cumo 0.4.3 → 0.5.1

data/ext/cumo/narray/gen/spec.rb

@@ -1,3 +1,5 @@
+# frozen_string_literal: true
+
 def_id "cast"
 def_id "eq"
 def_id "ne"
@@ -14,8 +16,8 @@ if is_float
   def_id "copysign"
 end
 if is_int
-  def_id "<<","left_shift"
-  def_id ">>","right_shift"
+  def_id "<<", "left_shift"
+  def_id ">>", "right_shift"
 end
 if is_comparable && !is_object
   def_id "gt"
@@ -42,13 +44,13 @@ if is_object
   def_id "nan?"
   def_id "infinite?"
   def_id "finite?"
-  def_id "==","eq"
-  def_id "!=","ne"
-  def_id ">" ,"gt"
-  def_id ">=","ge"
-  def_id "<" ,"lt"
-  def_id "<=","le"
-  def_id "<=>","ufo"
+  def_id "==", "eq"
+  def_id "!=", "ne"
+  def_id ">" , "gt"
+  def_id ">=", "ge"
+  def_id "<" , "lt"
+  def_id "<=", "le"
+  def_id "<=>", "ufo"
 end
 if (is_float || is_complex) && !is_object
   def_id "gemm"
@@ -119,18 +121,18 @@ def_method "store" do
   store_numeric
   store_from "Bit"
   if is_complex
-    store_from "DComplex","cumo_dcomplex","m_from_dcomplex"
-    store_from "SComplex","cumo_scomplex","m_from_scomplex"
+    store_from "DComplex", "cumo_dcomplex", "m_from_dcomplex"
+    store_from "SComplex", "cumo_scomplex", "m_from_scomplex"
   end
-  store_from "DFloat","double",   "m_from_real"
-  store_from "SFloat","float",    "m_from_real"
+  store_from "DFloat", "double",   "m_from_real"
+  store_from "SFloat", "float",    "m_from_real"
   store_from "Int64", "int64_t",  "m_from_int64"
   store_from "Int32", "int32_t",  "m_from_int32"
   store_from "Int16", "int16_t",  "m_from_sint"
   store_from "Int8",  "int8_t",   "m_from_sint"
-  store_from "UInt64","u_int64_t","m_from_uint64"
-  store_from "UInt32","u_int32_t","m_from_uint32"
-  store_from "UInt16","u_int16_t","m_from_sint"
+  store_from "UInt64", "u_int64_t", "m_from_uint64"
+  store_from "UInt32", "u_int32_t", "m_from_uint32"
+  store_from "UInt16", "u_int16_t", "m_from_sint"
   store_from "UInt8", "u_int8_t", "m_from_sint"
   store_from "RObject", "VALUE",  "m_num_to_data"
   store_array
@@ -144,6 +146,7 @@ def_singleton_method "cast"
 def_method "aref", op:"[]"
 def_method "aref_cpu"
 def_method "aset", op:"[]="
+def_method "at"
 
 def_method "coerce_cast"
 def_method "to_a"
@@ -167,15 +170,15 @@ if is_bit
   binary "xor", "^"
   binary "eq"
   bit_count "count_true"
-  def_alias "count_1","count_true"
-  def_alias "count","count_true"
+  def_alias "count_1", "count_true"
+  def_alias "count", "count_true"
   bit_count "count_false"
-  def_alias "count_0","count_false"
+  def_alias "count_0", "count_false"
   bit_count_cpu "count_true_cpu"
-  def_alias "count_1_cpu","count_true_cpu"
-  def_alias "count_cpu","count_true_cpu"
+  def_alias "count_1_cpu", "count_true_cpu"
+  def_alias "count_cpu", "count_true_cpu"
   bit_count_cpu "count_false_cpu"
-  def_alias "count_0_cpu","count_false_cpu"
+  def_alias "count_0_cpu", "count_false_cpu"
   bit_reduce "all?", 1
   bit_reduce "any?", 0
   def_method "none?", "none_p"
@@ -215,17 +218,17 @@ if is_complex
   unary2 "real", "rtype", "cRT"
   unary2 "imag", "rtype", "cRT"
   unary2 "arg",  "rtype", "cRT"
-  def_alias "angle","arg"
+  def_alias "angle", "arg"
   set2 "set_imag", "rtype", "cRT"
   set2 "set_real", "rtype", "cRT"
-  def_alias "imag=","set_imag"
-  def_alias "real=","set_real"
+  def_alias "imag=", "set_imag"
+  def_alias "real=", "set_real"
 else
   def_alias "conj", "view"
   def_alias "im", "view"
 end
 
