Menu
▾
▴
[X10-commits] SF.net SVN: x10:[13478] trunk
|
From: <spa...@us...> - 2010-03-20 07:20:06
|
Revision: 13478
http://x10.svn.sourceforge.net/x10/?rev=13478&view=rev
Author: sparksparkspark
Date: 2010-03-20 07:19:59 +0000 (Sat, 20 Mar 2010)
Log Message:
-----------
WIP on outstanding CUDA issues
Modified Paths:
--------------
trunk/x10.compiler/src/x10cuda/visit/CUDACodeGenerator.java
trunk/x10.runtime/src-cpp/x10aux/deserialization_dispatcher.h
trunk/x10.runtime/src-cpp/x10aux/network.cc
trunk/x10.runtime/x10rt/common/x10rt_cuda.cc
Modified: trunk/x10.compiler/src/x10cuda/visit/CUDACodeGenerator.java
===================================================================
--- trunk/x10.compiler/src/x10cuda/visit/CUDACodeGenerator.java 2010-03-20 05:45:37 UTC (rev 13477)
+++ trunk/x10.compiler/src/x10cuda/visit/CUDACodeGenerator.java 2010-03-20 07:19:59 UTC (rev 13478)
@@ -579,7 +579,7 @@
if (nodeHasCUDAAnnotation(block)) {
- inc.write("static x10_ulong "+SharedVarsMethods.DESERIALIZE_CUDA_METHOD+"("+DESERIALIZATION_BUFFER+" &__buf, x10aux::place __gpu, size_t &__blocks, size_t &__threads, size_t &__shm) {");
+ inc.write("static void "+SharedVarsMethods.DESERIALIZE_CUDA_METHOD+"("+DESERIALIZATION_BUFFER+" &__buf, x10aux::place __gpu, size_t &__blocks, size_t &__threads, size_t &__shm, size_t &argc, char *&argv, size_t &cmemc, char *&cmemv) {");
inc.newline(4); inc.begin(0);
inc.write(make_ref(cnamet)+" __this = "+cnamet+"::"+DESERIALIZE_METHOD+"<"+cnamet+">(__buf);");
@@ -647,10 +647,16 @@
inc.write(";");
inc.newline();
}
- inc.write("x10_ulong __remote_env = x10aux::remote_alloc(__gpu, sizeof(__env));"); inc.newline();
- inc.write("x10aux::cuda_put(__gpu, __remote_env, &__env, sizeof(__env));"); inc.newline();
- inc.write("return __remote_env;"); inc.end(); inc.newline();
- inc.write("}"); inc.newline(); inc.forceNewline();
+ if (true) {
+ inc.write("x10_ulong __remote_env = x10aux::remote_alloc(__gpu, sizeof(__env));"); inc.newline();
+ inc.write("x10aux::cuda_put(__gpu, __remote_env, &__env, sizeof(__env));"); inc.newline();
+ inc.write("::memcpy(argv, &__remote_env, sizeof (void*));"); inc.newline();
+ inc.write("argc = sizeof(void*);"); inc.end(); inc.newline();
+ } else {
+ inc.write("memcpy(argv, __env, sizeof(__env));"); inc.newline();
+ inc.write("argc = sizeof(__env);"); inc.end(); inc.newline();
+ }
+ inc.write("}"); inc.newline(); inc.forceNewline();
}
}
@@ -660,14 +666,8 @@
}
@Override
- public void visit(ArrayInit_c n) {
- // TODO Auto-generated method stub
- super.visit(n);
- }
-
- @Override
public void visit(Assert_c n) {
- // TODO Auto-generated method stub
+ assert !generatingKernel() : "Throwing exceptions not allowed in @CUDA code.";
super.visit(n);
}
@@ -685,7 +685,7 @@
@Override
public void visit(Await_c n) {
- // TODO Auto-generated method stub
+ assert !generatingKernel() : "Await not allowed in @CUDA code.";
super.visit(n);
}
@@ -715,7 +715,7 @@
@Override
public void visit(Catch_c n) {
- // TODO Auto-generated method stub
+ assert !generatingKernel() : "Catching exceptions not allowed in @CUDA code.";
super.visit(n);
}
@@ -726,14 +726,8 @@
}
@Override
- public void visit(ClassBody_c n) {
- // TODO Auto-generated method stub
- super.visit(n);
- }
-
- @Override
public void visit(ClosureCall_c c) {
- // TODO Auto-generated method stub
+ assert !generatingKernel() : "Closure calls not allowed in @CUDA code.";
super.visit(c);
}
@@ -744,18 +738,6 @@
}
@Override
- public void visit(ConstantDistMaker_c n) {
