tilework-cuda 0.18.4
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- checksums.yaml +7 -0
- data/LICENSE +21 -0
- data/ext/tilework/cuda/bridge.cpp +1236 -0
- data/ext/tilework/cuda/extconf.rb +105 -0
- data/ext/tilework/cuda/gpu_bridge.c +840 -0
- data/lib/tilework/backend/cuda.rb +32 -0
- metadata +66 -0
|
@@ -0,0 +1,105 @@
|
|
|
1
|
+
# frozen_string_literal: true
|
|
2
|
+
|
|
3
|
+
# mkmf does not support nvcc, so the Makefile is written directly.
|
|
4
|
+
#
|
|
5
|
+
# 1. CUDA Toolkit present -> link bridge.cpp with c++. cuBLAS / cuRAND /
|
|
6
|
+
# cuFFT / cuDNN are each optional: a library is linked (and
|
|
7
|
+
# TILEWORK_HAS_<LIB> defined) only when both its header and its
|
|
8
|
+
# shared library are found, so a Toolkit without cuDNN — the common
|
|
9
|
+
# case, since cuDNN is a separate download — still builds, and Ruby
|
|
10
|
+
# falls back for whatever is missing.
|
|
11
|
+
# 2. absent -> a no-op Makefile (so `bundle install` does not
|
|
12
|
+
# fail without CUDA; core's backend= turns the missing extension into a
|
|
13
|
+
# Tilework::DeviceError)
|
|
14
|
+
#
|
|
15
|
+
# The product goes to lib/tilework/cuda/bridge.<dlext> so it cannot collide
|
|
16
|
+
# with other backends.
|
|
17
|
+
|
|
18
|
+
require "rbconfig"
|
|
19
|
+
|
|
20
|
+
hdrdir = RbConfig::CONFIG["rubyhdrdir"]
|
|
21
|
+
archhdrdir = RbConfig::CONFIG["rubyarchhdrdir"]
|
|
22
|
+
dlext = RbConfig::CONFIG["DLEXT"] # linux: "so"
|
|
23
|
+
cc = RbConfig::CONFIG["CC"]
|
|
24
|
+
cxx = RbConfig::CONFIG["CXX"] || "c++"
|
|
25
|
+
srcdir = File.expand_path(__dir__)
|
|
26
|
+
|
|
27
|
+
cuda_home = ENV["CUDA_HOME"] || "/usr/local/cuda"
|
|
28
|
+
nvcc = "#{cuda_home}/bin/nvcc"
|
|
29
|
+
|
|
30
|
+
if File.exist?(nvcc)
|
|
31
|
+
# libcuda.so (Driver API) ships with the driver, not the CUDA Toolkit, so it
|
|
32
|
+
# may be missing from $CUDA_HOME/lib64 — typically on Jetson (aarch64), where
|
|
33
|
+
# it lives in /usr/lib/aarch64-linux-gnu. Only existing dirs are added to -L.
|
|
34
|
+
# /usr/lib64 is where Fedora/RHEL packages (cuDNN included) install.
|
|
35
|
+
lib_dirs = [
|
|
36
|
+
"#{cuda_home}/lib64",
|
|
37
|
+
"#{cuda_home}/lib64/stubs",
|
|
38
|
+
"/usr/lib/#{RbConfig::CONFIG["host_cpu"]}-linux-gnu",
|
|
39
|
+
"/usr/lib64",
|
|
40
|
+
].select { |d| Dir.exist?(d) }
|
|
41
|
+
ldflags = lib_dirs.map { |d| "-L#{d}" }.join(" ")
|
|
42
|
+
include_dirs = ["#{cuda_home}/include", "/usr/include"]
