0;
132       c2[1] = (uint8_t) vout2x1;
133 
134       a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
135       a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
136       a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
137 
138       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
139       c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
140       c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
141 
142       nc -= 2;
143     } else {
144       if (nc & 1) {
145         c0[0] = (uint8_t) vout0x0;
146         c1[0] = (uint8_t) vout1x0;
147         c2[0] = (uint8_t) vout2x0;
148       }
149 
150       nc = 0;
151     }
152   } while (nc != 0);
153 }
154