Skip to content

Commit

Permalink
Adjusted Clang format (#585)
Browse files Browse the repository at this point in the history 
Other: allow to experiment with untuned/default parameters
* Accept to specify a generic GPU (CUDA/HIP).
  • Loading branch information
@hfp
hfp authored Mar 17, 2022
1 parent f41dc4d commit 6248c55
Show file tree
Hide file tree
Showing 8 changed files with 151 additions and 116 deletions.
7 changes: 4 additions & 3 deletions .clang-format
View file
Original file line number Diff line number Diff line change
Expand Up @@ -3,11 +3,12 @@ AlignAfterOpenBracket: DontAlign
AlignEscapedNewlines: DontAlign
AlignTrailingComments: false
AllowShortCaseLabelsOnASingleLine: true
AllowShortIfStatementsOnASingleLine: WithoutElse
AllowShortIfStatementsOnASingleLine: AllIfsAndElse
AllowShortLoopsOnASingleLine: true
BraceWrapping:
BeforeCatch: true
BeforeElse: true
AfterControlStatement: MultiLine
BeforeCatch: true
BeforeElse: true
BreakBeforeBraces: Custom
ColumnLimit: 132
ConstructorInitializerIndentWidth: 0
Expand Down
82 changes: 43 additions & 39 deletions src/acc/cuda/Makefile
View file
Original file line number Diff line number Diff line change
Expand Up @@ -2,25 +2,27 @@
# It is for testing and comparison with other implementations.

MAKDIR := $(subst //,,$(dir $(firstword $(MAKEFILE_LIST)))/)
INCACC := $(wildcard $(MAKDIR)/*.h*) $(MAKDIR)/../acc.h
SRCACC := $(wildcard $(MAKDIR)/../cuda_hip/*.cpp) \
ACCDIR := $(MAKDIR)/..
DIRSMM := $(ACCDIR)/libsmm_acc
INCACC := $(wildcard $(MAKDIR)/*.h*) $(ACCDIR)/acc.h
SRCACC := $(wildcard $(ACCDIR)/cuda_hip/*.cpp) \
$(wildcard $(MAKDIR)/*.cpp) \
$(NULL)
OBJACC := $(SRCACC:.cpp=.o)

GPUSMM := $(wildcard $(MAKDIR)/../libsmm_acc/kernels/*.h*)
INCSMM := $(wildcard $(MAKDIR)/../libsmm_acc/*.h*) \
$(MAKDIR)/../libsmm_acc/smm_acc_kernels.h \
$(MAKDIR)/../libsmm_acc/parameters.h \
$(MAKDIR)/../acc_libsmm.h \
$(MAKDIR)/../acc_bench.h \
GPUSMM := $(wildcard $(DIRSMM)/kernels/*.h*)
INCSMM := $(wildcard $(DIRSMM)/*.h*) \
$(DIRSMM)/parameters.h \
$(DIRSMM)/smm_acc_kernels.h \
$(ACCDIR)/acc_libsmm.h \
$(ACCDIR)/acc_bench.h \
$(NULL)
SRCSMM := $(wildcard $(MAKDIR)/../libsmm_acc/*.cpp)
SRCSMM := $(wildcard $(DIRSMM)/*.cpp)
OBJSMM := $(SRCSMM:.cpp=.o)

INCALL := $(INCACC) $(INCSMM)

LIBXSMMROOT := $(wildcard $(MAKDIR)/../../../../libxsmm)
LIBXSMMROOT := $(wildcard $(ACCDIR)/../../../libxsmm)
ifeq (,$(LIBXSMMROOT))
LIBXSMMROOT := $(wildcard $(HOME)/libxsmm)
endif
Expand Down Expand Up @@ -63,7 +65,7 @@ else ifeq ($(WITH_GPU),P100)
else ifeq ($(WITH_GPU),V100)
ARCH_NUMBER = 70
else ifeq ($(WITH_GPU),A100)
# TODO: update when tuned parameters for A100 available
# TODO: update for A100 tuned parameters
override WITH_GPU := V100
ARCH_NUMBER = 80
else ifeq (,$(ARCH_NUMBER))
Expand Down Expand Up @@ -167,22 +169,22 @@ LDFLAGS += -lcudart -lcublas -lnvrtc -lcuda
CXXFLAGS += -std=c++11 $(CFLAGS)

