[BACKEND] Fix crash in reductions on i1 (#1996)

`getScratchSizeInBytes` was assuming that the size of all types in bits
is
a multiple of 8. If it is not, it would return 0. This caused a bug for
boolean
(i1) type, where the reduction lowering would attempt to use shared
memory,
which was not assigned to the op.

Fix this issue by setting the number of bytes per element to `ceil(bits
/ 8)`.

lib/Analysis/Utility.cpp

@@ -174,7 +174,7 @@ unsigned ReduceOpHelper::getScratchSizeInBytes() {
 
   unsigned bytesPerElem = 0;
   for (const auto &ty : srcElementTypes) {
-    bytesPerElem += ty.getIntOrFloatBitWidth() / 8;
+    bytesPerElem += ceil<unsigned>(ty.getIntOrFloatBitWidth(), 8);
   }
   return bytesPerElem * elems;
 }

test/Conversion/tritongpu_to_llvm.mlir

@@ -1380,3 +1380,19 @@ module attributes {"triton_gpu.compute-capability" = 80 : i32, "triton_gpu.num-c
     tt.return
   }
 }
+
+// -----
+#blocked = #triton_gpu.blocked<{sizePerThread = [8, 1], threadsPerWarp = [32, 1], warpsPerCTA = [1, 2], order = [1, 0], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
+#slice = #triton_gpu.slice<{dim = 1, parent = #blocked}>
+module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 2 : i32} {
+  // CHECK-LABEL: reduce_bools
+  tt.func public @reduce_bools(%arg: tensor<256x2xi1, #blocked>) {
+    // CHECK: llvm.mlir.addressof @global_smem
+    %24 = "tt.reduce"(%arg) <{axis = 1 : i32}> ({
+    ^bb0(%arg4: i1, %arg5: i1):
+      %48 = arith.ori %arg4, %arg5 : i1
+      tt.reduce.return %48 : i1
+    }) : (tensor<256x2xi1, #blocked>) -> tensor<256xi1, #slice>
+    tt.return
+  }
+}