-def_alias "conjugate","conj"
+def_alias "conjugate", "conj"
 
 # base_cond
 
@@ -278,9 +281,9 @@ if is_comparable
   cond_binary "lt"
   cond_binary "le"
   def_alias ">", "gt"
-  def_alias ">=","ge"
+  def_alias ">=", "ge"
   def_alias "<", "lt"
-  def_alias "<=","le"
+  def_alias "<=", "le"
   def_method "clip"
 end
 
@@ -296,32 +299,32 @@ end
 
 if is_int
   if is_unsigned
-    accum "sum","u_int64_t","cumo_cUInt64"
-    accum "prod","u_int64_t","cumo_cUInt64"
+    accum "sum", "u_int64_t", "cumo_cUInt64"
+    accum "prod", "u_int64_t", "cumo_cUInt64"
   else
-    accum "sum","int64_t","cumo_cInt64"
-    accum "prod","int64_t","cumo_cInt64"
+    accum "sum", "int64_t", "cumo_cInt64"
+    accum "prod", "int64_t", "cumo_cInt64"
   end
 else
-  accum "sum","dtype","cT"
-  accum "prod","dtype","cT"
+  accum "sum", "dtype", "cT"
+  accum "prod", "dtype", "cT"
 end
 
 if is_double_precision
-  accum "kahan_sum","dtype","cT"
+  accum "kahan_sum", "dtype", "cT"
 end
 
 if is_float
-  accum "mean","dtype","cT"
-  accum "stddev","rtype","cRT"
-  accum "var","rtype","cRT"
-  accum "rms","rtype","cRT"
+  accum "mean", "dtype", "cT"
+  accum "stddev", "rtype", "cRT"
+  accum "var", "rtype", "cRT"
+  accum "rms", "rtype", "cRT"
 end
 
 if is_comparable
-  accum "min","dtype","cT"
-  accum "max","dtype","cT"
-  accum "ptp","dtype","cT"
+  accum "min", "dtype", "cT"
+  accum "max", "dtype", "cT"
+  accum "ptp", "dtype", "cT"
   accum_index "max_index"
   accum_index "min_index"
   def_method "minmax"
@@ -333,8 +336,8 @@ if is_int && !is_object
   def_method "bincount"
 end
 
-cum "cumsum","add"
-cum "cumprod","mul"
+cum "cumsum", "add"
+cum "cumprod", "mul"
 
 # dot
 accum_binary "mulsum"
@@ -377,17 +380,17 @@ def_method "poly"
 
 if is_comparable && !is_object
   if is_float
-    qsort type_name,"dtype","*(dtype*)","_prnan"
-    qsort type_name,"dtype","*(dtype*)","_ignan"
+    qsort type_name, "dtype", "*(dtype*)", "_prnan"
+    qsort type_name, "dtype", "*(dtype*)", "_ignan"
   else
-    qsort type_name,"dtype","*(dtype*)"
+    qsort type_name, "dtype", "*(dtype*)"
   end
   def_method "sort"
   if is_float
-    qsort type_name+"_index","dtype*","**(dtype**)","_prnan"
-    qsort type_name+"_index","dtype*","**(dtype**)","_ignan"
+    qsort type_name + "_index", "dtype*", "**(dtype**)", "_prnan"
+    qsort type_name + "_index", "dtype*", "**(dtype**)", "_ignan"
   else
-    qsort type_name+"_index","dtype*","**(dtype**)"
+    qsort type_name + "_index", "dtype*", "**(dtype**)"
   end
   def_method "sort_index"
   def_method "median"
@@ -407,7 +410,7 @@ def_module do
   set ns_var: "cT"
   set class_name: cn
   set name: "#{nm}_math"
-  set full_module_name: fn+"::NMath"
+  set full_module_name: fn + "::NMath"
   set module_name: "Math"
   set module_var: "mTM"
 
