diff --git a/CMakeLists.txt b/CMakeLists.txt
index 0501e92baa1ee73ba77ecba6de6c5eb086aab5d4..66952aa9d463f50636f5c7edd2891a29f7727465 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -60,6 +60,8 @@ include(GenPkgConfig)
# Parameters/Options #
######################
+set(CHAMELEON_DEFINITIONS_LIST "")
+
# Add define for Fortran Mangling (should be defined somewhere else)
# ------------------------------------------------------------------
add_definitions(-DADD_)
@@ -686,6 +688,7 @@ endif()
###############################################################################
# Print Options #
#################
+get_directory_property( CHAMELEON_DEFINITIONS_LIST DIRECTORY ${CMAKE_SOURCE_DIR} COMPILE_DEFINITIONS )
include(PrintOpts)
###
diff --git a/runtime/starpu/codelets/codelet_zgemm.c b/runtime/starpu/codelets/codelet_zgemm.c
index ca94fdd206bf1d4e8748981b79cb0cd3a89e1bfc..8e52cb94b6268db109386e6957c39a5ac70709a5 100644
--- a/runtime/starpu/codelets/codelet_zgemm.c
+++ b/runtime/starpu/codelets/codelet_zgemm.c
@@ -121,27 +121,30 @@ static void cl_zgemm_cuda_func(void *descr[], void *cl_arg)
cuDoubleComplex beta;
cuDoubleComplex *C;
int ldc;
+ CUstream stream;
+ cublasHandle_t handle;
+ cublasStatus_t stat;
+ cublasOperation_t cublasTransA;
+ cublasOperation_t cublasTransB;
A = (const cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[0]);
B = (const cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[1]);
C = (cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[2]);
starpu_codelet_unpack_args(cl_arg, &transA, &transB, &m, &n, &k, &alpha, &lda, &ldb, &beta, &ldc);
- cublasHandle_t handle;
- cublasStatus_t stat = cublasCreate(&handle);
+ stat = cublasCreate(&handle);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("CUBLAS initialization failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- CUstream stream = starpu_cuda_get_local_stream();
+ stream = starpu_cuda_get_local_stream();
stat = cublasSetStream(handle, stream);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("cublasSetStream failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- cublasOperation_t cublasTransA;
if (transA == MorseNoTrans){
cublasTransA = CUBLAS_OP_N;
}else if(transA == MorseTrans){
@@ -151,7 +154,6 @@ static void cl_zgemm_cuda_func(void *descr[], void *cl_arg)
}else{
fprintf(stderr, "Error in cl_zgemm_cuda_func: bad transA parameter %d\n", transA);
}
- cublasOperation_t cublasTransB;
if (transB == MorseNoTrans){
cublasTransB = CUBLAS_OP_N;
}else if(transB == MorseTrans){
diff --git a/runtime/starpu/codelets/codelet_zgeqrt.c b/runtime/starpu/codelets/codelet_zgeqrt.c
index f22e424249237e57ebafe08cfa71ce5dfc2d8368..4171aa70e77fd6dd073fbe4d3795e875a8d74179 100644
--- a/runtime/starpu/codelets/codelet_zgeqrt.c
+++ b/runtime/starpu/codelets/codelet_zgeqrt.c
@@ -162,14 +162,17 @@ magma_zgemerge_gpu(magma_side_t side, magma_diag_t diag,
{
int i, j;
magmaDoubleComplex *cola, *colb;
+ CUstream stream;
cublasHandle_t handle;
- cublasStatus_t stat = cublasCreate(&handle);
+ cublasStatus_t stat;
+
+ stat = cublasCreate(&handle);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("CUBLAS initialization failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- CUstream stream = starpu_cuda_get_local_stream();
