vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

make : simplify nvcc flags

Browse source 
CUDA_CXXFLAGS and HOST_CXXFLAGS can no longer be overridden via the
command line, but NVCCFLAGS now can.
This commit is contained in:
Jared Van Bortel 2023-12-11 14:14:48 -05:00
parent 8a7b2fa528
commit 9b28f3413b
1 changed files with 38 additions and 44 deletions
Download patch file Download diff file Expand all files Collapse all files

82
Makefile
View file

@ -121,12 +121,12 @@ MK_CXXFLAGS = -std=c++11 -fPIC
# -Ofast tends to produce faster code, but may not be available for some compilers. # -Ofast tends to produce faster code, but may not be available for some compilers.
ifdef LLAMA_FAST ifdef LLAMA_FAST
MK_CFLAGS += -Ofast MK_CFLAGS += -Ofast
MK_HOST_CXXFLAGS += -Ofast HOST_CXXFLAGS += -Ofast
MK_CUDA_CXXFLAGS += -O3 MK_NVCCFLAGS += -O3
else else
MK_CFLAGS += -O3 MK_CFLAGS += -O3
MK_CXXFLAGS += -O3 MK_CXXFLAGS += -O3
endif endif
# clock_gettime came in POSIX.1b (1993) # clock_gettime came in POSIX.1b (1993)
@ -222,8 +222,8 @@ MK_CXXFLAGS += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
ifeq ($(CC_IS_CLANG), 1) ifeq ($(CC_IS_CLANG), 1)
# clang options # clang options
MK_CFLAGS += -Wunreachable-code-break -Wunreachable-code-return MK_CFLAGS += -Wunreachable-code-break -Wunreachable-code-return
MK_HOST_CXXFLAGS += -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi HOST_CXXFLAGS += -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi
ifneq '' '$(and $(CC_IS_LLVM_CLANG),$(filter 1,$(shell expr $(CC_VER) \>= 030800)))' ifneq '' '$(and $(CC_IS_LLVM_CLANG),$(filter 1,$(shell expr $(CC_VER) \>= 030800)))'
MK_CFLAGS += -Wdouble-promotion MK_CFLAGS += -Wdouble-promotion
@ -233,14 +233,14 @@ ifeq ($(CC_IS_CLANG), 1)
endif endif
else else
# gcc options # gcc options
MK_CFLAGS += -Wdouble-promotion MK_CFLAGS += -Wdouble-promotion
MK_HOST_CXXFLAGS += -Wno-array-bounds HOST_CXXFLAGS += -Wno-array-bounds
ifeq ($(shell expr $(CC_VER) \>= 070100), 1) ifeq ($(shell expr $(CC_VER) \>= 070100), 1)
MK_HOST_CXXFLAGS += -Wno-format-truncation HOST_CXXFLAGS += -Wno-format-truncation
endif endif
ifeq ($(shell expr $(CC_VER) \>= 080100), 1) ifeq ($(shell expr $(CC_VER) \>= 080100), 1)
MK_HOST_CXXFLAGS += -Wextra-semi HOST_CXXFLAGS += -Wextra-semi
endif endif
endif endif
@ -294,8 +294,8 @@ ifndef RISCV
ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64)) ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
# Use all CPU extensions that are available: # Use all CPU extensions that are available:
MK_CFLAGS += -march=native -mtune=native MK_CFLAGS += -march=native -mtune=native
MK_HOST_CXXFLAGS += -march=native -mtune=native HOST_CXXFLAGS += -march=native -mtune=native
# Usage AVX-only # Usage AVX-only
#MK_CFLAGS += -mfma -mf16c -mavx #MK_CFLAGS += -mfma -mf16c -mavx
@ -395,61 +395,59 @@ ifdef LLAMA_CUBLAS
MK_CPPFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include MK_CPPFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
MK_LDFLAGS += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib MK_LDFLAGS += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
OBJS += ggml-cuda.o OBJS += ggml-cuda.o
NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math MK_NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
ifdef LLAMA_CUDA_NVCC ifdef LLAMA_CUDA_NVCC
NVCC = $(LLAMA_CUDA_NVCC) NVCC = $(LLAMA_CUDA_NVCC)
else else
NVCC = nvcc NVCC = nvcc
endif #LLAMA_CUDA_NVCC endif #LLAMA_CUDA_NVCC
ifdef CUDA_DOCKER_ARCH ifdef CUDA_DOCKER_ARCH
NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH) MK_NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
else ifdef CUDA_POWER_ARCH else ifndef CUDA_POWER_ARCH
NVCCFLAGS += MK_NVCCFLAGS += -arch=native
else
NVCCFLAGS += -arch=native
endif # CUDA_DOCKER_ARCH endif # CUDA_DOCKER_ARCH
ifdef LLAMA_CUDA_FORCE_DMMV ifdef LLAMA_CUDA_FORCE_DMMV
NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV MK_NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
endif # LLAMA_CUDA_FORCE_DMMV endif # LLAMA_CUDA_FORCE_DMMV
ifdef LLAMA_CUDA_FORCE_MMQ ifdef LLAMA_CUDA_FORCE_MMQ
NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ MK_NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
endif # LLAMA_CUDA_FORCE_MMQ endif # LLAMA_CUDA_FORCE_MMQ
ifdef LLAMA_CUDA_DMMV_X ifdef LLAMA_CUDA_DMMV_X
NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X) MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
else else
NVCCFLAGS += -DGGML_CUDA_DMMV_X=32 MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
endif # LLAMA_CUDA_DMMV_X endif # LLAMA_CUDA_DMMV_X
ifdef LLAMA_CUDA_MMV_Y ifdef LLAMA_CUDA_MMV_Y
NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y) MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
else ifdef LLAMA_CUDA_DMMV_Y else ifdef LLAMA_CUDA_DMMV_Y
NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_DMMV_Y) # for backwards compatibility MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_DMMV_Y) # for backwards compatibility
else else
NVCCFLAGS += -DGGML_CUDA_MMV_Y=1 MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
endif # LLAMA_CUDA_MMV_Y endif # LLAMA_CUDA_MMV_Y
ifdef LLAMA_CUDA_F16 ifdef LLAMA_CUDA_F16
NVCCFLAGS += -DGGML_CUDA_F16 MK_NVCCFLAGS += -DGGML_CUDA_F16
endif # LLAMA_CUDA_F16 endif # LLAMA_CUDA_F16
ifdef LLAMA_CUDA_DMMV_F16 ifdef LLAMA_CUDA_DMMV_F16
NVCCFLAGS += -DGGML_CUDA_F16 MK_NVCCFLAGS += -DGGML_CUDA_F16
endif # LLAMA_CUDA_DMMV_F16 endif # LLAMA_CUDA_DMMV_F16
ifdef LLAMA_CUDA_KQUANTS_ITER ifdef LLAMA_CUDA_KQUANTS_ITER
NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER) MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
else else
NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2 MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
endif endif
ifdef LLAMA_CUDA_PEER_MAX_BATCH_SIZE ifdef LLAMA_CUDA_PEER_MAX_BATCH_SIZE
NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(LLAMA_CUDA_PEER_MAX_BATCH_SIZE) MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(LLAMA_CUDA_PEER_MAX_BATCH_SIZE)
else else
NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE
#ifdef LLAMA_CUDA_CUBLAS #ifdef LLAMA_CUDA_CUBLAS
# NVCCFLAGS += -DGGML_CUDA_CUBLAS # MK_NVCCFLAGS += -DGGML_CUDA_CUBLAS
#endif # LLAMA_CUDA_CUBLAS #endif # LLAMA_CUDA_CUBLAS
ifdef LLAMA_CUDA_CCBIN ifdef LLAMA_CUDA_CCBIN
NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN) MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
endif endif
ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
$(NVCC) $(NVCCFLAGS) -c $< -o $@ $(NVCC) $(BASE_CXXFLAGS) $(NVCCFLAGS) -Wno-pedantic -Xcompiler "$(HOST_CXXFLAGS)" -c $< -o $@
endif # LLAMA_CUBLAS endif # LLAMA_CUBLAS
ifdef LLAMA_CLBLAST ifdef LLAMA_CLBLAST
@ -512,15 +510,11 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h
endif # LLAMA_MPI endif # LLAMA_MPI
# combine build flags with cmdline overrides # combine build flags with cmdline overrides
override CFLAGS := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(CFLAGS) override CFLAGS := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(CFLAGS)
override CXXFLAGS := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS) BASE_CXXFLAGS := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
override CUDA_CXXFLAGS := $(MK_CUDA_CXXFLAGS) $(CUDA_CXXFLAGS) override CXXFLAGS := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS)
override HOST_CXXFLAGS := $(MK_HOST_CXXFLAGS) $(HOST_CXXFLAGS) override NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
override LDFLAGS := $(MK_LDFLAGS) $(LDFLAGS) override LDFLAGS := $(MK_LDFLAGS) $(LDFLAGS)
# save CXXFLAGS before we add host-only options
NVCCFLAGS := $(NVCCFLAGS) $(CXXFLAGS) $(CUDA_CXXFLAGS) -Wno-pedantic -Xcompiler "$(HOST_CXXFLAGS)"
override CXXFLAGS += $(HOST_CXXFLAGS)
# #
# Print build information # Print build information