@@ -433,14 +436,14 @@ def_module do
   math "atanh"
   math "sinc"
   if !is_c
-    math "atan2",2
-    math "hypot",2
+    math "atan2", 2
+    math "hypot", 2
     math "erf"
     math "erfc"
     math "log1p"
     math "expm1"
-    math "ldexp",2
-    math "frexp",1,"frexp"
+    math "ldexp", 2
+    math "frexp", 1, "frexp"
   end
 end
 end

data/ext/cumo/narray/gen/tmpl/accum.c

@@ -56,8 +56,8 @@ static void
 <% else %>
   @overload <%=name%>(axis:nil, keepdims:false)
 <% end %>
-  @param [Numeric,Array,Range] axis (keyword) Affected dimensions.
-  @param [TrueClass] keepdims (keyword) If true, the reduced axes are left in the result array as dimensions with size one.
+  @param [Numeric,Array,Range] axis  Performs <%=name%> along the axis.
+  @param [TrueClass] keepdims  If true, the reduced axes are left in the result array as dimensions with size one.
   @return [Cumo::<%=class_name%>] returns result of <%=name%>.
 */
 static VALUE

data/ext/cumo/narray/gen/tmpl/accum_binary.c

@@ -91,7 +91,7 @@ static VALUE
   @overload <%=op_map%>(other, axis:nil, keepdims:false)
 <% end %>
   @param [Cumo::NArray,Numeric] other
-  @param [Numeric,Array,Range] axis (keyword) Affected dimensions.
+  @param [Numeric,Array,Range] axis  Performs <%=name%> along the axis.
   @param [TrueClass] keepdims (keyword) If true, the reduced axes are left in the result array as dimensions with size one.
 <% if is_float %>
   @param [TrueClass] nan (keyword) If true, apply NaN-aware algorithm (avoid NaN if exists).

data/ext/cumo/narray/gen/tmpl/alloc_func.c

@@ -85,7 +85,7 @@ static const rb_data_type_t <%=type_name%>_data_type = {
     {0, <%=type_name%>_free, <%=type_name%>_memsize,},
     &cumo_na_data_type,
     &<%=type_name%>_info,
-    0, // flags
+    RUBY_TYPED_FROZEN_SHAREABLE, // flags
 };
 
 <% end %>

data/ext/cumo/narray/gen/tmpl/aref.c

@@ -21,35 +21,35 @@ static VALUE
 
   @example
       a = Cumo::DFloat.new(4,5).seq
-      => Cumo::DFloat#shape=[4,5]
-      [[0, 1, 2, 3, 4],
-       [5, 6, 7, 8, 9],
-       [10, 11, 12, 13, 14],
-       [15, 16, 17, 18, 19]]
+      # => Cumo::DFloat#shape=[4,5]
+      # [[0, 1, 2, 3, 4],
+      #  [5, 6, 7, 8, 9],
+      #  [10, 11, 12, 13, 14],
+      #  [15, 16, 17, 18, 19]]
 
       a[7]
-      => Cumo::DFloat#shape=[]
-      6.0
+      # => Cumo::DFloat#shape=[]
+      # 6.0
 
       a[1,1]
-      => Cumo::DFloat#shape=[]
-      6.0
+      # => Cumo::DFloat#shape=[]
+      # 6.0
 
