seemingly working fp32 implementation
This commit is contained in:
89
hw/unittest/tensor/Makefile
Normal file
89
hw/unittest/tensor/Makefile
Normal file
@@ -0,0 +1,89 @@
|
||||
DESTDIR ?= .
|
||||
RTL_DIR = ../../rtl
|
||||
DPI_DIR = $(abspath ../../dpi)
|
||||
SIM_DIR = ../../../sim
|
||||
THIRD_PARTY_DIR = $(abspath ../../../third_party)
|
||||
|
||||
CONFIGS +=
|
||||
PARAMS +=
|
||||
|
||||
CXXFLAGS += -std=c++17 -Wall -Wextra -Wfatal-errors -Wno-array-bounds
|
||||
CXXFLAGS += -fPIC -Wno-maybe-uninitialized
|
||||
CXXFLAGS += -fcoroutines
|
||||
CXXFLAGS += -I../../.. -I../../common -I../../../../sim/common
|
||||
CXXFLAGS += -I/$(THIRD_PARTY_DIR)/softfloat/source/include
|
||||
CXXFLAGS += -I/$(DPI_DIR)
|
||||
CXXFLAGS += $(CONFIGS)
|
||||
|
||||
LDFLAGS += $(THIRD_PARTY_DIR)/softfloat/build/Linux-x86_64-GCC/softfloat.a
|
||||
|
||||
# control RTL debug tracing states
|
||||
DBG_TRACE_FLAGS += -DDBG_TRACE_CACHE_BANK
|
||||
DBG_TRACE_FLAGS += -DDBG_TRACE_CACHE_MSHR
|
||||
DBG_TRACE_FLAGS += -DDBG_TRACE_CACHE_TAG
|
||||
DBG_TRACE_FLAGS += -DDBG_TRACE_CACHE_DATA
|
||||
|
||||
DBG_FLAGS += -DDEBUG_LEVEL=$(DEBUG) -DVCD_OUTPUT $(DBG_TRACE_FLAGS)
|
||||
|
||||
RTL_PKGS = $(RTL_DIR)/VX_gpu_pkg.sv
|
||||
|
||||
RTL_INCLUDE = -I$(RTL_DIR) -I$(DPI_DIR) -I$(RTL_DIR)/libs -I$(RTL_DIR)/fpu
|
||||
|
||||
# SRCS = cachesim.cpp testbench.cpp
|
||||
SRCS += $(DPI_DIR)/util_dpi.cpp
|
||||
SRCS += $(DPI_DIR)/float_dpi.cpp
|
||||
SRCS += $(SIM_DIR)/common/rvfloats.cpp
|
||||
SRCS += ./main.cpp
|
||||
|
||||
RTL_SRCS += $(RTL_DIR)/fpu/VX_tensor_core.sv
|
||||
RTL_SRCS += $(RTL_DIR)/fpu/VX_tensor_tb.sv
|
||||
|
||||
TOP = VX_tensor_tb
|
||||
|
||||
VL_FLAGS = --exe
|
||||
VL_FLAGS += --language 1800-2009 # -Wall -Wpedantic # --assert
|
||||
VL_FLAGS += -Wno-DECLFILENAME -Wno-REDEFMACRO
|
||||
VL_FLAGS += --x-initial unique --x-assign unique
|
||||
VL_FLAGS += -DSIMULATION -DSV_DPI
|
||||
VL_FLAGS += $(CONFIGS)
|
||||
VL_FLAGS += $(PARAMS)
|
||||
VL_FLAGS += $(RTL_INCLUDE)
|
||||
VL_FLAGS += $(RTL_PKGS)
|
||||
VL_FLAGS += --cc $(TOP) --top-module $(TOP)
|
||||
VL_FLAGS += --timing
|
||||
|
||||
# Enable Verilator multithreaded simulation
|
||||
THREADS ?= $(shell python -c 'import multiprocessing as mp; print(mp.cpu_count())')
|
||||
VL_FLAGS += -j $(THREADS)
|
||||
#VL_FLAGS += --threads $(THREADS)
|
||||
|
||||
# Debugigng
|
||||
ifdef DEBUG
|
||||
VL_FLAGS += --trace --trace-structs $(DBG_FLAGS)
|
||||
CXXFLAGS += -g -O0 $(DBG_FLAGS)
|
||||
else
|
||||
VL_FLAGS += -DNDEBUG
|
||||
CXXFLAGS += -O2 -DNDEBUG
|
||||
endif
|
||||
|
||||
# Enable perf counters
|
||||
ifdef PERF
|
||||
VL_FLAGS += -DPERF_ENABLE
|
||||
CXXFLAGS += -DPERF_ENABLE
|
||||
endif
|
||||
|
||||
PROJECT = tensor
|
||||
|
||||
all: $(DESTDIR)/$(PROJECT)
|
||||
|
||||
$(DESTDIR)/$(PROJECT): $(SRCS) $(RTL_SRCS)
|
||||
verilator --build $(VL_FLAGS) $(SRCS) -CFLAGS '$(CXXFLAGS)' -LDFLAGS '$(LDFLAGS)' -o ../$@
|
||||
|
||||
run: $(DESTDIR)/$(PROJECT)
|
||||
$(DESTDIR)/$(PROJECT)
|
||||
|
||||
waves: trace.vcd
|
||||
gtkwave -o trace.vcd
|
||||
|
||||
clean:
|
||||
rm -rf obj_dir $(DESTDIR)/$(PROJECT)
|
||||
Reference in New Issue
Block a user