AI/rocm-systems
2
0
Forc 0
Cód Saincheisteanna Iarratais Tarraingthe Gníomhartha Pacáistí Tionscadail Scaoileann Vicí Gníomhaíocht

This switches HIP from its currently convoluted macro + pfe based dispatch mechanism to a more natural one partially based on the existing module API. The basic idea is that HCC will always correctly emit __global__ functions: as empty-bodied stubs, on host, and as kernels, on device. It then becomes trivial to obtain the mangled name on host, at dispatch, from the function's address, and then to use the mangled name to retrieve the kernel. This should address all problems stemming from serialisation, dubious mismatches due to the manufactured functor, macro-isms et al. It also immediately enables support for generalised globals as a consequence of that being available in the module API. Finally, it will make debug much easier, since the actual names of the __global__ functions will automatically be used in traces etc. One detail is that due to how dispatch works now (hipLaunchKernel and hipLaunchKernelGGL are themselves variadic function templates which deduce the function type of the callee), in certain cases it may be necesssary to insert explicit casts to ensure that the variadic argument list selects a viable overload - this can be observed in some unit tests. Eventually we may be able to remove this limitation, but for now it does not appear terribly onerous. The code is not extremely HIPpie, nor is it fully optimised, but rather is intended as a starting point for the HIP team to make its own.

Brabhsáil Foinse 
[ROCm/clr commit: 28f87f7d2e]
Tá an tiomantas seo le fáil i:
Alex Voicu
2017-11-01 15:09:59 +00:00
tuismitheoir c1639486c6
tiomantas b738633bcb
D'athraigh 27 comhad le 1457 breiseanna agus 1352 scriosta
Íoslódáil an comhad paiste Íoslódáil Comhad Diff Leathnaigh gach comhaid Laghdaigh gach comhaid
projects/clr/hipamd/tests/src/runtimeApi/memory/hipMemcpy.cpp
+43 -34
Féach ar an gComhad
@@ -58,7 +58,7 @@ public:
void offset(int offset) { _offset = offset; };
int offset() const { return _offset; };
private:
T * _A_d;
T* _B_d;
@@ -72,7 +72,7 @@ private:
template<typename T>
DeviceMemory<T>::DeviceMemory(size_t numElements)
: _maxNumElements(numElements),
: _maxNumElements(numElements),
_offset(0)
{
T ** np = nullptr;
@@ -93,7 +93,7 @@ DeviceMemory<T>::~DeviceMemory ()
HipTest::freeArrays (_A_d, _B_d, _C_d, np, np, np, 0);
HIPCHECK (hipFree(_C_dd));
_C_dd = NULL;
};
@@ -125,7 +125,7 @@ public:
T * A_hh;
T* B_hh;
bool _usePinnedHost;
bool _usePinnedHost;
private:
size_t _maxNumElements;
@@ -165,11 +165,11 @@ HostMemory<T>::HostMemory(size_t numElements, bool usePinnedHost)
template<typename T>
void
HostMemory<T>::reset(size_t numElements, bool full)
HostMemory<T>::reset(size_t numElements, bool full)
{
// Initialize the host data:
for (size_t i=0; i<numElements; i++) {
(A_hh)[i] = 1097.0 + i;
(A_hh)[i] = 1097.0 + i;
(B_hh)[i] = 1492.0 + i; // Phi
if (full) {
@@ -213,8 +213,8 @@ template <typename T>
void memcpytest2(DeviceMemory<T> *dmem, HostMemory<T> *hmem, size_t numElements, bool useHostToHost, bool useDeviceToDevice, bool useMemkindDefault)
{
size_t sizeElements = numElements * sizeof(T);
printf ("test: %s<%s> size=%lu (%6.2fMB) usePinnedHost:%d, useHostToHost:%d, useDeviceToDevice:%d, useMemkindDefault:%d, offsets:dev:%+d host:+%d\n",
__func__,
printf ("test: %s<%s> size=%lu (%6.2fMB) usePinnedHost:%d, useHostToHost:%d, useDeviceToDevice:%d, useMemkindDefault:%d, offsets:dev:%+d host:+%d\n",
__func__,
TYPENAME(T),
sizeElements, sizeElements/1024.0/1024.0,
hmem->_usePinnedHost, useHostToHost, useDeviceToDevice, useMemkindDefault,
@@ -243,7 +243,16 @@ void memcpytest2(DeviceMemory<T> *dmem, HostMemory<T> *hmem, size_t numElements,
HIPCHECK ( hipMemcpy(dmem->B_d(), hmem->B_h(), sizeElements, useMemkindDefault ? hipMemcpyDefault : hipMemcpyHostToDevice));
}
hipLaunchKernel(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0, dmem->A_d(), dmem->B_d(), dmem->C_d(), numElements);
hipLaunchKernel(
HipTest::vectorADD,
dim3(blocks),
dim3(threadsPerBlock),
0,
0,
static_cast<const T*>(dmem->A_d()),
static_cast<const T*>(dmem->B_d()),
dmem->C_d(),
numElements);
if (useDeviceToDevice) {
// Do an extra device-to-device copy here to mix things up:
@@ -273,8 +282,8 @@ void memcpytest2_for_type(size_t numElements)
{
printSep();
DeviceMemory<T> memD(numElements);
HostMemory<T> memU(numElements, 0/*usePinnedHost*/);
DeviceMemory<T> memD(numElements);
HostMemory<T> memU(numElements, 0/*usePinnedHost*/);
HostMemory<T> memP(numElements, 1/*usePinnedHost*/);
for (int usePinnedHost =0; usePinnedHost<=1; usePinnedHost++) {
@@ -307,11 +316,11 @@ void memcpytest2_sizes(size_t maxElem=0)
maxElem = free/sizeof(T)/8;
}
printf (" device#%d: hipMemGetInfo: free=%zu (%4.2fMB) total=%zu (%4.2fMB) maxSize=%6.1fMB\n",
printf (" device#%d: hipMemGetInfo: free=%zu (%4.2fMB) total=%zu (%4.2fMB) maxSize=%6.1fMB\n",
deviceId, free, (float)(free/1024.0/1024.0), total, (float)(total/1024.0/1024.0), maxElem*sizeof(T)/1024.0/1024.0);
HIPCHECK ( hipDeviceReset() );
DeviceMemory<T> memD(maxElem);
HostMemory<T> memU(maxElem, 0/*usePinnedHost*/);
DeviceMemory<T> memD(maxElem);
HostMemory<T> memU(maxElem, 0/*usePinnedHost*/);
HostMemory<T> memP(maxElem, 1/*usePinnedHost*/);
for (size_t elem=1; elem<=maxElem; elem*=2) {
@@ -336,11 +345,11 @@ void memcpytest2_offsets(size_t maxElem, bool devOffsets, bool hostOffsets)
HIPCHECK(hipMemGetInfo(&free, &total));
printf (" device#%d: hipMemGetInfo: free=%zu (%4.2fMB) total=%zu (%4.2fMB) maxSize=%6.1fMB\n",
printf (" device#%d: hipMemGetInfo: free=%zu (%4.2fMB) total=%zu (%4.2fMB) maxSize=%6.1fMB\n",
deviceId, free, (float)(free/1024.0/1024.0), total, (float)(total/1024.0/1024.0), maxElem*sizeof(T)/1024.0/1024.0);
HIPCHECK ( hipDeviceReset() );
DeviceMemory<T> memD(maxElem);
HostMemory<T> memU(maxElem, 0/*usePinnedHost*/);
DeviceMemory<T> memD(maxElem);
HostMemory<T> memU(maxElem, 0/*usePinnedHost*/);
HostMemory<T> memP(maxElem, 1/*usePinnedHost*/);
size_t elem = maxElem / 2;
@@ -380,16 +389,16 @@ void multiThread_1(bool serialize, bool usePinnedHost)
{
printSep();
printf ("test: %s<%s> serialize=%d usePinnedHost=%d\n", __func__, TYPENAME(T), serialize, usePinnedHost);
DeviceMemory<T> memD(N);
HostMemory<T> mem1(N, usePinnedHost);
HostMemory<T> mem2(N, usePinnedHost);
DeviceMemory<T> memD(N);
HostMemory<T> mem1(N, usePinnedHost);
HostMemory<T> mem2(N, usePinnedHost);
std::thread t1 (memcpytest2<T>, &memD, &mem1, N, 0,0,0);
if (serialize) {
t1.join();
}
std::thread t2 (memcpytest2<T>,&memD, &mem2, N, 0,0,0);
if (serialize) {
t2.join();
@@ -427,21 +436,21 @@ int main(int argc, char *argv[])
// Some tests around the 64KB boundary which have historically shown issues:
printf ("\n\n=== tests&0x2 (64KB boundary)\n");
size_t maxElem = 32*1024*1024;
DeviceMemory<float> memD(maxElem);
HostMemory<float> memU(maxElem, 0/*usePinnedHost*/);
HostMemory<float> memP(maxElem, 0/*usePinnedHost*/);
DeviceMemory<float> memD(maxElem);
HostMemory<float> memU(maxElem, 0/*usePinnedHost*/);
HostMemory<float> memP(maxElem, 0/*usePinnedHost*/);
// These all pass:
memcpytest2<float>(&memD, &memP, 15*1024*1024, 0, 0, 0);
memcpytest2<float>(&memD, &memP, 16*1024*1024, 0, 0, 0);
memcpytest2<float>(&memD, &memP, 16*1024*1024+16*1024, 0, 0, 0);
memcpytest2<float>(&memD, &memP, 15*1024*1024, 0, 0, 0);
memcpytest2<float>(&memD, &memP, 16*1024*1024, 0, 0, 0);
memcpytest2<float>(&memD, &memP, 16*1024*1024+16*1024, 0, 0, 0);
// Just over 64MB:
memcpytest2<float>(&memD, &memP, 16*1024*1024+512*1024, 0, 0, 0);
memcpytest2<float>(&memD, &memP, 17*1024*1024+1024, 0, 0, 0);
memcpytest2<float>(&memD, &memP, 32*1024*1024, 0, 0, 0);
memcpytest2<float>(&memD, &memU, 32*1024*1024, 0, 0, 0);
memcpytest2<float>(&memD, &memP, 32*1024*1024, 1, 1, 0);
memcpytest2<float>(&memD, &memP, 32*1024*1024, 1, 1, 0);
memcpytest2<float>(&memD, &memP, 16*1024*1024+512*1024, 0, 0, 0);
memcpytest2<float>(&memD, &memP, 17*1024*1024+1024, 0, 0, 0);
memcpytest2<float>(&memD, &memP, 32*1024*1024, 0, 0, 0);
memcpytest2<float>(&memD, &memU, 32*1024*1024, 0, 0, 0);
memcpytest2<float>(&memD, &memP, 32*1024*1024, 1, 1, 0);
memcpytest2<float>(&memD, &memP, 32*1024*1024, 1, 1, 0);
}
@@ -464,7 +473,7 @@ int main(int argc, char *argv[])
// Simplest cases: serialize the threads, and also used pinned memory:
// This verifies that the sub-calls to memcpytest2 are correct.
multiThread_1<float>(true, true);
multiThread_1<float>(true, true);
// Serialize, but use unpinned memory to stress the unpinned memory xfer path.
multiThread_1<float>(true, false);