cumo 0.3.0 → 0.3.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: e2ed7641113820552c74682336a8a7a92cce11e8897a4b6e0d9f9d416431242d
-  data.tar.gz: b5e6a459c3bf096991587a40c7b1efeba3aa1b2004e37604d3adf5fc7dce68ee
+  metadata.gz: 8f07518545c0e4d72380382915462039151c132ee9827967e9d4c32604001f15
+  data.tar.gz: 94e5c261659e401d9a8016eb0a3493d9bce26abe19347602859f7a9fcc79014e
 SHA512:
-  metadata.gz: 86cdf6a12b38ebb87c672cc0cb87ed7a51a2b37fc5fdf0b21d386ac4446afa102952da69e8799130abef8420b2bf6cd09836d6d14832b8ca2d4693f0718275e8
-  data.tar.gz: 0f881077419ee9ca3c4b5b09fcf01e35be9884cca9d736745bdd45c5f8c3fab8b856a0c336f486ac12f26209d81f2128b8eefbcdbd3b240259ca21479dd13f41
+  metadata.gz: cf48d1e5452e2c871587da1f006b90f3de54c8dc911d4fa555bd54c7ea69741903ce000fd0b1fd38ba96fc46d65e79f24e3c9d858029df176cee7768e8de790b
+  data.tar.gz: 4c5d6bdc9ec211967458b9b4f3576a278b2d59011faf2ae5dd85979a4821496ad29db9eefd0fce9d0765f9ede38d42af2c8fc8dd250de41f9e22759ee9914a56

data/CHANGELOG.md

@@ -1,3 +1,11 @@
+# 0.3.1 (2019-04-16)
+
+Fixes:
+
+* Fix batch\_norm\_backward
+* Fix scalar.dot(scalar)
+* Fix clip
+
 # 0.3.0 (2019-04-10)
 
 Enhancements:

data/ext/cumo/include/cumo.h

@@ -10,8 +10,8 @@ extern "C" {
 #endif
 #endif
 
-#define CUMO_VERSION "0.3.0"
-#define CUMO_VERSION_CODE 30
+#define CUMO_VERSION "0.3.1"
+#define CUMO_VERSION_CODE 31
 
 bool cumo_compatible_mode_enabled_p();
 bool cumo_show_warning_enabled_p();

data/ext/cumo/include/cumo/cuda/cudnn.h

@@ -5,6 +5,7 @@
 #ifdef CUDNN_FOUND
 #include <cudnn.h>
 #endif // CUDNN_FOUND
+#include "cumo/narray.h"
 
 #if defined(__cplusplus)
 extern "C" {
@@ -81,6 +82,24 @@ cumo_cuda_cudnn_get_int_ary(int* int_ary, VALUE ary, size_t ndim, int default_va
     }
 }
 
+// VALUE is Ruby Array
+static inline size_t
+cumo_cuda_cudnn_get_int_axis(int* int_axis, VALUE axis)
+{
+    size_t axis_ndim;
+    Check_Type(axis, T_ARRAY);
+    axis_ndim = (size_t)(RARRAY_LEN(axis));
+    if (CUMO_NA_MAX_DIMENSION <= axis_ndim) {
+        rb_raise(rb_eArgError, "Size of axis must be smaller than %d, but was %d",
+                (int)CUMO_NA_MAX_DIMENSION, (int)axis_ndim);
+    }
+    for (size_t idim = 0; idim < axis_ndim; ++idim) {
+        int_axis[idim] = NUM2INT(rb_ary_entry(axis, (long)idim));
+    }
+    // TODO: check axis is sorted
+    return axis_ndim;
+}
+
 size_t
 cumo_cuda_cudnn_GetConvOutDim(
         size_t in_dim,

data/ext/cumo/narray/gen/tmpl/accum_binary_kernel.cu

@@ -36,6 +36,10 @@ void <%="cumo_#{type_name}_mulsum#{nan}_reduce_kernel_launch"%>(char *p1, char *
 {
     ssize_t s1_idx = s1 / sizeof(dtype);
     ssize_t s2_idx = s2 / sizeof(dtype);
+    if (n == 1) { // when n == 1, s1 and s3 could be 0
+        s1_idx = 1;
+        s2_idx = 1;
+    }
     thrust::device_ptr<dtype> p1_begin = thrust::device_pointer_cast((dtype*)p1);
     thrust::device_ptr<dtype> p1_end   = thrust::device_pointer_cast(((dtype*)p1) + n * s1_idx);
     thrust::device_ptr<dtype> p2_begin = thrust::device_pointer_cast((dtype*)p2);

