cumo 0.1.0

data/Rakefile

@@ -0,0 +1,43 @@
+require "bundler/gem_tasks"
+require "rake/testtask"
+
+Rake::TestTask.new(:test) do |t|
+  t.libs << "test"
+  t.libs << "lib"
+  t.test_files = FileList["test/**/*_test.rb"]
+end
+
+task :compile do
+  sh 'cd ext/cumo && ruby extconf.rb && make && make build-ctest'
+end
+
+task :ctest do
+  sh 'cd ext/cumo && ruby extconf.rb && make run-ctest'
+end
+
+task :clean do
+  sh 'cd ext/cumo && make clean'
+end
+
+task :docs do
+  dir = "ext/cumo"
+  srcs = %w[array.c data.c index.c math.c narray.c rand.c struct.c].map{|s| File.join(dir, "narray", s)}
+  srcs += %w[cublas.c driver.c nvrtc.c runtime.c memory_pool.cpp].map{|s| File.join(dir, "cuda", s) }
+  srcs << File.join(dir, "narray", "types/*.c")
+  srcs << "lib/cumo/narray/extra.rb"
+  sh "cd ext/cumo; ruby extconf.rb; make src"
+  sh "rm -rf docs .yardoc; yard doc -o docs -m markdown -r README.md #{srcs.join(' ')}"
+end
+task :doc => :docs
+
+task :gdb do
+  sh "gdb -x run.gdb --args ruby -I. ./test.rb"
+end
+
+task :default => [:clobber, :compile, :test]
+
+desc 'Open an irb session preloaded with the gem library'
+task :console do
+  sh 'irb -rubygems -I lib'
+end
+task :c => :console

data/bench/broadcast_fp32.rb

@@ -0,0 +1,138 @@
+require 'benchmark'
+require 'cumo/narray'
+
+num_iteration = 1000
+
+Benchmark.bm 20 do |r|
+  x = Cumo::SFloat.ones([1000,784])
+  y = Cumo::SFloat.ones([1000,784])
+  r.report "x.inplace + y" do
+    num_iteration.times do
+      x.inplace + y
+    end
+    Cumo::CUDA::Runtime.cudaDeviceSynchronize
+  end
+
+  x = Cumo::SFloat.ones([1000,784])
+  y = Cumo::SFloat.ones([1000,784])
+  r.report "x + y" do
+    num_iteration.times do
+      (x + y).free
+    end
+    Cumo::CUDA::Runtime.cudaDeviceSynchronize
+  end
+
+  x = Cumo::SFloat.ones([1000,784])
+  y = Cumo::SFloat.ones([1000,784])
+  r.report "x.inplace + 1.0" do
+    num_iteration.times do
+      x.inplace + 1.0
+    end
+    Cumo::CUDA::Runtime.cudaDeviceSynchronize
+  end
+
+  x = Cumo::SFloat.ones([1000,784])
+  z = Cumo::SFloat.ones([1000,1])
+  r.report "x.inplace + z" do
+    num_iteration.times do
+      x.inplace + z
+    end
+    Cumo::CUDA::Runtime.cudaDeviceSynchronize
+  end
+
+  x = Cumo::SFloat.ones([1000,784])
+  y = Cumo::SFloat.ones([1000,784])
+  r.report "x.inplace - y" do
+    num_iteration.times do
+      x.inplace - y
+    end
+    Cumo::CUDA::Runtime.cudaDeviceSynchronize
+  end
+
+  x = Cumo::SFloat.ones([1000,784])
+  y = Cumo::SFloat.ones([1000,784])
+  r.report "x.inplace - 1.0" do
+    num_iteration.times do
+      x.inplace - 1.0
