Merge pull request #159 from marty1885/apichange

Various updates

Etaler/Backends/OpenCLBackend.cpp

@@ -42,7 +42,9 @@ typedef struct __attribute__ ((packed)) _OpenCLView
 static void makeOpenCLView(const TensorImpl* x, OpenCLView* v)
 {
 	int dims = int(x->dimensions());
-	et_assert(dims <= OPENCL_TENSOR_MAX_DIMS, "Too much dimensions for OpenCL backend.");
+	et_check(dims <= OPENCL_TENSOR_MAX_DIMS
+		, "The OpenCL backend can only handle up to" + std::to_string(OPENCL_TENSOR_MAX_DIMS) + "D view"
+		"got " + std::to_string(dims) + "D");
 	auto stride = x->stride();
 	auto shape_stride = shapeToStride(x->shape());
 	for(int i=0;i<dims;i++) {

Etaler/Core/Shape.hpp

@@ -139,6 +139,8 @@ inline size_t unfold(const IdxType& index, const StrideType& stride)
 
 inline Shape shapeToStride(const Shape& shape)
 {
+	if(shape.empty())
+		return {};
 	Shape v;
 	v.resize(shape.size());
 	size_t acc = 1;
@@ -159,7 +161,7 @@ inline size_t unfoldIndex(const IdxType& index, const ShapeType& shape)
 template <typename ShapeType>
 inline Shape foldIndex(size_t index, const ShapeType& shape)
 {
-	assert(shape.size() != 0);
+	assert(shape.size() != 0 || index == 0);
 	svector<intmax_t> v = shapeToStride(shape);
 	Shape res;
 	res.resize(v.size());

Etaler/Core/SmallVector.hpp

@@ -960,6 +960,10 @@ void SmallVectorBase::grow_pod(void *FirstEl, size_t MinSizeInBytes,
   void *NewElts;
   if (BeginX == FirstEl) {
     NewElts = malloc(NewCapacityInBytes);
+    if(NewElts == NULL) {
+      fprintf(stderr, "SmallVectorBase::grow_pod failed: mad malloc");
+      abort();
+    }
 
     // Copy the elements over.  No need to run dtors on PODs.
     memcpy(NewElts, this->BeginX, CurSizeBytes);
@@ -982,4 +986,9 @@ namespace et
 	using svector = llvm_vecsmall::SmallVector<T, N>;
 }
 
+// Undef LLVM macros
+#undef LLVM_VECSMALL_NODISCARD
+#undef LLVM_VECSMALL_ATTRIBUTE_ALWAYS_INLINE
+#undef LLVM_VECSMALL_UNLIKELY
+
 #endif

Etaler/Core/Tensor.hpp

@@ -60,7 +60,7 @@ struct ETALER_EXPORT TensorIterator
 Tensor ETALER_EXPORT brodcast_to(const Tensor& t, Shape s);
 
 ETALER_EXPORT std::ostream& operator<< (std::ostream& os, const Tensor& t);
-std::string to_string(const Tensor& t);
+ETALER_EXPORT std::string to_string(const Tensor& t);
 
 using IndexList = svector<std::variant<Range, intmax_t, int, size_t, unsigned int>>;
 

README.md

@@ -88,10 +88,13 @@ Documents are avalible online on [Read the Docs](https://etaler.readthedocs.io/e
 | Linux      | Yes | Yes    |
 | OS X       | Yes | Yes    |
 | Windows    | Yes | Yes    |
+| FreeBSD    | Yes | *Yes   |
 
 * Build with GCC and libstdc++ on OS X 10.11.
 * Clang should work after OS X 10.14. See [BuildOnOSX.md](docs/BuildOnOSX.md)
 * Build with Visual Studio 2019 on Windows. See [BuildOnMSVC.md](docs/BuildOnMSVC.md)
+* OpenCL on FreeBSD is tested using POCL. Which has know bugs preventing Etaler to fully function on ARM.
+   * See [#151](https://github.com/etaler/Etaler/issues/151)
 
 ### Dependencies
 

docs/source/DeveloperNotes.md

@@ -23,6 +23,7 @@ This file documents some facts about the codebase that might be useful for it's
 * Nvidia's OpenCL implementation although very optimized, has piles upon piles of problems.
   * Use POCL w/ CUDA backend for debugging. POCL is a lot slower, but very stable.
   * Or use Intel/AMD's OpenCL SDK
+  * Mesa's `clover` [OpenCL SDK](https://cgit.freedesktop.org/mesa/mesa/log/?qt=grep&q=clover) works, but quite slow.
 
 ## Programing style