gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[PYTHON] Added TRITON_DEBUG_MODE which reallocates input tensors outside of the pytorch memory pool to spot out-of-bounds accesses more easily

Browse Source
This commit is contained in:
Philippe Tillet
2020-12-26 01:43:38 -05:00
parent c4fceeea49
commit c0bc7ed8b0
3 changed files with 27 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
python/examples/tutorials/vec_add.py
View File

@@ -14,7 +14,7 @@ __global__ void add(float* z, float* x, float* y, int N) {
bool check[TILE] = offset < N; bool check[TILE] = offset < N;
*?(check)pz = *?(check)px + *?(check)py; *pz = *px + *py;
} }
""" """
@@ -32,9 +32,8 @@ add = _add.apply
# test # test
torch.manual_seed(0) torch.manual_seed(0)
x = torch.rand(98432).cuda() x = torch.rand(900).cuda()
y = torch.rand(98432).cuda() y = torch.rand(900).cuda()
za = x + y za = x + y
zb = add(x, y) zb = add(x, y)
print(torch.allclose(za,zb)) print(torch.allclose(za,zb))

13
python/src/launch.cc
View File

@@ -57,7 +57,19 @@ void synchronize(int64_t dev_id) {
triton::driver::cu_stream stream(torch_get_cuda_stream(dev_id), false); triton::driver::cu_stream stream(torch_get_cuda_stream(dev_id), false);
stream.synchronize(); stream.synchronize();
} }
}
torch::Tensor raw_like(torch::Tensor x){
if(x.nbytes() == 0)
return torch::empty_like(x);
C10_CUDA_CHECK(cudaSetDevice(x.device().index()));
auto shape = x.sizes();
CUdeviceptr data;
triton::driver::dispatch::cuMemAlloc(&data, x.nbytes());
auto deleter = [data](void* ptr) { triton::driver::dispatch::cuMemFree_v2(data); };
auto ret = torch::from_blob((void*)data, shape, deleter, x.options());
ret.copy_(x);
return ret;
} }
void launch_kernel(int64_t op_id, int64_t dev_id, const std::string& args, void launch_kernel(int64_t op_id, int64_t dev_id, const std::string& args,
@@ -82,5 +94,6 @@ void launch_kernel(int64_t op_id, int64_t dev_id, const std::string& args,
static auto registry = torch::RegisterOperators() static auto registry = torch::RegisterOperators()
.op("triton::launch_kernel", &launch_kernel) .op("triton::launch_kernel", &launch_kernel)
.op("triton::raw_like", &raw_like)
.op("triton::cdiv_sum", &cdiv_sum) .op("triton::cdiv_sum", &cdiv_sum)
.op("triton::synchronize", &synchronize); .op("triton::synchronize", &synchronize);

10
python/triton/kernel.py
View File

@@ -89,6 +89,12 @@ class kernel:
return libtriton.get_fn_ptx((self.op_id, dev_id), opt) return libtriton.get_fn_ptx((self.op_id, dev_id), opt)
def __call__(self, *args, **kwargs): def __call__(self, *args, **kwargs):
if 'TRITON_DEBUG_MODE' in os.environ:
_args = args
args = [x for x in args]
for i in range(len(args)):
if isinstance(args[i], torch.Tensor):
args[i] = torch.ops.triton.raw_like(args[i])
for x in args: for x in args:
if isinstance(x, torch.Tensor): if isinstance(x, torch.Tensor):
device = x.device.index device = x.device.index
@@ -109,3 +115,7 @@ class kernel:
names = list(kwargs['constants'].keys()) if 'constants' in kwargs else [] names = list(kwargs['constants'].keys()) if 'constants' in kwargs else []
constants = list(kwargs['constants'].values()) if 'constants' in kwargs else [] constants = list(kwargs['constants'].values()) if 'constants' in kwargs else []
torch.ops.triton.launch_kernel(self.op_id, device, params, names, constants) torch.ops.triton.launch_kernel(self.op_id, device, params, names, constants)
if 'TRITON_DEBUG_MODE' in os.environ:
for i in range(len(args)):
if isinstance(args[i], torch.Tensor):
_args[i].copy_(args[i])