Improve float to string cast by ~20%-40% (#5401)

* Add cast f64 to string to benchmark

* Improve float to string performance using ryu

arrow-cast/Cargo.toml

@@ -52,6 +52,7 @@ lexical-core = { version = "^0.8", default-features = false, features = ["write-
 atoi = "2.0.0"
 comfy-table = { version = "7.0", optional = true, default-features = false }
 base64 = "0.21"
+ryu = "1.0.16"
 
 [dev-dependencies]
 criterion = { version = "0.5", default-features = false }

arrow-cast/src/display.rs

@@ -427,9 +427,23 @@ macro_rules! primitive_display {
     };
 }
 
+macro_rules! primitive_display_float {
+    ($($t:ty),+) => {
+        $(impl<'a> DisplayIndex for &'a PrimitiveArray<$t>
+        {
+            fn write(&self, idx: usize, f: &mut dyn Write) -> FormatResult {
+                let value = self.value(idx);
+                let mut buffer = ryu::Buffer::new();
+                f.write_str(buffer.format(value))?;
+                Ok(())
+            }
+        })+
+    };
+}
+
 primitive_display!(Int8Type, Int16Type, Int32Type, Int64Type);
 primitive_display!(UInt8Type, UInt16Type, UInt32Type, UInt64Type);
-primitive_display!(Float32Type, Float64Type);
+primitive_display_float!(Float32Type, Float64Type);
 
 impl<'a> DisplayIndex for &'a PrimitiveArray<Float16Type> {
     fn write(&self, idx: usize, f: &mut dyn Write) -> FormatResult {

arrow/benches/cast_kernels.rs

@@ -207,7 +207,9 @@ fn add_benchmark(c: &mut Criterion) {
     c.bench_function("cast f32 to string 512", |b| {
         b.iter(|| cast_array(&f32_array, DataType::Utf8))
     });
-
+    c.bench_function("cast f64 to string 512", |b| {
+        b.iter(|| cast_array(&f64_array, DataType::Utf8))
+    });
     c.bench_function("cast timestamp_ms to i64 512", |b| {
         b.iter(|| cast_array(&time_ms_array, DataType::Int64))
     });