+ stream = starpu_cuda_get_local_stream();
stat = cublasSetStream(handle, stream);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("cublasSetStream failed\n");
@@ -234,13 +237,19 @@ magma_zgemerge_gpu(magma_side_t side, magma_diag_t diag,
for(i=0; i<N; i++){
cola = A + i*LDA;
colb = B + i*LDB;
- cublasZcopy(i+1, cola, 1, colb, 1);
+ //cublasZcopy(i+1, cola, 1, colb, 1);
+ cudaMemcpy(colb , cola,
+ (i+1)*sizeof(cuDoubleComplex),
+ cudaMemcpyDeviceToDevice );
}
}else{
for(i=0; i<N; i++){
cola = A + i*LDA;
colb = B + i*LDB;
- cublasZcopy(M-i, cola + i, 1, colb + i, 1);
+ //cublasZcopy(M-i, cola + i, 1, colb + i, 1);
+ cudaMemcpy(colb+i , cola+i,
+ (M-i)*sizeof(cuDoubleComplex),
+ cudaMemcpyDeviceToDevice );
}
}
@@ -291,9 +300,13 @@ magma_zgeqrt_gpu( magma_int_t m, magma_int_t n, magma_int_t nb,
cudaMemset(dt_ref(0,0), 0, nb*n*sizeof(magmaDoubleComplex));
/* copy first panel of A on the host */
- cublasGetMatrix(m, min(nb,n), sizeof(magmaDoubleComplex),
- da_ref(0, 0), ldda,
- v, ldv);
+// cublasGetMatrix(m, min(nb,n), sizeof(magmaDoubleComplex),
+// da_ref(0, 0), ldda,
+// v, ldv);
+ /* copy first panel of A on the host */
+ cudaMemcpy( v, da_ref(0,0),
+ m*min(nb,n)*sizeof(magmaDoubleComplex),
+ cudaMemcpyDeviceToHost );
/* Use blocked code initially */
for (i = 0; i < k; i += nb) {
@@ -305,9 +318,13 @@ magma_zgeqrt_gpu( magma_int_t m, magma_int_t n, magma_int_t nb,
if (i>0){
/* copy panel of A from device to host */
- cublasGetMatrix(m, ib, sizeof(magmaDoubleComplex),
- da_ref(0, i), ldda,
- v, ldv);
+// cublasGetMatrix(m, ib, sizeof(magmaDoubleComplex),
+// da_ref(0, i), ldda,
+// v, ldv);
+ /* copy panel of A from device to host */
+ cudaMemcpy( v, da_ref(0,i),
+ m*ib*sizeof(magmaDoubleComplex),
+ cudaMemcpyDeviceToHost );
/* Apply H' to A(i:m,i+2*ib:n) from the left */
cols = n-old_i-2*old_ib;
diff --git a/runtime/starpu/codelets/codelet_zhemm.c b/runtime/starpu/codelets/codelet_zhemm.c
index e7c297dc5ce8ff90c876977db4437af58ae80f1f..1b4b1241b630ef62e11c2af655dc3fd4118e7adc 100644
--- a/runtime/starpu/codelets/codelet_zhemm.c
+++ b/runtime/starpu/codelets/codelet_zhemm.c
@@ -117,27 +117,30 @@ static void cl_zhemm_cuda_func(void *descr[], void *cl_arg)
cuDoubleComplex beta;
cuDoubleComplex *C;
int LDC;
+ CUstream stream;
+ cublasHandle_t handle;
+ cublasStatus_t stat;
+ cublasSideMode_t cublasSide;
+ cublasFillMode_t cublasUplo;
A = (const cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[0]);
B = (const cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[1]);
C = (cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[2]);
starpu_codelet_unpack_args(cl_arg, &side, &uplo, &M, &N, &alpha, &LDA, &LDB, &beta, &LDC);
- cublasHandle_t handle;
- cublasStatus_t stat = cublasCreate(&handle);
+ stat = cublasCreate(&handle);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("CUBLAS initialization failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- CUstream stream = starpu_cuda_get_local_stream();
+ stream = starpu_cuda_get_local_stream();
stat = cublasSetStream(handle, stream);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("cublasSetStream failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- cublasSideMode_t cublasSide;
