CFD/hypre
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Parprint; function name change of sync stream

Browse Source
This commit is contained in:
Ruipeng Li 2022-04-06 18:13:49 -07:00
parent 096ce6c1dd
commit 4421ebd83a
10 changed files with 47 additions and 68 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
src/seq_mv/csr_spgemm_device.c
View File

@ -49,7 +49,7 @@ hypreDevice_CSRSpGemm(hypre_CSRMatrix *A,
#endif
#ifdef HYPRE_SPGEMM_TIMING
hypre_ForceSyncCudaComputeStream(hypre_handle());
hypre_ForceSyncComputeStream(hypre_handle());
HYPRE_Real ta = hypre_MPI_Wtime();
#endif
@ -119,9 +119,9 @@ hypreDevice_CSRSpGemm(hypre_CSRMatrix *A,
}
#ifdef HYPRE_SPGEMM_TIMING
hypre_ForceSyncCudaComputeStream(hypre_handle());
hypre_ForceSyncComputeStream(hypre_handle());
HYPRE_Real tb = hypre_MPI_Wtime() - ta;
hypre_printf0("SpGemm time %f\n", tb);
HYPRE_SPGEMM_PRINT("SpGemm time %f\n", tb);
#endif
hypre_CSRMatrixNumNonzeros(C) = nnzC;

4
src/seq_mv/csr_spgemm_device.h
View File

@ -416,9 +416,9 @@ hypre_spgemm_get_num_groups_per_block()
}
#if defined(HYPRE_SPGEMM_PRINTF) || defined(HYPRE_SPGEMM_TIMING)
hypre_int hypre_printf0( const char * format, ... );
#define HYPRE_SPGEMM_PRINT(...) hypre_ParPrintf(hypre_MPI_COMM_WORLD, __VA_ARGS__)
#else
#define hypre_printf0(...)
#define HYPRE_SPGEMM_PRINT(...)
#endif
#endif /* HYPRE_USING_CUDA || defined(HYPRE_USING_HIP) */

14
src/seq_mv/csr_spgemm_device_cusparse.c
View File

@ -111,7 +111,7 @@ hypreDevice_CSRSpGemmCusparseGenericAPI(HYPRE_Int m,
#endif
#ifdef HYPRE_SPGEMM_TIMING
hypre_ForceSyncCudaComputeStream(hypre_handle());
hypre_ForceSyncComputeStream(hypre_handle());
t1 = hypre_MPI_Wtime();
#endif
@ -128,7 +128,7 @@ hypreDevice_CSRSpGemmCusparseGenericAPI(HYPRE_Int m,
spgemmDesc, &bufferSize1, dBuffer1) );
#ifdef HYPRE_SPGEMM_TIMING
hypre_ForceSyncCudaComputeStream(hypre_handle());
hypre_ForceSyncComputeStream(hypre_handle());
t2 = hypre_MPI_Wtime() - t1;
hypre_printf("WorkEst %f\n", t2);
#endif
@ -151,7 +151,7 @@ hypreDevice_CSRSpGemmCusparseGenericAPI(HYPRE_Int m,
spgemmDesc, &bufferSize2, dBuffer2) );
#ifdef HYPRE_SPGEMM_TIMING
hypre_ForceSyncCudaComputeStream(hypre_handle());
hypre_ForceSyncComputeStream(hypre_handle());
t2 = hypre_MPI_Wtime() - t1;
hypre_printf("Compute %f\n", t2);
#endif
@ -185,7 +185,7 @@ hypreDevice_CSRSpGemmCusparseGenericAPI(HYPRE_Int m,
spgemmDesc) );
#ifdef HYPRE_SPGEMM_TIMING
hypre_ForceSyncCudaComputeStream(hypre_handle());
hypre_ForceSyncComputeStream(hypre_handle());
t2 = hypre_MPI_Wtime() - t1;
hypre_printf("Copy %f\n", t2);
#endif
@ -263,7 +263,7 @@ hypreDevice_CSRSpGemmCusparseOldAPI(HYPRE_Int m,
hypre_SortCSRCusparse(k, n, nnzB, descr_B, d_ib, d_jb_sorted, d_b_sorted);
#ifdef HYPRE_SPGEMM_TIMING
hypre_ForceSyncCudaComputeStream(hypre_handle());
hypre_ForceSyncComputeStream(hypre_handle());
t2 = hypre_MPI_Wtime() - t1;
hypre_printf("sort %f\n", t2);
#endif
@ -297,7 +297,7 @@ hypreDevice_CSRSpGemmCusparseOldAPI(HYPRE_Int m,
}
#ifdef HYPRE_SPGEMM_TIMING
hypre_ForceSyncCudaComputeStream(hypre_handle());
hypre_ForceSyncComputeStream(hypre_handle());
t2 = hypre_MPI_Wtime() - t1;
hypre_printf("csrgemmNnz %f\n", t2);
#endif
@ -315,7 +315,7 @@ hypreDevice_CSRSpGemmCusparseOldAPI(HYPRE_Int m,
descr_C, d_c, d_ic, d_jc) );
#ifdef HYPRE_SPGEMM_TIMING
hypre_ForceSyncCudaComputeStream(hypre_handle());
hypre_ForceSyncComputeStream(hypre_handle());
t2 = hypre_MPI_Wtime() - t1;
hypre_printf("csrgemm %f\n", t2);
#endif

