Menu
▾
▴
[X10-commits] SF.net SVN: x10:[12118] trunk
|
From: <spa...@us...> - 2009-11-18 13:53:06
|
Revision: 12118
http://x10.svn.sourceforge.net/x10/?rev=12118&view=rev
Author: sparksparkspark
Date: 2009-11-18 13:52:57 +0000 (Wed, 18 Nov 2009)
Log Message:
-----------
1) Ignore wget failure when building x10rt if there is a tar already present
2) Implement automatic blocks/threads detection at x10rt level
Modified Paths:
--------------
trunk/x10.dist/samples/KMeansCUDA.x10
trunk/x10.runtime/src-cpp/x10aux/network.cc
trunk/x10.runtime/src-cpp/x10aux/network.h
trunk/x10.runtime/src-cpp/x10rt/common/x10rt_cuda.cc
trunk/x10.runtime/src-cpp/x10rt/common/x10rt_front.cc
trunk/x10.runtime/src-cpp/x10rt/common/x10rt_logical.cc
trunk/x10.runtime/src-cpp/x10rt/include/x10rt_cuda.h
trunk/x10.runtime/src-cpp/x10rt/include/x10rt_front.h
trunk/x10.runtime/src-cpp/x10rt/include/x10rt_logical.h
trunk/x10.runtime/src-cpp/x10rt/pgas/pgas.mk
Modified: trunk/x10.dist/samples/KMeansCUDA.x10
===================================================================
--- trunk/x10.dist/samples/KMeansCUDA.x10 2009-11-18 10:41:46 UTC (rev 12117)
+++ trunk/x10.dist/samples/KMeansCUDA.x10 2009-11-18 13:52:57 UTC (rev 12118)
@@ -80,7 +80,7 @@
val kernel_start_time = System.currentTimeMillis();
// classify kernel
- val blocks = 8, threads = 64;
+ val blocks = 4, threads = 256;
at (gpu) @CUDA {
for ((block) in 0..blocks-1) {
val clustercache = Rail.make[Float](num_clusters*4, clusters_copy);
Modified: trunk/x10.runtime/src-cpp/x10aux/network.cc
===================================================================
--- trunk/x10.runtime/src-cpp/x10aux/network.cc 2009-11-18 10:41:46 UTC (rev 12117)
+++ trunk/x10.runtime/src-cpp/x10aux/network.cc 2009-11-18 13:52:57 UTC (rev 12118)
@@ -31,6 +31,64 @@
volatile x10_long x10aux::serialized_bytes = 0;
volatile x10_long x10aux::deserialized_bytes = 0;
+
+const int cuda_kernel_cfgs[] = {
+ /*1024*/ 8, 128,
+ /*1024*/ 4, 256,
+ /*1024*/ 2, 512,
+
+ /*960*/ 5, 192,
+ /*960*/ 3, 320,
+
+ /*896*/ 7, 128,
+ /*896*/ 2, 448,
+
+ /*768*/ 6, 128,
+ /*768*/ 4, 192,
+ /*768*/ 3, 256,
+ /*768*/ 2, 384,
+
+ /*640*/ 5, 128,
+ /*640*/ 2, 320,
+
+ /*576*/ 3, 192,
+
+ /*512*/ 8, 64,
+ /*512*/ 4, 128,
+ /*512*/ 2, 256,
+ /*512*/ 1, 512,
+
+ /*448*/ 7, 64,
+ /*448*/ 1, 448,
+
+ /*384*/ 6, 64,
+ /*384*/ 3, 128,
+ /*384*/ 2, 192,
+ /*384*/ 1, 384,
+
+ /*320*/ 5, 64,
+ /*320*/ 1, 320,
+
+ /*256*/ 4, 64,
+ /*256*/ 2, 128,
+ /*256*/ 1, 256,
+
+ /*192*/ 3, 64,
+ /*192*/ 1, 192,
+
+ /*128*/ 2, 64,
+ /*128*/ 1, 128,
+
+ /*64*/ 1, 64,
+
+ 0 /* terminator */
+};
+
+void x10aux::blocks_threads (place p, msg_type t, int shm, int &bs, int &ts, const int *cfgs)
+{ x10rt_blocks_threads(p,t,shm,bs,ts,cfgs); }
+
+
+
void *kernel_put_finder (const x10rt_msg_params &p, x10rt_copy_sz)
{
x10aux::deserialization_buffer buf(static_cast<char*>(p.msg));
Modified: trunk/x10.runtime/src-cpp/x10aux/network.h
===================================================================
--- trunk/x10.runtime/src-cpp/x10aux/network.h 2009-11-18 10:41:46 UTC (rev 12117)