.PHONY: bench
bench: $(MAKDIR)/../acc_bench_smm $(MAKDIR)/../acc_bench_trans
bench: $(ACCDIR)/acc_bench_smm $(ACCDIR)/acc_bench_trans

.PHONY: all
all: bench $(MAKDIR)/../dbcsr_acc_test
all: bench $(ACCDIR)/dbcsr_acc_test

.PHONY: test
test: test-interface test-trans test-smm

.PHONY: test-interface
test-interface: $(MAKDIR)/../dbcsr_acc_test
test-interface: $(ACCDIR)/dbcsr_acc_test
@echo "--- DBCSR Backend Interface"
$(MAKDIR)/../dbcsr_acc_test
$(ACCDIR)/dbcsr_acc_test

.PHONY: test-trans
test-trans: bench
$(eval SHAPES = $(shell $(MAKDIR)/../acc_triplets.sh -k $(SPECID) -m $(MAXEXT) -n $(NTRANS) -a))
$(eval SHAPES = $(shell $(ACCDIR)/acc_triplets.sh -k $(SPECID) -m $(MAXEXT) -n $(NTRANS) -a))
@echo "--- DBCSR CUDA Transposes ($(words $(SHAPES)))"
@echo "LD_LIBRARY_PATH=${LD_LIBRARY_PATH}"
ifneq (,$(LD_PRELOAD))
Expand All @@ -191,16 +193,16 @@ endif
@echo "CXX: $$($(CXX) --version | head -n1)"
@echo "CC: $$($(CC) --version | head -n1)"
@echo "runtime libraries:"
@ldd $(MAKDIR)/../acc_bench_trans
@ldd $(ACCDIR)/acc_bench_trans
@echo "hostname: $$(hostname)"
@echo
@for SHAPE in $(SHAPES); do \
$(MAKDIR)/../acc_bench_trans $${SHAPE} || exit 1; \
$(ACCDIR)/acc_bench_trans $${SHAPE} || exit 1; \
echo; \
done

$(MAKDIR)/test-smm.log: bench
$(eval SHAPES = $(shell $(MAKDIR)/../acc_triplets.sh -k $(SPECID) -m $(MAXEXT) -n $(NSMMS)))
$(eval SHAPES = $(shell $(ACCDIR)/acc_triplets.sh -k $(SPECID) -m $(MAXEXT) -n $(NSMMS)))
@echo "--- DBCSR CUDA SMMs ($(words $(SHAPES)))"
@echo "LD_LIBRARY_PATH=${LD_LIBRARY_PATH}"
ifneq (,$(LD_PRELOAD))
Expand All @@ -209,11 +211,11 @@ endif
@echo "CXX: $$($(CXX) --version | head -n1)"
@echo "CC: $$($(CC) --version | head -n1)"
@echo "runtime libraries:"
@ldd $(MAKDIR)/../acc_bench_smm
@ldd $(ACCDIR)/acc_bench_smm
@echo "hostname: $$(hostname)"
@echo
@echo "$(SHAPES)" | xargs -n1 | \
(CHECK=$(if $(CHECK),$(CHECK),1) stdbuf --output=L $(MAKDIR)/../acc_bench_smm /dev/stdin \
(CHECK=$(if $(CHECK),$(CHECK),1) stdbuf --output=L $(ACCDIR)/acc_bench_smm /dev/stdin \
2>$(MAKDIR)/test-smm.err && rm $(MAKDIR)/test-smm.err) | tee $@
@if [ -s $(MAKDIR)/test-smm.err ]; then cat $(MAKDIR)/test-smm.err && exit 1; fi

Expand All @@ -227,42 +229,44 @@ endif
@echo "mean: $$(sed -n "/device:/p" $< | datamash -W -R 1 mean 4) GFLOPS/s"
endif

