[StreamExecutor] Executor add synchronous methods

[jhen]

Summary

Add Executor methods that block the host until completion. Since these
methods are host-synchronous, they don't require Stream arguments.

[jprice]

General question: if we have variants of these memcpy methods that take ElementCount parameters to allow for partial copies to/from the device allocations, should we also have variants with an Offset parameter as well to allow for partial copies that don't start at the origin?

streamexecutor/include/streamexecutor/Executor.h

71

If this is going to be backing onto something like cuMemHostRegister, don't we need the size of the allocation as well?

144

memory

streamexecutor/include/streamexecutor/PlatformInterfaces.h

146

registerHostMemory

[jlebar]

streamexecutor/include/streamexecutor/Executor.h

48

Should we take this by reference?

110

I wonder if we should assert that the two arrays' sizes are the same, in this case. Same for the H2D function.

[jhen]

• Respond to jprice's comments

[jhen]

In D23577#517138, @jprice wrote:

General question: if we have variants of these memcpy methods that take ElementCount parameters to allow for partial copies to/from the device allocations, should we also have variants with an Offset parameter as well to allow for partial copies that don't start at the origin?

Yes, I think allowing an offset argument is a good idea. I made that change to the Executor and Stream memcpy methods

streamexecutor/include/streamexecutor/Executor.h

48

The internal SE code uses the convention that a DeviceMemory<T>* is passed wherever a T* would be passed for host data and a const DeviceMemory<T>& would be passed wherever a const T* would be used for host data.

Do you think it would be better to pass DeviceMemory<T>& for T* and const DeviceMemory<T>& for const T*, or something like that?

71

You're totally right. Thanks for catching that.

110

To me that feels too restrictive. Without that check do you think it would be better to rename the method? Maybe synchronousMemcpyWholeSrc or something. I like it how it is now, but I don't want the interface to be confusing.

[jlebar]

streamexecutor/include/streamexecutor/Executor.h

48

Do you think it would be better to pass DeviceMemory<T>& for T* and const DeviceMemory<T>& for const T*, or something like that?

Yes, I think so, since a DeviceMemory<T> is essentially a T*. Or put another way, when you allocate host memory, you get a T*, and when you allocate device memory, you get a DeviceMemory<T>.

109

Hm, maybe we should match the arg order of memcpy, given that we're using that name.

110

To me that feels too restrictive.

I guess it's a question of what we're optimizing for.

It seems to me that much of the time, the dst will be of the same size as the src. In such cases, you'd want to know if you accidentally mismatched the sizes, because that's going to leave part of your dst uninitialized.

If you really want to copy into only part of dst, you can make that explicit by creating a new array slice:

vector<int> dst(100);
synchronousMemcpyD2H(src, makeMutableArrayRef(dst.data(), DeviceSrc.getElementCount()));

(makeMutableArrayRef doesn't exist at the moment, but that's easily fixed...)

Or alternatively, you could just call the other overload:

vector<int> dst(100);
synchronousMemcpyD2H(src, dst, DeviceSrc.getElementCount());

Both cases are explicit that we're doing something other than simply copying everything from dst to src.

[jprice]

streamexecutor/include/streamexecutor/Executor.h

102

Should be passing SrcElementOffset * sizeof(T) through to the platform executor as well now.

135

Missing offset argument as above.

166

Missing offset argument as above.

streamexecutor/include/streamexecutor/Stream.h

122–124

Missing offset argument as above.

163–165

Missing offset argument as above.

202–205

Missing offset argument as above.

[jhen]

• Add missing offset args in raw memcpys

[jhen]

streamexecutor/include/streamexecutor/Executor.h

102

Good catch! I promise to write a bunch of unit tests to catch this kind of thing once we have the interface decided.

[jhen]

• Reverse src/dst memcpy arg order

streamexecutor/include/streamexecutor/Executor.h

48

The thing I would really miss from the current interface is the way the user has to specify a DeviceMemory as a copy destination by taking its address before passing it as an argument to the copy function. I like how this makes it hard for the user to accidentally pass the source as the destination and vice versa. Is there a good way to keep this nice feature and match the model of DeviceMemory as a pointer more closely?

110

Okay, that sounds like a good reason to me.

[jlebar]

streamexecutor/include/streamexecutor/Executor.h

48

Is there a good way to keep this nice feature and match the model of DeviceMemory as a pointer more closely?

const-ness will get us some of the way there.

I could imaging designing an API that would let you do

MutableArrayRef<int> dst(...);
dst = CopyFromDevice(src);

But, in addition to requiring some substantial magic, that wouldn't work with our .then chaining.

[jlebar]

I had a thought about MemcpyD2H and friends last night. Not to make you change it again, but it's kind of weird that the function name says "D2H" but the args are (host_ptr, device_ptr). So long as it's called "memcpy", I think we probably should match memcpy's ordering, but what if we just called it "copy"?

That might somewhat speak to your concern about users getting the order wrong.

[jhen]

In D23577#518094, @jlebar wrote:

I had a thought about MemcpyD2H and friends last night. Not to make you change it again, but it's kind of weird that the function name says "D2H" but the args are (host_ptr, device_ptr). So long as it's called "memcpy", I think we probably should match memcpy's ordering, but what if we just called it "copy"?

That might somewhat speak to your concern about users getting the order wrong.

Yes, the D2H, etc names were the reason for the original (src, dst) argument ordering, so I think there is a potential for confusion with the current ordering.

For an alternative solution that really helps the user know what they are doing, how about introducing different wrapper classes for src and dst and requiring the user to wrap the argument before passing it to the copy method:

// Wrapper classes.
template <typename T> class CopySrc;
template <typename T> class CopyDst;

// Copy function decl.
template <typename T>
Error synchronousCopy(CopyDst<T> Src, CopySrc<T> Dst, size_t ElementCount);

// Helper functions "to" and "from" to construct wrapper instances.
template <typename T> CopySrc<T> from(const GlobalDeviceMemory<T> &DeviceSrc, size_t Offset = 0);
template <typename T> CopySrc<T> from(const T *HostSrc, size_t Offset = 0);
template <typename T> CopyDst<T> to(GlobalDeviceMemory<T> DeviceDst, size_t Offset = 0);
template <typename T> CopyDst<T> to(T *HostDst, size_t Offset = 0);

// User code that calls a copy.
Executor.synchronousCopy(to(DeviceMemory, DeviceElementOffset), from(HostPtr), ElementCount);

This seems to me like a pretty foolproof interface, and I like that it allows the user to specify the offset along with the pointer only if an offset is desired.

What do you think?

[jhen]

• Device memcpy uses DeviceArrayRef arg

[jhen]

The to and from functions ended up not being so elegant because ADL doesn't happen for the host arguments (llvm::ArrayRef, llvm::MutableArrayRef). So, instead, I made two classes just to wrap the device arrays: DeviceArrayRef and DeviceMutableArrayRef. I think this makes a nice symmetry between host and device and makes the user be explicit about which argument is src and which is dst.

The one thing I don't like is that the DeviceArrayRef constructor takes a size_t offset as its second argument, and this might get confusing because the llvm::ArrayRef takes a size_t size as its second argument.

Once we get this synchronous memcpy interface the way we like it, I'll also fix the other functions like free and asynchronous copies.

