Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Slab Hash version of GPU Coordinates Map Powered MinkowskiEngine #220

Open
wants to merge 1 commit into
base: master
Choose a base branch
from
Open

Slab Hash version of GPU Coordinates Map Powered MinkowskiEngine #220

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
44 changes: 41 additions & 3 deletions Makefile
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,7 +8,7 @@ Q ?= @
# CPU_ONLY := 1

CXX ?= g++
PYTHON ?= python
PYTHON ?= python3.6

EXTENSION_NAME := minkowski

Expand Down Expand Up @@ -66,17 +66,27 @@ ifneq ($(CPU_ONLY), 1)
endif

SRC_DIR := ./src
SRC_GPU_COORDS_MAP_DIR := ./src/3rdparty/gpu_coords_map/include
SRC_SLAB_HASH_DIR := ./src/3rdparty/gpu_coords_map/include/slab_hash
OBJ_DIR := ./objs
CPP_SRCS := $(wildcard $(SRC_DIR)/*.cpp)
CPP_SRCS_GPU_COORDS_MAP := $(wildcard $(SRC_GPU_COORDS_MAP_DIR)/*.cpp)
CPP_SRCS_SLAB_HASH:= $(wildcard $(SRC_SLAB_HASH_DIR)/*.cpp)
CU_SRCS := $(wildcard $(SRC_DIR)/*.cu)
CU_SRCS_GPU_COORDS_MAP := $(wildcard $(SRC_GPU_COORDS_MAP_DIR)/*.cu)
CU_SRCS_SLAB_HASH:= $(wildcard $(SRC_SLAB_HASH_DIR)/*.cu)
OBJS := $(patsubst $(SRC_DIR)/%.cpp,$(OBJ_DIR)/%.o,$(CPP_SRCS))
OBJS_GPU_COORDS_MAP := $(patsubst $(SRC_GPU_COORDS_MAP_DIR)/%.cpp,$(OBJ_DIR)/3rdparty/gpu_coords_map/include/%.o,$(CPP_SRCS_GPU_COORDS_MAP))
OBJS_SLAB_HASH := $(patsubst $(SRC_SLAB_HASH_DIR)/%.cpp,$(OBJ_DIR)/3rdparty/gpu_coords_map/include/slab_hash/%.o,$(CPP_SRCS_SLAB_HASH))
CU_OBJS := $(patsubst $(SRC_DIR)/%.cu,$(OBJ_DIR)/cuda/%.o,$(CU_SRCS))
CU_OBJS_GPU_COORDS_MAP := $(patsubst $(SRC_GPU_COORDS_MAP_DIR)/%.cu,$(OBJ_DIR)/3rdparty/gpu_coords_map/include/cuda/%.o,$(CU_SRCS_GPU_COORDS_MAP))
CU_OBJS_SLAB_HASH := $(patsubst $(SRC_SLAB_HASH_DIR)/%.cu,$(OBJ_DIR)/3rdparty/gpu_coords_map/include/slab_hash/cuda/%.o,$(CU_SRCS_SLAB_HASH))
STATIC_LIB := $(OBJ_DIR)/lib$(EXTENSION_NAME).a

# We will also explicitly add stdc++ to the link target.
LIBRARIES := stdc++ c10 caffe2 torch torch_python _C
ifneq ($(CPU_ONLY), 1)
LIBRARIES += cudart cublas cusparse caffe2_gpu c10_cuda
LIBRARIES += cudadevrt cudart cudadevrt cublas cudadevrt cusparse cudadevrt caffe2_gpu cudadevrt c10_cuda cudadevrt
CUDA_ARCH := -gencode arch=compute_30,code=sm_30 \
-gencode arch=compute_35,code=sm_35 \
-gencode=arch=compute_50,code=sm_50 \
Expand Down Expand Up @@ -118,6 +128,7 @@ ifeq ($(DEBUG), 1)
COMMON_FLAGS += -DDEBUG -g -O0
# https://gcoe-dresden.de/reaching-the-shore-with-a-fog-warning-my-eurohack-day-4-morning-session/
NVCCFLAGS := -g -G # -rdc true
# NVCCFLAGS := -g -G -rdc true
else
COMMON_FLAGS += -DNDEBUG -O3
endif
Expand All @@ -140,6 +151,7 @@ COMMON_FLAGS += $(foreach includedir,$(INCLUDE_DIRS),-I$(includedir)) \

