Read FIXED_LEN_BYTE_ARRAY as binary in parquet reader

cpp/src/io/parquet/page_data.cu

@@ -41,7 +41,7 @@ inline __device__ void gpuOutputString(volatile page_state_s* s,
                                        void* dstv)
 {
   auto [ptr, len] = gpuGetStringData(s, sb, src_pos);
-  if (s->dtype_len == 4) {
+  if (s->dtype_len == 4 and (s->col.data_type & 7) == BYTE_ARRAY) {
     // Output hash. This hash value is used if the option to convert strings to
     // categoricals is enabled. The seed value is chosen arbitrarily.
     uint32_t constexpr hash_seed = 33;
@@ -774,8 +774,12 @@ __global__ void __launch_bounds__(decode_block_size) gpuDecodePageData(
   if (s->dict_base) {
     out_thread0 = (s->dict_bits > 0) ? 64 : 32;
   } else {
-    out_thread0 =
-      ((s->col.data_type & 7) == BOOLEAN || (s->col.data_type & 7) == BYTE_ARRAY) ? 64 : 32;
+    switch (s->col.data_type & 7) {
+      case BOOLEAN: [[fallthrough]];
+      case BYTE_ARRAY: [[fallthrough]];
+      case FIXED_LEN_BYTE_ARRAY: out_thread0 = 64; break;
+      default: out_thread0 = 32;
+    }
   }
 
   PageNestingDecodeInfo* nesting_info_base = s->nesting_info;
@@ -812,7 +816,8 @@ __global__ void __launch_bounds__(decode_block_size) gpuDecodePageData(
         src_target_pos = gpuDecodeDictionaryIndices<false>(s, sb, src_target_pos, t & 0x1f).first;
       } else if ((s->col.data_type & 7) == BOOLEAN) {
         src_target_pos = gpuDecodeRleBooleans(s, sb, src_target_pos, t & 0x1f);
-      } else if ((s->col.data_type & 7) == BYTE_ARRAY) {
+      } else if ((s->col.data_type & 7) == BYTE_ARRAY or
+                 (s->col.data_type & 7) == FIXED_LEN_BYTE_ARRAY) {
         gpuInitStringDescriptors<false>(s, sb, src_target_pos, t & 0x1f);
       }
       if (t == 32) { *(volatile int32_t*)&s->dict_pos = src_target_pos; }
@@ -882,6 +887,8 @@ __global__ void __launch_bounds__(decode_block_size) gpuDecodePageData(
               }
               break;
           }
+        } else if (dtype == FIXED_LEN_BYTE_ARRAY) {
+          gpuOutputString(s, sb, val_src_pos, dst);
         } else if (dtype == INT96) {
           gpuOutputInt96Timestamp(s, sb, val_src_pos, static_cast<int64_t*>(dst));
         } else if (dtype_len == 8) {

cpp/src/io/parquet/page_decode.cuh

@@ -418,12 +418,20 @@ __device__ size_type gpuInitStringDescriptors(page_state_s volatile* s,
 
