half_to_float.cuh
1.89 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
#ifndef STIM_CUDA_HALF_TO_FLOAT_H
#define STIM_CUDA_HALF_TO_FLOAT_H
#include <iostream>
#include <cuda.h>
#include <stim/cuda/cudatools.h>
#include <stim/cuda/sharedmem.cuh>
#include <stim/cuda/cudatools/error.h>
#include "cuda_fp16.h"
__global__ void cuda_h2f(float* gpu_float, half* gpu_half, int x, int y, int z){
//calculate x,y,z coordinates for this thread
int xi = blockIdx.x * blockDim.x + threadIdx.x;
//find the grid size along y
int grid_y = y / blockDim.y;
int blockidx_y = blockIdx.y % grid_y;
int yi = blockidx_y * blockDim.y + threadIdx.y;
int zi = blockIdx.y / grid_y;
int i = zi * x * y + yi * x + xi;
if(xi >= x|| yi >= y || zi>= z) return;
gpu_float[i] = __half2float(gpu_half[i]);
}
void gpu_h2f(float* gpu_float, half* gpu_half, unsigned int x, unsigned int y, unsigned int z){
int max_threads = stim::maxThreadsPerBlock();
dim3 threads(sqrt (max_threads),sqrt (max_threads));
dim3 blocks(x / threads.x + 1, (y / threads.y + 1) * z);
//call the GPU kernel to determine the gradient
cuda_h2f <<< blocks, threads >>>(gpu_float, gpu_half, x, y, z);
}
void cpu_f2h(float* f_out, half* h_in, unsigned int x, unsigned int y, unsigned int z){
//calculate the number of pixels in the array
unsigned int pix = x* y* z;
//allocate memory on the GPU for the input half precision
half* gpu_half;
cudaMalloc(&gpu_half, pix * sizeof(half));
cudaMemcpy(gpu_half, h_in, pix * sizeof(half), cudaMemcpyHostToDevice);
//allocate memory on the GPU for the output float precision.
float* gpu_float;
cudaMalloc(&gpu_float, pix * sizeof(float));
//call the GPU version of this function
gpu_h2f(gpu_float, gpu_half, x, y, z);
cudaMemcpy(f_out, gpu_float, pix * sizeof(float), cudaMemcpyDeviceToHost);
//free allocated memory
cudaFree(gpu_float);
cudaFree(gpu_half);
}
#endif