GPU-Based-Matrix-Operations/Makefile at main · Olajide-Badejo/GPU-Based-Matrix-Operations · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
# ─────────────────────────────────────────────────────────────────────────────
# Makefile — GPU Matrix Ops
# Usage:
#   make            # build with optimisations
#   make DEBUG=1    # build with debug info / no optimisation
#   make run        # build + execute
#   make clean      # remove build artefacts
#   make profile    # build with Nsight-profiling flags
# ─────────────────────────────────────────────────────────────────────────────

NVCC := nvcc
CXX  := g++

# Detect SM version; fallback to 86 (Ampere RTX 30-series)
SM ?= 86

# ── Directories ───────────────────────────────────────────────────────────────
SRCDIR := src
INCDIR := include
BINDIR := bin
OBJDIR := obj

# ── Sources ───────────────────────────────────────────────────────────────────
CU_SRCS  := $(SRCDIR)/main.cu \
             $(SRCDIR)/matvec_kernels.cu \
             $(SRCDIR)/matmul_kernels.cu

CPP_SRCS := $(SRCDIR)/cpu_ops.cpp

CU_OBJS  := $(patsubst $(SRCDIR)/%.cu,  $(OBJDIR)/%.cu.o,  $(CU_SRCS))
CPP_OBJS := $(patsubst $(SRCDIR)/%.cpp, $(OBJDIR)/%.cpp.o, $(CPP_SRCS))

TARGET := $(BINDIR)/gpu_matrix_ops

# ── Flags ─────────────────────────────────────────────────────────────────────
COMMON_FLAGS := -I$(INCDIR)

ifdef DEBUG
NVCC_FLAGS := -g -G -O0 -arch=sm_$(SM) $(COMMON_FLAGS)
CXX_FLAGS  := -g -O0 $(COMMON_FLAGS)
else
NVCC_FLAGS := -O3 --use_fast_math -arch=sm_$(SM) \
              --expt-relaxed-constexpr \
              -lineinfo \
              --ptxas-options=-v \
              $(COMMON_FLAGS)
CXX_FLAGS  := -O3 -march=native $(COMMON_FLAGS)
endif

# FIX: removed -G from PROFILE flags.
# -G compiles device code with full debug instrumentation and disables all GPU
# optimisations, so profiling a -G binary gives meaningless timings that do not
# reflect real kernel performance. -lineinfo alone gives Nsight Compute and
# nvprof the source-to-PTX line mapping they need without touching optimisations.
ifdef PROFILE
NVCC_FLAGS += -lineinfo
endif

LINK_FLAGS := -lcudart

# ── Default target ────────────────────────────────────────────────────────────
.PHONY: all run clean profile info

all: $(TARGET)

$(TARGET): $(CU_OBJS) $(CPP_OBJS)
	@mkdir -p $(BINDIR)
	$(NVCC) $(NVCC_FLAGS) $^ -o $@ $(LINK_FLAGS)
	@echo ""
	@echo "  Built: $@"

$(OBJDIR)/%.cu.o: $(SRCDIR)/%.cu
	@mkdir -p $(OBJDIR)
	$(NVCC) $(NVCC_FLAGS) -dc -c $< -o $@

$(OBJDIR)/%.cpp.o: $(SRCDIR)/%.cpp
	@mkdir -p $(OBJDIR)
	$(CXX) $(CXX_FLAGS) -c $< -o $@

run: all
	@echo ""
	./$(TARGET)

profile: PROFILE=1
profile: all
	@echo "  Binary built with profiling flags."
	@echo "  Run with: ncu --set full ./$(TARGET)"
	@echo "  or      : nvprof ./$(TARGET)"

clean:
	rm -rf $(OBJDIR) $(BINDIR)
	@echo "  Cleaned."

info:
	@echo "  Target SM : sm_$(SM)"
	@echo "  NVCC      : $(shell which $(NVCC))"
	@echo "  CUDA ver  : $(shell $(NVCC) --version | tail -1)"