[jlebar]

I'm not crazy about how verbose simple operations become with the current
revision.

Just trying to play with this API:

se::GlobalDeviceMemory<int> DMem;
int *HPtr;
MutableArrayRef<int> HArrRef;

// These two operations seem like they're going to be the most common:

// Copy everything, assert sizes equal.
E.syncCopyD2H(DMem, HArrRef);

// Up to user to ensure that HPtr is big enough.  (We could leave off "length"
// and copy everything, but since we're passing a raw pointer here, seems like
// we should be explicit about the length.)
E.syncCopyD2H(DMem, HPtr, length);

// These two together suggest that we should have a third overload:

// Copy length elements, assert sizes are less than or equal to length.
E.syncCopyD2H(DMem, HArrRef, length);

These three are probably sufficient for everything other than copying starting
at a nonzero offset within DMem? I guess that's where makeMutableArrayRef
comes in. But since this is going to be an uncommon operation, I don't think
we want to uglify all of the other call sites to make this work.

What if we called our class GlobalDeviceMemorySlice<T>. Not calling it
"ArrayRef" would leave us free to modify its API somewhat:

• We could have an implicit constructor from GlobalDeviceMemory<T>, so we don't have to explicitly call makeMutableArrayRef.

• We wouldn't need mutable and non-mutable versions of this; we could just use const to differentiate.

• If it makes sense, we could even make it so that the way to construct a GlobalDeviceMemorySlice from a GlobalDeviceMemory is by calling functions on GlobalDeviceMemory. This would resolve the ambiguity around offset vs. length. Here's a not well-thought-through API:
  
  GlobalDeviceMemory<int> DMem; E.syncCopyD2H(DMem.drop_front(5).drop_back(42).take_first(11));

I suspect that simply copying slice(), drop_front(), and drop_back() from
ArrayRef would be pretty good.

[jhen]

• Many overloads and DeviceArraySlice

[jhen]

jlebar's last design suggestion seems reasonable to me. I implemented it (just for the D2H direction for now). If people like it, I will do the other directions too (H2D, D2D) and write a bunch of tests.

[jlebar]

streamexecutor/include/streamexecutor/DeviceMemory.h

101

If this must wrap specifically a GlobalDeviceMemory object, we should probably have those words in the name. Maybe "GlobalDeviceMemorySlice"? Or is that too long?

It also seems kind of weird to me to focus on "arrays" in the name, function names, and comments, since it's no more or less as much of an array as GlobalDeviceMemory, but we don't name/describe it in the same way.

106

We should assert that these are in range.

110

I think getOffset() and getCount() might be better names, but we should definitely be consistent with the names in GlobalDeviceMemory, as you've done here.

116

We should think about what exactly we want to be an error in these functions. I'm not sure that silently accepting out-of-range values makes sense.

131

Hm... Maybe ArrayRef's two-arg slice() function is more useful than take_first(). take_first(n) then just becomes slice(0, n), which is fairly clear I think? And I think the common operation will likely be "starting at offset k, take n elements", which it's not trivial to do using the functions here.

I don't think we need both the one-arg slice() and drop_front, as they're equivalent.

Sorry if I mislead you on this one. In that example, I meant just to illustrate how we'd chain calls together, not to suggest specifically the API. But I don't think I was so clear about that.

192

We could avoid having these functions entirely by having an asSlice() function, but I'm not sure that's better than this API. However, it would be nice not to reimplement these tricky functions. Could we just do e.g.

return DeviceArraySlice<ElemT>(*this).drop_front(DropCount);

?

streamexecutor/include/streamexecutor/Executor.h

156

My thought is that we'd have an implicit one-arg constructor in DeviceArraySlice<T> so we wouldn't need this overload.

DeviceArraySlice<T>::DeviceArraySlice<T>(const GlobalDeviceMemory<T>& Mem);

(In the same way, llvm::ArrayRef has an implicit one-arg constructor that accepts an std::vector<T>.)

[jhen]

• Respond to jlebar's comments

streamexecutor/include/streamexecutor/DeviceMemory.h

110

I renamed them to getOffset() and getCount(). I also renamed the GlobalDeviceMemory<T> method to getCount() for consistency.

116

ArrayRef uses assert, so I did that here too.

192

asSlice() sounds good to me. I put that in and removed all of these.

streamexecutor/include/streamexecutor/Executor.h

156

I tried this and unfortunately it doesn't seem to work perfectly. I think the template type inference is getting in the way of selecting the implicit constructor. For example, I have to call Executor.syncCopyD2H<int>(DeviceA, llvm::MutableArray<int>(Host)) in my test code, where I explicitly specify the function template argument for the syncCopyD2H function. Otherwise, I get a compile error that says it ignored the candidate template because it could not match GlobalDeviceMemorySlice agains GlobalDeviceMemory.

I removed the syncCopyD2H overloads for GlobalDeviceMemory for now, so that users have to explicitly supply the template argument or call asSlice(). Should I add back the overloads in order to get around this problem?

[jlebar]

I think this is probably good, modulo a few relatively minor things.

streamexecutor/include/streamexecutor/DeviceMemory.h

116

Nit, maybe we can reword this so it doesn't just repeat the function's name.

119

This is kind of confusing to me, I think because "element offset from base memory" isn't so idiomatic. Maybe we can reword this?

147

Why do we do std::min here instead of asserting?

streamexecutor/include/streamexecutor/Executor.h

100

Ohhh. *This* is why they're called getElementOffset and getElementCount.

I guess it's not so bad if we don't have any data structure that exposes the byte offset / byte count. But even if not I find this a kind of compelling reason for those names.

What do you think?

122

Maybe we can just create an arrayref and call one of the other overloads?

139

typo

141

I might actually suggest making this function the base overload; it seems like the most strongly-typed and general?

193

Should I add back the overloads in order to get around this problem?

Argh. That's the common case, so I guess, yes. I guess you only need it for the two-arg overloads and not the three-arg overload. :-/

[jhen]

• Respond to jlebar's comments 2

streamexecutor/include/streamexecutor/DeviceMemory.h

147

I missed that one (and it was buggy as it was anyway). Thanks for catching that!

streamexecutor/include/streamexecutor/Executor.h

100

Yes, I definitely prefer the getElementOffset and getElementCount names. Sorry about the confusion. I've changed all the names back now to include "Element".

193

I put the overloads back. I couldn't even get the one that took T* as its second parameter to work without the overload, so I had to put them all back. I even tried to define a conversion operator for GlobalDeviceMemory<T> to GlobalDeviceMemorySlice<T>, but that didn't work either.

[jlebar]

streamexecutor/include/streamexecutor/Executor.h

177

Well, this is a lot of overloads, but I guess it could be worse. :)

My only thought is that it would be nice not to duplicate so much of the comments. But other than that, I think you can go forth and do this for real.

[jhen]

• Add H2D and D2D functions and tests

[jhen]

• Memory "free" no longer takes wrapper pointer

[jhen]

Okay, I've got all the synchronousCopy functions in for all directions (D2H, H2D, D2D) and all the tests are in, so I think this CL is ready for final review.

streamexecutor/include/streamexecutor/Executor.h

177

I got rid of most of the comment duplication by referencing earlier functions in the comments for later functions.

