From 52199731a59f309e4fbf2e8117c39f7483b1c1cf Mon Sep 17 00:00:00 2001
From: "Kurt A. O'Hearn" <ohearnku@msu.edu>
Date: Fri, 1 Jul 2016 10:01:00 -0400
Subject: [PATCH] Update CUDA block calculations.
---
PG-PuReMD/src/cuda_utils.cu | 58 ++++++++++++++++++++++++++-----------
PG-PuReMD/src/reax_types.h | 40 +++++++++++++------------
2 files changed, 63 insertions(+), 35 deletions(-)
diff --git a/PG-PuReMD/src/cuda_utils.cu b/PG-PuReMD/src/cuda_utils.cu
index dcd8d61f..1ab8705b 100644
--- a/PG-PuReMD/src/cuda_utils.cu
+++ b/PG-PuReMD/src/cuda_utils.cu
@@ -1,18 +1,23 @@
#include "cuda_utils.h"
-extern "C" void cuda_malloc (void **ptr, int size, int memset, char *msg) {
+
+extern "C" void cuda_malloc (void **ptr, int size, int mem_set, char *msg) {
cudaError_t retVal = cudaSuccess;
retVal = cudaMalloc (ptr, size);
- if (retVal != cudaSuccess) {
+
+ if (retVal != cudaSuccess)
+ {
fprintf (stderr, "Failed to allocate memory on device for the res: %s... exiting with code: %d size: %d \n",
msg, retVal, size);
exit (-1);
}
- if (memset) {
+ if (mem_set)
+ {
retVal = cudaMemset (*ptr, 0, size);
+
if (retVal != cudaSuccess) {
fprintf (stderr, "Failed to memset memory on device for resource %s\n",
msg);
@@ -21,50 +26,66 @@ extern "C" void cuda_malloc (void **ptr, int size, int memset, char *msg) {
}
}
+
extern "C" void cuda_free (void *ptr, char *msg) {
cudaError_t retVal = cudaSuccess;
- if (!ptr) return;
+
+ if ( !ptr )
+ {
+ return;
+ }
retVal = cudaFree (ptr);
- if (retVal != cudaSuccess) {
+ if (retVal != cudaSuccess)
+ {
fprintf (stderr, "Failed to release memory on device for res %s... exiting with code %d -- Address %ld\n",
msg, retVal, (long int) ptr);
return;
}
}
+
extern "C" void cuda_memset (void *ptr, int data, size_t count, char *msg){
cudaError_t retVal = cudaSuccess;
retVal = cudaMemset (ptr, data, count);
- if (retVal != cudaSuccess) {
+
+ if (retVal != cudaSuccess)
+ {
fprintf (stderr, "Failed to memset memory on device for %s, cuda code %d\n",
msg, retVal);
exit (-1);
}
}
+
extern "C" void copy_host_device (void *host, void *dev, int size, enum cudaMemcpyKind dir, char *msg)
{
- cudaError_t retVal = cudaErrorNotReady;
+ cudaError_t retVal = cudaErrorNotReady;
if (dir == cudaMemcpyHostToDevice)
+ {
retVal = cudaMemcpy (dev, host, size, cudaMemcpyHostToDevice);
+ }
else
+ {
retVal = cudaMemcpy (host, dev, size, cudaMemcpyDeviceToHost);
+ }
- if (retVal != cudaSuccess) {
+ if (retVal != cudaSuccess)
+ {
fprintf (stderr, "could not copy resource %s from host to device: reason %d \n",
msg, retVal);
exit (-1);
}
}
+
extern "C" void copy_device (void *dest, void *src, int size, char *msg)
{
- cudaError_t retVal = cudaErrorNotReady;
+ cudaError_t retVal = cudaErrorNotReady;
retVal = cudaMemcpy (dest, src, size, cudaMemcpyDeviceToDevice);
if (retVal != cudaSuccess) {
@@ -74,31 +95,33 @@ extern "C" void copy_device (void *dest, void *src, int size, char *msg)
}
}
+
extern "C" void compute_blocks ( int *blocks, int *block_size, int count )
{
*block_size = CUDA_BLOCK_SIZE;
- *blocks = (count / CUDA_BLOCK_SIZE ) + (count % CUDA_BLOCK_SIZE == 0 ? 0 : 1);
+ *blocks = (int) CEIL((double) count / CUDA_BLOCK_SIZE);
}
+
extern "C" void compute_matvec_blocks ( int *blocks, int count )
{
- *blocks = ((count * MATVEC_KER_THREADS_PER_ROW) / MATVEC_BLOCK_SIZE) +
- (((count * MATVEC_KER_THREADS_PER_ROW) % MATVEC_BLOCK_SIZE) == 0 ? 0 : 1);