if (side == MorseLeft){
cublasSide = CUBLAS_SIDE_LEFT;
}else if (side == MorseRight){
@@ -145,7 +148,6 @@ static void cl_zhemm_cuda_func(void *descr[], void *cl_arg)
}else{
fprintf(stderr, "Error in cl_zhemm_cuda_func: bad side parameter %d\n", side);
}
- cublasFillMode_t cublasUplo;
if (uplo == MorseUpper){
cublasUplo = CUBLAS_FILL_MODE_UPPER;
}else if(uplo == MorseLower){
diff --git a/runtime/starpu/codelets/codelet_zher2k.c b/runtime/starpu/codelets/codelet_zher2k.c
index 4755ca80ef182c1b858770c5f60995f34ff30ec2..bd7da41cf61552d661b0ade9a5dc0e62277bc4ea 100644
--- a/runtime/starpu/codelets/codelet_zher2k.c
+++ b/runtime/starpu/codelets/codelet_zher2k.c
@@ -112,27 +112,30 @@ static void cl_zher2k_cuda_func(void *descr[], void *cl_arg)
double beta;
cuDoubleComplex *C;
int ldc;
+ CUstream stream;
+ cublasHandle_t handle;
+ cublasStatus_t stat;
+ cublasFillMode_t cublasUplo;
+ cublasOperation_t cublasTrans;
A = (const cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[0]);
B = (const cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[1]);
C = (cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[2]);
starpu_codelet_unpack_args(cl_arg, &uplo, &trans, &n, &k, &alpha, &lda, &ldb, &beta, &ldc);
- cublasHandle_t handle;
- cublasStatus_t stat = cublasCreate(&handle);
+ stat = cublasCreate(&handle);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("CUBLAS initialization failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- CUstream stream = starpu_cuda_get_local_stream();
+ stream = starpu_cuda_get_local_stream();
stat = cublasSetStream(handle, stream);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("cublasSetStream failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- cublasFillMode_t cublasUplo;
if (uplo == MorseUpper){
cublasUplo = CUBLAS_FILL_MODE_UPPER;
}else if(uplo == MorseLower){
@@ -142,8 +145,6 @@ static void cl_zher2k_cuda_func(void *descr[], void *cl_arg)
}else{
fprintf(stderr, "Error in cl_zher2k_cuda_func: bad uplo parameter %d\n", uplo);
}
-
- cublasOperation_t cublasTrans;
if (trans == MorseNoTrans){
cublasTrans = CUBLAS_OP_N;
}else if(trans == MorseTrans){
diff --git a/runtime/starpu/codelets/codelet_zherk.c b/runtime/starpu/codelets/codelet_zherk.c
index 18584b8636cb8a67211cfefb9b89eea6e70b5daa..44ae4f252a1ac2edcf76ae10fcd1838a3ea696fe 100644
--- a/runtime/starpu/codelets/codelet_zherk.c
+++ b/runtime/starpu/codelets/codelet_zherk.c
@@ -106,26 +106,29 @@ static void cl_zherk_cuda_func(void *descr[], void *cl_arg)
double beta;
cuDoubleComplex *C;
int ldc;
+ CUstream stream;
+ cublasHandle_t handle;
+ cublasStatus_t stat;
+ cublasFillMode_t cublasUplo;
+ cublasOperation_t cublasTrans;
A = (const cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[0]);
C = (cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[1]);
starpu_codelet_unpack_args(cl_arg, &uplo, &trans, &n, &k, &alpha, &lda, &beta, &ldc);
- cublasHandle_t handle;
- cublasStatus_t stat = cublasCreate(&handle);
+ stat = cublasCreate(&handle);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("CUBLAS initialization failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- CUstream stream = starpu_cuda_get_local_stream();
