#
# Arquiteturas de Alto Desempenho 2025/2026
#
# makefile for the first practical assignment (A1)
#
# makefile automatic variables:
#   $@ is the name of the target
#   $< is the name of the first prerequisite
#   $^ is the list of names of all prerequisites (without duplicates)
#

#
# CUDA installation directory --- /usr/local/cuda or $(CUDA_HOME)
#

CUDA_DIR = /opt/cuda


#
# OpenCL installation directory (for a NVidia graphics card, sama as CUDA)
#

OPENCL_DIR = $(CUDA_DIR)


#
# CUDA device architecture
#
#   GeForce GTX 1660 Ti --- sm_75
#   RTX A2000 Ada --------- sm_86
#   RTX A6000 Ada --------- sm_86
#   RTX 4070 -------------- sm_89
#

CUDA_ARCH = sm_86


#
# clean up
#

clean:
	rm -f sha1_tests
	rm -f sha1_cuda_test sha1_cuda_kernel.cubin
	rm -f coin_miner_cpu coin_miner_simd coin_miner_cuda coin_miner_cuda_kernel.cubin coin_miner_ocl
	rm -f coin_miner_wasm.js coin_miner_wasm.wasm
	rm -f benchmark
	rm -f a.out


#
# test the CUSTOM_SHA1_CODE macro
#

sha1_tests:	aad_sha1_cpu_tests.c aad_sha1.h aad_data_types.h aad_utilities.h makefile
	cc -march=native -Wall -Wshadow -Werror -O3 $< -o $@

sha1_cuda_test:	aad_sha1_cuda_test.c sha1_cuda_kernel.cubin aad_sha1.h aad_data_types.h aad_utilities.h aad_cuda_utilities.h makefile
	cc -march=native -Wall -Wshadow -Werror -O3  -I$(CUDA_DIR)/include $< -o $@ -lcuda


#
# compile the CUDA kernels
#

sha1_cuda_kernel.cubin:			aad_sha1_cuda_kernel.cu aad_sha1.h makefile
	nvcc -arch=$(CUDA_ARCH) --compiler-options -O2,-Wall -I$(CUDA_DIR)/include --cubin $< -o $@

all: sha1_tests sha1_cuda_test sha1_cuda_kernel.cubin


#
# DETI coin miners
#

coin_miner_cpu:	aad_coin_miner_cpu.c aad_sha1.h aad_sha1_cpu.h aad_data_types.h aad_utilities.h aad_vault.h makefile
	cc -march=native -Wall -Wshadow -Werror -O3 $< -o $@

coin_miner_simd:	aad_coin_miner_simd.c aad_sha1.h aad_sha1_cpu.h aad_data_types.h aad_utilities.h aad_vault.h makefile
	cc -march=native -Wall -Wshadow -Werror -fopenmp -mavx2 -O3 $< -o $@

coin_miner_cuda_kernel.cubin:	aad_coin_miner_cuda_kernel.cu aad_sha1.h makefile
	nvcc -arch=$(CUDA_ARCH) --compiler-options -O2,-Wall -I$(CUDA_DIR)/include --cubin $< -o $@

coin_miner_cuda:	aad_coin_miner_cuda.c coin_miner_cuda_kernel.cubin aad_sha1.h aad_sha1_cpu.h aad_data_types.h aad_utilities.h aad_vault.h aad_cuda_utilities.h makefile
	cc -march=native -Wall -Wshadow -Werror -O3 -I$(CUDA_DIR)/include $< -o $@ -lcuda

coin_miner_ocl:	aad_coin_miner_ocl.c aad_coin_miner_ocl_kernel.cl aad_sha1.h aad_sha1_cpu.h aad_sha1_ocl_kernel.cl aad_data_types.h aad_utilities.h aad_vault.h aad_ocl_utilities.h makefile
	cc -march=native -Wall -Wshadow -O3 $< -o $@ -lOpenCL

coin_miner_wasm:	aad_coin_miner_wasm.c aad_sha1.h aad_sha1_cpu.h aad_sha1_wasm.h aad_data_types.h aad_utilities.h aad_vault.h makefile
	emcc -O3 -flto -msimd128 -o coin_miner_wasm.js aad_coin_miner_wasm.c \
		-s WASM=1 \
		-s EXPORTED_FUNCTIONS='["_mine_coins_wasm","_get_statistics","_stop_mining","_reset_mining","_get_found_coin","_get_found_coins_count","_malloc","_free"]' \
		-s EXPORTED_RUNTIME_METHODS='["cwrap","ccall","getValue","setValue"]' \
		-s ALLOW_MEMORY_GROWTH=1 \
		-s MODULARIZE=1 \
		-s EXPORT_NAME='CoinMinerModule' \
		-s INITIAL_MEMORY=67108864

benchmark:	aad_benchmark.c aad_sha1.h aad_sha1_cpu.h aad_data_types.h aad_utilities.h makefile
	cc -march=native -Wall -Wshadow -Werror -O3 $< -o $@

miners: coin_miner_cpu coin_miner_simd coin_miner_wasm coin_miner_cuda coin_miner_ocl benchmark

all: 	sha1_tests sha1_cuda_test sha1_cuda_kernel.cubin \
		coin_miner_cpu coin_miner_simd coin_miner_wasm coin_miner_cuda coin_miner_cuda_kernel.cubin coin_miner_ocl \
		benchmark