18
src/seq_mv/csr_spgemm_device_numblocks.c
View File

@ -85,15 +85,15 @@ HYPRE_Int hypreDevice_CSRSpGemmBinnedGetMaxNumBlocks()
#endif
#if defined(HYPRE_SPGEMM_PRINTF)
hypre_printf0("=======================================================================\n");
hypre_printf0("SM count %d\n", multiProcessorCount);
hypre_printf0("Bin: "); for (HYPRE_Int i = 0; i < num_bins + 1; i++) { hypre_printf0("%5d ", i); } hypre_printf0("\n");
hypre_printf0("-----------------------------------------------------------------------\n");
hypre_printf0("Bdim: "); for (HYPRE_Int i = 0; i < num_bins + 1; i++) { hypre_printf0("%5d ", max_nblocks[2][i]); } hypre_printf0("\n");
hypre_printf0("Sym: "); for (HYPRE_Int i = 0; i < num_bins + 1; i++) { hypre_printf0("%5d ", max_nblocks[0][i]); } hypre_printf0("\n");
hypre_printf0("Bdim: "); for (HYPRE_Int i = 0; i < num_bins + 1; i++) { hypre_printf0("%5d ", max_nblocks[3][i]); } hypre_printf0("\n");
hypre_printf0("Num: "); for (HYPRE_Int i = 0; i < num_bins + 1; i++) { hypre_printf0("%5d ", max_nblocks[1][i]); } hypre_printf0("\n");
hypre_printf0("=======================================================================\n");
HYPRE_SPGEMM_PRINT("=======================================================================\n");
HYPRE_SPGEMM_PRINT("SM count %d\n", multiProcessorCount);
HYPRE_SPGEMM_PRINT("Bin: "); for (HYPRE_Int i = 0; i < num_bins + 1; i++) { HYPRE_SPGEMM_PRINT("%5d ", i); } HYPRE_SPGEMM_PRINT("\n");
HYPRE_SPGEMM_PRINT("-----------------------------------------------------------------------\n");
HYPRE_SPGEMM_PRINT("Bdim: "); for (HYPRE_Int i = 0; i < num_bins + 1; i++) { HYPRE_SPGEMM_PRINT("%5d ", max_nblocks[2][i]); } HYPRE_SPGEMM_PRINT("\n");
HYPRE_SPGEMM_PRINT("Sym: "); for (HYPRE_Int i = 0; i < num_bins + 1; i++) { HYPRE_SPGEMM_PRINT("%5d ", max_nblocks[0][i]); } HYPRE_SPGEMM_PRINT("\n");
HYPRE_SPGEMM_PRINT("Bdim: "); for (HYPRE_Int i = 0; i < num_bins + 1; i++) { HYPRE_SPGEMM_PRINT("%5d ", max_nblocks[3][i]); } HYPRE_SPGEMM_PRINT("\n");
HYPRE_SPGEMM_PRINT("Num: "); for (HYPRE_Int i = 0; i < num_bins + 1; i++) { HYPRE_SPGEMM_PRINT("%5d ", max_nblocks[1][i]); } HYPRE_SPGEMM_PRINT("\n");
HYPRE_SPGEMM_PRINT("=======================================================================\n");
#endif
return hypre_error_flag;