+ stream = starpu_cuda_get_local_stream();
stat = cublasSetStream(handle, stream);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("cublasSetStream failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- cublasFillMode_t cublasUplo;
if (uplo == MorseUpper){
cublasUplo = CUBLAS_FILL_MODE_UPPER;
}else if(uplo == MorseLower){
@@ -135,8 +138,6 @@ static void cl_zherk_cuda_func(void *descr[], void *cl_arg)
}else{
fprintf(stderr, "Error in cl_zherk_cuda_func: bad uplo parameter %d\n", uplo);
}
-
- cublasOperation_t cublasTrans;
if (trans == MorseNoTrans){
cublasTrans = CUBLAS_OP_N;
}else if(trans == MorseTrans){
diff --git a/runtime/starpu/codelets/codelet_zsymm.c b/runtime/starpu/codelets/codelet_zsymm.c
index ae1903daabf8b3bae8333d2bdfa31894c546697a..173d6d3b85562aea11b53a2f7214674e5d0ac0ec 100644
--- a/runtime/starpu/codelets/codelet_zsymm.c
+++ b/runtime/starpu/codelets/codelet_zsymm.c
@@ -117,27 +117,30 @@ static void cl_zsymm_cuda_func(void *descr[], void *cl_arg)
cuDoubleComplex beta;
cuDoubleComplex *C;
int LDC;
+ CUstream stream;
+ cublasHandle_t handle;
+ cublasStatus_t stat;
+ cublasSideMode_t cublasSide;
+ cublasFillMode_t cublasUplo;
A = (const cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[0]);
B = (const cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[1]);
C = (cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[2]);
starpu_codelet_unpack_args(cl_arg, &side, &uplo, &M, &N, &alpha, &LDA, &LDB, &beta, &LDC);
- cublasHandle_t handle;
- cublasStatus_t stat = cublasCreate(&handle);
+ stat = cublasCreate(&handle);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("CUBLAS initialization failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- CUstream stream = starpu_cuda_get_local_stream();
+ stream = starpu_cuda_get_local_stream();
stat = cublasSetStream(handle, stream);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("cublasSetStream failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- cublasSideMode_t cublasSide;
if (side == MorseLeft){
cublasSide = CUBLAS_SIDE_LEFT;
}else if (side == MorseRight){
@@ -145,7 +148,6 @@ static void cl_zsymm_cuda_func(void *descr[], void *cl_arg)
}else{
fprintf(stderr, "Error in cl_zsymm_cuda_func: bad side parameter %d\n", side);
}
- cublasFillMode_t cublasUplo;
if (uplo == MorseUpper){
cublasUplo = CUBLAS_FILL_MODE_UPPER;
}else if(uplo == MorseLower){
diff --git a/runtime/starpu/codelets/codelet_zsyr2k.c b/runtime/starpu/codelets/codelet_zsyr2k.c
index 2d99bdb82e8d11a97e699f71fb33b72fd834b6a3..2922de6fc788f852d80ee13756ee553274d52247 100644
--- a/runtime/starpu/codelets/codelet_zsyr2k.c
+++ b/runtime/starpu/codelets/codelet_zsyr2k.c
@@ -112,27 +112,30 @@ static void cl_zsyr2k_cuda_func(void *descr[], void *cl_arg)
cuDoubleComplex beta;
cuDoubleComplex *C;
int ldc;
+ CUstream stream;
+ cublasHandle_t handle;
+ cublasStatus_t stat;
+ cublasFillMode_t cublasUplo;
+ cublasOperation_t cublasTrans;
A = (const cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[0]);
B = (const cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[1]);
C = (cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[2]);