+    end
+    Cumo::CUDA::Runtime.cudaDeviceSynchronize
+  end
+
+  x = Cumo::SFloat.ones([1000,784])
+  z = Cumo::SFloat.ones([1000,1])
+  r.report "x.inplace - z" do
+    num_iteration.times do
+      x.inplace - z
+    end
+    Cumo::CUDA::Runtime.cudaDeviceSynchronize
+  end
+
+  x = Cumo::SFloat.ones([1000,784])
+  y = Cumo::SFloat.ones([1000,784])
+  r.report "x.inplace * y" do
+    num_iteration.times do
+      x.inplace * y
+    end
+    Cumo::CUDA::Runtime.cudaDeviceSynchronize
+  end
+
+  x = Cumo::SFloat.ones([1000,784])
+  y = Cumo::SFloat.ones([1000,784])
+  r.report "x.inplace * 1.0" do
+    num_iteration.times do
+      x.inplace * 1.0
+    end
+    Cumo::CUDA::Runtime.cudaDeviceSynchronize
+  end
+
+  x = Cumo::SFloat.ones([1000,784])
+  z = Cumo::SFloat.ones([1000,1])
+  r.report "x.inplace * z" do
+    num_iteration.times do
+      x.inplace * z
+    end
+    Cumo::CUDA::Runtime.cudaDeviceSynchronize
+  end
+
+  x = Cumo::SFloat.ones([1000,784])
+  y = Cumo::SFloat.ones([1000,784])
+  r.report "x.inplace / y" do
+    num_iteration.times do
+      x.inplace / y
+    end
+    Cumo::CUDA::Runtime.cudaDeviceSynchronize
+  end
+
+  x = Cumo::SFloat.ones([1000,784])
+  y = Cumo::SFloat.ones([1000,784])
+  r.report "x.inplace / 1.0" do
+    num_iteration.times do
+      x.inplace / 1.0
+    end
+    Cumo::CUDA::Runtime.cudaDeviceSynchronize
+  end
+
+  x = Cumo::SFloat.ones([1000,784])
+  z = Cumo::SFloat.ones([1000,1])
+  r.report "x.inplace / z" do
+    num_iteration.times do
+      x.inplace / z
+    end
+    Cumo::CUDA::Runtime.cudaDeviceSynchronize
+  end
+end
+
+#                            user     system      total        real
+# x.inplace + y          0.040000   0.000000   0.040000 (  0.039255)
+# x + y                  0.140000   0.520000   0.660000 (  0.689061)
+# x.inplace + 1.0        0.040000   0.000000   0.040000 (  0.040255)
+# x.inplace + z          0.090000   0.000000   0.090000 (  0.083663)
+# x.inplace - y          0.030000   0.010000   0.040000 (  0.039017)
+# x.inplace - 1.0        0.050000   0.000000   0.050000 (  0.053187)
+# x.inplace - z          0.080000   0.010000   0.090000 (  0.085909)
+# x.inplace * y          0.030000   0.010000   0.040000 (  0.038583)
+# x.inplace * 1.0        0.050000   0.010000   0.060000 (  0.051933)
+# x.inplace * z          0.080000   0.010000   0.090000 (  0.084889)
+# x.inplace / y          0.040000   0.000000   0.040000 (  0.038958)
+# x.inplace / 1.0        0.040000   0.010000   0.050000 (  0.050266)
+# x.inplace / z          0.080000   0.010000   0.090000 (  0.086685)