10
src/seq_mv/csr_spgemm_device_numer.c
View File

@ -38,7 +38,7 @@ hypreDevice_CSRSpGemmNumerWithRownnzUpperboundNoBin( HYPRE_Int m,
#ifdef HYPRE_SPGEMM_PRINTF
HYPRE_Int max_rc = HYPRE_THRUST_CALL(reduce, d_rc, d_rc + m, 0, thrust::maximum<HYPRE_Int>());
HYPRE_Int min_rc = HYPRE_THRUST_CALL(reduce, d_rc, d_rc + m, max_rc, thrust::minimum<HYPRE_Int>());
hypre_printf0("%s[%d]: max RC %d, min RC %d\n", __FILE__, __LINE__, max_rc, min_rc);
HYPRE_SPGEMM_PRINT("%s[%d]: max RC %d, min RC %d\n", __FILE__, __LINE__, max_rc, min_rc);
#endif
/* if rc contains exact rownnz: can allocate the final C=(ic,jc,c) directly;
@ -51,7 +51,7 @@ hypreDevice_CSRSpGemmNumerWithRownnzUpperboundNoBin( HYPRE_Int m,
hypre_create_ija(m, NULL, d_rc, d_ic, &d_jc, &d_c, &nnzC);
#ifdef HYPRE_SPGEMM_PRINTF
hypre_printf0("%s[%d]: nnzC %d\n", __FILE__, __LINE__, nnzC);
HYPRE_SPGEMM_PRINT("%s[%d]: nnzC %d\n", __FILE__, __LINE__, nnzC);
#endif
/* even with exact rownnz, still may need global hash, since shared hash is smaller than symbol */
@ -78,7 +78,7 @@ hypreDevice_CSRSpGemmNumerWithRownnzUpperboundNoBin( HYPRE_Int m,
const HYPRE_Int bs = q - p; \
if (bs) \
{ \
hypre_printf0("bin[%d]: %d rows, p %d, q %d\n", BIN, bs, p, q); \
HYPRE_SPGEMM_PRINT("bin[%d]: %d rows, p %d, q %d\n", BIN, bs, p, q); \
hypre_spgemm_numerical_with_rownnz<BIN2, SHMEM_HASH_SIZE, GROUP_SIZE, true> \
(bs, d_rind + p, k, n, GHASH, EXACT_ROWNNZ, d_ia, d_ja, d_a, d_ib, d_jb, d_b, d_rc, d_ic, d_jc, d_c); \
} \
@ -212,9 +212,9 @@ hypreDevice_CSRSpGemmNumerWithRownnzUpperbound( HYPRE_Int m,
}
#ifdef HYPRE_SPGEMM_TIMING
hypre_ForceSyncCudaComputeStream(hypre_handle());
hypre_ForceSyncComputeStream(hypre_handle());
HYPRE_Real t2 = hypre_MPI_Wtime() - t1;
hypre_printf0("SpGemmNumerical time %f\n", t2);
HYPRE_SPGEMM_PRINT("SpGemmNumerical time %f\n", t2);
#endif
#ifdef HYPRE_SPGEMM_NVTX

6
src/seq_mv/csr_spgemm_device_numer.h
View File

@ -443,11 +443,11 @@ hypre_spgemm_numerical_with_rownnz( HYPRE_Int m,
}
#ifdef HYPRE_SPGEMM_PRINTF
hypre_printf0("%s[%d], BIN[%d]: m %d k %d n %d, HASH %c, SHMEM_HASH_SIZE %d, GROUP_SIZE %d, "
HYPRE_SPGEMM_PRINT("%s[%d], BIN[%d]: m %d k %d n %d, HASH %c, SHMEM_HASH_SIZE %d, GROUP_SIZE %d, "
"exact_rownnz %d, need_ghash %d, ghash %p size %d\n",
__FILE__, __LINE__, BIN, m, k, n,
HASH_TYPE, SHMEM_HASH_SIZE, GROUP_SIZE, exact_rownnz, need_ghash, d_ghash_i, ghash_size);
hypre_printf0("kernel spec [%d %d %d] x [%d %d %d]\n", gDim.x, gDim.y, gDim.z, bDim.x, bDim.y, bDim.z);
HYPRE_SPGEMM_PRINT("kernel spec [%d %d %d] x [%d %d %d]\n", gDim.x, gDim.y, gDim.z, bDim.x, bDim.y, bDim.z);
#endif
#if defined(HYPRE_SPGEMM_DEVICE_USE_DSHMEM)
@ -579,7 +579,7 @@ hypreDevice_CSRSpGemmNumerPostCopy( HYPRE_Int m,
hypre_create_ija(m, NULL, d_rc, d_ic_new, &d_jc_new, &d_c_new, &nnzC_new);
#ifdef HYPRE_SPGEMM_PRINTF
hypre_printf0("%s[%d]: Post Copy: new nnzC %d\n", __FILE__, __LINE__, nnzC_new);
HYPRE_SPGEMM_PRINT("%s[%d]: Post Copy: new nnzC %d\n", __FILE__, __LINE__, nnzC_new);
#endif
/* copy to the final C */

