mmr/tkDNN
Archived
1
0
Fork 0
Code Issues 46 Pull Requests 12 Releases 4 Activity

Merge remote-tracking branch 'origin/master' into cnet

Browse Source
This commit is contained in:
Davide Sapienza
2021-07-22 17:03:23 +02:00
parent 2bb70859da f4e71a5a28
commit 0292e21c33
59 changed files with 7640 additions and 209 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/kernels/activation_leaky.cu
+4 -4
View File
@@ -1,7 +1,7 @@
#include "kernels.h"
__global__
void activation_leaky(dnnType *input, dnnType *output, int size) {
void activation_leaky(dnnType *input, dnnType *output, int size, float slope) {
int i = blockDim.x*blockIdx.x + threadIdx.x;
@@ -9,7 +9,7 @@ void activation_leaky(dnnType *input, dnnType *output, int size) {
if (input[i]>0)
output[i] = input[i];
else
output[i] = 0.1f*input[i];
output[i] = slope*input[i];
}
}
@@ -17,12 +17,12 @@ void activation_leaky(dnnType *input, dnnType *output, int size) {
/**
ELU activation function
*/
void activationLEAKYForward(dnnType* srcData, dnnType* dstData, int size, cudaStream_t stream)
void activationLEAKYForward(dnnType* srcData, dnnType* dstData, int size, float slope, cudaStream_t stream)
{
int blocks = (size+255)/256;
int threads = 256;
activation_leaky<<<blocks, threads, 0, stream>>>(srcData, dstData, size);
activation_leaky<<<blocks, threads, 0, stream>>>(srcData, dstData, size, slope);
}
src/kernels/postprocessing.cu
+19
View File
@@ -40,6 +40,25 @@ void sortAndTopKonDevice(dnnType *src_begin, int *idsrc, float *topk_scores, int
sortAndTopK_kernel<<<blocks, threads, 0>>>(src_begin, idsrc, topk_scores, topk_inds, topk_ys, topk_xs, size, K);
}
__global__
void maxElem_kernel(float *src_begin, float *dst_begin, const int n_classes, const int size){
int i = blockDim.x*blockIdx.x + threadIdx.x;
if (i > size)
return;
thrust::device_ptr<float> dPbeg ( &src_begin[i*n_classes] ) ;
thrust::device_ptr<float> dPend = dPbeg + n_classes;
thrust::device_ptr<float> result = thrust::max_element(thrust::device,dPbeg, dPend);
dst_begin[i] = result - dPbeg;
}
void maxElem(dnnType *src_begin, dnnType *dst_begin, const int c, const int h, const int w){
int blocks = (h*w)/32+1;
int threads = 32;
maxElem_kernel<<<blocks, threads, 0>>>(src_begin, dst_begin, c, h*w);
}
void topKxyclasses(int *ids_begin, int *ids_end, const int K, const int size, const int wh, int *clses, int *xs, int *ys){
thrust::transform(thrust::device, ids_begin, ids_end, thrust::make_constant_iterator(wh), clses, thrust::divides<int>());
thrust::transform(thrust::device, ids_begin, ids_end, thrust::make_constant_iterator(wh), ids_begin, thrust::modulus<int>());
src/kernels/resize.cu
+14 -27
View File
@@ -1,46 +1,33 @@
#include "kernels.h"
#include <stdio.h>
#define MIN(a,b) (((a)<(b))?(a):(b))
#define MAX(a,b) (((a)>(b))?(a):(b))
__global__ void resize_kernel( int i_N,float *x, int i_w, int i_h, int i_c,
__global__ void resize_kernel( int size,float *x, int i_w, int i_h, int i_c,
int o_w, int o_h, int o_c, int batch, float *out)
{
int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
if(i >= i_N) return;
int id = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
if(id >= size) return;
int out_index = i;
int out_w = i%o_w;
i = i/o_w;
int out_h = i%o_h;
i = i/o_h;
int out_c = i%o_c;
i = i/o_c;
int i = id % o_w;
id /= o_w;
