sanchayanmaity/gem5
1
0
Fork 0
Code Issues Pull requests Packages Projects Releases Wiki Activity

gpu-compute: use System cache line size in the GPU

Browse source
This commit is contained in:
Tony Gutierrez 2016-10-26 22:47:47 -04:00
parent 844fb845a5
commit aa7364276f
4 changed files with 9 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

3
src/gpu-compute/compute_unit.cc
View file

@ -75,7 +75,8 @@ ComputeUnit::ComputeUnit(const Params *p) : MemObject(p), fetchStage(p),
req_tick_latency(p->mem_req_latency * p->clk_domain->clockPeriod()), req_tick_latency(p->mem_req_latency * p->clk_domain->clockPeriod()),
resp_tick_latency(p->mem_resp_latency * p->clk_domain->clockPeriod()), resp_tick_latency(p->mem_resp_latency * p->clk_domain->clockPeriod()),
_masterId(p->system->getMasterId(name() + ".ComputeUnit")), _masterId(p->system->getMasterId(name() + ".ComputeUnit")),
lds(*p->localDataStore), globalSeqNum(0), wavefrontSize(p->wfSize), lds(*p->localDataStore), _cacheLineSize(p->system->cacheLineSize()),
globalSeqNum(0), wavefrontSize(p->wfSize),
kernelLaunchInst(new KernelLaunchStaticInst()) kernelLaunchInst(new KernelLaunchStaticInst())
{ {
/** /**

3
src/gpu-compute/compute_unit.hh
View file

@ -390,6 +390,8 @@ class ComputeUnit : public MemObject
int32_t int32_t
getRefCounter(const uint32_t dispatchId, const uint32_t wgId) const; getRefCounter(const uint32_t dispatchId, const uint32_t wgId) const;
int cacheLineSize() const { return _cacheLineSize; }
bool bool
sendToLds(GPUDynInstPtr gpuDynInst) __attribute__((warn_unused_result)); sendToLds(GPUDynInstPtr gpuDynInst) __attribute__((warn_unused_result));
@ -767,6 +769,7 @@ class ComputeUnit : public MemObject
uint64_t getAndIncSeqNum() { return globalSeqNum++; } uint64_t getAndIncSeqNum() { return globalSeqNum++; }
private: private:
const int _cacheLineSize;
uint64_t globalSeqNum; uint64_t globalSeqNum;
int wavefrontSize; int wavefrontSize;
GPUStaticInst *kernelLaunchInst; GPUStaticInst *kernelLaunchInst;

4
src/gpu-compute/fetch_unit.cc
View file

@ -132,10 +132,10 @@ FetchUnit::initiateFetch(Wavefront *wavefront)
// Since this is an instruction prefetch, if you're split then just finish // Since this is an instruction prefetch, if you're split then just finish
// out the current line. // out the current line.
unsigned block_size = RubySystem::getBlockSizeBytes(); int block_size = computeUnit->cacheLineSize();
// check for split accesses // check for split accesses
Addr split_addr = roundDown(vaddr + block_size - 1, block_size); Addr split_addr = roundDown(vaddr + block_size - 1, block_size);
unsigned size = block_size; int size = block_size;
if (split_addr > vaddr) { if (split_addr > vaddr) {
// misaligned access, just grab the rest of the line // misaligned access, just grab the rest of the line

4
src/gpu-compute/shader.cc
View file

@ -224,7 +224,7 @@ void
Shader::doFunctionalAccess(RequestPtr req, MemCmd cmd, void *data, Shader::doFunctionalAccess(RequestPtr req, MemCmd cmd, void *data,
bool suppress_func_errors, int cu_id) bool suppress_func_errors, int cu_id)
{ {
unsigned block_size = RubySystem::getBlockSizeBytes(); int block_size = cuList.at(cu_id)->cacheLineSize();
unsigned size = req->getSize(); unsigned size = req->getSize();
Addr tmp_addr; Addr tmp_addr;
@ -342,7 +342,7 @@ Shader::AccessMem(uint64_t address, void *ptr, uint32_t size, int cu_id,
{ {
uint8_t *data_buf = (uint8_t*)ptr; uint8_t *data_buf = (uint8_t*)ptr;
for (ChunkGenerator gen(address, size, RubySystem::getBlockSizeBytes()); for (ChunkGenerator gen(address, size, cuList.at(cu_id)->cacheLineSize());
!gen.done(); gen.next()) { !gen.done(); gen.next()) {
Request *req = new Request(0, gen.addr(), gen.size(), 0, Request *req = new Request(0, gen.addr(), gen.size(), 0,
cuList[0]->masterId(), 0, 0, 0); cuList[0]->masterId(), 0, 0, 0);