8
src/seq_mv/csr_spgemm_device_rowest.c
View File

@ -291,9 +291,9 @@ void csr_spmm_rownnz_cohen(HYPRE_Int M, HYPRE_Int K, HYPRE_Int N, HYPRE_Int *d_i
hypre_CurandUniformSingle(nsamples * N, d_V1, 0, 0, 0, 0);
#ifdef HYPRE_SPGEMM_TIMING
hypre_ForceSyncCudaComputeStream(hypre_handle());
hypre_ForceSyncComputeStream(hypre_handle());
t2 = hypre_MPI_Wtime() - t1;
hypre_printf0("Curand time %f\n", t2);
HYPRE_SPGEMM_PRINT("Curand time %f\n", t2);
#endif
dim3 gDim( (nsamples * N + bDim.z * HYPRE_WARP_SIZE - 1) / (bDim.z * HYPRE_WARP_SIZE) );
@ -407,9 +407,9 @@ hypreDevice_CSRSpGemmRownnzEstimate( HYPRE_Int m,
}
#ifdef HYPRE_SPGEMM_TIMING
hypre_ForceSyncCudaComputeStream(hypre_handle());
hypre_ForceSyncComputeStream(hypre_handle());
HYPRE_Real t2 = hypre_MPI_Wtime() - t1;
hypre_printf0("RownnzEst time %f\n", t2);
HYPRE_SPGEMM_PRINT("RownnzEst time %f\n", t2);
#endif
#ifdef HYPRE_PROFILE

18
src/seq_mv/csr_spgemm_device_symbl.c
View File

@ -17,7 +17,7 @@
const HYPRE_Int bs = q - p; \
if (bs) \
{ \
hypre_printf0("bin[%d]: %d rows, p %d, q %d\n", BIN, bs, p, q); \
HYPRE_SPGEMM_PRINT("bin[%d]: %d rows, p %d, q %d\n", BIN, bs, p, q); \
hypre_spgemm_symbolic_rownnz<BIN, SHMEM_HASH_SIZE, GROUP_SIZE, true> \
( bs, d_rind + p, k, n, GHASH, d_ia, d_ja, d_ib, d_jb, d_rc, CAN_FAIL, RF ); \
} \
@ -143,9 +143,9 @@ hypreDevice_CSRSpGemmRownnzUpperbound( HYPRE_Int m,
hypre_TFree(d_rf, HYPRE_MEMORY_DEVICE);
#ifdef HYPRE_SPGEMM_TIMING
hypre_ForceSyncCudaComputeStream(hypre_handle());
hypre_ForceSyncComputeStream(hypre_handle());
HYPRE_Real t2 = hypre_MPI_Wtime() - t1;
hypre_printf0("RownnzBound time %f\n", t2);
HYPRE_SPGEMM_PRINT("RownnzBound time %f\n", t2);
#endif
#ifdef HYPRE_SPGEMM_NVTX
@ -197,7 +197,7 @@ hypreDevice_CSRSpGemmRownnzNoBin( HYPRE_Int m,
if (num_failed_rows)
{
#ifdef HYPRE_SPGEMM_PRINTF
hypre_printf0("[%s, %d]: num of failed rows %d (%.2f)\n", __FILE__, __LINE__,
HYPRE_SPGEMM_PRINT("[%s, %d]: num of failed rows %d (%.2f)\n", __FILE__, __LINE__,
num_failed_rows, num_failed_rows / (m + 0.0) );
#endif
HYPRE_Int *d_rind = hypre_TAlloc(HYPRE_Int, num_failed_rows, HYPRE_MEMORY_DEVICE);
@ -257,9 +257,9 @@ hypreDevice_CSRSpGemmRownnzBinned( HYPRE_Int m,
hypre_CSRMatrixIntSpMVDevice(m, nnzA, 1, d_ia, d_ja, NULL, d_rind + 1, 0, d_rc);
#ifdef HYPRE_SPGEMM_TIMING
hypre_ForceSyncCudaComputeStream(hypre_handle());
hypre_ForceSyncComputeStream(hypre_handle());
HYPRE_Real t2 = hypre_MPI_Wtime() - t1;
hypre_printf0("RownnzEst time %f\n", t2);
HYPRE_SPGEMM_PRINT("RownnzEst time %f\n", t2);
#endif
#endif
@ -289,7 +289,7 @@ hypreDevice_CSRSpGemmRownnzBinned( HYPRE_Int m,
if (num_failed_rows)
{
#ifdef HYPRE_SPGEMM_PRINTF
hypre_printf0("[%s, %d]: num of failed rows %d (%.2f)\n", __FILE__, __LINE__,
HYPRE_SPGEMM_PRINT("[%s, %d]: num of failed rows %d (%.2f)\n", __FILE__, __LINE__,
num_failed_rows, num_failed_rows / (m + 0.0) );
#endif
HYPRE_Int *new_end =
@ -365,9 +365,9 @@ hypreDevice_CSRSpGemmRownnz( HYPRE_Int m,
}
#ifdef HYPRE_SPGEMM_TIMING
hypre_ForceSyncCudaComputeStream(hypre_handle());
hypre_ForceSyncComputeStream(hypre_handle());
HYPRE_Real t2 = hypre_MPI_Wtime() - t1;
hypre_printf0("Rownnz time %f\n", t2);
HYPRE_SPGEMM_PRINT("Rownnz time %f\n", t2);
#endif
#ifdef HYPRE_SPGEMM_NVTX