[jlebar]

streamexecutor/include/streamexecutor/Executor.h

93

Nit, can we make these two inequalities have the same direction?

streamexecutor/include/streamexecutor/PlatformInterfaces.h

160

Should we specify the device synchronicity semantics for these calls (and also in Executor, I guess)? The non-"synchronous" functions clearly are synchronous within a stream. But for these I have no idea what they're supposed to do.

streamexecutor/include/streamexecutor/Stream.h

111

I guess I expected that we'd match the Executor's interface, with the overloads and such. But maybe you plan to do that in a separate patch?

177

comma placement

[jhen]

• Update Stream copy interface

streamexecutor/include/streamexecutor/Stream.h

111

I was thinking of doing it in a separate patch, but changed my mind. Now these interfaces should be updated too.

[jprice]

Just some mild pedantry from me.

streamexecutor/include/streamexecutor/Executor.h

199

H2D

284

D2D

streamexecutor/include/streamexecutor/Stream.h

178

H2D

215

D2D

[jhen]

• Fix method names in error strings

[jhen]

In D23577#523434, @jprice wrote:

Just some mild pedantry from me.

:) Thanks for your help

[jlebar]

I think this is officially unwieldy. Which I accept the blame for. :)

It looks good to me, and I think if we want to change things, it will probably be a lot easier to do so once this is in.

streamexecutor/include/streamexecutor/PlatformInterfaces.h

161

Ah, but there can be no device work enqueued after the call starts but before it completes because this is host-synchronous! (Unless you can touch Streams on multiple host threads?)

So maybe the thing to say is simply that this does not block any ongoing device calls?

streamexecutor/include/streamexecutor/Stream.h

111

may

284

This review is getting so long I'm tempted to say it will be easier to check this in as-is and edit things like this later. But I think it would be nice to consider whether there's a way to reduce the comment copypasta -- there's a lot of it, and I think it obscures the API. I think in an earlier patch you did some doxygen "function group" magic?

Again, maybe better to save this for later.

[jhen]

• Clarify device synchronicity claims

[jhen]

In D23577#523872, @jlebar wrote:

I think this is officially unwieldy. Which I accept the blame for. :)

Yeah, it might be a bit of a maintenance problem, but I do think it will be much nicer for the user to use, so I think it's been an overall improvement.

streamexecutor/include/streamexecutor/Stream.h

284

I definitely think I can cut down the comments a lot and make the API more understandable in the process. As you suggested, I'll check this in now and then do that in the next CL.

[jprice]

Not sure what the protocol is for finding issues in patches that have already been committed - should I be posting to parallel_libs-dev instead?

This issue only showed up when I tried to actually write some code that uses all this stuff.

parallel-libs/trunk/streamexecutor/include/streamexecutor/Executor.h

44 ↗

(On Diff #69133)

There's a mismatch between the return types of these two allocateDeviceMemory methods. The PlatformExecutor version returns a GlobalDeviceMemoryBase, whereas this one tries to return a GlobalDeviceMemory<T>.

This causes compilation failure if this Executor::allocateDeviceMemory<T>() method is actually used. Because this method is templated, it needs to actually be called somewhere for the issue to show up, and it doesn't appear to be covered in the unit tests.

streamexecutor/include/streamexecutor/DeviceMemory.h

 /// The typed subclass will be used by user code, and the untyped base classes
 /// will be used for type-unsafe operations inside of StreamExecutor.
 ///
 //===----------------------------------------------------------------------===//
 
 #ifndef STREAMEXECUTOR_DEVICEMEMORY_H
 #define STREAMEXECUTOR_DEVICEMEMORY_H
 
+#include <algorithm>
 #include <cstddef>
 
 namespace streamexecutor {
 
 /// Wrapper around a generic global device memory allocation.
 ///
 /// This class represents a buffer of untyped bytes in the global memory space
 /// of a device. See GlobalDeviceMemory<T> for the corresponding type that
public:
   /// device virtual address space, but is platform-dependent.
   const void *getHandle() const { return Handle; }
 
 private:
   const void *Handle; // Platform-dependent value representing allocated memory.
   size_t ByteCount;   // Size in bytes of this allocation.
 };
 
+template <typename ElemT> class GlobalDeviceMemory;
+
+/// Reference to a slice of an array in device memory.
+///
+/// Contains a base array, an element count offset into that base array, and an
+/// element count for the size of the slice.
+template <typename ElemT> class DeviceArraySlice {
+public:
+  DeviceArraySlice(const GlobalDeviceMemory<ElemT> &BaseArray,
+                   size_t ElementOffset, size_t ElementCount)
+      : BaseArray(BaseArray), ElementOffset(ElementOffset),
+        ElementCount(ElementCount) {}
+
+  GlobalDeviceMemory<ElemT> &getBaseArray() { return BaseArray; }
+  size_t getElementOffset() const { return ElementOffset; }
+  size_t getElementCount() const { return ElementCount; }
+
+  /// Creates a slice of the array with the first DropCount elements removed.
+  DeviceArraySlice<ElemT> drop_front(size_t DropCount) {
+    return DeviceArraySlice<ElemT>(
+        BaseArray, ElementOffset + DropCount,
+        (DropCount > ElementCount) ? 0 : ElementCount - DropCount);
+  }
+
+  /// Creates a slice of the array with the last DropCount elements removed.
+  DeviceArraySlice<ElemT> drop_back(size_t DropCount) {
+    return DeviceArraySlice<ElemT>(
+        BaseArray, ElementOffset,
+        (DropCount > ElementCount) ? 0 : ElementCount - DropCount);
+  }
+
+  /// Creates a slice of the array that only contains the first TakeCount
+  /// elements.
+  DeviceArraySlice<ElemT> take_first(size_t TakeCount) {
+    return DeviceArraySlice<ElemT>(BaseArray, ElementOffset,
+                                   std::min(ElementCount, TakeCount));
+  }
+
+private:
+  GlobalDeviceMemory<ElemT> BaseArray;
+  size_t ElementOffset;
+  size_t ElementCount;
+};
+
 /// Typed wrapper around the "void *"-like GlobalDeviceMemoryBase class.
 ///
 /// For example, GlobalDeviceMemory<int> is a simple wrapper around
 /// GlobalDeviceMemoryBase that represents a buffer of integers stored in global
 /// device memory.
 template <typename ElemT>
 class GlobalDeviceMemory : public GlobalDeviceMemoryBase {
 public:
   /// Creates a typed area of GlobalDeviceMemory with a given opaque handle and
   /// the given element count.
   static GlobalDeviceMemory<ElemT> makeFromElementCount(const void *Handle,
                                                         size_t ElementCount) {
     return GlobalDeviceMemory<ElemT>(Handle, ElementCount);
   }
 
   /// Creates a typed device memory region from an untyped device memory region.
   ///
public:
 
   /// Copy-assignable like a pointer.
   GlobalDeviceMemory &operator=(const GlobalDeviceMemory &) = default;
 
   /// Returns the number of elements of type ElemT that constitute this
   /// allocation.
   size_t getElementCount() const { return getByteCount() / sizeof(ElemT); }
 