     while (pos < target_pos) {
       int len;
-      if (k + 4 <= dict_size) {
-        len = (cur[k]) | (cur[k + 1] << 8) | (cur[k + 2] << 16) | (cur[k + 3] << 24);
-        k += 4;
-        if (k + len > dict_size) { len = 0; }
+      if ((s->col.data_type & 7) == FIXED_LEN_BYTE_ARRAY) {
+        if (k < dict_size) {
+          len = s->dtype_len_in;
+        } else {
+          len = 0;
+        }
       } else {
-        len = 0;
+        if (k + 4 <= dict_size) {
+          len = (cur[k]) | (cur[k + 1] << 8) | (cur[k + 2] << 16) | (cur[k + 3] << 24);
+          k += 4;
+          if (k + len > dict_size) { len = 0; }
+        } else {
+          len = 0;
+        }
       }
       if constexpr (!sizes_only) {
         sb->dict_idx[rolling_index(pos)] = k;
@@ -1147,16 +1155,20 @@ inline __device__ bool setupLocalPageInfo(page_state_s* const s,
       }
       // Special check for downconversions
       s->dtype_len_in = s->dtype_len;
-      if (s->col.converted_type == DECIMAL && data_type == FIXED_LEN_BYTE_ARRAY) {
-        s->dtype_len = [dtype_len = s->dtype_len]() {
-          if (dtype_len <= sizeof(int32_t)) {
-            return sizeof(int32_t);
-          } else if (dtype_len <= sizeof(int64_t)) {
-            return sizeof(int64_t);
-          } else {
-            return sizeof(__int128_t);
-          }
-        }();
+      if (data_type == FIXED_LEN_BYTE_ARRAY) {
+        if (s->col.converted_type == DECIMAL) {
+          s->dtype_len = [dtype_len = s->dtype_len]() {
+            if (dtype_len <= sizeof(int32_t)) {
+              return sizeof(int32_t);
+            } else if (dtype_len <= sizeof(int64_t)) {
+              return sizeof(int64_t);
+            } else {
+              return sizeof(__int128_t);
+            }
+          }();
+        } else {
+          s->dtype_len = sizeof(string_index_pair);
+        }
       } else if (data_type == INT32) {
         if (dtype_len_out == 1) {
           // INT8 output
@@ -1212,7 +1224,7 @@ inline __device__ bool setupLocalPageInfo(page_state_s* const s,
               uint32_t len = idx < max_depth - 1 ? sizeof(cudf::size_type) : s->dtype_len;
               // if this is a string column, then dtype_len is a lie. data will be offsets rather
               // than (ptr,len) tuples.
-              if (data_type == BYTE_ARRAY && s->dtype_len != 4) { len = sizeof(cudf::size_type); }
+              if (is_string_col(s->col)) { len = sizeof(cudf::size_type); }
               nesting_info->data_out += (output_offset * len);
             }
             if (nesting_info->string_out != nullptr) {

cpp/src/io/parquet/page_string_decode.cu

@@ -595,39 +595,44 @@ __global__ void __launch_bounds__(preprocess_block_size) gpuComputePageStringSiz
     pp->num_valids = s->page.num_valids;
   }
 
-  // now process string info in the range [start_value, end_value)
-  // set up for decoding strings...can be either plain or dictionary
-  auto const& col          = s->col;
-  uint8_t const* data      = s->data_start;
-  uint8_t const* const end = s->data_end;
-  uint8_t const* dict_base = nullptr;
-  int dict_size            = 0;
-  size_t str_bytes         = 0;
-
-  switch (pp->encoding) {
-    case Encoding::PLAIN_DICTIONARY:
-    case Encoding::RLE_DICTIONARY:
-      // RLE-packed dictionary indices, first byte indicates index length in bits
-      if (col.str_dict_index) {
-        // String dictionary: use index
-        dict_base = reinterpret_cast<const uint8_t*>(col.str_dict_index);
-        dict_size = col.page_info[0].num_input_values * sizeof(string_index_pair);
-      } else {
-        dict_base = col.page_info[0].page_data;  // dictionary is always stored in the first page
-        dict_size = col.page_info[0].uncompressed_page_size;
-      }
+  auto const& col  = s->col;
+  size_t str_bytes = 0;
+  // short circuit for FIXED_LEN_BYTE_ARRAY
+  if ((col.data_type & 7) == FIXED_LEN_BYTE_ARRAY) {
+    str_bytes = pp->num_valids * s->dtype_len_in;
+  } else {
+    // now process string info in the range [start_value, end_value)
+    // set up for decoding strings...can be either plain or dictionary
+    uint8_t const* data      = s->data_start;
+    uint8_t const* const end = s->data_end;
+    uint8_t const* dict_base = nullptr;
+    int dict_size            = 0;
+
+    switch (pp->encoding) {
+      case Encoding::PLAIN_DICTIONARY:
+      case Encoding::RLE_DICTIONARY:
+        // RLE-packed dictionary indices, first byte indicates index length in bits
+        if (col.str_dict_index) {
+          // String dictionary: use index
+          dict_base = reinterpret_cast<const uint8_t*>(col.str_dict_index);
+          dict_size = col.page_info[0].num_input_values * sizeof(string_index_pair);
+        } else {
+          dict_base = col.page_info[0].page_data;  // dictionary is always stored in the first page
+          dict_size = col.page_info[0].uncompressed_page_size;
+        }
 