+ *blocks = (int) CEIL((double) count * MATVEC_KER_THREADS_PER_ROW / MATVEC_BLOCK_SIZE);
}
+
extern "C" void compute_nearest_pow_2 (int blocks, int *result)
{
- int power = 1;
- while (power < blocks) power *= 2;
-
- *result = power;
+ *result = (int) EXP2( CEIL( LOG2((double) blocks) ) );
}
+
void print_info ()
{
size_t total, free;
+
cudaMemGetInfo (&free, &total);
- if (cudaGetLastError () != cudaSuccess )
+
+ if ( cudaGetLastError () != cudaSuccess )
{
fprintf (stderr, "Error on the memory call \n");
return;
@@ -109,6 +132,7 @@ void print_info ()
free, free/(1024*1024), free/ (1024*1024*1024) );
}
+
extern "C" void print_device_mem_usage ()
{
print_info ();
diff --git a/PG-PuReMD/src/reax_types.h b/PG-PuReMD/src/reax_types.h
index d27c5539..504d8c21 100644
--- a/PG-PuReMD/src/reax_types.h
+++ b/PG-PuReMD/src/reax_types.h
@@ -85,12 +85,16 @@
#define FALSE 0
#define EXP exp
+#define EXP2 exp2
+#define LOG2 log2
#define SQRT sqrt
#define POW pow
#define ACOS acos
#define COS cos
#define SIN sin
#define TAN tan
+#define CEIL ceil
+#define FLOOR floor
#define SQR(x) ((x)*(x))
#define CUBE(x) ((x)*(x)*(x))
@@ -239,24 +243,24 @@
/******************* ENUMERATIONS *************************/
-enum geo_formats { CUSTOM, PDB, ASCII_RESTART, BINARY_RESTART, GF_N };
-
-enum restart_formats { WRITE_ASCII, WRITE_BINARY, RF_N };
-
-enum ensembles { NVE, bNVT, nhNVT, sNPT, iNPT, NPT, ens_N };
-
-enum lists { BONDS, OLD_BONDS, THREE_BODIES,
- HBONDS, FAR_NBRS, DBOS, DDELTAS, LIST_N
+enum geo_formats { CUSTOM = 0, PDB = 1, ASCII_RESTART = 2, BINARY_RESTART = 3, GF_N = 4 };
+
+enum restart_formats { WRITE_ASCII = 0, WRITE_BINARY = 1, RF_N = 2 };
+
+enum ensembles { NVE = 0, bNVT = 1, nhNVT = 2, sNPT = 3, iNPT = 4, NPT = 5, ens_N = 6 };
+
+enum lists { BONDS = 0, OLD_BONDS = 1, THREE_BODIES = 2,
+ HBONDS = 3, FAR_NBRS = 4, DBOS = 5, DDELTAS = 6, LIST_N = 7
};
-enum interactions { TYP_VOID, TYP_BOND, TYP_THREE_BODY,
- TYP_HBOND, TYP_FAR_NEIGHBOR, TYP_DBO, TYP_DDELTA, TYP_N
+enum interactions { TYP_VOID = 0, TYP_BOND = 1, TYP_THREE_BODY = 2,
+ TYP_HBOND = 3, TYP_FAR_NEIGHBOR = 4, TYP_DBO = 5, TYP_DDELTA = 6, TYP_N = 7
};
-enum message_tags { INIT, UPDATE, BNDRY, UPDATE_BNDRY,
- EXC_VEC1, EXC_VEC2, DIST_RVEC2, COLL_RVEC2,
- DIST_RVECS, COLL_RVECS, INIT_DESCS, ATOM_LINES,
- BOND_LINES, ANGLE_LINES, RESTART_ATOMS, TAGS_N
+enum message_tags { INIT = 0, UPDATE = 1, BNDRY = 2, UPDATE_BNDRY = 3,
+ EXC_VEC1 = 4, EXC_VEC2 = 5, DIST_RVEC2 = 6, COLL_RVEC2 = 7,
+ DIST_RVECS = 8, COLL_RVECS = 9, INIT_DESCS = 10, ATOM_LINES = 11,
+ BOND_LINES = 12, ANGLE_LINES = 13, RESTART_ATOMS = 14, TAGS_N = 15
};
enum errors { FILE_NOT_FOUND = -10, UNKNOWN_ATOM_TYPE = -11,
@@ -265,7 +269,7 @@ enum errors { FILE_NOT_FOUND = -10, UNKNOWN_ATOM_TYPE = -11,
INVALID_INPUT = -16, INVALID_GEO = -17
};
-enum exchanges { NONE, NEAR_EXCH, FULL_EXCH };
+enum exchanges { NONE = 0, NEAR_EXCH = 1, FULL_EXCH = 2 };
enum gcell_types { NO_NBRS = 0, NEAR_ONLY = 1, HBOND_ONLY = 2, FAR_ONLY = 4,
NEAR_HBOND = 3, NEAR_FAR = 5, HBOND_FAR = 6, FULL_NBRS = 7,
@@ -276,11 +280,11 @@ enum atoms { C_ATOM = 0, H_ATOM = 1, O_ATOM = 2, N_ATOM = 3,
S_ATOM = 4, SI_ATOM = 5, GE_ATOM = 6, X_ATOM = 7
};
-enum traj_methods { REG_TRAJ, MPI_TRAJ, TF_N };
+enum traj_methods { REG_TRAJ = 0, MPI_TRAJ = 1, TF_N = 2 };
-enum molecules { UNKNOWN, WATER };
+enum molecules { UNKNOWN = 0, WATER = 1 };
-enum list_on { TYP_HOST, TYP_DEVICE };
+enum list_on { TYP_HOST = 0, TYP_DEVICE = 1 };
--
GitLab