Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Fix ORC writer output corruption with string columns #7565

Merged
merged 4 commits into from
Mar 12, 2021
Merged

Fix ORC writer output corruption with string columns #7565

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
c9588bb
LoadNonNullIndices - fix condition under which we fetch the next veli…
vuule Mar 11, 2021
07ca641
add test
vuule Mar 11, 2021
c230cee
use block_size instead of 512
vuule Mar 11, 2021
15c354d
test comments
vuule Mar 11, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
36 changes: 21 additions & 15 deletions cpp/src/io/orc/dict_enc.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -70,24 +70,28 @@ static __device__ void LoadNonNullIndices(volatile dictinit_state_s *s,
Storage &temp_storage)
{
if (t == 0) { s->nnz = 0; }
for (uint32_t i = 0; i < s->chunk.num_rows; i += 512) {
for (uint32_t i = 0; i < s->chunk.num_rows; i += block_size) {
const uint32_t *valid_map = s->chunk.valid_map_base;
uint32_t is_valid, nz_pos;
if (t < 16) {
if (t < block_size / 32) {
if (!valid_map) {
s->scratch_red[t] = 0xffffffffu;
} else {
uint32_t row = s->chunk.start_row + i + t * 32;
uint32_t v = (row < s->chunk.start_row + s->chunk.num_rows)
? valid_map[(row + s->chunk.column_offset) / 32]
: 0;
if (row & 0x1f) {
uint32_t v1 = (row + 32 < s->chunk.start_row + s->chunk.num_rows)
? valid_map[((row + s->chunk.column_offset) / 32) + 1]
: 0;
v = __funnelshift_r(v, v1, row + s->chunk.column_offset);
uint32_t const row = s->chunk.start_row + i + t * 32;
auto const chunk_end = s->chunk.start_row + s->chunk.num_rows;

auto const valid_map_idx = (row + s->chunk.column_offset) / 32;
uint32_t valid = (row < chunk_end) ? valid_map[valid_map_idx] : 0;

auto const rows_in_next_word = (row + s->chunk.column_offset) & 0x1f;
if (rows_in_next_word != 0) {
auto const rows_in_current_word = 32 - rows_in_next_word;
// Read next word if any rows are within the chunk
uint32_t const valid_next =
(row + rows_in_current_word < chunk_end) ? valid_map[valid_map_idx + 1] : 0;
valid = __funnelshift_r(valid, valid_next, rows_in_next_word);
}
s->scratch_red[t] = v;
s->scratch_red[t] = valid;
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This whole block looks to be over optimized. It's pre-loading an offseted mask word into scratch_red only to have individual threads treating this shared memory as the new mask, and then doing an exclusive sum over the validity values.

This could be achieved using an iterator for validity, directly accessing the mask itself. and that iterator could be given an appropriate offset, which I believe would be the column offset + this dictionary chunk's start row.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Agreed. I also had a few ideas on what can be simplified around this code, but wanted to keep such changes separate from the actual fix.

}
}
__syncthreads();
Expand All @@ -109,7 +113,7 @@ static __device__ void LoadNonNullIndices(volatile dictinit_state_s *s,
* @param[in] chunks DictionaryChunk device array [rowgroup][column]
* @param[in] num_columns Number of columns
*/
// blockDim {512,1,1}
// blockDim {block_size,1,1}
template <int block_size>
__global__ void __launch_bounds__(block_size, 2)
gpuInitDictionaryIndices(DictionaryChunk *chunks, uint32_t num_columns)
Expand Down Expand Up @@ -411,9 +415,11 @@ void InitDictionaryIndices(DictionaryChunk *chunks,
uint32_t num_rowgroups,
rmm::cuda_stream_view stream)
{
dim3 dim_block(512, 1); // 512 threads per chunk
static constexpr int block_size = 512;
dim3 dim_block(block_size, 1);
dim3 dim_grid(num_columns, num_rowgroups);
gpuInitDictionaryIndices<512><<<dim_grid, dim_block, 0, stream.value()>>>(chunks, num_columns);
gpuInitDictionaryIndices<block_size>
<<<dim_grid, dim_block, 0, stream.value()>>>(chunks, num_columns);
}

/**
Expand Down
36 changes: 36 additions & 0 deletions cpp/tests/io/orc_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1025,4 +1025,40 @@ TEST_F(OrcStatisticsTest, Basic)
validate_statistics(stats.stripes_stats[0]);
}

TEST_F(OrcWriterTest, SlicedValidMask)
{
std::vector<const char*> strings;
// Need more than 32 elements to reproduce the issue
for (int i = 0; i < 34; ++i)
strings.emplace_back("a long string to make sure overflow affects the output");
// An element is null only to enforce the output column to be nullable
auto validity = cudf::detail::make_counting_transform_iterator(0, [](auto i) { return i != 32; });

column_wrapper<cudf::string_view> col{strings.begin(), strings.end(), validity};

std::vector<std::unique_ptr<column>> cols;
cols.push_back(col.release());

cudf_io::table_metadata expected_metadata;
expected_metadata.column_names.emplace_back("col_string");

// Bug tested here is easiest to reproduce when column_offset % 32 is 31
std::vector<cudf::size_type> indices{31, 34};
std::vector<cudf::column_view> sliced_col = cudf::slice(cols[0]->view(), indices);
cudf::table_view tbl{sliced_col};

auto filepath = temp_env->get_temp_filepath("OrcStrings.orc");
cudf_io::orc_writer_options out_opts =
cudf_io::orc_writer_options::builder(cudf_io::sink_info{filepath}, tbl)
.metadata(&expected_metadata);
cudf_io::write_orc(out_opts);

cudf_io::orc_reader_options in_opts =
cudf_io::orc_reader_options::builder(cudf_io::source_info{filepath}).use_index(false);
auto result = cudf_io::read_orc(in_opts);

CUDF_TEST_EXPECT_TABLES_EQUAL(tbl, result.tbl->view());
EXPECT_EQ(expected_metadata.column_names, result.metadata.column_names);
}

CUDF_TEST_PROGRAM_MAIN()