|
|
43
|
+
|
|
44
|
+
# header, -l name, macro
|
|
45
|
+
optional_libs = {
|
|
46
|
+
"cuBLAS" => ["cublas_v2.h", "cublas", "TILEWORK_HAS_CUBLAS"],
|
|
47
|
+
"cuRAND" => ["curand.h", "curand", "TILEWORK_HAS_CURAND"],
|
|
48
|
+
"cuFFT" => ["cufft.h", "cufft", "TILEWORK_HAS_CUFFT"],
|
|
49
|
+
"cuDNN" => ["cudnn.h", "cudnn", "TILEWORK_HAS_CUDNN"],
|
|
50
|
+
}
|
|
51
|
+
found = optional_libs.select do |_, (header, lib, _)|
|
|
52
|
+
include_dirs.any? { |d| File.file?("#{d}/#{header}") } &&
|
|
53
|
+
lib_dirs.any? { |d| File.exist?("#{d}/lib#{lib}.so") }
|
|
54
|
+
end
|
|
55
|
+
lib_cxxflags = found.values.map { |(_, _, macro)| " -D#{macro}" }.join
|
|
56
|
+
lib_ldlibs = found.values.map { |(_, lib, _)| " -l#{lib}" }.join
|
|
57
|
+
|
|
58
|
+
# bridge.cpp is pure host code with no __global__ (NVRTC compiles kernels at
|
|
59
|
+
# run time), so it builds with the host C++ compiler rather than nvcc. nvcc
|
|
60
|
+
# would only inject fatbin registration code and a libcudart dependency.
|
|
61
|
+
# nvcc's presence is used only to detect whether the CUDA Toolkit is installed.
|
|
62
|
+
makefile = <<~MAKEFILE
|
|
63
|
+
TARGET = bridge.#{dlext}
|
|
64
|
+
CC = #{cc}
|
|
65
|
+
CXX = #{cxx}
|
|
66
|
+
CFLAGS = -O2 -Wall -fPIC -DTILEWORK_BRIDGE_MODULE=\\\"CUDABridge\\\" -I#{hdrdir} -I#{archhdrdir}
|
|
67
|
+
CXXFLAGS = -O2 -Wall -fPIC -pthread#{lib_cxxflags} -I#{cuda_home}/include -I/usr/include
|
|
68
|
+
LDLIBS = #{ldflags} -lcuda -lnvrtc#{lib_ldlibs} -pthread
|
|
69
|
+
|
|
70
|
+
SRCDIR = #{srcdir}
|
|
71
|
+
|
|
72
|
+
all: $(TARGET)
|
|
73
|
+
|
|
74
|
+
gpu_bridge.o: $(SRCDIR)/gpu_bridge.c
|
|
75
|
+
\t$(CC) $(CFLAGS) -c $(SRCDIR)/gpu_bridge.c -o $@
|
|
76
|
+
|
|
77
|
+
bridge_cuda.o: $(SRCDIR)/bridge.cpp
|
|
78
|
+
\t$(CXX) $(CXXFLAGS) -c $(SRCDIR)/bridge.cpp -o $@
|
|
79
|
+
|
|
80
|
+
# Ruby symbols resolve when the extension loads, so they may stay undefined
|
|
81
|
+
# here (Linux shared objects allow that by default).
|
|
82
|
+
$(TARGET): gpu_bridge.o bridge_cuda.o
|
|
83
|
+
\t$(CXX) -shared gpu_bridge.o bridge_cuda.o $(LDLIBS) -o $@
|
|
84
|
+
|
|
85
|
+
clean:
|
|
86
|
+
\trm -f $(TARGET) gpu_bridge.o bridge_cuda.o
|
|
87
|
+
|
|
88
|
+
# gem install calls install with sitearchdir overridden. Install to the
|
|
89
|
+
# namespaced path rather than a flat one so require
|
|
90
|
+
# "tilework/cuda/bridge" resolves.
|
|
91
|
+
sitearchdir = .
|
|
92
|
+
install: $(TARGET)
|
|
93
|
+
\tmkdir -p $(sitearchdir)/tilework/cuda
|
|
94
|
+
\tcp $(TARGET) $(sitearchdir)/tilework/cuda/$(TARGET)
|
|
95
|
+
|
|
96
|
+
.PHONY: all clean install
|
|
97
|
+
MAKEFILE
|
|
98
|
+
|
|
99
|
+
File.write("Makefile", makefile)
|
|
100
|
+
status = optional_libs.keys.map { |name| "#{name} #{found.key?(name) ? "enabled" : "not found"}" }
|
|
101
|
+
puts "created Makefile (CUDA backend, #{cuda_home}; #{status.join(", ")})"
|
|
102
|
+
else
|
|
103
|
+
File.write("Makefile", "all:\ninstall:\nclean:\n.PHONY: all install clean\n")
|
|
104
|
+
puts "skipping tilework/cuda build (no CUDA toolchain found; set CUDA_HOME if it's installed elsewhere)"
|
|
105
|
+
end
|