added runtime api hipMemcpyFromSymbolAsync
Change-Id: Ibaf925faf0ba464dd0ed6c5ea74c224c2ce38889
[ROCm/clr commit: 3d348b2d81]
This commit is contained in:
@@ -1159,7 +1159,7 @@ hipError_t hipMemcpyToSymbol(const char* symbolName, const void *src, size_t siz
|
||||
*/
|
||||
hipError_t hipMemcpyToSymbolAsync(const char* symbolName, const void *src, size_t sizeBytes, size_t offset, hipMemcpyKind kind, hipStream_t stream);
|
||||
|
||||
|
||||
hipError_t hipMemcpyFromSymbolAsync(void *dst, const char* symbolName, size_t sizeBytes, size_t offset, hipMemcpyKind kind, hipStream_t stream);
|
||||
|
||||
/**
|
||||
* @brief Copy data from src to dst asynchronously.
|
||||
|
||||
@@ -1780,7 +1780,7 @@ 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)
|
||||
void ihipStream_t::lockedSymbolCopySync(hc::accelerator &acc, void* dst, void* src, size_t sizeBytes, unsigned kind)
|
||||
{
|
||||
if(kind == hipMemcpyHostToHost){
|
||||
acc.memcpy_symbol(dst, (void*)src, sizeBytes, Kalmar::hcMemcpyHostToHost);
|
||||
@@ -1796,11 +1796,18 @@ void ihipStream_t::lockedSymbolCopySync(hc::accelerator &acc, void* dst, const v
|
||||
}
|
||||
}
|
||||
|
||||
void ihipStream_t::lockedSymbolCopyAsync(hc::accelerator &acc, void* dst, const void* src, size_t sizeBytes, unsigned kind)
|
||||
void ihipStream_t::lockedSymbolCopyAsync(hc::accelerator &acc, void* dst, 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);
|
||||
if(kind == hipMemcpyHostToDevice) {
|
||||
hc::AmPointerInfo dstPtrInfo(NULL, dst, sizeBytes, acc, true, false);
|
||||
hc::am_memtracker_add(dst, dstPtrInfo);
|
||||
locked_getAv()->copy_async((void*)src, dst, sizeBytes);
|
||||
}
|
||||
if(kind == hipMemcpyDeviceToHost) {
|
||||
hc::AmPointerInfo srcPtrInfo(NULL, src, sizeBytes, acc, true, false);
|
||||
hc::am_memtracker_add(src, srcPtrInfo);
|
||||
locked_getAv()->copy_async((void*)src, dst, sizeBytes);
|
||||
}
|
||||
}
|
||||
|
||||
void ihipStream_t::locked_copyAsync(void* dst, const void* src, size_t sizeBytes, unsigned kind)
|
||||
@@ -1903,7 +1910,7 @@ 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
|
||||
|
||||
@@ -494,8 +494,8 @@ public:
|
||||
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);
|
||||
void lockedSymbolCopySync(hc::accelerator &acc, void *dst, void* src, size_t sizeBytes, unsigned kind);
|
||||
void lockedSymbolCopyAsync(hc::accelerator &acc, void *dst, void* src, size_t sizeBytes, unsigned kind);
|
||||
|
||||
//---
|
||||
// Member functions that begin with locked_ are thread-safe accessors - these acquire / release the critical mutex.
|
||||
|
||||
@@ -461,7 +461,7 @@ hipError_t hipMemcpyToSymbol(const char* symbolName, const void *src, size_t cou
|
||||
|
||||
if(kind == hipMemcpyHostToDevice || kind == hipMemcpyDeviceToHost || kind == hipMemcpyDeviceToDevice || kind == hipMemcpyHostToHost)
|
||||
{
|
||||
stream->lockedSymbolCopySync(acc, dst, src, count + offset, kind);
|
||||
stream->lockedSymbolCopySync(acc, dst, (void*)src, count + offset, kind);
|
||||
// acc.memcpy_symbol(dst, (void*)src, count+offset);
|
||||
} else {
|
||||
return ihipLogStatus(hipErrorInvalidValue);
|
||||
@@ -493,6 +493,44 @@ hipError_t hipMemcpyToSymbolAsync(const char* symbolName, const void *src, size_
|
||||
return ihipLogStatus(hipErrorInvalidSymbol);
|
||||
}
|
||||
|
||||
if (stream) {
|
||||
try {
|
||||
stream->lockedSymbolCopyAsync(acc, dst, (void*)src, count + offset, kind);
|
||||
}
|
||||
catch (ihipException ex) {
|
||||
e = ex._code;
|
||||
}
|
||||
} else {
|
||||
e = hipErrorInvalidValue;
|
||||
}
|
||||
|
||||
return ihipLogStatus(e);
|
||||
}
|
||||
|
||||
|
||||
hipError_t hipMemcpyFromSymbolAsync(void* dst, const char* symbolName, size_t count, size_t offset, hipMemcpyKind kind, hipStream_t stream)
|
||||
{
|
||||
HIP_INIT_CMD_API(symbolName, dst, count, offset, kind, stream);
|
||||
|
||||
if(symbolName == nullptr)
|
||||
{
|
||||
return ihipLogStatus(hipErrorInvalidSymbol);
|
||||
}
|
||||
|
||||
hipError_t e = hipSuccess;
|
||||
|
||||
auto ctx = ihipGetTlsDefaultCtx();
|
||||
|
||||
hc::accelerator acc = ctx->getDevice()->_acc;
|
||||
|
||||
void *src = acc.get_symbol_address(symbolName);
|
||||
tprintf(DB_MEM, " symbol '%s' resolved to address:%p\n", symbolName, src);
|
||||
|
||||
if(src == nullptr || dst == nullptr)
|
||||
{
|
||||
return ihipLogStatus(hipErrorInvalidSymbol);
|
||||
}
|
||||
|
||||
if (stream) {
|
||||
try {
|
||||
stream->lockedSymbolCopyAsync(acc, dst, src, count + offset, kind);
|
||||
@@ -505,8 +543,6 @@ hipError_t hipMemcpyToSymbolAsync(const char* symbolName, const void *src, size_
|
||||
}
|
||||
|
||||
return ihipLogStatus(e);
|
||||
|
||||
|
||||
}
|
||||
|
||||
//---
|
||||
|
||||
@@ -18,7 +18,7 @@ THE SOFTWARE.