starpu_codelet_unpack_args(cl_arg, &uplo, &trans, &n, &k, &alpha, &lda, &ldb, &beta, &ldc);
- cublasHandle_t handle;
- cublasStatus_t stat = cublasCreate(&handle);
+ stat = cublasCreate(&handle);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("CUBLAS initialization failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- CUstream stream = starpu_cuda_get_local_stream();
+ stream = starpu_cuda_get_local_stream();
stat = cublasSetStream(handle, stream);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("cublasSetStream failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- cublasFillMode_t cublasUplo;
if (uplo == MorseUpper){
cublasUplo = CUBLAS_FILL_MODE_UPPER;
}else if(uplo == MorseLower){
@@ -142,8 +145,6 @@ static void cl_zsyr2k_cuda_func(void *descr[], void *cl_arg)
}else{
fprintf(stderr, "Error in cl_zsyr2k_cuda_func: bad uplo parameter %d\n", uplo);
}
-
- cublasOperation_t cublasTrans;
if (trans == MorseNoTrans){
cublasTrans = CUBLAS_OP_N;
}else if(trans == MorseTrans){
diff --git a/runtime/starpu/codelets/codelet_zsyrk.c b/runtime/starpu/codelets/codelet_zsyrk.c
index 8803124586868e7629669f18ad9aa132a31a1c2b..0f69498a4e77cd61b5c9515560c4a6d7d6d770c2 100644
--- a/runtime/starpu/codelets/codelet_zsyrk.c
+++ b/runtime/starpu/codelets/codelet_zsyrk.c
@@ -107,26 +107,29 @@ static void cl_zsyrk_cuda_func(void *descr[], void *cl_arg)
cuDoubleComplex beta;
cuDoubleComplex *C;
int ldc;
+ CUstream stream;
+ cublasHandle_t handle;
+ cublasStatus_t stat;
+ cublasFillMode_t cublasUplo;
+ cublasOperation_t cublasTrans;
A = (const cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[0]);
C = (cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[1]);
starpu_codelet_unpack_args(cl_arg, &uplo, &trans, &n, &k, &alpha, &lda, &beta, &ldc);
- cublasHandle_t handle;
- cublasStatus_t stat = cublasCreate(&handle);
+ stat = cublasCreate(&handle);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("CUBLAS initialization failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- CUstream stream = starpu_cuda_get_local_stream();
+ stream = starpu_cuda_get_local_stream();
stat = cublasSetStream(handle, stream);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("cublasSetStream failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- cublasFillMode_t cublasUplo;
if (uplo == MorseUpper){
cublasUplo = CUBLAS_FILL_MODE_UPPER;
}else if(uplo == MorseLower){
@@ -136,8 +139,6 @@ static void cl_zsyrk_cuda_func(void *descr[], void *cl_arg)
}else{
fprintf(stderr, "Error in cl_zsyrk_cuda_func: bad uplo parameter %d\n", uplo);
}
-
- cublasOperation_t cublasTrans;
if (trans == MorseNoTrans){
cublasTrans = CUBLAS_OP_N;
}else if(trans == MorseTrans){
diff --git a/runtime/starpu/codelets/codelet_ztrmm.c b/runtime/starpu/codelets/codelet_ztrmm.c
index 497ca32ef4801416429d041159ee601a82fd41b9..510005d40d382230dcf86d242e4f28d69e06c5d0 100644
--- a/runtime/starpu/codelets/codelet_ztrmm.c
+++ b/runtime/starpu/codelets/codelet_ztrmm.c
@@ -111,26 +111,31 @@ static void cl_ztrmm_cuda_func(void *descr[], void *cl_arg)
int LDA;
cuDoubleComplex *B;
int LDB;
+ CUstream stream;
+ cublasHandle_t handle;
+ cublasStatus_t stat;