-      // FIXME: need to return an error condition...this won't actually do anything
-      if (s->dict_bits > 32 || !dict_base) { CUDF_UNREACHABLE("invalid dictionary bit size"); }
-
-      str_bytes = totalDictEntriesSize(
-        data, dict_base, s->dict_bits, dict_size, (end - data), start_value, end_value);
-      break;
-    case Encoding::PLAIN:
-      dict_size = static_cast<int32_t>(end - data);
-      str_bytes = is_bounds_pg ? totalPlainEntriesSize(data, dict_size, start_value, end_value)
-                               : dict_size - sizeof(int) * (pp->num_input_values - pp->num_nulls);
-      break;
+        // FIXME: need to return an error condition...this won't actually do anything
+        if (s->dict_bits > 32 || !dict_base) { CUDF_UNREACHABLE("invalid dictionary bit size"); }
+
+        str_bytes = totalDictEntriesSize(
+          data, dict_base, s->dict_bits, dict_size, (end - data), start_value, end_value);
+        break;
+      case Encoding::PLAIN:
+        dict_size = static_cast<int32_t>(end - data);
+        str_bytes = is_bounds_pg ? totalPlainEntriesSize(data, dict_size, start_value, end_value)
+                                 : dict_size - sizeof(int) * pp->num_valids;
+        break;
+    }
   }
 
   if (t == 0) {

cpp/src/io/parquet/parquet_gpu.hpp

@@ -432,8 +432,9 @@ struct EncPage {
  */
 constexpr bool is_string_col(ColumnChunkDesc const& chunk)
 {
-  return (chunk.data_type & 7) == BYTE_ARRAY and (chunk.data_type >> 3) != 4 and
-         chunk.converted_type != DECIMAL;
+  return ((chunk.data_type & 7) == BYTE_ARRAY and (chunk.data_type >> 3) != 4 and
+          chunk.converted_type != DECIMAL) or
+         ((chunk.data_type & 7) == FIXED_LEN_BYTE_ARRAY and chunk.converted_type != DECIMAL);
 }
 
 /**

cpp/src/io/parquet/reader_impl.cpp

@@ -382,9 +382,15 @@ table_with_metadata reader::impl::read_chunk_internal(
 
   // Create the final output cudf columns.
   for (size_t i = 0; i < _output_buffers.size(); ++i) {
-    auto const metadata = _reader_column_schema.has_value()
-                            ? std::make_optional<reader_column_schema>((*_reader_column_schema)[i])
-                            : std::nullopt;
+    auto metadata      = _reader_column_schema.has_value()
+                           ? std::make_optional<reader_column_schema>((*_reader_column_schema)[i])
+                           : std::nullopt;
+    auto const& schema = _metadata->get_schema(_output_column_schemas[i]);
+    // FIXED_LEN_BYTE_ARRAY always read as binary
+    if (schema.type == FIXED_LEN_BYTE_ARRAY and schema.converted_type != DECIMAL) {
+      metadata = std::make_optional<reader_column_schema>();
+      metadata->set_convert_binary_to_strings(false);
+    }
     // Only construct `out_metadata` if `_output_metadata` has not been cached.
     if (!_output_metadata) {
       column_name_info& col_name = out_metadata.schema_info[i];

python/cudf/cudf/tests/test_parquet.py

@@ -2518,6 +2518,15 @@ def test_parquet_reader_binary_decimal(datadir):
     assert_eq(expect, got)
 
 
+def test_parquet_reader_fixed_bin(datadir):
+    fname = datadir / "fixed_len_byte_array.parquet"
+
+    expect = pd.read_parquet(fname)
+    got = cudf.read_parquet(fname).to_pandas()
+
+    assert_eq(expect, got)
+
+
 def test_parquet_reader_rle_boolean(datadir):
     fname = datadir / "rle_boolean_encoding.parquet"