- // TODO Auto-generated method stub
- super.visit(n);
- }
-
- @Override
- public void visit(ConstructorDecl_c dec) {
- // TODO Auto-generated method stub
- super.visit(dec);
- }
-
- @Override
public void visit(Do_c n) {
// TODO Auto-generated method stub
super.visit(n);
@@ -780,12 +762,6 @@
}
@Override
- public void visit(FieldDecl_c dec) {
- // TODO Auto-generated method stub
- super.visit(dec);
- }
-
- @Override
public void visit(FloatLit_c n) {
// TODO Auto-generated method stub
super.visit(n);
@@ -822,12 +798,6 @@
}
@Override
- public void visit(Import_c n) {
- // TODO Auto-generated method stub
- super.visit(n);
- }
-
- @Override
public void visit(Initializer_c n) {
// TODO Auto-generated method stub
super.visit(n);
@@ -877,60 +847,12 @@
}
@Override
- public void visit(LocalClassDecl_c n) {
- // TODO Auto-generated method stub
- super.visit(n);
- }
-
- @Override
- public void visit(LocalDecl_c dec) {
- // TODO Auto-generated method stub
- super.visit(dec);
- }
-
- @Override
- public void visit(MethodDecl_c dec) {
- // TODO Auto-generated method stub
- super.visit(dec);
- }
-
- @Override
- public void visit(Node n) {
- // TODO Auto-generated method stub
- super.visit(n);
- }
-
- @Override
public void visit(NullLit_c n) {
// TODO Auto-generated method stub
super.visit(n);
}
@Override
- public void visit(PackageNode_c n) {
- // TODO Auto-generated method stub
- super.visit(n);
- }
-
- @Override
- public void visit(ParExpr_c n) {
- // TODO Auto-generated method stub
- super.visit(n);
- }
-
- @Override
- public void visit(PropertyDecl_c n) {
- // TODO Auto-generated method stub
- super.visit(n);
- }
-
- @Override
- public void visit(RegionMaker_c n) {
- // TODO Auto-generated method stub
- super.visit(n);
- }
-
- @Override
public void visit(Return_c ret) {
// TODO Auto-generated method stub
super.visit(ret);
@@ -962,13 +884,13 @@
@Override
public void visit(Throw_c n) {
- // TODO Auto-generated method stub
+ assert !generatingKernel() : "Throwing exceptions not allowed in @CUDA code.";
super.visit(n);
}
@Override
public void visit(Try_c n) {
- // TODO Auto-generated method stub
+ assert !generatingKernel() : "Catching exceptions not allowed in @CUDA code.";
super.visit(n);
}
@@ -1004,7 +926,8 @@
@Override
public void visit(X10Call_c n) {
- // TODO Auto-generated method stub
+ // In fact they are allowed, as long as they are implemented with @Native
+ //assert !generatingKernel() : "Calling functions not allowed in @CUDA code.";
super.visit(n);
}
@@ -1028,7 +951,7 @@
@Override
public void visit(X10Instanceof_c n) {
- // TODO Auto-generated method stub
+ assert !generatingKernel() : "Runtime types not available in @CUDA code.";
super.visit(n);
}
Modified: trunk/x10.runtime/src-cpp/x10aux/deserialization_dispatcher.h
===================================================================
--- trunk/x10.runtime/src-cpp/x10aux/deserialization_dispatcher.h 2010-03-20 05:45:37 UTC (rev 13477)
+++ trunk/x10.runtime/src-cpp/x10aux/deserialization_dispatcher.h 2010-03-20 07:19:59 UTC (rev 13478)
@@ -26,8 +26,8 @@
typedef ref<x10::lang::Object> (*Deserializer)(deserialization_buffer &buf);
template<> inline const char *typeName<Deserializer>() { return "Deserializer"; }
- typedef x10_ulong (*CUDAPre)(deserialization_buffer &buf, place p,
- size_t &blocks, size_t &threads, size_t &shm);
+ typedef void (*CUDAPre)(deserialization_buffer &buf, place p,
+ size_t &blocks, size_t &threads, size_t &shm, size_t &argc, char *&argv, size_t &cmemc, char *&cmemv);