CXXFLAGS += -fopenmp -fPIC -fwrapv -std=c++14 $(COMMON_FLAGS) $(WARNINGS)
NVCCFLAGS += -std=c++14 -ccbin=$(CXX) -Xcompiler -fPIC $(COMMON_FLAGS)
NVCCFLAGS += -rdc true
LINKFLAGS += -pthread -fPIC $(WARNINGS) -Wl,-rpath=$(PYTHON_LIB_DIR) -Wl,--no-as-needed -Wl,--sysroot=/
LDFLAGS += $(foreach librarydir,$(LIBRARY_DIRS),-L$(librarydir)) \
$(foreach library,$(LIBRARIES),-l$(library))
Expand All @@ -148,7 +160,7 @@ ifeq ($(CPU_ONLY), 1)
ALL_OBJS := $(OBJS)
CXXFLAGS += -DCPU_ONLY
else
ALL_OBJS := $(OBJS) $(CU_OBJS)
ALL_OBJS := $(OBJS) $(OBJS_GPU_COORDS_MAP) $(OBJS_SLAB_HASH) $(CU_OBJS) $(CU_OBJS_GPU_COORDS_MAP) $(CU_OBJS_SLAB_HASH)
endif

all: $(STATIC_LIB)
Expand All @@ -157,8 +169,19 @@ all: $(STATIC_LIB)
$(OBJ_DIR):
@ mkdir -p $@
@ mkdir -p $@/cuda
@ mkdir -p $@/3rdparty/gpu_coords_map/include/cuda
@ mkdir -p $@/3rdparty/gpu_coords_map/include/slab_hash/cuda

$(OBJ_DIR)/%.o: $(SRC_DIR)/%.cpp | $(OBJ_DIR)
@ echo CXX $<
@ echo $(CXXFLAGS)
$(Q)$(CXX) $< $(CXXFLAGS) -c -o $@

$(OBJ_DIR)/3rdparty/gpu_coords_map/include/%.o: $(SRC_GPU_COORDS_MAP_DIR)/%.cpp | $(OBJ_DIR)
@ echo CXX $<
$(Q)$(CXX) $< $(CXXFLAGS) -c -o $@

$(OBJ_DIR)/3rdparty/gpu_coords_map/include/slab_hash/%.o: $(SRC_SLAB_HASH_DIR)/%.cpp | $(OBJ_DIR)
@ echo CXX $<
$(Q)$(CXX) $< $(CXXFLAGS) -c -o $@

Expand All @@ -168,8 +191,23 @@ $(OBJ_DIR)/cuda/%.o: $(SRC_DIR)/%.cu | $(OBJ_DIR)
-odir $(@D)
$(Q)$(NVCC) $(NVCCFLAGS) $(CUDA_ARCH) -c $< -o $@

$(OBJ_DIR)/3rdparty/gpu_coords_map/include/cuda/%.o: $(SRC_GPU_COORDS_MAP_DIR)/%.cu | $(OBJ_DIR)
@ echo NVCC $<
$(Q)$(NVCC) $(NVCCFLAGS) $(CUDA_ARCH) -M $< -o ${@:.o=.d} \
-odir $(@D)
$(Q)$(NVCC) $(NVCCFLAGS) $(CUDA_ARCH) -c $< -o $@

$(OBJ_DIR)/3rdparty/gpu_coords_map/include/slab_hash/cuda/%.o: $(SRC_SLAB_HASH_DIR)/%.cu| $(OBJ_DIR)
@ echo NVCC $<
$(Q)$(NVCC) $(NVCCFLAGS) $(CUDA_ARCH) -M $< -o ${@:.o=.d} \
-odir $(@D)
$(Q)$(NVCC) $(NVCCFLAGS) $(CUDA_ARCH) -c $< -o $@

$(STATIC_LIB): $(ALL_OBJS) | $(OBJ_DIR)
$(RM) -f $(STATIC_LIB)
@ echo $(LINKFLAGS)
@ echo $(LDFLAGS)
@ echo $(CXXFLAGS)
@ echo LD -o $@
ar rc $(STATIC_LIB) $(ALL_OBJS)

Expand Down
25 changes: 18 additions & 7 deletions MinkowskiEngine/MinkowskiCoords.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,7 +35,8 @@
if 'OMP_NUM_THREADS' in os.environ:
CPU_COUNT = int(os.environ['OMP_NUM_THREADS'])