       a[1..3,1]
-      => Cumo::DFloat#shape=[3]
-      [6, 11, 16]
+      # => Cumo::DFloat#shape=[3]
+      # [6, 11, 16]
 
       a[1,[1,3,4]]
-      => Cumo::DFloat#shape=[3]
-      [6, 8, 9]
+      # => Cumo::DFloat#shape=[3]
+      # [6, 8, 9]
 
       a[true,2].fill(99)
       a
-      => Cumo::DFloat#shape=[4,5]
-      [[0, 1, 99, 3, 4],
-       [5, 6, 99, 8, 9],
-       [10, 11, 99, 13, 14],
-       [15, 16, 99, 18, 19]]
+      # => Cumo::DFloat#shape=[4,5]
+      # [[0, 1, 99, 3, 4],
+      #  [5, 6, 99, 8, 9],
+      #  [10, 11, 99, 13, 14],
+      #  [15, 16, 99, 18, 19]]
  */
 static VALUE
 <%=c_func(-1)%>(int argc, VALUE *argv, VALUE self)

data/ext/cumo/narray/gen/tmpl/aset.c

@@ -10,31 +10,31 @@
 
   @example
       a = Cumo::DFloat.new(3,4).seq
-      => Cumo::DFloat#shape=[3,4]
-      [[0, 1, 2, 3],
-       [4, 5, 6, 7],
-       [8, 9, 10, 11]]
+      # => Cumo::DFloat#shape=[3,4]
+      # [[0, 1, 2, 3],
+      #  [4, 5, 6, 7],
+      #  [8, 9, 10, 11]]
 
       a[1,2]=99
       a
-      => Cumo::DFloat#shape=[3,4]
-      [[0, 1, 2, 3],
-       [4, 5, 99, 7],
-       [8, 9, 10, 11]]
+      # => Cumo::DFloat#shape=[3,4]
+      # [[0, 1, 2, 3],
+      #  [4, 5, 99, 7],
+      #  [8, 9, 10, 11]]
 
       a[1,[0,2]] = [101,102]
       a
-      => Cumo::DFloat#shape=[3,4]
-      [[0, 1, 2, 3],
-       [101, 5, 102, 7],
-       [8, 9, 10, 11]]
+      # => Cumo::DFloat#shape=[3,4]
+      # [[0, 1, 2, 3],
+      #  [101, 5, 102, 7],
+      #  [8, 9, 10, 11]]
 
       a[1,true]=99
       a
-      => Cumo::DFloat#shape=[3,4]
-      [[0, 1, 2, 3],
-       [99, 99, 99, 99],
-       [8, 9, 10, 11]]
+      # => Cumo::DFloat#shape=[3,4]
+      # [[0, 1, 2, 3],
+      #  [99, 99, 99, 99],
+      #  [8, 9, 10, 11]]
 
 */
 static VALUE

data/ext/cumo/narray/gen/tmpl/at.c

@@ -0,0 +1,34 @@
+/*
+  Multi-dimensional array indexing.
+  Same as [] for one-dimensional NArray.
+  Similar to numpy's tuple indexing, i.e., `a[[1,2,..],[3,4,..]]`
+  @overload at(*indices)
+  @param [Numeric,Range,etc] *indices  Multi-dimensional Index Arrays.
+  @return [Cumo::NArray::<%=class_name%>] one-dimensional NArray view.
+
+  @example
+      x = Cumo::DFloat.new(3,3,3).seq
+      => Cumo::DFloat#shape=[3,3,3]
+       [[[0, 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]]]
+
+      x.at([0,1,2],[0,1,2],[-1,-2,-3])
+      => Cumo::DFloat(view)#shape=[3]
+       [2, 13, 24]
+ */
+static VALUE
+<%=c_func(-1)%>(int argc, VALUE *argv, VALUE self)
+{
+    int result_nd;
+    size_t pos;
+
+    result_nd = cumo_na_get_result_dimension(self, argc, argv, sizeof(dtype), &pos);
+    return cumo_na_at_main(argc, argv, self, 0, result_nd, pos);
+}

data/ext/cumo/narray/gen/tmpl/batch_norm.c

@@ -157,8 +157,11 @@ static VALUE
     status = cumo_cuda_cudnn_CreateTensorDescriptor(&x_desc, x_cont, cudnn_dtype);
     if (status != CUDNN_STATUS_SUCCESS) goto BATCH_NORM_ERROR;
 