data/bench/cumo_bench.rb

@@ -0,0 +1,193 @@
+require 'cumo/narray'
+require 'benchmark'
+
+NUM = (ARGV.first || 100).to_i
+
+a = Cumo::Float32.new(10).seq(1)
+b = Cumo::Float32.new(10).seq(10,10)
+c = a + b
+c.free
+
+def elementwise
+  puts 'element-wise'
+  Benchmark.bm do |r|
+    a = Cumo::Float32.new(10000).seq(1)
+    b = Cumo::Float32.new(10000).seq(10,10)
+    (a + b).free # warm up
+    r.report('10**4') do
+      NUM.times do
+        (a + b).free
+        Cumo::CUDA::Runtime.cudaDeviceSynchronize
+      end
+    end
+
+    a = Cumo::Float32.new(100000).seq(1)
+    b = Cumo::Float32.new(100000).seq(10,10)
+    (a + b).free # warm up
+    r.report('10**5') do
+      NUM.times do
+        (a + b).free
+        Cumo::CUDA::Runtime.cudaDeviceSynchronize
+      end
+    end
+
+    a = Cumo::Float32.new(1000000).seq(1)
+    b = Cumo::Float32.new(1000000).seq(10,10)
+    (a + b).free # warm up
+    r.report('10**6') do
+      NUM.times do
+        (a + b).free
+        Cumo::CUDA::Runtime.cudaDeviceSynchronize
+      end
+    end
+
+    a = Cumo::Float32.new(10000000).seq(1)
+    b = Cumo::Float32.new(10000000).seq(10,10)
+    (a + b).free # warm up
+    r.report('10**7') do
+      NUM.times do
+        (a + b).free
+        Cumo::CUDA::Runtime.cudaDeviceSynchronize
+      end
+    end
+
+    a = Cumo::Float32.new(100000000).seq(1)
+    b = Cumo::Float32.new(100000000).seq(10,10)
+    (a + b).free # warm up
+    r.report('10**8') do
+      NUM.times do
+        (a + b).free
+        Cumo::CUDA::Runtime.cudaDeviceSynchronize
+      end
+    end
+  end
+end
+
+def reduction
+  puts 'reduction'
+  Benchmark.bm do |r|
+    a = Cumo::Float32.new(10000).seq(1)
+    r.report('10**4') do
+      NUM.times do
+        (a.sum).free
+        Cumo::CUDA::Runtime.cudaDeviceSynchronize
+      end
+    end
+
+    a = Cumo::Float32.new(100000).seq(1)
+    r.report('10**5') do
+      NUM.times do
+        (a.sum).free
+        Cumo::CUDA::Runtime.cudaDeviceSynchronize
+      end
+    end
+
+    a = Cumo::Float32.new(1000000).seq(1)
+    r.report('10**6') do
+      NUM.times do
+        (a.sum).free
+        Cumo::CUDA::Runtime.cudaDeviceSynchronize
+      end
+    end
+
+    a = Cumo::Float32.new(10000000).seq(1)
+    r.report('10**7') do
+      NUM.times do
+        (a.sum).free
+        Cumo::CUDA::Runtime.cudaDeviceSynchronize
+      end
+    end
+
+    a = Cumo::Float32.new(100000000).seq(1)
+    r.report('10**8') do
+      NUM.times do
+        (a.sum).free
+        Cumo::CUDA::Runtime.cudaDeviceSynchronize
+      end
+    end
+  end
+end
+
+def dot
+  num = 3
+  puts 'dot'
+  Benchmark.bm do |r|
+    a = Cumo::Float32.new(100,100).seq(1)
+    b = Cumo::Float32.new(100,100).seq(10,10)
+    a.dot(b).free # warm up
+    r.report('10**4') do
+      num.times do
+        a.dot(b).free
+        Cumo::CUDA::Runtime.cudaDeviceSynchronize
+      end
+    end
+
+    a = Cumo::Float32.new(100,1000).seq(1)
+    b = Cumo::Float32.new(1000,100).seq(10,10)
+    a.dot(b).free # warm up
+    r.report('10**5') do
+      num.times do
+        a.dot(b).free
+        Cumo::CUDA::Runtime.cudaDeviceSynchronize
+      end
+    end
+
+    a = Cumo::Float32.new(100,10000).seq(1)
+    b = Cumo::Float32.new(10000,100).seq(10,10)
+    a.dot(b).free # warm up
+    r.report('10**6') do
+      num.times do
+        a.dot(b).free
+        Cumo::CUDA::Runtime.cudaDeviceSynchronize
+      end
+    end
+
+    a = Cumo::Float32.new(100,100000).seq(1)
+    b = Cumo::Float32.new(100000,100).seq(10,10)
+    a.dot(b).free # warm up
+    r.report('10**7') do
+      num.times do
+        a.dot(b).free
+        Cumo::CUDA::Runtime.cudaDeviceSynchronize
+      end
+    end
+
+    a = Cumo::Float32.new(100,1000000).seq(1)
+    b = Cumo::Float32.new(1000000,100).seq(10,10)
+    a.dot(b).free # warm up
+    r.report('10**8') do
+      num.times do
+        a.dot(b).free
+        Cumo::CUDA::Runtime.cudaDeviceSynchronize
+      end
+    end
+  end
+end
+
+elementwise
+reduction
+dot
+
+# Tesla V100-SXM2...
+#
+# element-wise
+#        user     system      total        real
+# 10**4  0.000000   0.000000   0.000000 (  0.005769)
+# 10**5  0.010000   0.000000   0.010000 (  0.006609)
+# 10**6  0.000000   0.010000   0.010000 (  0.010313)
+# 10**7  0.040000   0.010000   0.050000 (  0.050986)
+# 10**8  0.310000   0.130000   0.440000 (  0.449699)
+# reduction
+#        user     system      total        real
+# 10**4  0.010000   0.000000   0.010000 (  0.009484)
+# 10**5  0.020000   0.010000   0.030000 (  0.022071)
+# 10**6  0.100000   0.050000   0.150000 (  0.152070)
+# 10**7  1.150000   0.600000   1.750000 (  1.754977)
+# 10**8 11.720000   5.750000  17.470000 ( 17.470990)
+# dot
+#        user     system      total        real
+# 10**4  0.000000   0.000000   0.000000 (  0.000351)
+# 10**5  0.000000   0.000000   0.000000 (  0.000838)
+# 10**6  0.000000   0.000000   0.000000 (  0.002702)
+# 10**7  0.020000   0.010000   0.030000 (  0.024650)
+# 10**8  0.180000   0.060000   0.240000 (  0.245101)