int j = id % o_h;
id /= o_h;
int k = id % o_c;
id /= o_c;
int b = id % batch;
//copying last column/last row as padding
int in_index = ((i*i_c + MIN(out_c,i_c-1))*i_h + MIN(out_h,i_h-1))*i_w + MIN(out_w, i_w-1);
out[out_index] = x[in_index];
int out_index = i + o_w*(j + o_h*(k + o_c*b));
int add_index = i/(o_w/i_w) + i_w*(j/(o_h/i_h) + i_h*(k + i_c*b));
out[out_index] = x[add_index];
}
void resizeForward( dnnType* srcData, dnnType* dstData, int n, int i_c, int i_h, int i_w,
int o_c, int o_h, int o_w, cudaStream_t stream )
{
int i_size = n*i_c*i_h*i_w;
int o_size = n*o_c*o_h*o_w;
int blocks = (o_size+255)/256;
int threads = 256;
if(i_c == o_c && i_h == o_h && i_w == o_w )
{
checkCuda(cudaMemcpy(dstData, srcData, i_size*sizeof(dnnType), cudaMemcpyDeviceToDevice));
}
else
{
checkCuda(cudaMemset(dstData, 0, o_size*sizeof(dnnType)));
resize_kernel<<<blocks, threads, 0, stream>>>(o_size, srcData, i_w, i_h, i_c, o_w, o_h, o_c, n, dstData);
// printDeviceVector(i_size, srcData);
// printDeviceVector(o_size, dstData);
}
resize_kernel<<<blocks, threads, 0, stream>>>(o_size, srcData, i_w, i_h, i_c, o_w, o_h, o_c, n, dstData);
}
src/kernels/shortcut.cu
+48 -15
View File
@@ -21,27 +21,60 @@ __global__ void shortcut_kernel(int size, int minw, int minh, int minc, int stri
//out[out_index] += add[add_index];
}
__global__ void shortcut_mul_kernel(int size, int minw, int minh, int minc, int sample, int batch,
int w1, int h1, int c1, dnnType *mul,
int w2, int h2, int c2, float s1, float s2, dnnType *out)
{
int id = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
if (id >= size) return;
int i = id % minw;
id /= minw;
int j = id % minh;
id /= minh;
int k = id % minc;
id /= minc;
int b = id % batch;
int out_index = i*sample + w1*(j*sample + h1*(k + c1*b));
out[out_index] = out[out_index] * mul[k + c2*b];
}
void shortcutForward(dnnType* srcData, dnnType* dstData, int n1, int c1, int h1, int w1, int s1,
int n2, int c2, int h2, int w2, int s2,
cudaStream_t stream)
bool mul, cudaStream_t stream)
{
assert(n1 == n2);
int batch = n1;
int minw = (w1 < w2) ? w1 : w2;
int minh = (h1 < h2) ? h1 : h2;
int minc = (c1 < c2) ? c1 : c2;
if(!mul){
int minw = (w1 < w2) ? w1 : w2;
int minh = (h1 < h2) ? h1 : h2;
int minc = (c1 < c2) ? c1 : c2;
int stride = w1/w2;
int sample = w2/w1;
assert(stride == h1/h2);
assert(sample == h2/h1);
if(stride < 1) stride = 1;
if(sample < 1) sample = 1;
int stride = w1/w2;
int sample = w2/w1;
assert(stride == h1/h2);
assert(sample == h2/h1);
if(stride < 1) stride = 1;
if(sample < 1) sample = 1;
int size = batch * minw * minh * minc;
int blocks = (size+255)/256;
int threads = 256;
shortcut_kernel<<<blocks, threads, 0, stream>>>(size, minw, minh, minc, stride, sample, batch,
w1, h1, c1, srcData, w2, h2, c2, s1, s2, dstData);
}
else{
int minw = w1;
int minh = h1;
int minc = c1;
int sample = 1;
int size = batch * minw * minh * minc;
int blocks = (size+255)/256;
int threads = 256;
shortcut_kernel<<<blocks, threads, 0, stream>>>(size, minw, minh, minc, stride, sample, batch,
w1, h1, c1, srcData, w2, h2, c2, s1, s2, dstData);
int size = batch * minw * minh * minc;
int blocks = (size+255)/256;
int threads = 256;
shortcut_mul_kernel<<<blocks, threads, 0, stream>>>(size, minw, minh, minc, sample, batch,
w1, h1, c1, srcData, w2, h2, c2, s1, s2, dstData);
}
}