_memory_manager_backend = MemoryManagerBackend.PYTORCH
#_memory_manager_backend = MemoryManagerBackend.PYTORCH
_memory_manager_backend = MemoryManagerBackend.CUDA


def set_memory_manager_backend(backend: MemoryManagerBackend):
Expand Down Expand Up @@ -102,7 +103,8 @@ class CoordsManager():
def __init__(self,
num_threads: int = -1,
memory_manager_backend: MemoryManagerBackend = None,
D: int = -1):
D: int = -1,
device: str = 'cuda'):
if D < 1:
raise ValueError(f"Invalid dimension {D}")
self.D = D
Expand All @@ -111,7 +113,9 @@ def __init__(self,
if memory_manager_backend is None:
global _memory_manager_backend
memory_manager_backend = _memory_manager_backend
coords_man = MEB.CoordsManager(num_threads, memory_manager_backend)
coords_man = MEB.CoordsManager(num_threads, memory_manager_backend) \
if device == 'cpu' else \
MEB.GPUCoordsManager(D, 0, memory_manager_backend)
self.CPPCoordsManager = coords_man

def initialize(self,
Expand All @@ -120,14 +124,15 @@ def initialize(self,
force_creation: bool = False,
force_remap: bool = False,
allow_duplicate_coords: bool = False,
return_inverse: bool = False) -> torch.LongTensor:
return_inverse: bool = False) -> torch.IntTensor:
assert isinstance(coords_key, CoordsKey)
unique_index = torch.LongTensor()
inverse_mapping = torch.LongTensor()
# TODO(ljm): Adjust cpu interface from long to int acoordingly
unique_index = torch.IntTensor()
inverse_mapping = torch.IntTensor()
self.CPPCoordsManager.initializeCoords(
coords, unique_index, inverse_mapping, coords_key.CPPCoordsKey,
force_creation, force_remap, allow_duplicate_coords, return_inverse)
return unique_index, inverse_mapping
return unique_index.long(), inverse_mapping.long()

def create_coords_key(self,
coords: torch.IntTensor,
Expand Down Expand Up @@ -171,6 +176,9 @@ def stride(self,
def reduce(self):
origin_key = CoordsKey(self.D)
origin_key.setTensorStride(convert_to_int_list(0, self.D))
# TODO(ljm): Get batch_size by createOriginCoords
# TODO(ljm): Find a better way to get batch_size
# Notice(ljm): It can be concluded that the batch indices are contigous by GetCoordsAt
origin_key.setKey(self.CPPCoordsManager.createOriginCoords(self.D))
return origin_key

Expand Down Expand Up @@ -322,6 +330,9 @@ def get_kernel_map(self,
is_transpose,
is_pool)

kernel_map[0] = kernel_map[0].long()
kernel_map[1] = kernel_map[1].long()

return kernel_map

def get_coords_map(self, in_key_or_tensor_strides,
Expand Down
9 changes: 7 additions & 2 deletions MinkowskiEngine/SparseTensor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -229,6 +229,7 @@ def __init__(
of the current sparse tensor. By default, it is 1.

"""
print(coords)
assert isinstance(feats,
torch.Tensor), "Features must be a torch.Tensor"
assert feats.ndim == 2, f"The feature should be a matrix, The input feature is an order-{feats.ndim} tensor."
Expand All @@ -254,12 +255,14 @@ def __init__(
assert isinstance(coords, torch.Tensor), \
"Coordinate must be of type torch.Tensor"

print(isinstance(coords, torch.IntTensor))
if not isinstance(coords, torch.IntTensor):
warnings.warn(
'Coords implicitly converted to torch.IntTensor. ' +
'To remove this warning, use `.int()` to convert the ' +
'coords into an torch.IntTensor')
coords = torch.floor(coords).int()
print(isinstance(coords, torch.IntTensor))
# coords = torch.floor(coords).int()

if coords.device.type != 'cpu':
warnings.warn(
Expand All @@ -283,7 +286,9 @@ def __init__(
if _global_coords_man is None:
_global_coords_man = CoordsManager(
memory_manager_backend=memory_manager_backend,
D=coords.size(1) - 1)
D=coords.size(1) - 1,
device=coords.device.type if coords is not None else 'cuda')
# TODO(ljm): handle device when coords is None
coords_manager = _global_coords_man
else:
assert coords is not None, "Initial coordinates must be given"
Expand Down
Loading