Add support for decimal types in ORC writer

cpp/src/io/orc/orc.cpp

@@ -227,7 +227,7 @@ void ProtobufWriter::put_row_index_entry(int32_t present_blk,
   if (data_blk >= 0) { sz += put_uint(data_blk); }
   if (data_ofs >= 0) {
     sz += put_uint(data_ofs);
-    if (kind != STRING && kind != FLOAT && kind != DOUBLE) {
+    if (kind != STRING && kind != FLOAT && kind != DOUBLE && kind != DECIMAL) {
       putb(0);  // RLE run pos always zero (assumes RLE aligned with row index boundaries)
       sz++;
       if (kind == BOOLEAN) {
@@ -293,8 +293,8 @@ size_t ProtobufWriter::write(const SchemaType &s)
   w.field_packed_uint(2, s.subtypes);
   w.field_repeated_string(3, s.fieldNames);
   // w.field_uint(4, s.maximumLength);
-  // w.field_uint(5, s.precision);
-  // w.field_uint(6, s.scale);
+  if (s.precision) w.field_uint(5, *s.precision);
+  if (s.scale) w.field_uint(6, *s.scale);
   return w.value();
 }
 

cpp/src/io/orc/orc.h

@@ -56,10 +56,10 @@ struct SchemaType {
   TypeKind kind = INVALID_TYPE_KIND;  // the kind of this type
   std::vector<uint32_t> subtypes;  // the type ids of any subcolumns for list, map, struct, or union
   std::vector<std::string> fieldNames;  // the list of field names for struct
-  uint32_t maximumLength =
-    0;  // optional: the maximum length of the type for varchar or char in UTF-8 characters
-  uint32_t precision = 0;  // optional: the precision and scale for decimal
-  uint32_t scale     = 0;
+  std::optional<uint32_t>
+    maximumLength;  // the maximum length of the type for varchar or char in UTF-8 characters
+  std::optional<uint32_t> precision;  // the precision for decimal
+  std::optional<uint32_t> scale;      // the scale for decimal
 };
 
 struct UserMetadataItem {

cpp/src/io/orc/orc_gpu.h

@@ -99,8 +99,7 @@ struct ColumnDesc {
   uint8_t encoding_kind;                   // column encoding kind (orc::ColumnEncodingKind)
   uint8_t type_kind;                       // column data type (orc::TypeKind)
   uint8_t dtype_len;      // data type length (for types that can be mapped to different sizes)
-  uint8_t decimal_scale;  // number of fractional decimal digits for decimal type (bit 7 set if
-                          // converting to float64)
+  int32_t decimal_scale;  // number of fractional decimal digits for decimal type
   int32_t ts_clock_rate;  // output timestamp clock frequency (0=default, 1000=ms, 1000000000=ns)
 };
 
@@ -122,9 +121,10 @@ struct EncChunk {
   uint8_t encoding_kind;  // column encoding kind (orc::ColumnEncodingKind)
   uint8_t type_kind;      // column data type (orc::TypeKind)
   uint8_t dtype_len;      // data type length
-  uint8_t scale;          // scale for decimals or timestamps
+  int32_t scale;          // scale for decimals or timestamps
 
   uint32_t *dict_index;  // dictionary index from row index
+  device_span<uint32_t> decimal_offsets;
   column_device_view *leaf_column;
 };
 