+    status = cudnnCreateTensorDescriptor(&bn_desc);
+    if (status != CUDNN_STATUS_SUCCESS) goto BATCH_NORM_ERROR;
+
     mode = cumo_cuda_cudnn_GetBatchNormMode(axis_ndim, int_axis);
-    status = cumo_cuda_cudnn_CreateBNTensorDescriptor(&bn_desc, x_desc, mode);
+    status = cudnnDeriveBNTensorDescriptor(bn_desc, x_desc, mode);
     if (status != CUDNN_STATUS_SUCCESS) goto BATCH_NORM_ERROR;
     // TODO: bn_desc may return another type, and may need to cast gamma, beta, mean, var
 
@@ -193,7 +196,7 @@ BATCH_NORM_ERROR:
 }
 
 #else // CUDNN_FOUND
-VALUE cumo_cuda_eCUDNNError;
+#include "cumo/cuda/cudnn.h"
 
 static VALUE
 <%=c_func(-1)%>(int argc, VALUE argv[], VALUE self)

data/ext/cumo/narray/gen/tmpl/batch_norm_backward.c

@@ -134,8 +134,11 @@ static VALUE
     status = cumo_cuda_cudnn_CreateTensorDescriptor(&x_desc, x_cont, cudnn_dtype);
     if (status != CUDNN_STATUS_SUCCESS) goto BATCH_NORM_BACKWARD_ERROR;
 
+    status = cudnnCreateTensorDescriptor(&bn_desc);
+    if (status != CUDNN_STATUS_SUCCESS) goto BATCH_NORM_BACKWARD_ERROR;
+
     mode = cumo_cuda_cudnn_GetBatchNormMode(axis_ndim, int_axis);
-    status = cumo_cuda_cudnn_CreateBNTensorDescriptor(&bn_desc, x_desc, mode);
+    status = cudnnDeriveBNTensorDescriptor(bn_desc, x_desc, mode);
     if (status != CUDNN_STATUS_SUCCESS) goto BATCH_NORM_BACKWARD_ERROR;
     // TODO: bn_desc may return another type, and may need to cast gamma, gy, mean, var
 
@@ -178,11 +181,11 @@ BATCH_NORM_BACKWARD_ERROR:
 }
 
 #else // CUDNN_FOUND
-VALUE cumo_cuda_eCudnnError;
+#include "cumo/cuda/cudnn.h"
 
 static VALUE
 <%=c_func(-1)%>(int argc, VALUE argv[], VALUE self)
 {
-    rb_raise(cumo_cuda_eCudnnError, "cuDNN is not available");
+    rb_raise(cumo_cuda_eCUDNNError, "cuDNN is not available");
 }
 #endif // CUDNN_FOUND

data/ext/cumo/narray/gen/tmpl/bincount.c

@@ -116,22 +116,22 @@ static VALUE
     otherwise returns UInt32 or UInt64 depending on the size along last axis.
   @example
     Cumo::Int32[0..4].bincount
-    => Cumo::UInt32#shape=[5]
-       [1, 1, 1, 1, 1]
+    # => Cumo::UInt32#shape=[5]
+    #    [1, 1, 1, 1, 1]
 
