1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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_igemm_minmax_fp32_ukernel_1x2__wasm_fmagic(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_igemm_minmax_fp32_ukernel_1x2__wasm_fmagic(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const uint8_t**restrict a,
23 const void*restrict w,
24 uint8_t*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const uint8_t* zero,
29 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31 assert(mr != 0);
32 assert(mr <= 1);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(ks != 0);
36 assert(ks % (1 * sizeof(void*)) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 uint8_t* c0 = c;
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 w = (const void*) ((const int32_t*) w + 2);
48
49 size_t p = ks;
50 do {
51 const uint8_t* restrict a0 = a[0];
52 assert(a0 != NULL);
53 if XNN_UNPREDICTABLE(a0 != zero) {
54 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
55 }
56 a += 1;
57
58 size_t k = kc;
59 do {
60 const int32_t va0 = (int32_t) (uint32_t) *a0++;
61
62 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
63 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
64 w = (const void*) ((const uint8_t*) w + 2);
65
66 vacc0x0 += va0 * vb0;
67 vacc0x1 += va0 * vb1;
68
69 k -= sizeof(uint8_t);
70 } while (k != 0);
71 p -= 1 * sizeof(void*);
72 } while (p != 0);
73
74 float vfpacc0x0 = (float) vacc0x0;
75 float vfpacc0x1 = (float) vacc0x1;
76
77 const float vscale = params->fp32_scalar_fmagic.scale;
78 vfpacc0x0 *= vscale;
79 vfpacc0x1 *= vscale;
80
81 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
82 vfpacc0x0 = __builtin_wasm_max_f32(vfpacc0x0, voutput_min_less_zero_point);
83 vfpacc0x1 = __builtin_wasm_max_f32(vfpacc0x1, 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
89 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
90 vfpacc0x0 += vmagic_bias;
91 vfpacc0x1 += vmagic_bias;
92
93 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
94 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
95 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
96
97 if XNN_LIKELY(nc >= 2) {
98 c0[0] = (uint8_t) vout0x0;
99 c0[1] = (uint8_t) vout0x1;
100
101 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
102
103 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
104 nc -= 2;
105 } else {
106 if (nc & 1) {
107 c0[0] = (uint8_t) vout0x0;
108 }
109
110 nc = 0;
111 }
112 } while (nc != 0);
113 }
114