+  /// Creates a slice of the array with the first DropCount elements removed.
+  DeviceArraySlice<ElemT> drop_front(size_t DropCount) {
+    return DeviceArraySlice<ElemT>(
+        *this, DropCount,
+        (DropCount > getElementCount()) ? 0 : getElementCount() - DropCount);
+  }
+
+  /// Creates a slice of the array with the last DropCount elements removed.
+  DeviceArraySlice<ElemT> drop_back(size_t DropCount) {
+    return DeviceArraySlice<ElemT>(
+        *this, 0,
+        (DropCount > getElementCount()) ? 0 : getElementCount() - DropCount);
+  }
+
+  /// Creates a slice of the array that only contains the first TakeCount
+  /// elements.
+  DeviceArraySlice<ElemT> take_first(size_t TakeCount) {
+    return DeviceArraySlice<ElemT>(*this, 0,
+                                   std::min(getElementCount(), TakeCount));
+  }
+
 private:
   /// Constructs a GlobalDeviceMemory instance from an opaque handle and an
   /// element count.
   ///
   /// This constructor is not public because there is a potential for confusion
   /// between the size of the buffer in bytes and the size of the buffer in
   /// elements.
   ///

streamexecutor/include/streamexecutor/Executor.h

 /// The Executor class which represents a single device of a specific platform.
 ///
 //===----------------------------------------------------------------------===//
 
 #ifndef STREAMEXECUTOR_EXECUTOR_H
 #define STREAMEXECUTOR_EXECUTOR_H
 
 #include "streamexecutor/KernelSpec.h"
+#include "streamexecutor/PlatformInterfaces.h"
 #include "streamexecutor/Utils/Error.h"
 
 namespace streamexecutor {
 
 class KernelInterface;
-class PlatformExecutor;
 class Stream;
 
 class Executor {
 public:
   explicit Executor(PlatformExecutor *PExecutor);
   virtual ~Executor();
 
   /// Gets the kernel implementation for the underlying platform.
   virtual Expected<std::unique_ptr<KernelInterface>>
   getKernelImplementation(const MultiKernelLoaderSpec &Spec) {
     // TODO(jhen): Implement this.
     return nullptr;
   }
 
   Expected<std::unique_ptr<Stream>> createStream();
 
+  /// Allocates an array of ElementCount entries of type T in device memory.
+  template <typename T>
+  Expected<GlobalDeviceMemory<T>> allocateDeviceMemory(size_t ElementCount) {
+    return PExecutor->allocateDeviceMemory(ElementCount * sizeof(T));
+  }
+
+  /// Frees memory previously allocated with allocateDeviceMemory.
+  template <typename T> Error freeDeviceMemory(GlobalDeviceMemory<T> *Memory) {
+    return PExecutor->freeDeviceMemory(Memory);
+  }
+
+  /// Allocates an array of ElementCount entries of type T in host memory.
+  ///
+  /// Host memory allocated by this function can be used for asynchronous memory
+  /// copies on streams. See Stream::thenMemcpyD2H and Stream::thenMemcpyH2D.
+  template <typename T> Expected<T *> allocateHostMemory(size_t ElementCount) {
+    return PExecutor->allocateHostMemory(ElementCount * sizeof(T));
+  }
+
+  /// Frees memory previously allocated with allocateHostMemory.
+  template <typename T> Error freeHostMemory(T *Memory) {
+    return PExecutor->freeHostMemory(Memory);
+  }
+
+  /// Registers a previously allocated host array of type T for asynchronous
+  /// memory operations.
+  ///
+  /// Host memory registered by this function can be used for asynchronous
+  /// memory copies on streams. See Stream::thenMemcpyD2H and
+  /// Stream::thenMemcpyH2D.
+  template <typename T>
+  Error registerHostMemory(T *Memory, size_t ElementCount) {
+    return PExecutor->registerHostMemory(Memory, ElementCount * sizeof(T));
+  }
+
+  /// Unregisters host memory previously registered by registerHostMemory.
+  template <typename T> Error unregisterHostMemory(T *Memory) {
+    return PExecutor->unregisterHostMemory(Memory);
+  }
+
+  /// Host-synchronously copies a slice of an array of elements of type T from
+  /// host to device memory.
+  ///
+  /// Returns an error if the Src and Dst sizes do not match.
+  ///
+  /// The calling host thread is blocked until the copy completes. Can be used
+  /// with any host memory, the host memory does not have to be allocated with
+  /// allocateHostMemory or registered with registerHostMemory.
+  template <typename T>
+  Error synchronousCopyD2H(DeviceArraySlice<T> Src,
+                           llvm::MutableArrayRef<T> Dst) {
+    if (Src.getElementCount() != Dst.size())
+      return make_error("array size mismatch during copy D2H: device source "
+                        "element count = " +
+                        llvm::Twine(Src.getElementCount()) +
+                        ", host destination element count = " +
+                        llvm::Twine(Dst.size()));
+    return PExecutor->synchronousCopyD2H(
+        Src.getBaseArray(), Src.getElementOffset() * sizeof(T), Dst.data(), 0,
+        Src.getElementCount() * sizeof(T));
+  }
+
+  /// Host-synchronously copies a slice of an array of elements of type T from
+  /// host to device memory.
+  ///
+  /// Returns an error if ElementCount is too large for the source slice.
+  ///
+  /// The calling host thread is blocked until the copy completes. Can be used
+  /// with any host memory, the host memory does not have to be allocated with
+  /// allocateHostMemory or registered with registerHostMemory.
+  template <typename T>
+  Error synchronousCopyD2H(DeviceArraySlice<T> Src, T *Dst,
+                           size_t ElementCount) {
+    if (Src.getElementOffset() + ElementCount > Src.getElementCount())
+      return make_error(
+          "copying too many elements, " + llvm::Twine(ElementCount) +
+          ", from a device array of element count " +
+          llvm::Twine(Src.getElementCount()) + " at element offset " +
+          llvm::Twine(Src.getElementOffset()));
+    return PExecutor->synchronousCopyD2H(
+        Src.getBaseArray(), Src.getElementOffset() * sizeof(T), Dst, 0,
+        Src.getElementCount() * sizeof(T));
+  }
+
+  /// Host-synchronously copies a slice of an array of elements of type T from
+  /// host to device memory.
+  ///
+  /// Returns an error if ElementCount is too large for the source slice or the
+  /// destination.
+  ///
+  /// The calling host thread is blocked until the copy completes. Can be used
+  /// with any host memory, the host memory does not have to be allocated with
+  /// allocateHostMemory or registered with registerHostMemory.
+  template <typename T>
+  Error synchronousCopyD2H(DeviceArraySlice<T> Src,
+                           llvm::MutableArrayRef<T> Dst, size_t ElementCount) {
+    if (Dst.size() < ElementCount)
+      return make_error(
+          "copyint too many elements, " + llvm::Twine(ElementCount) +
+          ", to a host array of element count " + llvm::Twine(Dst.size()));
+    return synchronousCopyD2H(Src, Dst.data(), ElementCount);
+  }
+
+  /// Host-synchronously copies a slice of an array of elements of type T from
+  /// host to device memory.
+  ///
+  /// Returns an error if the Src and Dst sizes do not match.
+  ///
+  /// The calling host thread is blocked until the copy completes. Can be used
+  /// with any host memory, the host memory does not have to be allocated with
+  /// allocateHostMemory or registered with registerHostMemory.
+  template <typename T>
+  Error synchronousCopyD2H(GlobalDeviceMemory<T> Src,
+                           llvm::MutableArrayRef<T> Dst) {
+    return synchronousCopyD2H(Src.drop_front(0), Dst);
+  }
+
+  /// Host-synchronously copies a slice of an array of elements of type T from
+  /// host to device memory.
+  ///
+  /// Returns an error if ElementCount is too large for the source slice.
+  ///
+  /// The calling host thread is blocked until the copy completes. Can be used
+  /// with any host memory, the host memory does not have to be allocated with
+  /// allocateHostMemory or registered with registerHostMemory.
+  template <typename T>
+  Error synchronousCopyD2H(GlobalDeviceMemory<T> Src, T *Dst,
+                           size_t ElementCount) {
+    return synchronousCopyD2H(Src.drop_front(0), Dst, ElementCount);
+  }
+
+  /// Host-synchronously copies a slice of an array of elements of type T from
+  /// host to device memory.
+  ///
+  /// Returns an error if ElementCount is too large for the source slice or the
+  /// destination.
+  ///
+  /// The calling host thread is blocked until the copy completes. Can be used
+  /// with any host memory, the host memory does not have to be allocated with
+  /// allocateHostMemory or registered with registerHostMemory.
+  template <typename T>
+  Error synchronousCopyD2H(GlobalDeviceMemory<T> Src,
+                           llvm::MutableArrayRef<T> Dst, size_t ElementCount) {
+    return synchronousCopyD2H(Src.drop_front(0), Dst, ElementCount);
+  }
+
 private:
   PlatformExecutor *PExecutor;
 };
 