     Cumo::Int32[0, 1, 1, 3, 2, 1, 7].bincount
-    => Cumo::UInt32#shape=[8]
-       [1, 3, 1, 1, 0, 0, 0, 1]
+    # => Cumo::UInt32#shape=[8]
+    #    [1, 3, 1, 1, 0, 0, 0, 1]
 
     x = Cumo::Int32[0, 1, 1, 3, 2, 1, 7, 23]
     x.bincount.size == x.max+1
-    => true
+    # => true
 
     w = Cumo::DFloat[0.3, 0.5, 0.2, 0.7, 1.0, -0.6]
     x = Cumo::Int32[0, 1, 1, 2, 2, 2]
     x.bincount(w)
-    => Cumo::DFloat#shape=[3]
-       [0.3, 0.7, 1.1]
+    # => Cumo::DFloat#shape=[3]
+    #    [0.3, 0.7, 1.1]
 
 */
 static VALUE

data/ext/cumo/narray/gen/tmpl/clip.c

@@ -75,28 +75,24 @@ static void
 
   @example
       a = Cumo::Int32.new(10).seq
-      p a.clip(1,8)
-      # Cumo::Int32#shape=[10]
-      # [1, 1, 2, 3, 4, 5, 6, 7, 8, 8]
-
-      p a
-      # Cumo::Int32#shape=[10]
+      # => Cumo::Int32#shape=[10]
       # [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
 
-      p a.inplace.clip(3,6)
-      # Cumo::Int32(view)#shape=[10]
-      # [3, 3, 3, 3, 4, 5, 6, 6, 6, 6]
+      a.clip(1,8)
+      # => Cumo::Int32#shape=[10]
+      # [1, 1, 2, 3, 4, 5, 6, 7, 8, 8]
 
-      p a
-      # Cumo::Int32#shape=[10]
+      a.inplace.clip(3,6)
+      a
+      # => Cumo::Int32#shape=[10]
       # [3, 3, 3, 3, 4, 5, 6, 6, 6, 6]
 
-      p a = Cumo::Int32.new(10).seq
-      # Cumo::Int32#shape=[10]
+      b = Cumo::Int32.new(10).seq
+      # => Cumo::Int32#shape=[10]
       # [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
 
-      p a.clip([3,4,1,1,1,4,4,4,4,4], 8)
-      # Cumo::Int32#shape=[10]
+      b.clip([3,4,1,1,1,4,4,4,4,4], 8)
+      # => Cumo::Int32#shape=[10]
       # [3, 4, 2, 3, 4, 5, 6, 7, 8, 8]
 */
 static VALUE

data/ext/cumo/narray/gen/tmpl/conv.c

@@ -206,7 +206,7 @@ CONV_ERROR:
 }
 
 #else // CUDNN_FOUND
-VALUE cumo_cuda_eCUDNNError;
+#include "cumo/cuda/cudnn.h"
 
 static VALUE
 <%=c_func(-1)%>(int argc, VALUE argv[], VALUE self)

data/ext/cumo/narray/gen/tmpl/conv_grad_w.c

@@ -95,6 +95,7 @@ static VALUE
     CUMO_CUDA_CUDNN_CHECK_DIM_EQ(sizet_w_shape[0], ngy->shape[1]);
     CUMO_CUDA_CUDNN_CHECK_DIM_EQ(sizet_w_shape[1], nx->shape[1]);
 
+#if !defined(NDEBUG)
     {
         // shape check of gy
         size_t *y_shape = ngy->shape;
@@ -105,6 +106,7 @@ static VALUE
                     x_shape[i + 2], sizet_w_shape[i + 2], int_stride[i], int_pad[i]));
         }
     }
+#endif
 
     x_cont = cumo_na_as_contiguous_array(x);
     gy_cont = cumo_na_as_contiguous_array(gy);
@@ -173,7 +175,7 @@ CONV_GRAD_W_ERROR:
 }
 