+++ trunk/x10.runtime/src-cpp/x10aux/network.h 2009-11-18 13:52:57 UTC (rev 12118)
@@ -36,7 +36,10 @@
inline x10_boolean is_cuda (place p) { return x10rt_is_cuda(p); }
inline void event_probe (void) { x10rt_probe(); }
+ extern const int cuda_cfgs[];
+ void blocks_threads (place p, msg_type t, int shm, int &bs, int &ts, const int *cfgs=cuda_cfgs);
+
inline x10_ulong remote_alloc (place p, size_t sz) {
_X_(ANSI_BOLD<<ANSI_X10RT<<"Remote alloc: "<<ANSI_RESET
<<"size "<<sz<<" to place: "<<p);
Modified: trunk/x10.runtime/src-cpp/x10rt/common/x10rt_cuda.cc
===================================================================
--- trunk/x10.runtime/src-cpp/x10rt/common/x10rt_cuda.cc 2009-11-18 10:41:46 UTC (rev 12117)
+++ trunk/x10.runtime/src-cpp/x10rt/common/x10rt_cuda.cc 2009-11-18 13:52:57 UTC (rev 12118)
@@ -238,8 +238,9 @@
};
};
- static void ensure_initialized (void)
+ void ensure_initialized (void)
{
+ // only do once per process
static int done = 0;
if (!done) {
CU_SAFE(cuInit(0));
@@ -247,6 +248,16 @@
}
}
+ bool stream_ready (CUstream s)
+ {
+ CUresult r = cuStreamQuery(s);
+ if (r==CUDA_ERROR_NOT_READY) return false;
+ CU_SAFE(r);
+ return true;
+ }
+
+ int round_up (int x, int y) { return (x + (y-1)) / y * y; }
+
}
struct x10rt_cuda_ctx {
@@ -509,20 +520,20 @@
pthread_mutex_lock(&big_lock_of_doom);
if (ctx->cbs.arrc <= p.type) {
- fprintf(stderr,"X10RT: Kernel %llu is invalid.\n", (unsigned long long)p.type);
+ fprintf(stderr,"X10RT: async %lu is invalid.\n", (unsigned long)p.type);
abort();
}
- if (ctx->cbs[p.type].kernel_cbs.pre == NULL) {
- fprintf(stderr,"X10RT: Kernel %llu has no 'pre' registered.\n", (unsigned long long)p.type);
- abort();
- }
if (ctx->cbs[p.type].kernel_cbs.kernel == NULL) {
- fprintf(stderr,"X10RT: Kernel %llu has no kernel registered.\n",(unsigned long long)p.type);
+ fprintf(stderr,"X10RT: async %lu is not a CUDA kernel.\n",(unsigned long)p.type);
abort();
}
+ if (ctx->cbs[p.type].kernel_cbs.pre == NULL) {
+ fprintf(stderr,"X10RT: CUDA Kernel %lu has no 'pre' registered.\n", (unsigned long)p.type);
+ abort();
+ }
if (ctx->cbs[p.type].kernel_cbs.post == NULL) {
- fprintf(stderr,"X10RT: Kernel %llu has no 'post' registered.\n",(unsigned long long)p.type);
+ fprintf(stderr,"X10RT: CUDA Kernel %lu has no 'post' registered.\n",(unsigned long)p.type);
abort();
}
@@ -539,15 +550,69 @@
}
+
+void x10rt_cuda_blocks_threads (x10rt_cuda_ctx *ctx, x10rt_msg_type type, int dyn_shm,
+ int &blocks, int &threads, const int *cfg)
+{
#ifdef ENABLE_CUDA
-static bool stream_ready (CUstream s)
-{
- CUresult r = cuStreamQuery(s);
- if (r==CUDA_ERROR_NOT_READY) return false;
- CU_SAFE(r);
- return true;
+ if (ctx->cbs.arrc <= type) {
+ fprintf(stderr,"X10RT: async %lu is invalid.\n", (unsigned long)type);
+ abort();
+ }
+ if (ctx->cbs[type].kernel_cbs.kernel == NULL) {
+ fprintf(stderr,"X10RT: async %lu is not a CUDA kernel.\n",(unsigned long)type);
+ abort();
+ }
+ CUfunction k = ctx->cbs[type].kernel_cbs.kernel;
+
+ pthread_mutex_lock(&big_lock_of_doom);
+ CU_SAFE(cuCtxPushCurrent(ctx->ctx));
+
+ int mps, max_shm;