$(MAKDIR)/../libsmm_acc/parameters.h: $(MAKDIR)/Makefile $(MAKDIR)/../libsmm_acc/generate_parameters.py $(MAKDIR)/../libsmm_acc/parameters/parameters_$(WITH_GPU).json
@cd $(MAKDIR)/../libsmm_acc && $(PYTHON) ../libsmm_acc/generate_parameters.py --gpu_version=$(WITH_GPU) --base_dir=../libsmm_acc/parameters
PARDIR := $(DIRSMM)/parameters
PARAMS := $(wildcard $(PARDIR)/parameters_$(WITH_GPU).json)
$(DIRSMM)/parameters.h: $(MAKDIR)/Makefile $(DIRSMM)/generate_parameters.py $(PARAMS)
@cd $(DIRSMM) && $(PYTHON) ../libsmm_acc/generate_parameters.py --gpu_version=$(WITH_GPU) --base_dir=../libsmm_acc/parameters

$(MAKDIR)/../libsmm_acc/smm_acc_kernels.h: $(GPUSMM) $(MAKDIR)/Makefile $(MAKDIR)/../libsmm_acc/generate_kernels.py $(MAKDIR)/../libsmm_acc/parameters/parameters_$(WITH_GPU).json
@cd $(MAKDIR)/../libsmm_acc && $(PYTHON) ../libsmm_acc/generate_kernels.py ../libsmm_acc/kernels
$(DIRSMM)/smm_acc_kernels.h: $(GPUSMM) $(MAKDIR)/Makefile $(DIRSMM)/generate_kernels.py $(PARAMS)
@cd $(DIRSMM) && $(PYTHON) ../libsmm_acc/generate_kernels.py ../libsmm_acc/kernels

$(MAKDIR)/../dbcsr_acc.a: $(OBJACC) $(MAKDIR)/../libsmm_acc/libsmm_acc_init.o
$(ACCDIR)/dbcsr_acc.a: $(OBJACC) $(DIRSMM)/libsmm_acc_init.o
$(AR) -rs $@ $^

$(MAKDIR)/../dbcsr_acc_smm.a: $(OBJSMM)
$(ACCDIR)/dbcsr_acc_smm.a: $(OBJSMM)
$(AR) -rs $@ $^

%.o: %.cpp $(INCALL) $(MAKDIR)/Makefile
$(CXX) $(CXXFLAGS) $(CFLAGS_XSMM) -c $< -o $@

$(MAKDIR)/acc_bench_smm.o: $(MAKDIR)/../acc_bench_smm.c $(MAKDIR)/Makefile
$(MAKDIR)/acc_bench_smm.o: $(ACCDIR)/acc_bench_smm.c $(MAKDIR)/Makefile
ifneq (0,$(LIBXSMM))
$(CC) $(CFLAGS) $(CFLAGS_XSMM) -c $< -o $@
else
$(CC) $(CFLAGS) -c $< -o $@
endif
$(MAKDIR)/../acc_bench_smm: $(MAKDIR)/acc_bench_smm.o $(MAKDIR)/../dbcsr_acc.a $(MAKDIR)/../dbcsr_acc_smm.a
$(ACCDIR)/acc_bench_smm: $(MAKDIR)/acc_bench_smm.o $(ACCDIR)/dbcsr_acc.a $(ACCDIR)/dbcsr_acc_smm.a
$(CXX) $^ $(LDFLAGS) -o $@

$(MAKDIR)/acc_bench_trans.o: $(MAKDIR)/../acc_bench_trans.c $(MAKDIR)/Makefile
$(MAKDIR)/acc_bench_trans.o: $(ACCDIR)/acc_bench_trans.c $(MAKDIR)/Makefile
ifneq (0,$(LIBXSMM))
$(CC) $(CFLAGS) $(CFLAGS_XSMM) -c $< -o $@
else
$(CC) $(CFLAGS) -c $< -o $@
endif
$(MAKDIR)/../acc_bench_trans: $(MAKDIR)/acc_bench_trans.o $(MAKDIR)/../dbcsr_acc.a $(MAKDIR)/../dbcsr_acc_smm.a
$(ACCDIR)/acc_bench_trans: $(MAKDIR)/acc_bench_trans.o $(ACCDIR)/dbcsr_acc.a $(ACCDIR)/dbcsr_acc_smm.a
$(CXX) $^ $(LDFLAGS) -o $@

