Replace agg_count_distinct_bitmap_gpu definition with a C++ one

Author: kurapov-peter

omniscidb/QueryEngine/Compiler/genx.cpp

@@ -18,6 +18,7 @@ double atomic_min_double(GENERIC_ADDR_SPACE double* addr, const double val);
 double atomic_min_float(GENERIC_ADDR_SPACE float* addr, const float val);
 double atomic_max_double(GENERIC_ADDR_SPACE double* addr, const double val);
 double atomic_max_float(GENERIC_ADDR_SPACE float* addr, const float val);
+void atomic_or(GENERIC_ADDR_SPACE int32_t* addr, const int32_t val);
 GENERIC_ADDR_SPACE int64_t* declare_dynamic_shared_memory();
 
 void sync_threadblock();
@@ -105,4 +106,38 @@ const GENERIC_ADDR_SPACE int64_t* init_shared_mem(
   sync_threadblock();
   return shared_groups_buffer;
 }
+
+void agg_count_distinct_bitmap_gpu(GENERIC_ADDR_SPACE int64_t* agg,
+                                   const int64_t val,
+                                   const int64_t min_val,
+                                   const int64_t base_dev_addr,
+                                   const int64_t base_host_addr,
+                                   const uint64_t sub_bitmap_count,
+                                   const uint64_t bitmap_bytes) {
+  const uint64_t bitmap_idx = val - min_val;
+  const uint32_t byte_idx = bitmap_idx >> 3;
+  const uint32_t word_idx = byte_idx >> 2;
+  const uint32_t byte_word_idx = byte_idx & 3;
+  const int64_t host_addr = *agg;
+  GENERIC_ADDR_SPACE int32_t* bitmap =
+      (GENERIC_ADDR_SPACE int32_t*)(base_dev_addr + host_addr - base_host_addr +
+                                    (get_thread_index() & (sub_bitmap_count - 1)) *
+                                        bitmap_bytes);
+  switch (byte_word_idx) {
+    case 0:
+      atomic_or(&bitmap[word_idx], 1 << (bitmap_idx & 7));
+      break;
+    case 1:
+      atomic_or(&bitmap[word_idx], 1 << ((bitmap_idx & 7) + 8));
+      break;
+    case 2:
+      atomic_or(&bitmap[word_idx], 1 << ((bitmap_idx & 7) + 16));
+      break;
+    case 3:
+      atomic_or(&bitmap[word_idx], 1 << ((bitmap_idx & 7) + 24));
+      break;
+    default:
+      break;
+  }
+}
 }

omniscidb/QueryEngine/Compiler/genx.ll

@@ -273,60 +273,6 @@ define double @atomic_max_double(double addrspace(4)* %addr, double noundef %val
     ret double %old.cst
 }
 
-
-define void @agg_count_distinct_bitmap_gpu(i64 addrspace(4)* %agg, i64 noundef %val, i64 noundef %min_val, i64 noundef %base_dev_addr, i64 noundef %base_host_addr, i64 noundef %sub_bitmap_count, i64 noundef %bitmap_bytes) {
-    %bitmap_idx = sub nsw i64 %val, %min_val
-    %bitmap_idx.i32 = trunc i64 %bitmap_idx to i32
-    %byte_idx.i64 = lshr i64 %bitmap_idx, 3
-    %byte_idx = trunc i64 %byte_idx.i64 to i32
-    %word_idx = lshr i32 %byte_idx, 2
-    %word_idx.i64 = zext i32 %word_idx to i64
-    %byte_word_idx = and i32 %byte_idx, 3
-    %host_addr = load i64, i64 addrspace(4)* %agg
-    %sub_bm_m1 = sub i64 %sub_bitmap_count, 1
-    %tid = call i64 @get_thread_index()
-    %sub_tid = and i64 %sub_bm_m1, %tid
-    %rhs = mul i64 %sub_tid, %bitmap_bytes
-    %base_dev = add nsw i64 %base_dev_addr, %host_addr
-    %lhs = sub nsw i64 %base_dev, %base_host_addr
-    %bitmap_addr = add i64 %lhs, %rhs
-    %bitmap = inttoptr i64 %bitmap_addr to i32 addrspace(4)*
-    switch i32 %byte_word_idx, label %.exit [
-        i32 0, label %.c0
-        i32 1, label %.c1
-        i32 2, label %.c2
-        i32 3, label %.c3
-   ]
-
-.c0:
-    %btwi0 = getelementptr inbounds i32, i32 addrspace(4)* %bitmap, i64 %word_idx.i64
-    %res0 = and i32 %bitmap_idx.i32, 7
-    br label %.default
-.c1:
-    %btwi1 = getelementptr inbounds i32, i32 addrspace(4)* %bitmap, i64 %word_idx.i64
-    %btidx71 = and i32 %bitmap_idx.i32, 7
-    %res1 = or i32 %btidx71, 8
-    br label %.default
-.c2:
-    %btwi2 = getelementptr inbounds i32, i32 addrspace(4)* %bitmap, i64 %word_idx.i64
-    %btidx72 = and i32 %bitmap_idx.i32, 7
-    %res2 = or i32 %btidx72, 16
-    br label %.default
-.c3:
-    %btwi3 = getelementptr inbounds i32, i32 addrspace(4)* %bitmap, i64 %word_idx.i64
-    %btidx73 = and i32 %bitmap_idx.i32, 7
-    %res3 = or i32 %btidx73, 24
-    br label %.default
-.default:
-    %res = phi i32 [ %res0, %.c0 ], [ %res1, %.c1 ], [ %res2, %.c2], [ %res3, %.c3 ]
-    %arg0 = phi i32 addrspace(4)* [ %btwi0, %.c0 ], [ %btwi1, %.c1 ], [ %btwi2, %.c2], [ %btwi3, %.c3 ]
-    %arg1 = shl nuw i32 1, %res
-    tail call void @atomic_or(i32 addrspace(4)* %arg0, i32 noundef %arg1)
-    br label %.exit
-.exit:
-    ret void
-}
-
 define void @write_back_non_grouped_agg(i64 addrspace(4)* %input_buffer, i64 addrspace(4)* %output_buffer, i32 noundef %agg_idx) {
     %tid = call i64 @get_thread_index()
     %agg_idx.i64 = sext i32 %agg_idx to i64