This commit is contained in:
Evghenii
2014-03-05 16:16:51 +01:00
parent 644118cd17
commit 335d36211c
2 changed files with 10 additions and 8 deletions

View File

@@ -112,7 +112,7 @@ int main(int argc, char * argv[])
for (int i = 0; i < n; i++) for (int i = 0; i < n; i++)
assert(A[j*n+i].first == j && A[j*n+i].second == i); assert(A[j*n+i].first == j && A[j*n+i].second == i);
fprintf(stderr, " tranpose done in %g msec :: BW= % GB/s\n", fprintf(stderr, " tranpose done in %g msec :: BW= %g GB/s\n",
dt , 2*m*n*sizeof(int)*2/dt*1e3/1e9); dt , 2*m*n*sizeof(int)*2/dt*1e3/1e9);

View File

@@ -86,7 +86,6 @@ void transpose_serial(uniform T A[], const uniform int m, const uniform int n)
} }
#endif #endif
static uniform T * uniform tmpAll = NULL;
static uniform int * uniform joverb = NULL; static uniform int * uniform joverb = NULL;
static uniform int * uniform iovera = NULL; static uniform int * uniform iovera = NULL;
static uniform int a,b,c; static uniform int a,b,c;
@@ -94,8 +93,6 @@ static uniform int a,b,c;
static static
void transpose_init(const uniform int m, const uniform int n, const uniform int nTask) void transpose_init(const uniform int m, const uniform int n, const uniform int nTask)
{ {
const uniform int tmpSize = max(m,n) * programCount * nTask;
tmpAll = uniform new uniform T [tmpSize];
joverb = uniform new uniform int[n]; joverb = uniform new uniform int[n];
iovera = uniform new uniform int[m]; iovera = uniform new uniform int[m];
@@ -113,7 +110,6 @@ void transpose_finalize()
{ {
delete iovera; delete iovera;
delete joverb; delete joverb;
delete tmpAll;
} }
task task
@@ -123,7 +119,7 @@ void transpose_step1(uniform T A[], const uniform int m, const uniform int n)
const uniform int nibeg = taskIndex * n_per_task; const uniform int nibeg = taskIndex * n_per_task;
const uniform int niend = min(nibeg + n_per_task, n); const uniform int niend = min(nibeg + n_per_task, n);
uniform T * uniform tmp = tmpAll + max(m,n)*taskIndex; uniform T * uniform tmp = uniform new uniform T[m];
for (uniform int j = nibeg; j < niend; j++) for (uniform int j = nibeg; j < niend; j++)
{ {
@@ -134,6 +130,8 @@ void transpose_step1(uniform T A[], const uniform int m, const uniform int n)
foreach (i = 0 ... m) foreach (i = 0 ... m)
A[base + i] = tmp[i]; A[base + i] = tmp[i];
} }
delete tmp;
} }
task task
@@ -143,7 +141,7 @@ void transpose_step2(uniform T A[], const uniform int m, const uniform int n)
const uniform int mibeg = taskIndex * m_per_task; const uniform int mibeg = taskIndex * m_per_task;
const uniform int miend = min(mibeg + m_per_task, m); const uniform int miend = min(mibeg + m_per_task, m);
uniform T * uniform tmp = tmpAll + max(m,n)*programCount * taskIndex; uniform T * uniform tmp = uniform new uniform T[n*programCount];
uniform T (*uniform tmp2D)[programCount] = (uniform T (*uniform)[programCount])tmp; uniform T (*uniform tmp2D)[programCount] = (uniform T (*uniform)[programCount])tmp;
foreach (i = mibeg ... miend) foreach (i = mibeg ... miend)
@@ -153,6 +151,8 @@ void transpose_step2(uniform T A[], const uniform int m, const uniform int n)
for (uniform int j = 0; j < n; j++) for (uniform int j = 0; j < n; j++)
A[j*m + i] = tmp2D[j][programIndex]; A[j*m + i] = tmp2D[j][programIndex];
} }
delete tmp;
} }
task task
@@ -162,7 +162,7 @@ void transpose_step3(uniform T A[], const uniform int m, const uniform int n)
const uniform int nibeg = taskIndex * n_per_task; const uniform int nibeg = taskIndex * n_per_task;
const uniform int niend = min(nibeg + n_per_task, n); const uniform int niend = min(nibeg + n_per_task, n);
uniform T * uniform tmp = tmpAll + max(m,n)*taskIndex; uniform T * uniform tmp = uniform new uniform T[m];
for (uniform int j = nibeg; j < niend; j++) for (uniform int j = nibeg; j < niend; j++)
{ {
@@ -172,6 +172,8 @@ void transpose_step3(uniform T A[], const uniform int m, const uniform int n)
foreach (i = 0 ... m) foreach (i = 0 ... m)
A[base + i] = tmp[i]; A[base + i] = tmp[i];
} }
delete tmp;
} }
export export