$(MAKDIR)/dbcsr_acc_test.o: $(MAKDIR)/../../../tests/dbcsr_acc_test.c $(MAKDIR)/Makefile
$(CC) $(CFLAGS) -I$(MAKDIR)/../.. -c $< -o $@
$(MAKDIR)/../dbcsr_acc_test: $(MAKDIR)/dbcsr_acc_test.o $(MAKDIR)/../dbcsr_acc.a $(MAKDIR)/../dbcsr_acc_smm.a
$(MAKDIR)/dbcsr_acc_test.o: $(ACCDIR)/../../tests/dbcsr_acc_test.c $(MAKDIR)/Makefile
$(CC) $(CFLAGS) -I$(ACCDIR)/.. -c $< -o $@
$(ACCDIR)/dbcsr_acc_test: $(MAKDIR)/dbcsr_acc_test.o $(ACCDIR)/dbcsr_acc.a $(ACCDIR)/dbcsr_acc_smm.a
$(CXX) $^ $(LDFLAGS) -o $@

.PHONY: clean
Expand All @@ -271,13 +275,13 @@ clean:
@rm -f $(MAKDIR)/dbcsr_acc_test.o
@rm -f $(MAKDIR)/acc_bench_trans.o
@rm -f $(MAKDIR)/acc_bench_smm.o
@rm -f $(MAKDIR)/../libsmm_acc/parameters.h
@rm -f $(MAKDIR)/../libsmm_acc/smm_acc_kernels.h
@rm -f $(DIRSMM)/parameters.h
@rm -f $(DIRSMM)/smm_acc_kernels.h
@rm -f $(MAKDIR)/test-smm.err

.PHONY: realclean
realclean: clean
@rm -f $(MAKDIR)/../dbcsr_acc.a $(MAKDIR)/../dbcsr_acc_smm.a
@rm -f $(MAKDIR)/../acc_bench_smm $(MAKDIR)/../acc_bench_trans
@rm -f $(MAKDIR)/../dbcsr_acc_test
@rm -f $(ACCDIR)/dbcsr_acc.a $(ACCDIR)/dbcsr_acc_smm.a
@rm -f $(ACCDIR)/acc_bench_smm $(ACCDIR)/acc_bench_trans
@rm -f $(ACCDIR)/dbcsr_acc_test
@rm -f $(MAKDIR)/test-smm.log
39 changes: 24 additions & 15 deletions src/acc/libsmm_acc/generate_parameters.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -20,28 +20,37 @@

# ===============================================================================
def main(gpu_version, base_dir):
# Read existing parameters
print("GPU version: {}".format(gpu_version))
param_fn = path.join(base_dir, "parameters_{}.json".format(gpu_version))
with open(param_fn) as f:
all_kernels = [params_dict_to_kernel(**params) for params in json.load(f)]
print(
"About to process {:,} kernels from file {}".format(len(all_kernels), param_fn)
)

# Read GPU properties (warp size)
gpu_props_fn = path.join(base_dir, "../kernels/gpu_properties.json")
arch_code = gpu_architectures[path.basename(param_fn)]
with open(gpu_props_fn) as f:
gpu_warp_size = json.load(f)[arch_code]["Threads_/_Warp"]
try: # Read existing parameters
param_fn = path.join(base_dir, "parameters_{}.json".format(gpu_version))
with open(param_fn) as f:
print("GPU version: {}".format(gpu_version))
all_kernels = [params_dict_to_kernel(**params) for params in json.load(f)]
print(
"About to process {:,} kernels from file {}".format(
len(all_kernels), param_fn
)
)
except: # noqa: E722
all_kernels = []
pass

try: # Read GPU properties (warp size)
gpu_props_fn = path.join(base_dir, "../kernels/gpu_properties.json")
arch_code = gpu_architectures[path.basename(param_fn)]
with open(gpu_props_fn) as f:
gpu_warp_size = json.load(f)[arch_code]["Threads_/_Warp"]
except: # noqa: E722
gpu_warp_size = 32
pass
print("GPU warp size: {}".format(gpu_warp_size))

# Construct output
out, all_pars = write_parameters_file(all_kernels, gpu_warp_size)

# Write to c++ header-file
file_h = "parameters.h"
print("Found {:,} kernels in file {}".format(len(all_kernels), param_fn))
if all_kernels:
print("Found {:,} kernels in file {}".format(len(all_kernels), param_fn))
print("Printing them to file {}".format(file_h))
with open(file_h, "w") as f:
f.write(out)
Expand Down
Loading

0 comments on commit 6248c55

Please sign in to comment.