cpp/src/io/orc/reader_impl.cu

@@ -434,7 +434,7 @@ table_with_metadata reader::impl::read(size_type skip_rows,
       // sign of the scale is changed since cuDF follows c++ libraries like CNL
       // which uses negative scaling, but liborc and other libraries
       // follow positive scaling.
-      auto const scale = -static_cast<int32_t>(_metadata->ff.types[col].scale);
+      auto const scale = -static_cast<int32_t>(_metadata->ff.types[col].scale.value_or(0));
       column_types.emplace_back(col_type, scale);
     } else {
       column_types.emplace_back(col_type);
@@ -526,7 +526,7 @@ table_with_metadata reader::impl::read(size_type skip_rows,
         chunk.num_rows      = stripe_info->numberOfRows;
         chunk.encoding_kind = stripe_footer->columns[_selected_columns[j]].kind;
         chunk.type_kind     = _metadata->ff.types[_selected_columns[j]].kind;
-        chunk.decimal_scale = _metadata->ff.types[_selected_columns[j]].scale;
+        chunk.decimal_scale = _metadata->ff.types[_selected_columns[j]].scale.value_or(0);
         chunk.rowgroup_id   = num_rowgroups;
         chunk.dtype_len     = (column_types[j].id() == type_id::STRING)
                             ? sizeof(std::pair<const char *, size_t>)

cpp/src/io/orc/stripe_enc.cu

@@ -781,6 +781,9 @@ __global__ void __launch_bounds__(block_size)
               s->lengths.u32[nz_idx]                  = value.size_bytes();
             }
             break;
+            // Reusing the lengths array for the scale stream
+            // Note: can be written in a faster manner, given that all values are equal
+          case DECIMAL: s->lengths.u32[nz_idx] = zigzag(s->chunk.scale); break;
           default: break;
         }
       }
@@ -814,7 +817,7 @@ __global__ void __launch_bounds__(block_size)
         uint32_t nz = s->buf.u32[511];
         s->nnz += nz;
         s->numvals += nz;
-        s->numlengths += (s->chunk.type_kind == TIMESTAMP ||
+        s->numlengths += (s->chunk.type_kind == TIMESTAMP || s->chunk.type_kind == DECIMAL ||
                           (s->chunk.type_kind == STRING && s->chunk.encoding_kind != DICTIONARY_V2))
                            ? nz
                            : 0;
@@ -865,6 +868,17 @@ __global__ void __launch_bounds__(block_size)
               n = s->numvals;
             }
             break;
+          case DECIMAL: {
+            if (valid) {
+              uint64_t const zz_val = (s->chunk.leaf_column->type().id() == type_id::DECIMAL32)
+                                        ? zigzag(s->chunk.leaf_column->element<int32_t>(row))
+                                        : zigzag(s->chunk.leaf_column->element<int64_t>(row));
+              auto const offset =
+                (row == s->chunk.start_row) ? 0 : s->chunk.decimal_offsets[row - 1];
+              StoreVarint(s->stream.data_ptrs[CI_DATA] + offset, zz_val);
+            }
+            n = s->numvals;
+          } break;
           default: n = s->numvals; break;
         }
         __syncthreads();
@@ -878,6 +892,7 @@ __global__ void __launch_bounds__(block_size)
             n = IntegerRLE<CI_DATA2, uint64_t, false, 0x3ff, block_size>(
               s, s->lengths.u64, s->nnz - s->numlengths, s->numlengths, flush, t, temp_storage.u64);
             break;
+          case DECIMAL:
           case STRING:
             n = IntegerRLE<CI_DATA2, uint32_t, false, 0x3ff, block_size>(
               s, s->lengths.u32, s->nnz - s->numlengths, s->numlengths, flush, t, temp_storage.u32);
@@ -893,7 +908,9 @@ __global__ void __launch_bounds__(block_size)
   __syncthreads();
   if (t <= CI_PRESENT && s->stream.ids[t] >= 0) {
     // Update actual compressed length
-    streams[col_id][group_id].lengths[t] = s->strm_pos[t];
+    // (not needed for decimal data, whose exact size is known before encode)
+    if (!(t == CI_DATA && s->chunk.type_kind == DECIMAL))
+      streams[col_id][group_id].lengths[t] = s->strm_pos[t];
     if (!s->stream.data_ptrs[t]) {
       streams[col_id][group_id].data_ptrs[t] =
         static_cast<uint8_t *>(const_cast<void *>(s->chunk.leaf_column->head())) +