cumo 0.4.3 → 0.5.1

data/bench/broadcast_fp32.rb

@@ -1,11 +1,13 @@
+# frozen_string_literal: true
+
 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])
+  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
@@ -13,8 +15,8 @@ Benchmark.bm 20 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([1000,784])
-  y = Cumo::SFloat.ones([1000,784])
+  x = Cumo::SFloat.ones([1000, 784])
+  y = Cumo::SFloat.ones([1000, 784])
   r.report "x + y" do
     num_iteration.times do
       (x + y).free
@@ -22,8 +24,8 @@ Benchmark.bm 20 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([1000,784])
-  y = Cumo::SFloat.ones([1000,784])
+  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
@@ -31,8 +33,8 @@ Benchmark.bm 20 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([1000,784])
-  z = Cumo::SFloat.ones([1000,1])
+  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
@@ -40,8 +42,8 @@ Benchmark.bm 20 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([1000,784])
-  y = Cumo::SFloat.ones([1000,784])
+  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
@@ -49,8 +51,8 @@ Benchmark.bm 20 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([1000,784])
-  y = Cumo::SFloat.ones([1000,784])
+  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
@@ -58,8 +60,8 @@ Benchmark.bm 20 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([1000,784])
-  z = Cumo::SFloat.ones([1000,1])
+  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
@@ -67,8 +69,8 @@ Benchmark.bm 20 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([1000,784])
-  y = Cumo::SFloat.ones([1000,784])
+  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
@@ -76,8 +78,8 @@ Benchmark.bm 20 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([1000,784])
-  y = Cumo::SFloat.ones([1000,784])
+  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
@@ -85,8 +87,8 @@ Benchmark.bm 20 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([1000,784])
-  z = Cumo::SFloat.ones([1000,1])
+  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
@@ -94,8 +96,8 @@ Benchmark.bm 20 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([1000,784])
-  y = Cumo::SFloat.ones([1000,784])
+  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
@@ -103,8 +105,8 @@ Benchmark.bm 20 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([1000,784])
-  y = Cumo::SFloat.ones([1000,784])
+  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
@@ -112,8 +114,8 @@ Benchmark.bm 20 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([1000,784])
-  z = Cumo::SFloat.ones([1000,1])
+  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

data/bench/cumo_bench.rb

@@ -1,3 +1,5 @@
+# frozen_string_literal: true
+
 require 'cumo/narray'
 require 'benchmark'
 
@@ -5,7 +7,7 @@ NUM = (ARGV.first || 100).to_i
 
 # warm up
 a = Cumo::Float32.new(10).seq(1)
-b = Cumo::Float32.new(10).seq(10,10)
+b = Cumo::Float32.new(10).seq(10, 10)
 c = a + b
 c.free
 
@@ -14,7 +16,7 @@ def elementwise(num = nil)
   puts "elementwise(#{num})"
   Benchmark.bm do |r|
     a = Cumo::Float32.new(10000).seq(1)
-    b = Cumo::Float32.new(10000).seq(10,10)
+    b = Cumo::Float32.new(10000).seq(10, 10)
     (a + b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**4') do
@@ -25,7 +27,7 @@ def elementwise(num = nil)
     end
 
     a = Cumo::Float32.new(100000).seq(1)
-    b = Cumo::Float32.new(100000).seq(10,10)
+    b = Cumo::Float32.new(100000).seq(10, 10)
     (a + b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**5') do
@@ -36,7 +38,7 @@ def elementwise(num = nil)
     end
 
     a = Cumo::Float32.new(1000000).seq(1)
-    b = Cumo::Float32.new(1000000).seq(10,10)
+    b = Cumo::Float32.new(1000000).seq(10, 10)
     (a + b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**6') do
@@ -47,7 +49,7 @@ def elementwise(num = nil)
     end
 
     a = Cumo::Float32.new(10000000).seq(1)
-    b = Cumo::Float32.new(10000000).seq(10,10)
+    b = Cumo::Float32.new(10000000).seq(10, 10)
     (a + b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**7') do
@@ -58,7 +60,7 @@ def elementwise(num = nil)
     end
 
     a = Cumo::Float32.new(100000000).seq(1)
-    b = Cumo::Float32.new(100000000).seq(10,10)
+    b = Cumo::Float32.new(100000000).seq(10, 10)
     (a + b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**8') do
@@ -130,8 +132,8 @@ def dot(num = nil)
   num ||= 1
   puts "dot(#{num})"
   Benchmark.bm do |r|
-    a = Cumo::Float32.new(100,100).seq(1)
-    b = Cumo::Float32.new(100,100).seq(10,10)
+    a = Cumo::Float32.new(100, 100).seq(1)
+    b = Cumo::Float32.new(100, 100).seq(10, 10)
     a.dot(b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**4') do
@@ -141,8 +143,8 @@ def dot(num = nil)
       end
     end
 
