GPU binning and compression. (#3319)

* GPU binning and compression.

- binning and index compression are done inside the DeviceShard constructor
- in case of a DMatrix with multiple row batches, it is first converted into a single row batch

src/common/compressed_iterator.h

@@ -8,6 +8,10 @@
 #include <cstddef>
 #include <algorithm>
 
+#ifdef __CUDACC__
+#include "device_helpers.cuh"
+#endif
+
 namespace xgboost {
 namespace common {
 
@@ -96,6 +100,23 @@ class CompressedBufferWriter {
       }
     }
   }
+
+#ifdef __CUDACC__
+  __device__ void AtomicWriteSymbol
+    (CompressedByteT* buffer, uint64_t symbol, size_t offset) {
+    size_t ibit_start = offset * symbol_bits_;
+    size_t ibit_end = (offset + 1) * symbol_bits_ - 1;
+    size_t ibyte_start = ibit_start / 8, ibyte_end = ibit_end / 8;
+
+    symbol <<= 7 - ibit_end % 8;
+    for (ptrdiff_t ibyte = ibyte_end; ibyte >= (ptrdiff_t)ibyte_start; --ibyte) {
+      dh::AtomicOrByte(reinterpret_cast<unsigned int*>(buffer + detail::kPadding),
+                   ibyte, symbol & 0xff);
+      symbol >>= 8;
+    }
+  }
+#endif
+
   template <typename IterT>
   void Write(CompressedByteT *buffer, IterT input_begin, IterT input_end) {
     uint64_t tmp = 0;

src/common/device_helpers.cuh

@@ -122,6 +122,14 @@ inline size_t AvailableMemory(int device_idx) {
   return device_free;
 }
 
+inline size_t TotalMemory(int device_idx) {
+  size_t device_free = 0;
+  size_t device_total = 0;
+  safe_cuda(cudaSetDevice(device_idx));
+  dh::safe_cuda(cudaMemGetInfo(&device_free, &device_total));
+  return device_total;
+}
+
 /**
  * \fn  inline int max_shared_memory(int device_idx)
  *
@@ -155,6 +163,12 @@ inline void CheckComputeCapability() {
   }
 }
 
+
+DEV_INLINE void AtomicOrByte(unsigned int* __restrict__ buffer, size_t ibyte, unsigned char b) {
+  atomicOr(&buffer[ibyte / sizeof(unsigned int)], (unsigned int)b << (ibyte % (sizeof(unsigned int)) * 8));
+}
+
+
 /*
  * Range iterator
  */