 } // namespace streamexecutor
 
 #endif // STREAMEXECUTOR_EXECUTOR_H

streamexecutor/include/streamexecutor/PlatformInterfaces.h

public:
   /// Launches a kernel on the given stream.
   virtual Error launch(PlatformStreamHandle *S, BlockDimensions BlockSize,
                        GridDimensions GridSize, const KernelBase &Kernel,
                        const PackedKernelArgumentArrayBase &ArgumentArray) {
     return make_error("launch not implemented for platform " + getName());
   }
 
   /// Copies data from the device to the host.
-  virtual Error memcpyD2H(PlatformStreamHandle *S,
+  ///
+  /// HostDst should have been allocated by allocateHostMemory or registered
+  /// with registerHostMemory.
+  virtual Error memcpyD2H(PlatformStreamHandle *S, void *HostDst,
                           const GlobalDeviceMemoryBase &DeviceSrc,
-                          void *HostDst, size_t ByteCount) {
+                          size_t ByteCount, size_t SrcByteOffset = 0) {
     return make_error("memcpyD2H not implemented for platform " + getName());
   }
 
   /// Copies data from the host to the device.
-  virtual Error memcpyH2D(PlatformStreamHandle *S, const void *HostSrc,
-                          GlobalDeviceMemoryBase *DeviceDst, size_t ByteCount) {
+  ///
+  /// HostSrc should have been allocated by allocateHostMemory or registered
+  /// with registerHostMemory.
+  virtual Error memcpyH2D(PlatformStreamHandle *S,
+                          GlobalDeviceMemoryBase *DeviceDst,
+                          const void *HostSrc, size_t ByteCount,
+                          size_t DstByteOffset = 0) {
     return make_error("memcpyH2D not implemented for platform " + getName());
   }
 
   /// Copies data from one device location to another.
   virtual Error memcpyD2D(PlatformStreamHandle *S,
+                          GlobalDeviceMemoryBase *DeviceDst,
                           const GlobalDeviceMemoryBase &DeviceSrc,
-                          GlobalDeviceMemoryBase *DeviceDst, size_t ByteCount) {
+                          size_t ByteCount, size_t SrcByteOffset = 0,
+                          size_t DstByteOffset = 0) {
     return make_error("memcpyD2D not implemented for platform " + getName());
   }
 
   /// Blocks the host until the given stream completes all the work enqueued up
   /// to the point this function is called.
   virtual Error blockHostUntilDone(PlatformStreamHandle *S) {
     return make_error("blockHostUntilDone not implemented for platform " +
                       getName());
   }
+
+  /// Allocates untyped device memory of a given size in bytes.
+  virtual Expected<GlobalDeviceMemoryBase>
+  allocateDeviceMemory(size_t ByteCount) {
+    return make_error("allocateDeviceMemory not implemented for platform " +
+                      getName());
+  }
+
+  /// Frees device memory previously allocated by allocateDeviceMemory.
+  virtual Error freeDeviceMemory(GlobalDeviceMemoryBase *Memory) {
+    return make_error("freeDeviceMemory not implemented for platform " +
+                      getName());
+  }
+
+  /// Allocates untyped host memory of a given size in bytes.
+  ///
+  /// Host memory allocated via this method is suitable for use with memcpyH2D
+  /// and memcpyD2H.
+  virtual Expected<void *> allocateHostMemory(size_t ByteCount) {
+    return make_error("allocateHostMemory not implemented for platform " +
+                      getName());
+  }
+
+  /// Frees host memory allocated by allocateHostMemory.
+  virtual Error freeHostMemory(void *Memory) {
+    return make_error("freeHostMemory not implemented for platform " +
+                      getName());
+  }
+
+  /// Registers previously allocated host memory so it can be used with
+  /// memcpyH2D and memcpyD2H.
+  virtual Error registerHostMemory(void *Memory, size_t ByteCount) {
+    return make_error("registerHostMemory not implemented for platform " +
+                      getName());
+  }
+
+  /// Unregisters host memory previously registered with registerHostMemory.
+  virtual Error unregisterHostMemory(void *Memory) {
+    return make_error("unregisterHostMemory not implemented for platform " +
+                      getName());
+  }
+
+  /// Copies the given number of bytes from device memory to host memory.
+  ///
+  /// Blocks the calling host thread until the copy is completed. Can operate on
+  /// any host memory, not just registered host memory or host memory allocated
+  /// by allocateHostMemory.
+  virtual Error synchronousCopyD2H(const GlobalDeviceMemoryBase &DeviceSrc,
+                                   size_t SrcByteOffset, void *HostDst,
+                                   size_t DstByteOffset, size_t ByteCount) {
+    return make_error("synchronousCopyD2H not implemented for platform " +
+                      getName());
+  }
+
+  /// Copies the given number of bytes from host memory to device memory.
+  ///
+  /// Blocks the calling host thread until the copy is completed. Can operate on
+  /// any host memory, not just registered host memory or host memory allocated
+  /// by allocateHostMemory.
+  virtual Error synchronousCopyH2D(const void *HostSrc, size_t SrcByteOffset,
+                                   GlobalDeviceMemoryBase DeviceDst,
+                                   size_t DstByteOffset, size_t ByteCount) {
+    return make_error("synchronousCopyH2D not implemented for platform " +
+                      getName());
+  }
+
+  /// Copies the given number of bytes from one location to another in device
+  /// memory.
+  virtual Error synchronousCopyD2D(GlobalDeviceMemoryBase DeviceDst,
+                                   size_t DstByteOffset,
+                                   const GlobalDeviceMemoryBase &DeviceSrc,
+                                   size_t SrcByteOffset, size_t ByteCount) {
+    return make_error("synchronousCopyD2D not implemented for platform " +
+                      getName());
+  }
 };
 