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

@@ -88,12 +88,7 @@ static VALUE
         double_decay = NUM2DBL(decay);
     }
     if (axis != Qnil) {
-        Check_Type(axis, T_ARRAY);
-        axis_ndim = (size_t)(RARRAY_LEN(axis));
-        for (size_t idim = 0; idim < axis_ndim; ++idim) {
-            int_axis[idim] = NUM2INT(rb_ary_entry(axis, (long)idim));
-        }
-        // TODO: check axis is sorted
+        axis_ndim = cumo_cuda_cudnn_get_int_axis(int_axis, axis);
     }
 
     CumoGetNArray(x, nx);

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

@@ -55,7 +55,7 @@ static VALUE
     size_t axis_ndim = 1;
 
     rb_scan_args(argc, argv, "2:", &gamma, &gy, &kw_hash);
-    rb_get_kwargs(kw_hash, kw_table, 0, 8, opts);
+    rb_get_kwargs(kw_hash, kw_table, 0, 7, opts);
     mean = cumo_cuda_cudnn_option_value(opts[0], Qnil);
     inv_std = cumo_cuda_cudnn_option_value(opts[1], Qnil);
     eps = cumo_cuda_cudnn_option_value(opts[2], Qnil);
@@ -74,12 +74,7 @@ static VALUE
         double_eps = NUM2DBL(eps);
     }
     if (axis != Qnil) {
-        Check_Type(axis, T_ARRAY);
-        axis_ndim = (size_t)(RARRAY_LEN(axis));
-        for (size_t idim = 0; idim < axis_ndim; ++idim) {
-            int_axis[idim] = NUM2INT(rb_ary_entry(axis, (long)idim));
-        }
-        // TODO: check axis is sorted
+        axis_ndim = cumo_cuda_cudnn_get_int_axis(int_axis, axis);
     }
 
     CumoGetNArray(x, nx);
@@ -135,11 +130,11 @@ static VALUE
     gbeta_ptr = cumo_na_get_offset_pointer_for_write(gbeta);
 
     status = cumo_cuda_cudnn_CreateTensorDescriptor(&x_desc, x_cont, cudnn_dtype);
-    if (status != CUDNN_STATUS_SUCCESS) goto BATCH_NORM_ERROR;
+    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);
-    if (status != CUDNN_STATUS_SUCCESS) goto BATCH_NORM_ERROR;
+    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
 
     handle = cumo_cuda_cudnn_handle();
@@ -164,9 +159,9 @@ static VALUE
             double_eps,
             mean_ptr,
             inv_std_ptr);
-    if (status != CUDNN_STATUS_SUCCESS) goto BATCH_NORM_ERROR;
+    if (status != CUDNN_STATUS_SUCCESS) goto BATCH_NORM_BACKWARD_ERROR;
 
-BATCH_NORM_ERROR:
+BATCH_NORM_BACKWARD_ERROR:
     if (x_desc) cudnnDestroyTensorDescriptor(x_desc);
     if (bn_desc) cudnnDestroyTensorDescriptor(bn_desc);
     cumo_cuda_cudnn_check_status(status);

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

@@ -11,6 +11,7 @@ static void
     CUMO_INIT_PTR(lp, 2, p3, s3);
     CUMO_INIT_PTR(lp, 3, p4, s4);
     CUMO_SHOW_SYNCHRONIZE_FIXME_WARNING_ONCE("<%=name%>", "<%=type_name%>");
+    cumo_cuda_runtime_check_status(cudaDeviceSynchronize());
     for (; i--;) {
         CUMO_GET_DATA_STRIDE(p1,s1,dtype,x);
         CUMO_GET_DATA_STRIDE(p2,s2,dtype,min);
@@ -34,6 +35,7 @@ static void
     CUMO_INIT_PTR(lp, 1, p2, s2);
     CUMO_INIT_PTR(lp, 2, p3, s3);
     CUMO_SHOW_SYNCHRONIZE_FIXME_WARNING_ONCE("<%=name%>_min", "<%=type_name%>");
+    cumo_cuda_runtime_check_status(cudaDeviceSynchronize());
     for (; i--;) {
         CUMO_GET_DATA_STRIDE(p1,s1,dtype,x);
         CUMO_GET_DATA_STRIDE(p2,s2,dtype,min);
@@ -54,6 +56,7 @@ static void
     CUMO_INIT_PTR(lp, 1, p2, s2);
     CUMO_INIT_PTR(lp, 2, p3, s3);
     CUMO_SHOW_SYNCHRONIZE_FIXME_WARNING_ONCE("<%=name%>_max", "<%=type_name%>");
+    cumo_cuda_runtime_check_status(cudaDeviceSynchronize());
     for (; i--;) {
         CUMO_GET_DATA_STRIDE(p1,s1,dtype,x);
         CUMO_GET_DATA_STRIDE(p2,s2,dtype,max);