 #else // CUDNN_FOUND
-VALUE cumo_cuda_eCUDNNError;
+#include "cumo/cuda/cudnn.h"
 
 static VALUE
 <%=c_func(-1)%>(int argc, VALUE argv[], VALUE self)

data/ext/cumo/narray/gen/tmpl/conv_transpose.c

@@ -234,7 +234,7 @@ CONV_TRANSPOSE_ERROR:
 }
 
 #else // CUDNN_FOUND
-VALUE cumo_cuda_eCUDNNError;
+#include "cumo/cuda/cudnn.h"
 
 static VALUE
 <%=c_func(-1)%>(int argc, VALUE argv[], VALUE self)

data/ext/cumo/narray/gen/tmpl/cum.c

@@ -30,7 +30,7 @@ static void
 /*
   <%=name%> of self.
   @overload <%=name%>(axis:nil, nan:false)
-  @param [Numeric,Array,Range] axis  Affected dimensions.
+  @param [Numeric,Array,Range] axis  Performs <%=name%> along the axis.
   @param [TrueClass] nan  If true, apply NaN-aware algorithm (avoid NaN if exists).
   @return [Cumo::<%=class_name%>] <%=name%> of self.
 */

data/ext/cumo/narray/gen/tmpl/each.c

@@ -34,8 +34,10 @@ static void
   passing that element as a parameter.
   @overload <%=name%>
   @return [Cumo::NArray] self
-  For a block {|x| ... }
-  @yield [x]  x is element of NArray.
+  For a block `{|x| ... }`,
+  @yieldparam [Numeric] x  an element of NArray.
+  @see #each_with_index
+  @see #map
 */
 static VALUE
 <%=c_func(0)%>(VALUE self)

data/ext/cumo/narray/gen/tmpl/each_with_index.c

@@ -55,9 +55,12 @@ static void
   Invokes the given block once for each element of self,
   passing that element and indices along each axis as parameters.
   @overload <%=name%>
+  For a block `{|x,i,j,...| ... }`,
+  @yieldparam [Numeric] x  an element
+  @yieldparam [Integer] i,j,...  multitimensional indices
   @return [Cumo::NArray] self
-  For a block {|x,i,j,...| ... }
-  @yield [x,i,j,...]  x is an element, i,j,... are multidimensional indices.
+  @see #each
+  @see #map_with_index
 */
 static VALUE
 <%=c_func(0)%>(VALUE self)

data/ext/cumo/narray/gen/tmpl/fixed_batch_norm.c

@@ -106,8 +106,11 @@ static VALUE
     status = cumo_cuda_cudnn_CreateTensorDescriptor(&x_desc, x_cont, cudnn_dtype);
     if (status != CUDNN_STATUS_SUCCESS) goto FIXED_BATCH_NORM_ERROR;
 
+    status = cudnnCreateTensorDescriptor(&bn_desc);
+    if (status != CUDNN_STATUS_SUCCESS) goto FIXED_BATCH_NORM_ERROR;
+
     mode = cumo_cuda_cudnn_GetBatchNormMode(axis_ndim, int_axis);
-    status = cumo_cuda_cudnn_CreateBNTensorDescriptor(&bn_desc, x_desc, mode);
+    status = cudnnDeriveBNTensorDescriptor(bn_desc, x_desc, mode);
     if (status != CUDNN_STATUS_SUCCESS) goto FIXED_BATCH_NORM_ERROR;
     // TODO: bn_desc may return another type, and may need to cast gamma, beta, mean, var
 
@@ -139,7 +142,7 @@ FIXED_BATCH_NORM_ERROR:
 }
 
 #else // CUDNN_FOUND
-VALUE cumo_cuda_eCUDNNError;
+#include "cumo/cuda/cudnn.h"
 
 static VALUE
 <%=c_func(-1)%>(int argc, VALUE argv[], VALUE self)

data/ext/cumo/narray/gen/tmpl/init_class.c

@@ -14,6 +14,7 @@
     rb_hash_aset(hCast, rb_cArray,   cT);
     <% for x in upcast %>
     <%= x %><% end %>
+    rb_obj_freeze(hCast);
 
     <% @children.each do |m| %>
     <%= m.init_def %><% end %>