 } // namespace streamexecutor
 
 #endif // STREAMEXECUTOR_PLATFORMINTERFACES_H

streamexecutor/include/streamexecutor/Stream.h

public:
 
   /// Entrain onto the stream a memcpy of a given number of elements from a
   /// device source to a host destination.
   ///
   /// HostDst must be a pointer to host memory allocated by
   /// Executor::allocateHostMemory or otherwise allocated and then
   /// registered with Executor::registerHostMemory.
   template <typename T>
-  Stream &thenMemcpyD2H(const GlobalDeviceMemory<T> &DeviceSrc,
-                        llvm::MutableArrayRef<T> HostDst, size_t ElementCount) {
-    if (ElementCount > DeviceSrc.getElementCount())
+  Stream &thenMemcpyD2H(llvm::MutableArrayRef<T> HostDst,
+                        const GlobalDeviceMemory<T> &DeviceSrc,
+                        size_t ElementCount, size_t SrcElementOffset = 0) {
+    if (ElementCount + SrcElementOffset > DeviceSrc.getElementCount())
       setError("copying too many elements, " + llvm::Twine(ElementCount) +
-               ", from device memory array of size " +
+               ", at element offset " + llvm::Twine(SrcElementOffset) +
+               " from device memory array of element count " +
                llvm::Twine(DeviceSrc.getElementCount()));
     else if (ElementCount > HostDst.size())
       setError("copying too many elements, " + llvm::Twine(ElementCount) +
-               ", to host array of size " + llvm::Twine(HostDst.size()));
+               ", to host array of element count " +
+               llvm::Twine(HostDst.size()));
     else
-      setError(PExecutor->memcpyD2H(ThePlatformStream.get(), DeviceSrc,
-                                    HostDst.data(), ElementCount * sizeof(T)));
+      setError(PExecutor->memcpyD2H(ThePlatformStream.get(), HostDst.data(),
+                                    DeviceSrc, ElementCount * sizeof(T),
+                                    SrcElementOffset * sizeof(T)));
     return *this;
   }
 
   /// Same as thenMemcpyD2H above, but copies the entire source to the
   /// destination.
   template <typename T>
-  Stream &thenMemcpyD2H(const GlobalDeviceMemory<T> &DeviceSrc,
-                        llvm::MutableArrayRef<T> HostDst) {
-    return thenMemcpyD2H(DeviceSrc, HostDst, DeviceSrc.getElementCount());
+  Stream &thenMemcpyD2H(llvm::MutableArrayRef<T> HostDst,
+                        const GlobalDeviceMemory<T> &DeviceSrc) {
+    if (DeviceSrc.getElementCount() != HostDst.size()) {
+      setError("thenMemcpyD2H device source element count, " +
+               llvm::Twine(DeviceSrc.getElementCount()) +
+               " ,does not match host destination element count, " +
+               llvm::Twine(HostDst.size()));
+      return *this;
+    }
+    return thenMemcpyD2H(HostDst, DeviceSrc, DeviceSrc.getElementCount());
   }
 
   /// Entrain onto the stream a memcpy of a given number of elements from a host
   /// source to a device destination.
   ///
   /// HostSrc must be a pointer to host memory allocated by
   /// Executor::allocateHostMemory or otherwise allocated and then
   /// registered with Executor::registerHostMemory.
   template <typename T>
-  Stream &thenMemcpyH2D(llvm::ArrayRef<T> HostSrc,
-                        GlobalDeviceMemory<T> *DeviceDst, size_t ElementCount) {
+  Stream &thenMemcpyH2D(GlobalDeviceMemory<T> *DeviceDst,
+                        llvm::ArrayRef<T> HostSrc, size_t ElementCount,
+                        size_t DstElementOffset = 0) {
     if (ElementCount > HostSrc.size())
       setError("copying too many elements, " + llvm::Twine(ElementCount) +
-               ", from host array of size " + llvm::Twine(HostSrc.size()));
-    else if (ElementCount > DeviceDst->getElementCount())
+               ", from host array of element count " +
+               llvm::Twine(HostSrc.size()));
+    else if (DstElementOffset + ElementCount > DeviceDst->getElementCount())
       setError("copying too many elements, " + llvm::Twine(ElementCount) +
-               ", to device memory array of size " +
-               llvm::Twine(DeviceDst->getElementCount()));
+               ", to device memory array of element count " +
+               llvm::Twine(DeviceDst->getElementCount()) +
+               " at element offset " + llvm::Twine(DstElementOffset));
     else
-      setError(PExecutor->memcpyH2D(ThePlatformStream.get(), HostSrc.data(),
-                                    DeviceDst, ElementCount * sizeof(T)));
+      setError(PExecutor->memcpyH2D(ThePlatformStream.get(), DeviceDst,
+                                    HostSrc.data(), ElementCount * sizeof(T),
+                                    DstElementOffset * sizeof(T)));
     return *this;
   }
 
   /// Same as thenMemcpyH2D above, but copies the entire source to the
   /// destination.
   template <typename T>
-  Stream &thenMemcpyH2D(llvm::ArrayRef<T> HostSrc,
-                        GlobalDeviceMemory<T> *DeviceDst) {
-    return thenMemcpyH2D(HostSrc, DeviceDst, HostSrc.size());
+  Stream &thenMemcpyH2D(GlobalDeviceMemory<T> *DeviceDst,
+                        llvm::ArrayRef<T> HostSrc) {
+    if (HostSrc.size() != DeviceDst->getElementCount()) {
+      setError("thenMemcpyH2D host source element count, " +
+               llvm::Twine(HostSrc.size()) +
+               " ,does not match device destination element count, " +
+               llvm::Twine(DeviceDst->getElementCount()));
+      return *this;
+    }
+    return thenMemcpyH2D(DeviceDst, HostSrc, HostSrc.size());
   }
 
   /// Entrain onto the stream a memcpy of a given number of elements from a
   /// device source to a device destination.
   template <typename T>
-  Stream &thenMemcpyD2D(const GlobalDeviceMemory<T> &DeviceSrc,
-                        GlobalDeviceMemory<T> *DeviceDst, size_t ElementCount) {
-    if (ElementCount > DeviceSrc.getElementCount())
+  Stream &thenMemcpyD2D(GlobalDeviceMemory<T> *DeviceDst,
+                        const GlobalDeviceMemory<T> &DeviceSrc,
+                        size_t ElementCount, size_t SrcElementOffset = 0,
+                        size_t DstElementOffset = 0) {
+    if (SrcElementOffset + ElementCount > DeviceSrc.getElementCount())
       setError("copying too many elements, " + llvm::Twine(ElementCount) +
-               ", from device memory array of size " +
+               ", at element offset " + llvm::Twine(SrcElementOffset) +
+               " from device memory array of element count " +
                llvm::Twine(DeviceSrc.getElementCount()));
-    else if (ElementCount > DeviceDst->getElementCount())
+    else if (DstElementOffset + ElementCount > DeviceDst->getElementCount())
       setError("copying too many elements, " + llvm::Twine(ElementCount) +
-               ", to device memory array of size " +
-               llvm::Twine(DeviceDst->getElementCount()));
+               ", to device memory array of element count " +
+               llvm::Twine(DeviceDst->getElementCount()) +
+               " at element offset " + llvm::Twine(DstElementOffset));
     else
-      setError(PExecutor->memcpyD2D(ThePlatformStream.get(), DeviceSrc,
-                                    DeviceDst, ElementCount * sizeof(T)));
+      setError(PExecutor->memcpyD2D(ThePlatformStream.get(), DeviceDst,
+                                    DeviceSrc, ElementCount * sizeof(T),
+                                    SrcElementOffset * sizeof(T),
+                                    DstElementOffset * sizeof(T)));
     return *this;
   }
 