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

@@ -134,13 +134,13 @@ static VALUE
     y_ptr = cumo_na_get_offset_pointer_for_write(y);
 
     status = cumo_cuda_cudnn_CreateTensorDescriptor(&x_desc, x_cont, cudnn_dtype);
-    if (status != CUDNN_STATUS_SUCCESS) goto CONV_ERROR;
+    if (status != CUDNN_STATUS_SUCCESS) goto CONV_TRANSPOSE_ERROR;
     status = cumo_cuda_cudnn_CreateTensorDescriptor(&y_desc, y, cudnn_dtype);
-    if (status != CUDNN_STATUS_SUCCESS) goto CONV_ERROR;
+    if (status != CUDNN_STATUS_SUCCESS) goto CONV_TRANSPOSE_ERROR;
     status = cumo_cuda_cudnn_CreateFilterDescriptor(&w_desc, w_cont, cudnn_dtype);
-    if (status != CUDNN_STATUS_SUCCESS) goto CONV_ERROR;
+    if (status != CUDNN_STATUS_SUCCESS) goto CONV_TRANSPOSE_ERROR;
     status = cumo_cuda_cudnn_CreateConvolutionDescriptor(&conv_desc, ndim, int_stride, int_pad, cudnn_dtype);
-    if (status != CUDNN_STATUS_SUCCESS) goto CONV_ERROR;
+    if (status != CUDNN_STATUS_SUCCESS) goto CONV_TRANSPOSE_ERROR;
 
     handle = cumo_cuda_cudnn_handle();
 
@@ -160,7 +160,7 @@ static VALUE
             int_pad,
             ndim,
             cudnn_dtype);
-    if (status != CUDNN_STATUS_SUCCESS) goto CONV_ERROR;
+    if (status != CUDNN_STATUS_SUCCESS) goto CONV_TRANSPOSE_ERROR;
     algo = perf_result.algo;
     workspace_size = perf_result.memory;
 
@@ -179,7 +179,7 @@ static VALUE
             (void*)&beta,
             y_desc,
             (void*)y_ptr);
-    if (status != CUDNN_STATUS_SUCCESS) goto CONV_ERROR;
+    if (status != CUDNN_STATUS_SUCCESS) goto CONV_TRANSPOSE_ERROR;
 
     if (b != Qnil) {
         size_t new_shape[CUMO_NA_MAX_DIMENSION];
@@ -208,7 +208,7 @@ static VALUE
         // restore b.shape
         nb_cont->ndim = b_ndim;
         nb_cont->shape = b_shape;
-        if (status != CUDNN_STATUS_SUCCESS) goto CONV_ERROR;
+        if (status != CUDNN_STATUS_SUCCESS) goto CONV_TRANSPOSE_ERROR;
 
         status = cudnnAddTensor(
                     handle,
@@ -218,10 +218,10 @@ static VALUE
                     (void*)&alpha,
                     y_desc,
                     (void*)y_ptr);
-        if (status != CUDNN_STATUS_SUCCESS) goto CONV_ERROR;
+        if (status != CUDNN_STATUS_SUCCESS) goto CONV_TRANSPOSE_ERROR;
     }
 
-CONV_ERROR:
+CONV_TRANSPOSE_ERROR:
     if (x_desc) cudnnDestroyTensorDescriptor(x_desc);
     if (y_desc) cudnnDestroyTensorDescriptor(y_desc);
     if (b_desc) cudnnDestroyTensorDescriptor(b_desc);

