aaron/ispc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

+options

Browse Source
This commit is contained in:
Evghenii
2013-11-15 17:59:04 +01:00
parent 95d6647dce
commit a2d12517e7
3 changed files with 12 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
examples_cuda/cuda_ispc.h
View File

@@ -151,6 +151,7 @@ static CUmodule loadModule(
#endif #endif
{ {
// Load the PTX from the string myPtx (64-bit) // Load the PTX from the string myPtx (64-bit)
if (print_log)
fprintf(stderr, "Loading ptx..\n"); fprintf(stderr, "Loading ptx..\n");
myErr = cuLinkAddData(*lState, CU_JIT_INPUT_PTX, (void*)module, strlen(module)+1, 0, 0, 0, 0); myErr = cuLinkAddData(*lState, CU_JIT_INPUT_PTX, (void*)module, strlen(module)+1, 0, 0, 0, 0);
myErr = cuLinkAddFile(*lState, CU_JIT_INPUT_LIBRARY, cudadevrt_lib, 0,0,0); myErr = cuLinkAddFile(*lState, CU_JIT_INPUT_LIBRARY, cudadevrt_lib, 0,0,0);
@@ -177,6 +178,7 @@ static CUmodule loadModule(
// Destroy the linker invocation // Destroy the linker invocation
checkCudaErrors(cuLinkDestroy(*lState)); checkCudaErrors(cuLinkDestroy(*lState));
if (print_log)
fprintf(stderr, " loadModule took %g ms \n", 1e3*(rtc() - t0)); fprintf(stderr, " loadModule took %g ms \n", 1e3*(rtc() - t0));
return cudaModule; return cudaModule;
} }

4
examples_cuda/options/options.cpp
View File

@@ -143,6 +143,7 @@ int main(int argc, char *argv[]) {
// Binomial options, serial implementation // Binomial options, serial implementation
// //
double binomial_serial = 1e30; double binomial_serial = 1e30;
#if 0
for (int i = 0; i < 3; ++i) { for (int i = 0; i < 3; ++i) {
reset_and_start_timer(); reset_and_start_timer();
const double t0 = rtc(); const double t0 = rtc();
@@ -159,6 +160,7 @@ int main(int argc, char *argv[]) {
printf("\t\t\t\t(%.2fx speedup from ISPC, %.2fx speedup from ISPC + tasks)\n", printf("\t\t\t\t(%.2fx speedup from ISPC, %.2fx speedup from ISPC + tasks)\n",
binomial_serial / binomial_ispc, binomial_serial / binomial_tasks); binomial_serial / binomial_ispc, binomial_serial / binomial_tasks);
#endif
// //
// Black-Scholes options pricing model, ispc implementation, 1 thread // Black-Scholes options pricing model, ispc implementation, 1 thread
@@ -200,6 +202,7 @@ int main(int argc, char *argv[]) {
// Black-Scholes options pricing model, serial implementation // Black-Scholes options pricing model, serial implementation
// //
double bs_serial = 1e30; double bs_serial = 1e30;
#if 0
for (int i = 0; i < 3; ++i) { for (int i = 0; i < 3; ++i) {
reset_and_start_timer(); reset_and_start_timer();
const double t0 = rtc(); const double t0 = rtc();
@@ -216,6 +219,7 @@ int main(int argc, char *argv[]) {
printf("\t\t\t\t(%.2fx speedup from ISPC, %.2fx speedup from ISPC + tasks)\n", printf("\t\t\t\t(%.2fx speedup from ISPC, %.2fx speedup from ISPC + tasks)\n",
bs_serial / bs_ispc, bs_serial / bs_ispc_tasks); bs_serial / bs_ispc, bs_serial / bs_ispc_tasks);
#endif
return 0; return 0;
} }

6
examples_cuda/options/options.ispc
View File

@@ -76,10 +76,12 @@ export void
black_scholes_ispc_tasks(uniform float Sa[], uniform float Xa[], uniform float Ta[], black_scholes_ispc_tasks(uniform float Sa[], uniform float Xa[], uniform float Ta[],
uniform float ra[], uniform float va[], uniform float ra[], uniform float va[],
uniform float result[], uniform int count) { uniform float result[], uniform int count) {
uniform int nTasks = max((int)64, (int)count/16384); uniform int nTasks = 2048; //count/16384; //max((int)64, (int)count/16384);
launch[nTasks] bs_task(Sa, Xa, Ta, ra, va, result, count); launch[nTasks] bs_task(Sa, Xa, Ta, ra, va, result, count);
} }
/********/
export void export void
black_scholes_ispc(uniform float Sa[], uniform float Xa[], uniform float Ta[], black_scholes_ispc(uniform float Sa[], uniform float Xa[], uniform float Ta[],
@@ -149,6 +151,6 @@ binomial_put_ispc_tasks(uniform float Sa[], uniform float Xa[],
uniform float Ta[], uniform float ra[], uniform float Ta[], uniform float ra[],
uniform float va[], uniform float result[], uniform float va[], uniform float result[],
uniform int count) { uniform int count) {
uniform int nTasks = max((int)64, (int)count/16384); uniform int nTasks = 2048; //count/16384; //max((int)64, (int)count/16384);
launch[nTasks] binomial_task(Sa, Xa, Ta, ra, va, result, count); launch[nTasks] binomial_task(Sa, Xa, Ta, ra, va, result, count);
} }