   /// Same as thenMemcpyD2D above, but copies the entire source to the
   /// destination.
   template <typename T>
-  Stream &thenMemcpyD2D(const GlobalDeviceMemory<T> &DeviceSrc,
-                        GlobalDeviceMemory<T> *DeviceDst) {
-    return thenMemcpyD2D(DeviceSrc, DeviceDst, DeviceSrc.getElementCount());
+  Stream &thenMemcpyD2D(GlobalDeviceMemory<T> *DeviceDst,
+                        const GlobalDeviceMemory<T> &DeviceSrc) {
+    if (DeviceSrc.getElementCount() != DeviceDst->getElementCount()) {
+      setError("thenMemcpyH2D device source element count, " +
+               llvm::Twine(DeviceSrc.getElementCount()) +
+               " ,does not match device destination element count, " +
+               llvm::Twine(DeviceDst->getElementCount()));
+      return *this;
+    }
+    return thenMemcpyD2D(DeviceDst, DeviceSrc, DeviceSrc.getElementCount());
   }
 
   /// Blocks the host code, waiting for the operations entrained on the stream
   /// (enqueued up to this point in program execution) to complete.
   ///
   /// Returns true if there are no errors on the stream.
   bool blockHostUntilDone() {
     Error E = PExecutor->blockHostUntilDone(ThePlatformStream.get());
private:
   llvm::Optional<std::string> ErrorMessage;
 
   Stream(const Stream &) = delete;
   void operator=(const Stream &) = delete;
 };
 
 } // namespace streamexecutor
 
 #endif // STREAMEXECUTOR_STREAM_H

streamexecutor/include/streamexecutor/Utils/Error.h

 
 #include "llvm/Support/Error.h"
 
 namespace streamexecutor {
 
 using llvm::consumeError;
 using llvm::Error;
 using llvm::Expected;
-using llvm::StringRef;
+using llvm::Twine;
 
 // Makes an Error object from an error message.
-Error make_error(StringRef Message);
+Error make_error(Twine Message);
 
 // Consumes the input error and returns its error message.
 //
 // Assumes the input was created by the make_error function above.
 std::string consumeAndGetMessage(Error &&E);
 
 } // namespace streamexecutor
 
 #endif // STREAMEXECUTOR_UTILS_ERROR_H

streamexecutor/lib/Utils/Error.cpp

 };
 
 char StreamExecutorError::ID = 0;
 
 } // namespace
 
 namespace streamexecutor {
 
-Error make_error(StringRef Message) {
-  return llvm::make_error<StreamExecutorError>(Message);
+Error make_error(Twine Message) {
+  return llvm::make_error<StreamExecutorError>(Message.str());
 }
 
 std::string consumeAndGetMessage(Error &&E) {
   if (!E) {
     return "success";
   }
   std::string Message;
   llvm::handleAllErrors(std::move(E),
                         [&Message](const StreamExecutorError &SEE) {
                           Message = SEE.getErrorMessage();
                         });
   return Message;
 }
 
 } // namespace streamexecutor

streamexecutor/lib/unittests/CMakeLists.txt

 add_executable(
+    executor_test
+    ExecutorTest.cpp)
+target_link_libraries(
+    executor_test
+    streamexecutor
+    ${GTEST_BOTH_LIBRARIES}
+    ${CMAKE_THREAD_LIBS_INIT})
+add_test(ExecutorTest executor_test)
+
+add_executable(
     kernel_test
     KernelTest.cpp)
 target_link_libraries(
     kernel_test
     streamexecutor
     ${GTEST_BOTH_LIBRARIES}
     ${CMAKE_THREAD_LIBS_INIT})
 add_test(KernelTest kernel_test)

streamexecutor/lib/unittests/ExecutorTest.cpp

+//===-- ExecutorTest.cpp - Tests for Executor -----------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+///
+/// \file
+/// This file contains the unit tests for Executor code.
+///
+//===----------------------------------------------------------------------===//
+
+#include <cstdlib>
+#include <cstring>
+
+#include "streamexecutor/Executor.h"
+#include "streamexecutor/PlatformInterfaces.h"
+
+#include "gtest/gtest.h"
+
+namespace {
+
+namespace se = ::streamexecutor;
+
+class MockPlatformExecutor : public se::PlatformExecutor {
+public:
+  ~MockPlatformExecutor() override {}
+
+  std::string getName() const override { return "MockPlatformExecutor"; }
+
+  se::Expected<std::unique_ptr<se::PlatformStreamHandle>>
+  createStream() override {
+    return se::make_error("not implemented");
+  }
+
+  se::Expected<se::GlobalDeviceMemoryBase>
+  allocateDeviceMemory(size_t ByteCount) override {
+    return se::GlobalDeviceMemoryBase(std::malloc(ByteCount));
+  }
+
+  se::Error freeDeviceMemory(se::GlobalDeviceMemoryBase *Memory) override {
+    std::free(const_cast<void *>(Memory->getHandle()));
+    return se::Error::success();
+  }
+
+  se::Expected<void *> allocateHostMemory(size_t ByteCount) override {
+    return std::malloc(ByteCount);
+  }
+
+  se::Error freeHostMemory(void *Memory) override {
+    std::free(Memory);
+    return se::Error::success();
+  }
+
+  se::Error synchronousCopyD2H(const se::GlobalDeviceMemoryBase &DeviceSrc,
+                               size_t SrcByteOffset, void *HostDst,
+                               size_t DstByteOffset,
+                               size_t ByteCount) override {
+    std::memcpy(static_cast<char *>(HostDst) + DstByteOffset,
+                static_cast<const char *>(DeviceSrc.getHandle()) +
+                    SrcByteOffset,
+                ByteCount);
+    return se::Error::success();
+  }
+
+  se::Error synchronousCopyH2D(const void *HostSrc, size_t SrcByteOffset,
+                               se::GlobalDeviceMemoryBase DeviceDst,
+                               size_t DstByteOffset,
+                               size_t ByteCount) override {
+    std::memcpy(static_cast<char *>(const_cast<void *>(DeviceDst.getHandle())) +
+                    DstByteOffset,
+                static_cast<const char *>(HostSrc) + SrcByteOffset, ByteCount);
+    return se::Error::success();
+  }
+
+  se::Error synchronousCopyD2D(se::GlobalDeviceMemoryBase DeviceDst,
+                               size_t DstByteOffset,
+                               const se::GlobalDeviceMemoryBase &DeviceSrc,
+                               size_t SrcByteOffset,
+                               size_t ByteCount) override {
+    std::memcpy(static_cast<char *>(const_cast<void *>(DeviceDst.getHandle())) +
+                    DstByteOffset,
+                static_cast<const char *>(DeviceSrc.getHandle()) +
+                    SrcByteOffset,
+                ByteCount);
+    return se::Error::success();
+  }
+};
+
+/// Test fixture to hold objects used by tests.
+class ExecutorTest : public ::testing::Test {
+public:
+  ExecutorTest()
+      : HostA{0, 1, 2, 3, 4},
+        DeviceA(se::GlobalDeviceMemory<int>::makeFromElementCount(HostA, 5)),
+        HostB{5, 6, 7, 8, 9},
+        DeviceB(se::GlobalDeviceMemory<int>::makeFromElementCount(HostB, 5)),
+        Host{10, 11, 12, 13, 14}, Executor(&PExecutor) {}
+
+  // Device memory is backed by host arrays.
+  int HostA[5];
+  se::GlobalDeviceMemory<int> DeviceA;
+  int HostB[5];
+  se::GlobalDeviceMemory<int> DeviceB;
+
+  // Host memory to be used as actual host memory.
+  int Host[5];
+
+  MockPlatformExecutor PExecutor;
+  se::Executor Executor;
+};
+
+TEST_F(ExecutorTest, SyncCopyD2HFullArrayEqualSize) {
+  EXPECT_FALSE(static_cast<bool>(
+      Executor.synchronousCopyD2H(DeviceA, llvm::MutableArrayRef<int>(Host))));
+  for (int i = 0; i < 5; ++i) {
+    EXPECT_EQ(HostA[i], Host[i]);
+  }
+}
+
+TEST_F(ExecutorTest, SyncCopyD2HSliceEqualSize) {
+  EXPECT_FALSE(static_cast<bool>(
+      Executor.synchronousCopyD2H(DeviceA.drop_front(1).take_first(2),
+                                  llvm::MutableArrayRef<int>(Host + 1, 2))));
+  for (int i = 1; i < 3; ++i) {
+    EXPECT_EQ(HostA[i], Host[i]);
+  }
+}
+
+TEST_F(ExecutorTest, SyncCopyD2HTooManyElements) {
+  se::Error E = Executor.synchronousCopyD2H(
+      DeviceA, llvm::MutableArrayRef<int>(Host), 20);
+  EXPECT_TRUE(static_cast<bool>(E));
+  se::consumeError(std::move(E));
+}
+
+} // namespace