data/bench/numo_bench.rb

@@ -0,0 +1,138 @@
+require 'numo/narray'
+require 'benchmark'
+
+NUM = (ARGV.first || 100).to_i
+
+# warm up
+a = Numo::Float32.new(10).seq(1)
+b = Numo::Float32.new(10).seq(10,10)
+c = a + b
+
+def elementwise
+  puts 'element-wise'
+  Benchmark.bm do |r|
+    a = Numo::Float32.new(10000).seq(1)
+    b = Numo::Float32.new(10000).seq(10,10)
+    r.report('10**4') do
+      NUM.times { (a + b) }
+    end
+
+    a = Numo::Float32.new(100000).seq(1)
+    b = Numo::Float32.new(100000).seq(10,10)
+    r.report('10**5') do
+      NUM.times { (a + b) }
+    end
+
+    a = Numo::Float32.new(1000000).seq(1)
+    b = Numo::Float32.new(1000000).seq(10,10)
+    r.report('10**6') do
+      NUM.times { (a + b) }
+    end
+
+    a = Numo::Float32.new(10000000).seq(1)
+    b = Numo::Float32.new(10000000).seq(10,10)
+    r.report('10**7') do
+      NUM.times { (a + b) }
+    end
+
+    a = Numo::Float32.new(100000000).seq(1)
+    b = Numo::Float32.new(100000000).seq(10,10)
+    r.report('10**8') do
+      NUM.times { (a + b) }
+    end
+  end
+end
+
+def reduction
+  puts 'reduction'
+  Benchmark.bm do |r|
+    a = Numo::Float32.new(10000).seq(1)
+    r.report('10**4') do
+      NUM.times { (a.sum) }
+    end
+
+    a = Numo::Float32.new(100000).seq(1)
+    r.report('10**5') do
+      NUM.times { (a.sum) }
+    end
+
+    a = Numo::Float32.new(1000000).seq(1)
+    r.report('10**6') do
+      NUM.times { (a.sum) }
+    end
+
+    a = Numo::Float32.new(10000000).seq(1)
+    r.report('10**7') do
+      NUM.times { (a.sum) }
+    end
+
+    a = Numo::Float32.new(100000000).seq(1)
+    r.report('10**8') do
+      NUM.times { (a.sum) }
+    end
+  end
+end
+
+def dot
+  num = 3
+  puts 'dot'
+  Benchmark.bm do |r|
+    a = Numo::Float32.new(100,100).seq(1)
+    b = Numo::Float32.new(100,100).seq(10,10)
+    r.report('10**4') do
+      num.times { a.dot(b) }
+    end
+
+    a = Numo::Float32.new(100,1000).seq(1)
+    b = Numo::Float32.new(1000,100).seq(10,10)
+    r.report('10**5') do
+      num.times { a.dot(b) }
+    end
+
+    a = Numo::Float32.new(100,10000).seq(1)
+    b = Numo::Float32.new(10000,100).seq(10,10)
+    r.report('10**6') do
+      num.times { a.dot(b) }
+    end
+
+    a = Numo::Float32.new(100,100000).seq(1)
+    b = Numo::Float32.new(100000,100).seq(10,10)
+    r.report('10**7') do
+      num.times { a.dot(b) }
+    end
+
+    a = Numo::Float32.new(100,1000000).seq(1)
+    b = Numo::Float32.new(1000000,100).seq(10,10)
+    r.report('10**8') do
+      num.times { a.dot(b) }
+    end
+  end
+end
+
+elementwise
+reduction
+dot
+
+# Intel(R) Xeon(R) CPU E5-2686 v4 @ 2.30GHz
+#
+# element-wise
+#        user     system      total        real
+# 10**4  0.010000   0.000000   0.010000 (  0.002212)
+# 10**5  0.000000   0.020000   0.020000 (  0.021604)
+# 10**6  0.060000   0.060000   0.120000 (  0.120241)
+# 10**7  0.980000   0.890000   1.870000 (  1.874592)
+# 10**8  9.530000   8.520000  18.050000 ( 18.054087)
+# reduction
+#        user     system      total        real
+# 10**4  0.000000   0.000000   0.000000 (  0.001313)
+# 10**5  0.010000   0.000000   0.010000 (  0.011400)
+# 10**6  0.110000   0.000000   0.110000 (  0.111674)
+# 10**7  1.120000   0.000000   1.120000 (  1.127018)
+# 10**8 11.770000   0.010000  11.780000 ( 11.770858)
+# dot
+#        user     system      total        real
+# 10**4  0.000000   0.000000   0.000000 (  0.003935)
+# 10**5  0.040000   0.000000   0.040000 (  0.037682)
+# 10**6  0.380000   0.000000   0.380000 (  0.377312)
+# 10**7  3.790000   0.000000   3.790000 (  3.792297)
+# 10**8 38.820000   0.000000  38.820000 ( 38.816987)