mirror/darknet
1
0
Fork 0
mirror of https://github.com/pjreddie/darknet.git synced 2023-08-10 21:13:14 +03:00
Code Issues Projects Releases Wiki Activity

CUDA so fast

Browse Source
This commit is contained in:
Joseph Redmon
2015-01-22 16:38:24 -08:00
parent 4ac78c8926
commit 809f924db2
57 changed files with 1116 additions and 2181 deletions
Download Patch File Download Diff File Expand all files Collapse all files

75
src/softmax_layer.c
View File

@ -1,5 +1,6 @@
#include "softmax_layer.h"
#include "mini_blas.h"
#include "blas.h"
#include "cuda.h"
#include <float.h>
#include <math.h>
#include <stdlib.h>
@ -15,8 +16,8 @@ softmax_layer *make_softmax_layer(int batch, int inputs)
layer->delta = calloc(inputs*batch, sizeof(float));
layer->jacobian = calloc(inputs*inputs*batch, sizeof(float));
#ifdef GPU
layer->output_cl = cl_make_array(layer->output, inputs*batch);
layer->delta_cl = cl_make_array(layer->delta, inputs*batch);
layer->output_gpu = cuda_make_array(layer->output, inputs*batch);
layer->delta_gpu = cuda_make_array(layer->delta, inputs*batch);
#endif
return layer;
}
@ -49,71 +50,3 @@ void backward_softmax_layer(const softmax_layer layer, float *delta)
}
}
#ifdef GPU
void pull_softmax_layer_output(const softmax_layer layer)
{
cl_read_array(layer.output_cl, layer.output, layer.inputs*layer.batch);
}
cl_kernel get_softmax_forward_kernel()
{
static int init = 0;
static cl_kernel kernel;
if(!init){
kernel = get_kernel("src/softmax_layer.cl", "forward", 0);
init = 1;
}
return kernel;
}
void forward_softmax_layer_gpu(const softmax_layer layer, cl_mem input)
{
cl_kernel kernel = get_softmax_forward_kernel();
cl_command_queue queue = cl.queue;
cl_uint i = 0;
cl.error = clSetKernelArg(kernel, i++, sizeof(layer.inputs), (void*) &layer.inputs);
cl.error = clSetKernelArg(kernel, i++, sizeof(input), (void*) &input);
cl.error = clSetKernelArg(kernel, i++, sizeof(layer.output_cl), (void*) &layer.output_cl);
check_error(cl);
const size_t global_size[] = {layer.batch};
cl.error = clEnqueueNDRangeKernel(queue, kernel, 1, 0, global_size, 0, 0, 0, 0);
check_error(cl);
/*
cl_read_array(layer.output_cl, layer.output, layer.inputs*layer.batch);
int z;
for(z = 0; z < layer.inputs*layer.batch; ++z) printf("%f,",layer.output[z]);
*/
}
void backward_softmax_layer_gpu(const softmax_layer layer, cl_mem delta)
{
copy_ongpu(layer.batch*layer.inputs, layer.delta_cl, 1, delta, 1);
}
#endif
/* This is if you want softmax w/o log-loss classification. You probably don't.
int i,j,b;
for(b = 0; b < layer.batch; ++b){
for(i = 0; i < layer.inputs; ++i){
for(j = 0; j < layer.inputs; ++j){
int d = (i==j);
layer.jacobian[b*layer.inputs*layer.inputs + i*layer.inputs + j] =
layer.output[b*layer.inputs + i] * (d - layer.output[b*layer.inputs + j]);
}
}
}
for(b = 0; b < layer.batch; ++b){
int M = layer.inputs;
int N = 1;
int K = layer.inputs;
float *A = layer.jacobian + b*layer.inputs*layer.inputs;
float *B = layer.delta + b*layer.inputs;
float *C = delta + b*layer.inputs;
gemm(0,0,M,N,K,1,A,K,B,N,0,C,N);
}
*/