+ CU_SAFE(cuDeviceGetAttribute(&mps, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, ctx->hw));
+ CU_SAFE(cuDeviceGetAttribute(&max_shm,CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK,ctx->hw));
+
+ CUdevprop prop;
+ CU_SAFE(cuDeviceGetProperties(&prop, ctx->hw));
+ int max_regs = prop.regsPerBlock;
+
+ int major, minor;
+ CU_SAFE(cuDeviceComputeCapability(&major, &minor, ctx->hw));
+
+
+ int static_shm, regs;
+ CU_SAFE(cuFuncGetAttribute(&static_shm, CU_FUNC_ATTRIBUTE_SHARED_SIZE_BYTES, k));
+ CU_SAFE(cuFuncGetAttribute(®s, CU_FUNC_ATTRIBUTE_NUM_REGS, k));
+
+ CU_SAFE(cuCtxPopCurrent(NULL));
+ pthread_mutex_unlock(&big_lock_of_doom);
+
+ // round up to 512 bytes (the granularity of shm allocation)
+ int shm = round_up(dyn_shm + static_shm, 512);
+
+ int alloc_size = (minor>=2) ? 512 : 256;
+ int max_threads = (minor>=2) ? 1024 : 768;
+
+ while (*cfg) {
+ int b = *(cfg++);
+ int t = *(cfg++);
+ if (b*shm > max_shm) continue;
+ if (t*b > max_threads) continue;
+ int block_regs = round_up(regs*round_up(t,64), alloc_size);
+ if (b*block_regs > max_regs) continue;
+ blocks = b * mps;
+ threads = t;
+ return;
+ }
+
+ blocks = 0;
+ threads = 0;
+
+#else
+ (void)ctx; (void)msg_type; (void)dyn_shm; (void)blocks; (void)threads; (void)cfg;
+ abort();
+#endif
}
-#endif
void x10rt_cuda_probe (x10rt_cuda_ctx *ctx)
Modified: trunk/x10.runtime/src-cpp/x10rt/common/x10rt_front.cc
===================================================================
--- trunk/x10.runtime/src-cpp/x10rt/common/x10rt_front.cc 2009-11-18 10:41:46 UTC (rev 12117)
+++ trunk/x10.runtime/src-cpp/x10rt/common/x10rt_front.cc 2009-11-18 13:52:57 UTC (rev 12118)
@@ -96,6 +96,11 @@
{ x10rt_lgl_remote_op_fence(); }
+void x10rt_blocks_threads (x10rt_place d, x10rt_msg_type type, int dyn_shm,
+ int &blocks, int &threads, const int *cfg)
+{ x10rt_lgl_blocks_threads (d, type, dyn_shm, blocks, threads, cfg); }
+
+
void x10rt_probe (void)
{ x10rt_lgl_probe(); }
Modified: trunk/x10.runtime/src-cpp/x10rt/common/x10rt_logical.cc
===================================================================
--- trunk/x10.runtime/src-cpp/x10rt/common/x10rt_logical.cc 2009-11-18 10:41:46 UTC (rev 12117)
+++ trunk/x10.runtime/src-cpp/x10rt/common/x10rt_logical.cc 2009-11-18 13:52:57 UTC (rev 12118)
@@ -705,6 +705,36 @@
x10rt_net_remote_op_fence();
}
+void x10rt_lgl_blocks_threads (x10rt_place d, x10rt_msg_type type, int dyn_shm,
+ int &blocks, int &threads, const int *cfg)
+{
+ assert(d < x10rt_lgl_nplaces());
+
+ if (d < x10rt_lgl_nhosts()) {
+ blocks = 8; threads = 1;
+ } else if (x10rt_lgl_parent(d) == x10rt_lgl_here()) {
+ // local accelerator
+ switch (x10rt_lgl_type(d)) {
+ case X10RT_LGL_CUDA: {
+ x10rt_cuda_ctx *cctx = static_cast<x10rt_cuda_ctx*>(g.accel_ctxs[g.index[d]]);
+ x10rt_cuda_blocks_threads(cctx, type, dyn_shm, blocks, threads, cfg);
+ } break;
+ case X10RT_LGL_SPE: {
+ blocks = 8; threads = 1;
+ } break;
+ default: {
+ fprintf(stderr,"Place %lu has invalid type %d in remote_op_xor.\n",
+ d, x10rt_lgl_type(d));
+ abort();
+ }
+ }
+ } else {
+ fprintf(stderr,"Routing of remote ops still unsupported.\n");
