1 files changed, 38 insertions, 0 deletions
diff --git a/math/aarch64/experimental/atanf_common.h b/math/aarch64/experimental/atanf_common.h
new file mode 100644
index 000000000000..3e6542047309
--- /dev/null
+++ b/math/aarch64/experimental/atanf_common.h
@@ -0,0 +1,38 @@
+/*
+ * Single-precision polynomial evaluation function for scalar
+ * atan(x) and atan2(y,x).
+ *
+ * Copyright (c) 2021-2024, Arm Limited.
+ * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
+ */
+
+#ifndef PL_MATH_ATANF_COMMON_H
+#define PL_MATH_ATANF_COMMON_H
+
+#include "math_config.h"
+#include "poly_scalar_f32.h"
+
+/* Polynomial used in fast atanf(x) and atan2f(y,x) implementations
+   The order 7 polynomial P approximates (atan(sqrt(x))-sqrt(x))/x^(3/2).  */
+static inline float
+eval_poly (float z, float az, float shift)
+{
+  /* Use 2-level Estrin scheme for P(z^2) with deg(P)=7. However,
+     a standard implementation using z8 creates spurious underflow
+     in the very last fma (when z^8 is small enough).
+     Therefore, we split the last fma into a mul and and an fma.
+     Horner and single-level Estrin have higher errors that exceed
+     threshold.  */
+  float z2 = z * z;
+  float z4 = z2 * z2;
+
+  /* Then assemble polynomial.  */
+  float y = fmaf (
+      z4, z4 * pairwise_poly_3_f32 (z2, z4, __atanf_poly_data.poly + 4),
+      pairwise_poly_3_f32 (z2, z4, __atanf_poly_data.poly));
+  /* Finalize:
+     y = shift + z * P(z^2).  */
+  return fmaf (y, z2 * az, az) + shift;
+}
+
+#endif // PL_MATH_ATANF_COMMON_H