1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/scalar.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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/gemm.h>
13 #include <xnnpack/math.h>
14
15
xnn_f32_gemminc_minmax_ukernel_1x4__wasm(size_t mr,size_t nc,size_t kc,const float * restrict a,size_t a_stride,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,const float * restrict acc,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_f32_gemminc_minmax_ukernel_1x4__wasm(
17 size_t mr,
18 size_t nc,
19 size_t kc,
20 const float* restrict a,
21 size_t a_stride,
22 const float* restrict w,
23 float* restrict c,
24 size_t cm_stride,
25 size_t cn_stride,
26 const float*restrict acc,
27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 1);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(float) == 0);
34 assert(a != NULL);
35 assert(w != NULL);
36 assert(c != NULL);
37 assert(acc != NULL);
38
39 const float* a0 = a;
40 float* c0 = c;
41
42 const float vmin = params->scalar.min;
43 const float vmax = params->scalar.max;
44 do {
45 float vacc00 = acc[0];
46 float vacc01 = acc[1];
47 float vacc02 = acc[2];
48 float vacc03 = acc[3];
49 acc += 4;
50
51 size_t k = kc;
52 do {
53 const float va0 = *a0++;
54
55 const float vb0 = w[0];
56 const float vb1 = w[1];
57 const float vb2 = w[2];
58 const float vb3 = w[3];
59 w += 4;
60
61 vacc00 += va0 * vb0;
62 vacc01 += va0 * vb1;
63 vacc02 += va0 * vb2;
64 vacc03 += va0 * vb3;
65
66 k -= sizeof(float);
67 } while (k != 0);
68
69 vacc00 = __builtin_wasm_max_f32(vacc00, vmin);
70 vacc01 = __builtin_wasm_max_f32(vacc01, vmin);
71 vacc02 = __builtin_wasm_max_f32(vacc02, vmin);
72 vacc03 = __builtin_wasm_max_f32(vacc03, vmin);
73
74 vacc00 = __builtin_wasm_min_f32(vacc00, vmax);
75 vacc01 = __builtin_wasm_min_f32(vacc01, vmax);
76 vacc02 = __builtin_wasm_min_f32(vacc02, vmax);
77 vacc03 = __builtin_wasm_min_f32(vacc03, vmax);
78
79 if XNN_LIKELY(nc >= 4) {
80 c0[0] = vacc00;
81 c0[1] = vacc01;
82 c0[2] = vacc02;
83 c0[3] = vacc03;
84 c0 = (float*) ((uintptr_t) c0 + cn_stride);
85
86 a0 = (const void*) ((uintptr_t) a0 - kc);
87
88 nc -= 4;
89 } else {
90 if (nc & 2) {
91 c0[0] = vacc00;
92 c0[1] = vacc01;
93 vacc00 = vacc02;
94 c0 += 2;
95 }
96 if (nc & 1) {
97 c0[0] = vacc00;
98 }
99
100 nc = 0;
101 }
102 } while (nc != 0);
103 }
104