+ abort();
+ }
+}
+
+
void x10rt_lgl_probe (void)
{
x10rt_net_probe();
Modified: trunk/x10.runtime/src-cpp/x10rt/include/x10rt_cuda.h
===================================================================
--- trunk/x10.runtime/src-cpp/x10rt/include/x10rt_cuda.h 2009-11-18 10:41:46 UTC (rev 12117)
+++ trunk/x10.runtime/src-cpp/x10rt/include/x10rt_cuda.h 2009-11-18 13:52:57 UTC (rev 12118)
@@ -28,6 +28,9 @@
void x10rt_cuda_send_put (x10rt_cuda_ctx *ctx, x10rt_msg_params &, void *buf, x10rt_copy_sz len);
+void x10rt_cuda_blocks_threads (x10rt_cuda_ctx *ctx, x10rt_msg_type type, int dyn_shm,
+ int &blocks, int &threads, const int *cfg);
+
void *x10rt_cuda_device_alloc (x10rt_cuda_ctx *ctx, size_t sz);
void x10rt_cuda_device_free (x10rt_cuda_ctx *ctx, void *ptr);
Modified: trunk/x10.runtime/src-cpp/x10rt/include/x10rt_front.h
===================================================================
--- trunk/x10.runtime/src-cpp/x10rt/include/x10rt_front.h 2009-11-18 10:41:46 UTC (rev 12117)
+++ trunk/x10.runtime/src-cpp/x10rt/include/x10rt_front.h 2009-11-18 13:52:57 UTC (rev 12118)
@@ -43,6 +43,9 @@
void x10rt_remote_op_fence (void);
+void x10rt_blocks_threads (x10rt_place d, x10rt_msg_type type, int dyn_shm,
+ int &blocks, int &threads, const int *cfg);
+
void x10rt_probe (void);
void x10rt_finalize (void);
Modified: trunk/x10.runtime/src-cpp/x10rt/include/x10rt_logical.h
===================================================================
--- trunk/x10.runtime/src-cpp/x10rt/include/x10rt_logical.h 2009-11-18 10:41:46 UTC (rev 12117)
+++ trunk/x10.runtime/src-cpp/x10rt/include/x10rt_logical.h 2009-11-18 13:52:57 UTC (rev 12118)
@@ -58,6 +58,9 @@
void x10rt_lgl_remote_op_fence (void);
+void x10rt_lgl_blocks_threads (x10rt_place d, x10rt_msg_type type, int dyn_shm,
+ int &blocks, int &threads, const int *cfg);
+
void x10rt_lgl_probe (void);
void x10rt_lgl_finalize (void);
Modified: trunk/x10.runtime/src-cpp/x10rt/pgas/pgas.mk
===================================================================
--- trunk/x10.runtime/src-cpp/x10rt/pgas/pgas.mk 2009-11-18 10:41:46 UTC (rev 12117)
+++ trunk/x10.runtime/src-cpp/x10rt/pgas/pgas.mk 2009-11-18 13:52:57 UTC (rev 12118)
@@ -101,7 +101,7 @@
$(AR) $(ARFLAGS) $@ $(COMMON_OBJS)
else
$(SOCKETS_TGZ).phony:
- $(WGET) -N "http://dist.codehaus.org/x10/binaryReleases/svn head/$(SOCKETS_TGZ)"
+ -$(WGET) -N "http://dist.codehaus.org/x10/binaryReleases/svn head/$(SOCKETS_TGZ)"
$(SOCKETS_TGZ): $(SOCKETS_TGZ).phony
@@ -142,7 +142,7 @@
$(AR) $(ARFLAGS) $@ $(COMMON_OBJS)
else
$(LAPI_TGZ).phony:
- $(WGET) -N "http://dist.codehaus.org/x10/binaryReleases/svn head/$(LAPI_TGZ)"
+ -$(WGET) -N "http://dist.codehaus.org/x10/binaryReleases/svn head/$(LAPI_TGZ)"
$(LAPI_TGZ): $(LAPI_TGZ).phony
@@ -179,7 +179,7 @@
$(AR) $(ARFLAGS) $@ $(COMMON_OBJS)
else
$(BGP_TGZ).phony:
- $(WGET) -N "http://dist.codehaus.org/x10/binaryReleases/svn head/$(BGP_TGZ)"
+ -$(WGET) -N "http://dist.codehaus.org/x10/binaryReleases/svn head/$(BGP_TGZ)"
$(BGP_TGZ): $(BGP_TGZ).phony
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