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

@@ -94,11 +94,11 @@ static VALUE
     gx_ptr = cumo_na_get_offset_pointer_for_write(gx);
 
     status = cumo_cuda_cudnn_CreateTensorDescriptor(&x_desc, x_cont, cudnn_dtype);
-    if (status != CUDNN_STATUS_SUCCESS) goto POOLING_ERROR;
+    if (status != CUDNN_STATUS_SUCCESS) goto POOLING_BACKAWARD_ERROR;
     status = cumo_cuda_cudnn_CreateTensorDescriptor(&y_desc, y, cudnn_dtype);
-    if (status != CUDNN_STATUS_SUCCESS) goto POOLING_ERROR;
+    if (status != CUDNN_STATUS_SUCCESS) goto POOLING_BACKAWARD_ERROR;
     status = cumo_cuda_cudnn_CreatePoolingDescriptor(&pool_desc, int_mode, ndim, int_kernel_size, int_stride, int_pad);
-    if (status != CUDNN_STATUS_SUCCESS) goto POOLING_ERROR;
+    if (status != CUDNN_STATUS_SUCCESS) goto POOLING_BACKAWARD_ERROR;
 
     handle = cumo_cuda_cudnn_handle();
     status = cudnnPoolingBackward(
@@ -114,9 +114,9 @@ static VALUE
             (void*)&beta,
             x_desc,
             (void*)gx_ptr);
-    if (status != CUDNN_STATUS_SUCCESS) goto POOLING_ERROR;
+    if (status != CUDNN_STATUS_SUCCESS) goto POOLING_BACKAWARD_ERROR;
 
-POOLING_ERROR:
+POOLING_BACKAWARD_ERROR:
     if (x_desc) cudnnDestroyTensorDescriptor(x_desc);
     if (y_desc) cudnnDestroyTensorDescriptor(y_desc);
     if (pool_desc) cudnnDestroyPoolingDescriptor(pool_desc);

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

@@ -98,11 +98,11 @@ static VALUE
     y_ptr = cumo_na_get_offset_pointer_for_write(y);
 
     status = cumo_cuda_cudnn_CreateTensorDescriptor(&x_desc, x_cont, cudnn_dtype);
-    if (status != CUDNN_STATUS_SUCCESS) goto POOLING_ERROR;
+    if (status != CUDNN_STATUS_SUCCESS) goto POLLING_FORWARD_ERROR;
     status = cumo_cuda_cudnn_CreateTensorDescriptor(&y_desc, y, cudnn_dtype);
-    if (status != CUDNN_STATUS_SUCCESS) goto POOLING_ERROR;
+    if (status != CUDNN_STATUS_SUCCESS) goto POLLING_FORWARD_ERROR;
     status = cumo_cuda_cudnn_CreatePoolingDescriptor(&pool_desc, int_mode, ndim, int_kernel_size, int_stride, int_pad);
-    if (status != CUDNN_STATUS_SUCCESS) goto POOLING_ERROR;
+    if (status != CUDNN_STATUS_SUCCESS) goto POLLING_FORWARD_ERROR;
 
     handle = cumo_cuda_cudnn_handle();
     status = cudnnPoolingForward(
@@ -114,9 +114,9 @@ static VALUE
             (void*)&beta,
             y_desc,
             (void*)y_ptr);
-    if (status != CUDNN_STATUS_SUCCESS) goto POOLING_ERROR;
+    if (status != CUDNN_STATUS_SUCCESS) goto POLLING_FORWARD_ERROR;
 
-POOLING_ERROR:
+POLLING_FORWARD_ERROR:
     if (x_desc) cudnnDestroyTensorDescriptor(x_desc);
     if (y_desc) cudnnDestroyTensorDescriptor(y_desc);
     if (pool_desc) cudnnDestroyPoolingDescriptor(pool_desc);

data/lib/cumo/narray/extra.rb

@@ -1133,6 +1133,9 @@ module Cumo
 
     def dot(b)
       t = self.class::UPCAST[b.class]
+      if self.ndim == 0 and b.ndim == 0
+        return self * b
+      end
       if [SFloat, DFloat, SComplex, DComplex].include?(t)
         b = self.class.asarray(b)
         case self.ndim

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: cumo
 version: !ruby/object:Gem::Version
-  version: 0.3.0
+  version: 0.3.1
 platform: ruby
 authors:
 - Naotoshi Seo
 autorequire: 
 bindir: exe
 cert_chain: []
-date: 2019-04-10 00:00:00.000000000 Z
+date: 2019-04-16 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: numo-narray