+ cublasSideMode_t cublasSide;
+ cublasFillMode_t cublasUplo;
+ cublasOperation_t cublasTransA;
+ cublasDiagType_t cublasDiag;
A = (const cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[0]);
B = (cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[1]);
starpu_codelet_unpack_args(cl_arg, &side, &uplo, &transA, &diag, &M, &N, &alpha, &LDA, &LDB);
- cublasHandle_t handle;
- cublasStatus_t stat = cublasCreate(&handle);
+ stat = cublasCreate(&handle);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("CUBLAS initialization failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- CUstream stream = starpu_cuda_get_local_stream();
+ stream = starpu_cuda_get_local_stream();
stat = cublasSetStream(handle, stream);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("cublasSetStream failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- cublasSideMode_t cublasSide;
if (side == MorseLeft){
cublasSide = CUBLAS_SIDE_LEFT;
}else if (side == MorseRight){
@@ -138,7 +143,6 @@ static void cl_ztrmm_cuda_func(void *descr[], void *cl_arg)
}else{
fprintf(stderr, "Error in cl_ztrmm_cuda_func: bad side parameter %d\n", side);
}
- cublasFillMode_t cublasUplo;
if (uplo == MorseUpper){
cublasUplo = CUBLAS_FILL_MODE_UPPER;
}else if(uplo == MorseLower){
@@ -148,7 +152,6 @@ static void cl_ztrmm_cuda_func(void *descr[], void *cl_arg)
}else{
fprintf(stderr, "Error in cl_ztrmm_cuda_func: bad uplo parameter %d\n", uplo);
}
- cublasOperation_t cublasTransA;
if (transA == MorseNoTrans){
cublasTransA = CUBLAS_OP_N;
}else if(transA == MorseTrans){
@@ -158,7 +161,6 @@ static void cl_ztrmm_cuda_func(void *descr[], void *cl_arg)
}else{
fprintf(stderr, "Error in cl_ztrmm_cuda_func: bad transA parameter %d\n", transA);
}
- cublasDiagType_t cublasDiag;
if (diag == MorseNonUnit){
cublasDiag = CUBLAS_DIAG_NON_UNIT;
}else if(diag == MorseUnit){
diff --git a/runtime/starpu/codelets/codelet_ztrsm.c b/runtime/starpu/codelets/codelet_ztrsm.c
index cca877b5635e4cf71d6479078011d1f6dd748c35..0b100e81a01d055a030dd5f0be086ff5894d20d6 100644
--- a/runtime/starpu/codelets/codelet_ztrsm.c
+++ b/runtime/starpu/codelets/codelet_ztrsm.c
@@ -111,26 +111,31 @@ static void cl_ztrsm_cuda_func(void *descr[], void *cl_arg)
int lda;
cuDoubleComplex *B;
int ldb;
+ CUstream stream;
+ cublasHandle_t handle;
+ cublasStatus_t stat;
+ cublasSideMode_t cublasSide;
+ cublasFillMode_t cublasUplo;
+ cublasOperation_t cublasTransA;
+ cublasDiagType_t cublasDiag;
A = (const cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[0]);
B = (cuDoubleComplex *)STARPU_MATRIX_GET_PTR(descr[1]);
starpu_codelet_unpack_args(cl_arg, &side, &uplo, &transA, &diag, &m, &n, &alpha, &lda, &ldb);
- cublasHandle_t handle;
- cublasStatus_t stat = cublasCreate(&handle);
+ stat = cublasCreate(&handle);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("CUBLAS initialization failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- CUstream stream = starpu_cuda_get_local_stream();
+ stream = starpu_cuda_get_local_stream();
stat = cublasSetStream(handle, stream);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("cublasSetStream failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- cublasSideMode_t cublasSide;