-    a = Cumo::Float32.new(100,1000).seq(1)
-    b = Cumo::Float32.new(1000,100).seq(10,10)
+    a = Cumo::Float32.new(100, 1000).seq(1)
+    b = Cumo::Float32.new(1000, 100).seq(10, 10)
     a.dot(b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**5') do
@@ -152,8 +154,8 @@ def dot(num = nil)
       end
     end
 
-    a = Cumo::Float32.new(100,10000).seq(1)
-    b = Cumo::Float32.new(10000,100).seq(10,10)
+    a = Cumo::Float32.new(100, 10000).seq(1)
+    b = Cumo::Float32.new(10000, 100).seq(10, 10)
     a.dot(b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**6') do
@@ -163,8 +165,8 @@ def dot(num = nil)
       end
     end
 
-    a = Cumo::Float32.new(100,100000).seq(1)
-    b = Cumo::Float32.new(100000,100).seq(10,10)
+    a = Cumo::Float32.new(100, 100000).seq(1)
+    b = Cumo::Float32.new(100000, 100).seq(10, 10)
     a.dot(b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**7') do
@@ -174,8 +176,8 @@ def dot(num = nil)
       end
     end
 
-    a = Cumo::Float32.new(100,1000000).seq(1)
-    b = Cumo::Float32.new(1000000,100).seq(10,10)
+    a = Cumo::Float32.new(100, 1000000).seq(1)
+    b = Cumo::Float32.new(1000000, 100).seq(10, 10)
     a.dot(b).free # warm up
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
     r.report('10**8') do

data/bench/numo_bench.rb

@@ -1,3 +1,5 @@
+# frozen_string_literal: true
+
 require 'numo/narray'
 require 'benchmark'
 
@@ -5,7 +7,7 @@ NUM = (ARGV.first || 100).to_i
 
 # warm up
 a = Numo::Float32.new(10).seq(1)
-b = Numo::Float32.new(10).seq(10,10)
+b = Numo::Float32.new(10).seq(10, 10)
 c = a + b
 
 def elementwise(num = nil)
@@ -13,31 +15,31 @@ def elementwise(num = nil)
   puts "elementwise(#{num})"
   Benchmark.bm do |r|
     a = Numo::Float32.new(10000).seq(1)
-    b = Numo::Float32.new(10000).seq(10,10)
+    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)
+    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)
+    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)
+    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)
+    b = Numo::Float32.new(100000000).seq(10, 10)
     r.report('10**8') do
       NUM.times { (a + b) }
     end
@@ -79,32 +81,32 @@ def dot(num = nil)
   num ||= 1
   puts "dot(#{num})"
   Benchmark.bm do |r|
-    a = Numo::Float32.new(100,100).seq(1)
-    b = Numo::Float32.new(100,100).seq(10,10)
+    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)
+    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)
+    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)
+    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)
+    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

data/bench/reduction_fp32.rb

@@ -1,3 +1,5 @@
+# frozen_string_literal: true
+
 require 'benchmark'
 require 'cumo/narray'
 
@@ -5,7 +7,7 @@ num_iteration = 100
 Cumo::CUDA::Runtime.cudaDeviceSynchronize
 
 Benchmark.bm 30 do |r|
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.sum" do
     num_iteration.times do
       x.sum
@@ -13,7 +15,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.sum(axis: 0)" do
     num_iteration.times do
       x.sum(axis: 0)
@@ -21,7 +23,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.sum(axis: 1)" do
     num_iteration.times do
       x.sum(axis: 1)
@@ -29,7 +31,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.sum(keepdims: true)" do
     num_iteration.times do
       x.sum(keepdims: true)
@@ -37,7 +39,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.sum(axis: 0, keepdims: true)" do
     num_iteration.times do
       x.sum(axis: 0, keepdims: true)
@@ -45,7 +47,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.sum(axis: 1, keepdims: true)" do
     num_iteration.times do
       x.sum(axis: 1, keepdims: true)
@@ -53,7 +55,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.max" do
     num_iteration.times do
       x.max
@@ -61,7 +63,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.max(axis: 0)" do
     num_iteration.times do
       x.max(axis: 0)
@@ -69,7 +71,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.max(axis: 1)" do
     num_iteration.times do
       x.max(axis: 1)
@@ -77,7 +79,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.max(keepdims: true)" do
     num_iteration.times do
       x.max(keepdims: true)
@@ -85,7 +87,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.max(axis: 0, keepdims: true)" do
     num_iteration.times do
       x.max(axis: 0, keepdims: true)
@@ -93,7 +95,7 @@ Benchmark.bm 30 do |r|
     Cumo::CUDA::Runtime.cudaDeviceSynchronize
   end
 
