Rewrite sparse dmatrix using callbacks. (#7092)

- Reduce dependency on dmlc parsers and provide an interface for users to load data by themselves.
- Remove use of threaded iterator and IO queue.
- Remove `page_size`.
- Make sure the number of pages in memory is bounded.
- Make sure the cache can not be violated.
- Provide an interface for internal algorithms to process data asynchronously.

tests/cpp/tree/gpu_hist/test_gradient_based_sampler.cu

@@ -27,7 +27,7 @@ void VerifySampling(size_t page_size,
   }
   gpair.SetDevice(0);
 
-  BatchParam param{0, 256, page_size};
+  BatchParam param{0, 256};
   auto page = (*dmat->GetBatches<EllpackPage>(param).begin()).Impl();
   if (page_size != 0) {
     EXPECT_NE(page->n_rows, kRows);
@@ -82,7 +82,7 @@ TEST(GradientBasedSampler, NoSamplingExternalMemory) {
   auto gpair = GenerateRandomGradients(kRows);
   gpair.SetDevice(0);
 
-  BatchParam param{0, 256, kPageSize};
+  BatchParam param{0, 256};
   auto page = (*dmat->GetBatches<EllpackPage>(param).begin()).Impl();
   EXPECT_NE(page->n_rows, kRows);
 

tests/cpp/tree/gpu_hist/test_histogram.cu

@@ -15,7 +15,7 @@ void TestDeterministicHistogram(bool is_dense, int shm_size) {
 
   float sparsity = is_dense ? 0.0f : 0.5f;
   auto matrix = RandomDataGenerator(kRows, kCols, sparsity).GenerateDMatrix();
-  BatchParam batch_param{0, static_cast<int32_t>(kBins), 0};
+  BatchParam batch_param{0, static_cast<int32_t>(kBins)};
 
   for (auto const& batch : matrix->GetBatches<EllpackPage>(batch_param)) {
     auto* page = batch.Impl();
@@ -116,7 +116,7 @@ void TestGPUHistogramCategorical(size_t num_categories) {
   auto x = GenerateRandomCategoricalSingleColumn(kRows, num_categories);
   auto cat_m = GetDMatrixFromData(x, kRows, 1);
   cat_m->Info().feature_types.HostVector().push_back(FeatureType::kCategorical);
-  BatchParam batch_param{0, static_cast<int32_t>(kBins), 0};
+  BatchParam batch_param{0, static_cast<int32_t>(kBins)};
   tree::RowPartitioner row_partitioner(0, kRows);
   auto ridx = row_partitioner.GetRows(0);
   dh::device_vector<GradientPairPrecise> cat_hist(num_categories);

tests/cpp/tree/test_gpu_hist.cu

@@ -152,7 +152,6 @@ TEST(GpuHist, ApplySplit) {
   BatchParam bparam;
   bparam.gpu_id = 0;
   bparam.max_bin = 3;
-  bparam.gpu_page_size = 0;
 
   for (auto& ellpack : m->GetBatches<EllpackPage>(bparam)){
     auto impl = ellpack.Impl();
@@ -291,9 +290,13 @@ void TestHistogramIndexImpl() {
   // Extract the device maker from the histogram makers and from that its compressed
   // histogram index
   const auto &maker = hist_maker.maker;
+  auto grad = GenerateRandomGradients(kNRows);
+  grad.SetDevice(0);
+  maker->Reset(&grad, hist_maker_dmat.get(), kNCols);
   std::vector<common::CompressedByteT> h_gidx_buffer(maker->page->gidx_buffer.HostVector());
 
   const auto &maker_ext = hist_maker_ext.maker;
+  maker_ext->Reset(&grad, hist_maker_ext_dmat.get(), kNCols);
   std::vector<common::CompressedByteT> h_gidx_buffer_ext(maker_ext->page->gidx_buffer.HostVector());
 
   ASSERT_EQ(maker->page->Cuts().TotalBins(), maker_ext->page->Cuts().TotalBins());
@@ -365,7 +368,7 @@ void UpdateTree(HostDeviceVector<GradientPair>* gpair, DMatrix* dmat,
     // Loop over the batches and count the records
     int64_t batch_count = 0;
     int64_t row_count = 0;
-    for (const auto& batch : dmat->GetBatches<EllpackPage>({0, max_bin, gpu_page_size})) {
+    for (const auto& batch : dmat->GetBatches<EllpackPage>({0, max_bin})) {
       EXPECT_LT(batch.Size(), dmat->Info().num_row_);
       batch_count++;
       row_count += batch.Size();
@@ -386,7 +389,6 @@ void UpdateTree(HostDeviceVector<GradientPair>* gpair, DMatrix* dmat,
 
   tree::GPUHistMakerSpecialised<GradientPairPrecise> hist_maker;
   GenericParameter generic_param(CreateEmptyGenericParam(0));
-  generic_param.gpu_page_size = gpu_page_size;
   hist_maker.Configure(args, &generic_param);
 
   hist_maker.Update(gpair, dmat, {tree});