4
src/seq_mv/csr_spgemm_device_symbl.h
View File

@ -365,11 +365,11 @@ hypre_spgemm_symbolic_rownnz( HYPRE_Int m,
}
#ifdef HYPRE_SPGEMM_PRINTF
hypre_printf0("%s[%d], BIN[%d]: m %d k %d n %d, HASH %c, SHMEM_HASH_SIZE %d, GROUP_SIZE %d, "
HYPRE_SPGEMM_PRINT("%s[%d], BIN[%d]: m %d k %d n %d, HASH %c, SHMEM_HASH_SIZE %d, GROUP_SIZE %d, "
"can_fail %d, need_ghash %d, ghash %p size %d\n",
__FILE__, __LINE__, BIN, m, k, n,
HASH_TYPE, SHMEM_HASH_SIZE, GROUP_SIZE, can_fail, need_ghash, d_ghash_i, ghash_size);
hypre_printf0("kernel spec [%d %d %d] x [%d %d %d]\n", gDim.x, gDim.y, gDim.z, bDim.x, bDim.y, bDim.z);
HYPRE_SPGEMM_PRINT("kernel spec [%d %d %d] x [%d %d %d]\n", gDim.x, gDim.y, gDim.z, bDim.x, bDim.y, bDim.z);
#endif
#if defined(HYPRE_SPGEMM_DEVICE_USE_DSHMEM)

27
src/seq_mv/csr_spgemm_device_util.c
View File

@ -226,7 +226,7 @@ HYPRE_Int hypre_SpGemmCreateBins( HYPRE_Int m,
HYPRE_Int *h_bin_ptr )
{
#ifdef HYPRE_SPGEMM_TIMING
hypre_ForceSyncCudaComputeStream(hypre_handle());
hypre_ForceSyncComputeStream(hypre_handle());
HYPRE_Real t1 = hypre_MPI_Wtime();
#endif
@ -265,34 +265,13 @@ HYPRE_Int hypre_SpGemmCreateBins( HYPRE_Int m,
hypre_TFree(d_bin_ptr, HYPRE_MEMORY_DEVICE);
#ifdef HYPRE_SPGEMM_TIMING
hypre_ForceSyncCudaComputeStream(hypre_handle());
hypre_ForceSyncComputeStream(hypre_handle());
HYPRE_Real t2 = hypre_MPI_Wtime() - t1;
hypre_printf0("%s[%d]: Binning time %f\n", __FILE__, __LINE__, t2);
HYPRE_SPGEMM_PRINT("%s[%d]: Binning time %f\n", __FILE__, __LINE__, t2);
#endif
return hypre_error_flag;
}
#if defined(HYPRE_SPGEMM_PRINTF) || defined(HYPRE_SPGEMM_TIMING)
hypre_int hypre_printf0( const char * format, ... )
{
HYPRE_Int myid;
hypre_MPI_Comm_rank(hypre_MPI_COMM_WORLD, &myid );
hypre_int ret = 0;
if (!myid)
{
va_list args;
va_start (args, format);
ret = vprintf (format, args);
va_end (args);
}
fflush(stdout);
return ret;
}
#endif
#endif