-  x = Cumo::SFloat.ones([500,500])
+  x = Cumo::SFloat.ones([500, 500])
   r.report "x.max(axis: 1, keepdims: true)" do
     num_iteration.times do
       x.max(axis: 1, keepdims: true)

data/bin/console

@@ -1,4 +1,5 @@
 #!/usr/bin/env ruby
+# frozen_string_literal: true
 
 require "bundler/setup"
 require "cumo"

data/cumo.gemspec

@@ -1,9 +1,9 @@
-# coding: utf-8
+# frozen_string_literal: true
+
 lib = File.expand_path("../lib", __FILE__)
 $LOAD_PATH.unshift(lib) unless $LOAD_PATH.include?(lib)
 
 cumo_version = File.read(File.join(__dir__, "ext/cumo/include/cumo.h")).match(/CUMO_VERSION "([^"]+)"/)[1]
-numo_narray_version = File.read(File.join(__dir__, "numo-narray-version")).strip
 
 Gem::Specification.new do |spec|
   spec.name          = "cumo"
@@ -16,17 +16,14 @@ Gem::Specification.new do |spec|
   spec.homepage      = "https://github.com/sonots/cumo"
   spec.license       = "BSD-3-Clause"
 
-  spec.files         = `git ls-files -z`.split("\x0").reject do |f|
-    f.match(%r{^(test|spec|features)/})
+  spec.required_ruby_version = ">= 3.0.0"
+
+  spec.files = `git ls-files -z`.split("\x0").reject do |f|
+    f.match(%r{^(test|spec|features|docker)/})
   end
   spec.test_files    = `git ls-files -- {test,spec,features}/*`.split("\n")
   spec.bindir        = "exe"
   spec.executables   = spec.files.grep(%r{^exe/}) { |f| File.basename(f) }
   spec.require_paths = ["lib"]
   spec.extensions    = ["ext/cumo/extconf.rb"]
-
-  spec.add_runtime_dependency "numo-narray", numo_narray_version
-
-  spec.add_development_dependency "bundler", "~> 1.15"
-  spec.add_development_dependency "rake", "~> 10.0"
 end

data/docker-build.sh

@@ -0,0 +1,4 @@
+#!/bin/bash
+
+script_dir=$(cd $(dirname ${BASH_SOURCE:-$0}); pwd)
+docker build -t cumo-dev ${script_dir}

data/docker-launch.sh

@@ -0,0 +1,4 @@
+#!/bin/bash
+
+project_dir="$(cd $(dirname ${BASH_SOURCE:-$0}); pwd)/"
+docker run --gpus all -v $project_dir:/workspace -it cumo-dev bash

data/docs/src-tree.md

@@ -6,7 +6,7 @@
     * Technically, it is not possible to use CRuby API such as `VALUE` in .cu files.
         * CRuby API is not callable from CUDA kernel because they do not have `__device__` modifier.
         * nvcc does not support `#include RUBY_EXTCONF_H`, so can not include `ruby.h`.
-    * (RULE) It is allowed to use C++14 codes in .cu files.
+    * (RULE) It is allowed to use C++17 codes in .cu files.
 * Rest of `*.{h,c}` files are for host (CPU).
     * Call C wrapper functions defined in .cu files.
     * It can use CRuby API.

data/ext/cumo/cuda/cudnn.c

@@ -50,7 +50,7 @@ cumo_cuda_cudnn_handle()
   @return [Boolean] Returns true if cuDNN is available
  */
 static VALUE
-rb_cudnn_available_p()
+rb_cudnn_available_p(VALUE self)
 {
 #if CUDNN_FOUND
     return Qtrue;
@@ -72,7 +72,7 @@ Init_cumo_cuda_cudnn(void)
     rb_define_const(mCUDA, "Cudnn", mCUDNN); // alias
     eCUDNNError = rb_define_class_under(mCUDA, "CUDNNError", rb_eStandardError);
 
-    rb_define_singleton_method(mCUDNN, "available?", RUBY_METHOD_FUNC(rb_cudnn_available_p), 0);
+    rb_define_singleton_method(mCUDNN, "available?", rb_cudnn_available_p, 0);
 #ifdef CUDNN_FOUND
     rb_define_const(mCUDNN, "CUDNN_POOLING_MAX", INT2NUM(CUDNN_POOLING_MAX));
     rb_define_const(mCUDNN, "CUDNN_POOLING_MAX_DETERMINISTIC", INT2NUM(CUDNN_POOLING_MAX_DETERMINISTIC));

