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_2x2__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_2x2__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 <= 2);
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
43 const int32_t vb_zero_point = params->fp32_scalar_fmagic.kernel_zero_point;
44 do {
45 int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
46 int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
47 int32_t vacc1x0 = vacc0x0;
48 int32_t vacc1x1 = vacc0x1;
49 w = (const void*) ((const int32_t*) w + 2);
50
51 size_t k = kc;
52 do {
53 const int32_t va0 = (int32_t) (uint32_t) *a0++;
54 const int32_t va1 = (int32_t) (uint32_t) *a1++;
55
56 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
57 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
58 w = (const void*) ((const uint8_t*) w + 2);
59
60 vacc0x0 += va0 * vb0;
61 vacc0x1 += va0 * vb1;
62 vacc1x0 += va1 * vb0;
63 vacc1x1 += va1 * vb1;
64
65 k -= sizeof(uint8_t);
66 } while (k != 0);
67
68 float vfpacc0x0 = (float) vacc0x0;
69 float vfpacc0x1 = (float) vacc0x1;
70 float vfpacc1x0 = (float) vacc1x0;
71 float vfpacc1x1 = (float) vacc1x1;
72
73 const float vscale = params->fp32_scalar_fmagic.scale;
74 vfpacc0x0 *= vscale;
75 vfpacc0x1 *= vscale;
76 vfpacc1x0 *= vscale;
77 vfpacc1x1 *= vscale;
78
79 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
80 vfpacc0x0 = __builtin_wasm_max_f32(vfpacc0x0, voutput_min_less_zero_point);
81 vfpacc0x1 = __builtin_wasm_max_f32(vfpacc0x1, voutput_min_less_zero_point);
82 vfpacc1x0 = __builtin_wasm_max_f32(vfpacc1x0, voutput_min_less_zero_point);
83 vfpacc1x1 = __builtin_wasm_max_f32(vfpacc1x1, voutput_min_less_zero_point);
84
85 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
86 vfpacc0x0 = __builtin_wasm_min_f32(vfpacc0x0, voutput_max_less_zero_point);
87 vfpacc0x1 = __builtin_wasm_min_f32(vfpacc0x1, voutput_max_less_zero_point);
88 vfpacc1x0 = __builtin_wasm_min_f32(vfpacc1x0, voutput_max_less_zero_point);
89 vfpacc1x1 = __builtin_wasm_min_f32(vfpacc1x1, voutput_max_less_zero_point);
90
91 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
92 vfpacc0x0 += vmagic_bias;
93 vfpacc0x1 += vmagic_bias;
94 vfpacc1x0 += vmagic_bias;
95 vfpacc1x1 += vmagic_bias;
96
97 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
98 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
99 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
100 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0) - vmagic_bias_less_output_zero_point;
101 int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1) - vmagic_bias_less_output_zero_point;
102
103 if XNN_LIKELY(nc >= 2) {
104 c0[0] = (uint8_t) vout0x0;
105 c0[1] = (uint8_t) vout0x1;
106 c1[0] = (uint8_t) vout1x0;
107 c1[1] = (uint8_t) vout1x1;
108
109 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
110 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
111
112 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
113 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
114
115 nc -= 2;
116 } else {
117 if (nc & 1) {
118 c0[0] = (uint8_t) vout0x0;
119 c1[0] = (uint8_t) vout1x0;
120 }
121
122 nc = 0;
123 }
124 } while (nc != 0);
125 }
126