template<> inline const char *typeName<CUDAPre>() { return "CUDAPre"; }
typedef void *(*BufferFinder)(deserialization_buffer &buf, x10_int len);
Modified: trunk/x10.runtime/src-cpp/x10aux/network.cc
===================================================================
--- trunk/x10.runtime/src-cpp/x10aux/network.cc 2010-03-20 05:45:37 UTC (rev 13477)
+++ trunk/x10.runtime/src-cpp/x10aux/network.cc 2010-03-20 07:19:59 UTC (rev 13478)
@@ -294,16 +294,14 @@
// note: high bytes thrown away in implicit conversion
serialization_id_t sid = x10aux::DeserializationDispatcher::getSerializationId(p->type);
x10aux::CUDAPre pre = x10aux::DeserializationDispatcher::getCUDAPre(sid);
- x10_ulong env = pre(buf, p->dest_place, *blocks, *threads, *shm);
+ pre(buf, p->dest_place, *blocks, *threads, *shm, *argc, *argv, *cmemc, *cmemv);
assert(buf.consumed() <= p->len);
- *argv = (char*)(size_t)env;
- *argc = sizeof(void*);
}
static void cuda_post (const x10rt_msg_params *p, void *env)
{
_X_(ANSI_X10RT<<"Receiving a kernel post callback, deserialising..."<<ANSI_RESET);
- remote_free(p->dest_place, (x10_ulong)(size_t)env);
+ //remote_free(p->dest_place, (x10_ulong)(size_t)env);
x10aux::deserialization_buffer buf(static_cast<char*>(p->msg));
x10aux::ref<x10::lang::Reference> fs = buf.read<x10aux::ref<x10::lang::Reference> >();
x10aux::ref<x10::lang::Runtime> rt = x10::lang::PlaceLocalHandle_methods<x10aux::ref<x10::lang::Runtime> >::apply(x10::lang::Runtime::FMGL(runtime));
Modified: trunk/x10.runtime/x10rt/common/x10rt_cuda.cc
===================================================================
--- trunk/x10.runtime/x10rt/common/x10rt_cuda.cc 2010-03-20 05:45:37 UTC (rev 13477)
+++ trunk/x10.runtime/x10rt/common/x10rt_cuda.cc 2010-03-20 07:19:59 UTC (rev 13478)
@@ -22,7 +22,7 @@
// TODO: fine grained synchronisation, lock free datastructures
pthread_mutex_t big_lock_of_doom;
- static inline void DEBUG(const char *fmt, ...) {
+ inline void DEBUG(const char *fmt, ...) {
(void) fmt;
va_list ap;
va_start(ap, fmt);
@@ -108,7 +108,7 @@
/* }}} */
- static size_t dma_slice_sz (void) {
+ size_t dma_slice_sz (void) {
static size_t sz = 0;
if (sz == 0) {
const char *env_var = "X10RT_CUDA_DMA_SLICE";
@@ -275,6 +275,7 @@
void *pinned_mem2;
void *front;
void *back;
+ size_t commit;
op_queue<x10rt_cuda_kernel> kernel_q;
op_queue<x10rt_cuda_copy> dma_q;
Table<x10rt_functions> cbs;
@@ -433,7 +434,6 @@
#endif
}
-
void *x10rt_cuda_device_alloc (x10rt_cuda_ctx *ctx,
size_t len)
{
@@ -441,6 +441,8 @@
pthread_mutex_lock(&big_lock_of_doom);
CU_SAFE(cuCtxPushCurrent(ctx->ctx));
CUdeviceptr ptr;
+ ctx->commit += len;
+ //fprintf(stderr,"CUDA committed memory: %llu bytes\n", (unsigned long long)ctx->commit);
CU_SAFE(cuMemAlloc(&ptr, len));
CU_SAFE(cuCtxPopCurrent(NULL));
pthread_mutex_unlock(&big_lock_of_doom);
@@ -684,7 +686,8 @@
CUfunction k = ctx->cbs[type].kernel_cbs.kernel;
// y and z params we leave as 1, as threads can vary from 1 to 512
CU_SAFE(cuFuncSetBlockShape(k, kop->threads, 1, 1));
- CU_SAFE(cuParamSetv(k, 0, &kop->argv, kop->argc));
+ //fprintf(stderr,"%p<<<%d,%d,%d>>> argc: %d argv: %p\n", (void*)k, kop->blocks, kop->threads, kop->shm, kop->argc, *(void**)kop->argv);
+ CU_SAFE(cuParamSetv(k, 0, &kop->argv[0], kop->argc));
CU_SAFE(cuParamSetSize(k, kop->argc));
CU_SAFE(cuFuncSetSharedSize(k, kop->shm));
CU_SAFE(cuLaunchGridAsync(k, kop->blocks, 1, ctx->kernel_q.stream));
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