if (side == MorseLeft){
cublasSide = CUBLAS_SIDE_LEFT;
}else if (side == MorseRight){
@@ -138,7 +143,6 @@ static void cl_ztrsm_cuda_func(void *descr[], void *cl_arg)
}else{
fprintf(stderr, "Error in cl_ztrsm_cuda_func: bad side parameter %d\n", side);
}
- cublasFillMode_t cublasUplo;
if (uplo == MorseUpper){
cublasUplo = CUBLAS_FILL_MODE_UPPER;
}else if(uplo == MorseLower){
@@ -148,7 +152,6 @@ static void cl_ztrsm_cuda_func(void *descr[], void *cl_arg)
}else{
fprintf(stderr, "Error in cl_ztrsm_cuda_func: bad uplo parameter %d\n", uplo);
}
- cublasOperation_t cublasTransA;
if (transA == MorseNoTrans){
cublasTransA = CUBLAS_OP_N;
}else if(transA == MorseTrans){
@@ -158,7 +161,6 @@ static void cl_ztrsm_cuda_func(void *descr[], void *cl_arg)
}else{
fprintf(stderr, "Error in cl_ztrsm_cuda_func: bad transA parameter %d\n", transA);
}
- cublasDiagType_t cublasDiag;
if (diag == MorseNonUnit){
cublasDiag = CUBLAS_DIAG_NON_UNIT;
}else if(diag == MorseUnit){
diff --git a/runtime/starpu/codelets/codelet_ztsmqr.c b/runtime/starpu/codelets/codelet_ztsmqr.c
index 8cfa12803a19361918145499979d0732df07f5bb..e88ca291d05880f79ded733d97daf0c942ef48a8 100644
--- a/runtime/starpu/codelets/codelet_ztsmqr.c
+++ b/runtime/starpu/codelets/codelet_ztsmqr.c
@@ -245,7 +245,12 @@ magma_zparfb_gpu(magma_side_t side, magma_trans_t trans,
magma_trans_t transW;
magma_trans_t transA2;
cublasHandle_t handle;
- cublasStatus_t stat = cublasCreate(&handle);
+ cublasStatus_t stat;
+ cublasOperation_t cublasTrans;
+ cublasOperation_t cublasTransW;
+ cublasOperation_t cublasTransA2;
+
+ stat = cublasCreate(&handle);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("CUBLAS initialization failed\n");
assert( stat == CUBLAS_STATUS_SUCCESS );
@@ -257,7 +262,6 @@ magma_zparfb_gpu(magma_side_t side, magma_trans_t trans,
assert( stat == CUBLAS_STATUS_SUCCESS );
}
- cublasOperation_t cublasTrans;
if (trans == MagmaNoTrans){
cublasTrans = CUBLAS_OP_N;
}else if(trans == MagmaTrans){
@@ -329,7 +333,6 @@ magma_zparfb_gpu(magma_side_t side, magma_trans_t trans,
transW = storev == MorseColumnwise ? MagmaConjTrans : MagmaNoTrans;
transA2 = storev == MorseColumnwise ? MagmaNoTrans : MagmaConjTrans;
- cublasOperation_t cublasTransW;
if (transW == MagmaNoTrans){
cublasTransW = CUBLAS_OP_N;
}else if(transW == MagmaTrans){
@@ -339,7 +342,6 @@ magma_zparfb_gpu(magma_side_t side, magma_trans_t trans,
}else{
fprintf(stderr, "Error in magma_zparfb_gpu: bad transW parameter %d\n", transW);
}
- cublasOperation_t cublasTransA2;
if (transA2 == MagmaNoTrans){
cublasTransA2 = CUBLAS_OP_N;
}else if(transA2 == MagmaTrans){
@@ -358,7 +360,6 @@ magma_zparfb_gpu(magma_side_t side, magma_trans_t trans,
(const cuDoubleComplex *) &zone,
(cuDoubleComplex*)WORK /* K*N1 */, LDWORK);
- WORKC = NULL;
if (WORKC == NULL) {
/* W = op(T) * W */
cublasZtrmm( handle,
@@ -437,7 +438,6 @@ magma_zparfb_gpu(magma_side_t side, magma_trans_t trans,
transW = storev == MorseColumnwise ? MagmaNoTrans : MagmaConjTrans;
transA2 = storev == MorseColumnwise ? MagmaConjTrans : MagmaNoTrans;
- cublasOperation_t cublasTransW;
if (transW == MagmaNoTrans){
cublasTransW = CUBLAS_OP_N;
}else if(transW == MagmaTrans){
