fixed symbol memcpy issue
Change-Id: I89d7401be51d194bcbf771020ba66e3d3b6a18f8
[ROCm/clr commit: 0d4e6ae60a]
Tento commit je obsažen v:
@@ -1780,6 +1780,28 @@ void ihipStream_t::locked_copySync(void* dst, const void* src, size_t sizeBytes,
|
||||
}
|
||||
}
|
||||
|
||||
void ihipStream_t::lockedSymbolCopySync(hc::accelerator &acc, void* dst, const void* src, size_t sizeBytes, unsigned kind)
|
||||
{
|
||||
if(kind == hipMemcpyHostToHost){
|
||||
acc.memcpy_symbol(dst, (void*)src, sizeBytes, Kalmar::hcMemcpyHostToHost);
|
||||
}
|
||||
if(kind == hipMemcpyHostToDevice){
|
||||
acc.memcpy_symbol(dst, (void*)src, sizeBytes);
|
||||
}
|
||||
if(kind == hipMemcpyDeviceToDevice){
|
||||
acc.memcpy_symbol(dst, (void*)src, sizeBytes, Kalmar::hcMemcpyDeviceToDevice);
|
||||
}
|
||||
if(kind == hipMemcpyDeviceToHost){
|
||||
acc.memcpy_symbol(dst, (void*)src, sizeBytes, Kalmar::hcMemcpyDeviceToHost);
|
||||
}
|
||||
}
|
||||
|
||||
void ihipStream_t::lockedSymbolCopyAsync(hc::accelerator &acc, void* dst, const void* src, size_t sizeBytes, unsigned kind)
|
||||
{
|
||||
hc::AmPointerInfo dstPtrInfo(NULL, dst, sizeBytes, acc, true, false);
|
||||
hc::am_memtracker_add(dst, dstPtrInfo);
|
||||
locked_getAv()->copy_async((void*)src, dst, sizeBytes);
|
||||
}
|
||||
|
||||
void ihipStream_t::locked_copyAsync(void* dst, const void* src, size_t sizeBytes, unsigned kind)
|
||||
{
|
||||
@@ -1881,6 +1903,7 @@ void ihipStream_t::locked_copyAsync(void* dst, const void* src, size_t sizeBytes
|
||||
LockedAccessor_StreamCrit_t crit(_criticalData);
|
||||
|
||||
this->ensureHaveQueue(crit);
|
||||
|
||||
#if USE_COPY_EXT_V2
|
||||
crit->_av.copy_ext(src, dst, sizeBytes, hcCopyDir, srcPtrInfo, dstPtrInfo, copyDevice ? ©Device->getDevice()->_acc : nullptr, forceUnpinnedCopy);
|
||||
#else
|
||||
|
||||
@@ -273,7 +273,7 @@ extern void recordApiTrace(std::string *fullStr, const std::string &apiStr);
|
||||
API_TRACE(0, __VA_ARGS__);
|
||||
|
||||
|
||||
// Like above, but will trace with DB_CMD.
|
||||
// Like above, but will trace with DB_CMD.
|
||||
// Replace HIP_INIT_API with this call inside important APIs that launch work on the GPU:
|
||||
// kernel launches, copy commands, memory sets, etc.
|
||||
#define HIP_INIT_CMD_API(...) \
|
||||
@@ -479,8 +479,6 @@ private:
|
||||
typedef ihipStreamCriticalBase_t<StreamMutex> ihipStreamCritical_t;
|
||||
typedef LockedAccessor<ihipStreamCritical_t> LockedAccessor_StreamCrit_t;
|
||||
|
||||
|
||||
|
||||
//---
|
||||
// Internal stream structure.
|
||||
class ihipStream_t {
|
||||
@@ -494,10 +492,10 @@ public:
|
||||
|
||||
// kind is hipMemcpyKind
|
||||
void locked_copySync (void* dst, const void* src, size_t sizeBytes, unsigned kind, bool resolveOn = true);
|
||||
|
||||
|
||||
void locked_copyAsync(void* dst, const void* src, size_t sizeBytes, unsigned kind);
|
||||
|
||||
void lockedSymbolCopySync(hc::accelerator &acc, void *dst, const void* src, size_t sizeBytes, unsigned kind);
|
||||
void lockedSymbolCopyAsync(hc::accelerator &acc, void *dst, const void* src, size_t sizeBytes, unsigned kind);
|
||||
|
||||
//---
|
||||
// Member functions that begin with locked_ are thread-safe accessors - these acquire / release the critical mutex.