|
||||
*/
|
||||
|
||||
/* HIT_START
|
||||
* BUILD: %t %s
|
||||
* BUILD: %t %s
|
||||
* RUN: %t
|
||||
* HIT_END
|
||||
*/
|
||||
@@ -32,44 +32,53 @@ THE SOFTWARE.
|
||||
#define SIZE 1024*4
|
||||
|
||||
#ifdef __HIP_PLATFORM_HCC__
|
||||
__attribute__((address_space(1))) int global[NUM];
|
||||
__attribute__((address_space(1))) int globalIn[NUM];
|
||||
__attribute__((address_space(1))) int globalOut[NUM];
|
||||
#endif
|
||||
|
||||
#ifdef __HIP_PLATFORM_NVCC__
|
||||
__device__ int global[NUM];
|
||||
__device__ int globalIn[NUM];
|
||||
__device__ int globalOut[NUM];
|
||||
#endif
|
||||
|
||||
__global__ void Assign(hipLaunchParm lp, int* Out)
|
||||
{
|
||||
int tid = hipThreadIdx_x + hipBlockIdx_x * hipBlockDim_x;
|
||||
Out[tid] = global[tid];
|
||||
Out[tid] = globalIn[tid];
|
||||
globalOut[tid] = globalIn[tid];
|
||||
}
|
||||
|
||||
int main()
|
||||
{
|
||||
int *A, *Am, *B, *Ad;
|
||||
int *A, *Am, *B, *Ad, *C, *Cm;
|
||||
A = new int[NUM];
|
||||
B = new int[NUM];
|
||||
C = new int[NUM];
|
||||
for(unsigned i=0;i<NUM;i++) {
|
||||
A[i] = -1*i;
|
||||
B[i] = 0;
|
||||
C[i] = 0;
|
||||
}
|
||||
|
||||
hipMalloc((void**)&Ad, SIZE);
|
||||
hipHostMalloc((void**)&Am, SIZE);
|
||||
hipHostMalloc((void**)&Cm, SIZE);
|
||||
for(unsigned i=0;i<NUM;i++) {
|
||||
Am[i] = -1*i;
|
||||
Cm[i] = 0;
|
||||
}
|
||||
|
||||
hipStream_t stream;
|
||||
hipStreamCreate(&stream);
|
||||
hipMemcpyToSymbolAsync(HIP_SYMBOL(global), Am, SIZE, 0, hipMemcpyHostToDevice, stream);
|
||||
hipMemcpyToSymbolAsync(HIP_SYMBOL(globalIn), 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);
|
||||
|
||||
hipMemcpyFromSymbolAsync(Cm, HIP_SYMBOL(globalOut), SIZE, 0, hipMemcpyDeviceToHost, stream);
|
||||
hipStreamSynchronize(stream);
|
||||
for(unsigned i=0;i<NUM;i++) {
|
||||
assert(A[i] == B[i]);
|
||||
assert(Am[i] == B[i]);
|
||||
assert(Am[i] == Cm[i]);
|
||||
}
|
||||
|
||||
for(unsigned i=0;i<NUM;i++) {
|
||||
@@ -77,11 +86,12 @@ int main()
|
||||
B[i] = 0;
|
||||
}
|
||||
|
||||
hipMemcpyToSymbol(HIP_SYMBOL(global), A, SIZE, 0, hipMemcpyHostToDevice);
|
||||
hipMemcpyToSymbol(HIP_SYMBOL(globalIn), A, SIZE, 0, hipMemcpyHostToDevice);
|
||||
hipLaunchKernel(Assign, dim3(1,1,1), dim3(NUM,1,1), 0, 0, Ad);
|
||||
hipMemcpy(B, Ad, SIZE, hipMemcpyDeviceToHost);
|
||||
for(unsigned i=0;i<NUM;i++) {
|
||||
assert(A[i] == B[i]);
|
||||
}
|
||||
|
||||
passed();
|
||||
}
|
||||
|
||||
Verwijs in nieuw issue
Block a user