@@ -447,7 +447,6 @@ magma_zparfb_gpu(magma_side_t side, magma_trans_t trans,
}else{
fprintf(stderr, "Error in magma_zparfb_gpu: bad transW parameter %d\n", transW);
}
- cublasOperation_t cublasTransA2;
if (transA2 == MagmaNoTrans){
cublasTransA2 = CUBLAS_OP_N;
}else if(transA2 == MagmaTrans){
@@ -466,7 +465,6 @@ magma_zparfb_gpu(magma_side_t side, magma_trans_t trans,
(const cuDoubleComplex *) &zone,
(cuDoubleComplex*)WORK /* M1*K */, LDWORK);
- WORKC = NULL;
if (WORKC == NULL) {
/* W = W * op(T) */
cublasZtrmm( handle,
@@ -548,7 +546,6 @@ magma_zparfb_gpu(magma_side_t side, magma_trans_t trans,
magmaDoubleComplex *WORK, magma_int_t LDWORK,
magmaDoubleComplex *WORKC, magma_int_t LDWORKC,
CUstream stream)
-
{
#if defined(PRECISION_z) || defined(PRECISION_c)
cuDoubleComplex zzero = make_cuDoubleComplex(0.0, 0.0);
diff --git a/runtime/starpu/codelets/codelet_ztsqrt.c b/runtime/starpu/codelets/codelet_ztsqrt.c
index 0aa41ebdfef0fd13ae1834241c2b455179cadd1d..9eedc071b00c1bac821a21741b27df3723c8fc22 100644
--- a/runtime/starpu/codelets/codelet_ztsqrt.c
+++ b/runtime/starpu/codelets/codelet_ztsqrt.c
@@ -213,11 +213,17 @@ magma_ztsqrt2_gpu( magma_int_t m, magma_int_t n, magma_int_t nb,
cublasGetMatrix(nb, nb, sizeof(magmaDoubleComplex),
da1_ref(0, 0), ldda1,
d, ldd);
+// cudaMemcpy( d, da1_ref(0,0),
+// nb*nb*sizeof(cuDoubleComplex),
+// cudaMemcpyDeviceToHost );
/* copy first panel of A2 from device to host: da2 -> a2 */
- cublasGetMatrix(m, nb, sizeof(magmaDoubleComplex),
- da2_ref(0, 0), ldda2,
- a2, lda2);
+// cublasGetMatrix(m, nb, sizeof(magmaDoubleComplex),
+// da2_ref(0, 0), ldda2,
+// a2, lda2);
+ cudaMemcpy( a2, da2_ref(0, 0),
+ m*nb*sizeof(cuDoubleComplex),
+ cudaMemcpyDeviceToHost );
/* This is only blocked code for now */
for (i = 0; i < n; i += nb) {
@@ -234,11 +240,17 @@ magma_ztsqrt2_gpu( magma_int_t m, magma_int_t n, magma_int_t nb,
cublasGetMatrix(ib, ib, sizeof(magmaDoubleComplex),
da1_ref(i, i), ldda1,
d, ldd);
+// cudaMemcpy( d, da1_ref(i,i),
+// ib*ib*sizeof(cuDoubleComplex),
+// cudaMemcpyDeviceToHost );
/* copy panel of A2 from device to host: da2 -> a2 */
cublasGetMatrix(rows, ib, sizeof(magmaDoubleComplex),
da2_ref(0, i), ldda2,
a2, lda2);
+// cudaMemcpy( a2, da2_ref(0,i),
+// rows*ib*sizeof(cuDoubleComplex),
+// cudaMemcpyDeviceToHost );
/* Apply H' to A(i:m,i+2*ib:n) from the left */
cols = n-old_i-2*old_ib;
@@ -268,16 +280,25 @@ magma_ztsqrt2_gpu( magma_int_t m, magma_int_t n, magma_int_t nb,
cublasSetMatrix(rows, ib, sizeof(magmaDoubleComplex),
a2, lda2,
da2_ref(0, i), ldda2);
+// cudaMemcpy( da2_ref(0,i), a2,
+// rows*ib*sizeof(cuDoubleComplex),
+// cudaMemcpyHostToDevice );
/* Send the triangular factor T from hwork to the GPU */
cublasSetMatrix(ib, ib, sizeof(magmaDoubleComplex),
t, ldt,
dt_ref(0, i), lddt);
+// cudaMemcpy( dt_ref(0,i), t,
+// ib*ib*sizeof(cuDoubleComplex),
+// cudaMemcpyHostToDevice );
/* get back the diag tile in A1 from host to device: d -> da1 */
cublasSetMatrix(ib, ib, sizeof(magmaDoubleComplex),
d, ldd,
da1_ref(i, i), ldda1);
+// cudaMemcpy( da1_ref(i, i), d,
+// ib*ib*sizeof(cuDoubleComplex),
+// cudaMemcpyHostToDevice );
/* tsmqr update on one panel forward (look ahead 1) */
if (i + ib < n) {