|
||||
|
||||
@@ -449,17 +449,23 @@ hipError_t hipMemcpyToSymbol(const char* symbolName, const void *src, size_t cou
|
||||
|
||||
hc::accelerator acc = ctx->getDevice()->_acc;
|
||||
|
||||
void *ptr = acc.get_symbol_address(symbolName);
|
||||
tprintf(DB_MEM, " symbol '%s' resolved to address:%p\n", symbolName, ptr);
|
||||
void *dst = acc.get_symbol_address(symbolName);
|
||||
tprintf(DB_MEM, " symbol '%s' resolved to address:%p\n", symbolName, dst);
|
||||
|
||||
if(ptr == nullptr)
|
||||
if(dst == nullptr)
|
||||
{
|
||||
return ihipLogStatus(hipErrorInvalidSymbol);
|
||||
}
|
||||
|
||||
hipStream_t stream = ihipSyncAndResolveStream(hipStreamNull);
|
||||
|
||||
stream->locked_copySync(ptr, src, count + offset, kind);
|
||||
if(kind == hipMemcpyHostToDevice || kind == hipMemcpyDeviceToHost || kind == hipMemcpyDeviceToDevice || kind == hipMemcpyHostToHost)
|
||||
{
|
||||
stream->lockedSymbolCopySync(acc, dst, src, count + offset, kind);
|
||||
// acc.memcpy_symbol(dst, (void*)src, count+offset);
|
||||
} else {
|
||||
return ihipLogStatus(hipErrorInvalidValue);
|
||||
}
|
||||
|
||||
return ihipLogStatus(hipSuccess);
|
||||
}
|
||||
@@ -479,17 +485,17 @@ hipError_t hipMemcpyToSymbolAsync(const char* symbolName, const void *src, size_
|
||||
|
||||
hc::accelerator acc = ctx->getDevice()->_acc;
|
||||
|
||||
void *ptr = acc.get_symbol_address(symbolName);
|
||||
tprintf(DB_MEM, " symbol '%s' resolved to address:%p\n", symbolName, ptr);
|
||||
void *dst = acc.get_symbol_address(symbolName);
|
||||
tprintf(DB_MEM, " symbol '%s' resolved to address:%p\n", symbolName, dst);
|
||||
|
||||
if(ptr == nullptr)
|
||||
if(dst == nullptr)
|
||||
{
|
||||
return ihipLogStatus(hipErrorInvalidSymbol);
|
||||
}
|
||||
|
||||
if (stream) {
|
||||
try {
|
||||
stream->locked_copyAsync(ptr, src, count + offset, kind);
|
||||
stream->lockedSymbolCopyAsync(acc, dst, src, count + offset, kind);
|
||||
}
|
||||
catch (ihipException ex) {
|
||||
e = ex._code;
|
||||
|
||||
@@ -47,7 +47,7 @@ __global__ void Assign(hipLaunchParm lp, int* Out)
|
||||
|
||||
int main()
|
||||
{
|
||||
int *A, *B, *Ad;
|
||||
int *A, *Am, *B, *Ad;
|
||||
A = new int[NUM];
|
||||
B = new int[NUM];
|
||||
for(unsigned i=0;i<NUM;i++) {
|
||||
@@ -56,10 +56,14 @@ int main()
|
||||
}
|
||||
|
||||
hipMalloc((void**)&Ad, SIZE);
|
||||
hipHostMalloc((void**)&Am, SIZE);
|
||||
for(unsigned i=0;i<NUM;i++) {
|
||||
Am[i] = -1*i;
|
||||
}
|
||||
|
||||
hipStream_t stream;
|
||||
hipStreamCreate(&stream);
|
||||
hipMemcpyToSymbolAsync(HIP_SYMBOL(global), A, SIZE, 0, hipMemcpyHostToDevice, stream);
|
||||
hipMemcpyToSymbolAsync(HIP_SYMBOL(global), Am, SIZE, 0, hipMemcpyHostToDevice, stream);
|
||||
hipStreamSynchronize(stream);
|
||||
hipLaunchKernel(Assign, dim3(1,1,1), dim3(NUM,1,1), 0, 0, Ad);
|
||||
hipMemcpy(B, Ad, SIZE, hipMemcpyDeviceToHost);
|
||||
|
||||
Odkázat v novém úkolu
Zablokovat Uživatele