data/ext/cumo/cuda/cudnn_impl.cpp

@@ -74,6 +74,25 @@ cumo_cuda_cudnn_CreateTensorDescriptor(
         status = cudnnSetTensor4dDescriptor(
                 *desc, CUDNN_TENSOR_NCHW, cudnn_dtype, shape[0], shape[1], shape[2], shape[3]);
     }
+    else if (ndim < 4) {
+        // cuDNN 9 fix: Force 4D (N, C, H, W)
+        int pad_shape[4] = {1, 1, 1, 1};
+
+        if (ndim == 1) {
+            // 1D: arrays are treated as "Channel" (1, C, 1, 1)
+            pad_shape[1] = (int)(shape[0]);
+        } else {
+            // 2D: [N, C] -> [N, C, 1, 1]
+            // 3D: [N, C, H] -> [N, C, H, 1]
+            for (int idim = 0; idim < ndim; ++idim) {
+                pad_shape[idim] = (int)(shape[idim]);
+            }
+        }
+
+        status = cudnnSetTensor4dDescriptor(
+                *desc, CUDNN_TENSOR_NCHW, cudnn_dtype,
+                pad_shape[0], pad_shape[1], pad_shape[2], pad_shape[3]);
+    }
     else {
         int int_shape[CUMO_NA_MAX_DIMENSION];
         for (int idim = 0; idim < ndim; ++idim) {
@@ -514,8 +533,11 @@ cumo_cuda_cudnn_FindConvolutionBackwardFilterAlgorithm(
 // TODO(sonots): Support other than 4, 5 dimensional arrays by reshaping into 4-dimensional arrays as Chainer does.
 cudnnBatchNormMode_t
 cumo_cuda_cudnn_GetBatchNormMode(size_t ndim, int* axis) {
-    if (ndim == 1 && axis[0] == 0) {  // (1, channels, (depth, )height, width)
-        return CUDNN_BATCHNORM_PER_ACTIVATION;
+    if (ndim == 1) {
+        return CUDNN_BATCHNORM_SPATIAL;
+    }
+    if (ndim == 2) {
+        return CUDNN_BATCHNORM_SPATIAL;
     }
     if ((ndim == 3 && axis[0] == 0 && axis[1] == 2 && axis[2] == 3) ||
         (ndim == 4 && axis[0] == 0 && axis[1] == 2 && axis[2] == 3 && axis[3] == 4)) {  // (1, channels, (1, )1, 1)
@@ -533,7 +555,7 @@ cumo_cuda_cudnn_CreateBNTensorDescriptor(
 {
     cudnnStatus_t status = CUDNN_STATUS_SUCCESS;
     status = cudnnCreateTensorDescriptor(desc);
-    if (status = CUDNN_STATUS_SUCCESS) return status;
+    if (status == CUDNN_STATUS_SUCCESS) return status;
 
     status = cudnnDeriveBNTensorDescriptor(*desc, x_desc, mode);
     return status;

data/ext/cumo/cuda/driver.c

@@ -33,7 +33,11 @@ rb_cuCtxCreate(VALUE self, VALUE flags, VALUE dev)
     CUcontext _pctx;
     CUresult status;
 
+#if defined(CUDA_VERSION) && CUDA_VERSION >= 13000
+    status = cuCtxCreate(&_pctx, NULL, _flags, _dev);
+#else
     status = cuCtxCreate(&_pctx, _flags, _dev);
+#endif
 
     check_status(status);
     return SIZET2NUM((size_t)_pctx);
@@ -418,5 +422,9 @@ Init_cumo_cuda_driver()
 
     cuInit(0);
     cuDeviceGet(&cuDevice, 0);
+#if defined(CUDA_VERSION) && CUDA_VERSION >= 13000
+    cuCtxCreate(&context, NULL, 0, cuDevice);
+#else
     cuCtxCreate(&context, 0, cuDevice);
+#endif
 }

data/ext/cumo/cumo.c

@@ -114,13 +114,17 @@ Init_cumo()
     const char* env;
     VALUE mCumo;
 
+#ifdef HAVE_RB_EXT_RACTOR_SAFE
+    rb_ext_ractor_safe(true);
+#endif
+
     mCumo = rb_define_module("Cumo");
 
     rb_define_const(mCumo, "VERSION", rb_str_new2(CUMO_VERSION));
 
-    rb_define_singleton_method(mCumo, "enable_compatible_mode", RUBY_METHOD_FUNC(rb_enable_compatible_mode), 0);
-    rb_define_singleton_method(mCumo, "disable_compatible_mode", RUBY_METHOD_FUNC(rb_disable_compatible_mode), 0);
-    rb_define_singleton_method(mCumo, "compatible_mode_enabled?", RUBY_METHOD_FUNC(rb_compatible_mode_enabled_p), 0);
+    rb_define_singleton_method(mCumo, "enable_compatible_mode", rb_enable_compatible_mode, 0);
+    rb_define_singleton_method(mCumo, "disable_compatible_mode", rb_disable_compatible_mode, 0);
+    rb_define_singleton_method(mCumo, "compatible_mode_enabled?", rb_compatible_mode_enabled_p, 0);
 
     // default is false
     env = getenv("CUMO_COMPATIBLE_MODE");