streamexecutor/lib/unittests/StreamTest.cpp

public:
 
   std::string getName() const override { return "MockPlatformExecutor"; }
 
   se::Expected<std::unique_ptr<se::PlatformStreamHandle>>
   createStream() override {
     return nullptr;
   }
 
-  se::Error memcpyD2H(se::PlatformStreamHandle *,
+  se::Error memcpyD2H(se::PlatformStreamHandle *, void *HostDst,
                       const se::GlobalDeviceMemoryBase &DeviceSrc,
-                      void *HostDst, size_t ByteCount) override {
-    std::memcpy(HostDst, DeviceSrc.getHandle(), ByteCount);
+                      size_t ByteCount, size_t SrcByteOffset) override {
+    std::memcpy(HostDst, static_cast<const char *>(DeviceSrc.getHandle()) +
+                             SrcByteOffset,
+                ByteCount);
     return se::Error::success();
   }
 
-  se::Error memcpyH2D(se::PlatformStreamHandle *, const void *HostSrc,
+  se::Error memcpyH2D(se::PlatformStreamHandle *,
                       se::GlobalDeviceMemoryBase *DeviceDst,
-                      size_t ByteCount) override {
-    std::memcpy(const_cast<void *>(DeviceDst->getHandle()), HostSrc, ByteCount);
+                      const void *HostSrc, size_t ByteCount,
+                      size_t DstByteOffset) override {
+    std::memcpy(
+        static_cast<char *>(const_cast<void *>(DeviceDst->getHandle())) +
+            DstByteOffset,
+        HostSrc, ByteCount);
     return se::Error::success();
   }
 
   se::Error memcpyD2D(se::PlatformStreamHandle *,
-                      const se::GlobalDeviceMemoryBase &DeviceSrc,
                       se::GlobalDeviceMemoryBase *DeviceDst,
-                      size_t ByteCount) override {
-    std::memcpy(const_cast<void *>(DeviceDst->getHandle()),
-                DeviceSrc.getHandle(), ByteCount);
+                      const se::GlobalDeviceMemoryBase &DeviceSrc,
+                      size_t ByteCount, size_t SrcByteOffset,
+                      size_t DstByteOffset) override {
+    std::memcpy(
+        static_cast<char *>(const_cast<void *>(DeviceDst->getHandle())) +
+            SrcByteOffset,
+        static_cast<const char *>(DeviceSrc.getHandle()) + DstByteOffset,
+        ByteCount);
     return se::Error::success();
   }
 };
 
 /// Test fixture to hold objects used by tests.
 class StreamTest : public ::testing::Test {
 public:
   StreamTest()
protected:
   // Host memory to be used as actual host memory.
   int Host[10];
 
   MockPlatformExecutor PExecutor;
   se::Stream Stream;
 };
 
 TEST_F(StreamTest, MemcpyCorrectSize) {
-  Stream.thenMemcpyH2D(llvm::ArrayRef<int>(Host), &DeviceA);
+  Stream.thenMemcpyH2D(&DeviceA, llvm::ArrayRef<int>(Host));
   EXPECT_TRUE(Stream.isOK());
 
-  Stream.thenMemcpyD2H(DeviceA, llvm::MutableArrayRef<int>(Host));
+  Stream.thenMemcpyD2H(llvm::MutableArrayRef<int>(Host), DeviceA);
   EXPECT_TRUE(Stream.isOK());
 
-  Stream.thenMemcpyD2D(DeviceA, &DeviceB);
+  Stream.thenMemcpyD2D(&DeviceB, DeviceA);
   EXPECT_TRUE(Stream.isOK());
 }
 
 TEST_F(StreamTest, MemcpyH2DTooManyElements) {
-  Stream.thenMemcpyH2D(llvm::ArrayRef<int>(Host), &DeviceA, 20);
+  Stream.thenMemcpyH2D(&DeviceA, llvm::ArrayRef<int>(Host), 20);
   EXPECT_FALSE(Stream.isOK());
 }
 
 TEST_F(StreamTest, MemcpyD2HTooManyElements) {
-  Stream.thenMemcpyD2H(DeviceA, llvm::MutableArrayRef<int>(Host), 20);
+  Stream.thenMemcpyD2H(llvm::MutableArrayRef<int>(Host), DeviceA, 20);
   EXPECT_FALSE(Stream.isOK());
 }
 
 TEST_F(StreamTest, MemcpyD2DTooManyElements) {
-  Stream.thenMemcpyD2D(DeviceA, &DeviceB, 20);
+  Stream.thenMemcpyD2D(&DeviceB, DeviceA, 20);
   EXPECT_FALSE(Stream.isOK());
 }
 
 } // namespace