xref: /aosp_15_r20/external/XNNPACK/src/x32-packx/x4-scalar.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Copyright 2019 Google LLC
2*4bdc9457SAndroid Build Coastguard Worker //
3*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
4*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
5*4bdc9457SAndroid Build Coastguard Worker 
6*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
7*4bdc9457SAndroid Build Coastguard Worker 
8*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/packx.h>
9*4bdc9457SAndroid Build Coastguard Worker 
10*4bdc9457SAndroid Build Coastguard Worker 
xnn_x32_packx_ukernel_4x__scalar(size_t m,size_t k,const uint32_t * restrict x,size_t x_stride,uint32_t * restrict y)11*4bdc9457SAndroid Build Coastguard Worker void xnn_x32_packx_ukernel_4x__scalar(
12*4bdc9457SAndroid Build Coastguard Worker     size_t m,
13*4bdc9457SAndroid Build Coastguard Worker     size_t k,
14*4bdc9457SAndroid Build Coastguard Worker     const uint32_t* restrict x,
15*4bdc9457SAndroid Build Coastguard Worker     size_t x_stride,
16*4bdc9457SAndroid Build Coastguard Worker     uint32_t* restrict y)
17*4bdc9457SAndroid Build Coastguard Worker {
18*4bdc9457SAndroid Build Coastguard Worker   assert(m != 0);
19*4bdc9457SAndroid Build Coastguard Worker   assert(k != 0);
20*4bdc9457SAndroid Build Coastguard Worker 
21*4bdc9457SAndroid Build Coastguard Worker   const float* x0 = (const float*) x;
22*4bdc9457SAndroid Build Coastguard Worker   const float* x1 = (const float*) ((uintptr_t) x0 + x_stride);
23*4bdc9457SAndroid Build Coastguard Worker   if (m < 2) {
24*4bdc9457SAndroid Build Coastguard Worker     x1 = x0;
25*4bdc9457SAndroid Build Coastguard Worker   }
26*4bdc9457SAndroid Build Coastguard Worker   const float* x2 = (const float*) ((uintptr_t) x1 + x_stride);
27*4bdc9457SAndroid Build Coastguard Worker   if (m <= 2) {
28*4bdc9457SAndroid Build Coastguard Worker     x2 = x1;
29*4bdc9457SAndroid Build Coastguard Worker   }
30*4bdc9457SAndroid Build Coastguard Worker   const float* x3 = (const float*) ((uintptr_t) x2 + x_stride);
31*4bdc9457SAndroid Build Coastguard Worker   if (m != 4) {
32*4bdc9457SAndroid Build Coastguard Worker     x3 = x2;
33*4bdc9457SAndroid Build Coastguard Worker   }
34*4bdc9457SAndroid Build Coastguard Worker 
35*4bdc9457SAndroid Build Coastguard Worker   float*restrict y_f32 = (float*) y;
36*4bdc9457SAndroid Build Coastguard Worker 
37*4bdc9457SAndroid Build Coastguard Worker   do {
38*4bdc9457SAndroid Build Coastguard Worker     const float vx0 = *x0++;
39*4bdc9457SAndroid Build Coastguard Worker     const float vx1 = *x1++;
40*4bdc9457SAndroid Build Coastguard Worker     const float vx2 = *x2++;
41*4bdc9457SAndroid Build Coastguard Worker     const float vx3 = *x3++;
42*4bdc9457SAndroid Build Coastguard Worker 
43*4bdc9457SAndroid Build Coastguard Worker     y_f32[0] = vx0;
44*4bdc9457SAndroid Build Coastguard Worker     y_f32[1] = vx1;
45*4bdc9457SAndroid Build Coastguard Worker     y_f32[2] = vx2;
46*4bdc9457SAndroid Build Coastguard Worker     y_f32[3] = vx3;
47*4bdc9457SAndroid Build Coastguard Worker     y_f32 += 4;
48*4bdc9457SAndroid Build Coastguard Worker   } while (--k != 0);
49*4bdc9457SAndroid Build Coastguard Worker }
50