pl/math: Implement SVE sinpif/cospif

Implemented single precision SVE sinpi and cospi functions.

Accuracy:
_ZGVsMxv_sinpif 2.48 ULP
_ZGVsMxv_cospif 2.60 ULP

3 files changed, 126 insertions, 0 deletions

diff --git a/pl/math/include/mathlib.h b/pl/math/include/mathlib.h
index c252232..5b39ba7 100644
--- a/pl/math/include/mathlib.h
+++ b/pl/math/include/mathlib.h
@@ -156,6 +156,7 @@ svfloat64x2_t _ZGVsMxv_cexpi (svfloat64_t, svbool_t);
 svfloat32_t _ZGVsMxv_coshf (svfloat32_t, svbool_t);
 svfloat64_t _ZGVsMxv_cosh (svfloat64_t, svbool_t);
 svfloat32_t _ZGVsMxv_cosf (svfloat32_t, svbool_t);
+svfloat32_t _ZGVsMxv_cospif (svfloat32_t, svbool_t);
 svfloat64_t _ZGVsMxv_cos (svfloat64_t, svbool_t);
 svfloat32_t _ZGVsMxv_erff (svfloat32_t, svbool_t);
 svfloat64_t _ZGVsMxv_erf (svfloat64_t, svbool_t);
@@ -186,6 +187,7 @@ svfloat64_t _ZGVsMxvv_pow (svfloat64_t, svfloat64_t, svbool_t);
 svfloat32_t _ZGVsMxv_sinhf (svfloat32_t, svbool_t);
 svfloat64_t _ZGVsMxv_sinh (svfloat64_t, svbool_t);
 svfloat32_t _ZGVsMxv_sinf (svfloat32_t, svbool_t);
+svfloat32_t _ZGVsMxv_sinpif (svfloat32_t, svbool_t);
 svfloat64_t _ZGVsMxv_sin (svfloat64_t, svbool_t);
 svfloat32_t _ZGVsMxv_tanhf (svfloat32_t, svbool_t);
 svfloat64_t _ZGVsMxv_tanh (svfloat64_t, svbool_t);
diff --git a/pl/math/sv_cospif_2u6.c b/pl/math/sv_cospif_2u6.c
new file mode 100644
index 0000000..33ab8e0
--- /dev/null
+++ b/pl/math/sv_cospif_2u6.c
@@ -0,0 +1,65 @@
+/*
+ * Single-precision SVE cospi(x) function.
+ *
+ * Copyright (c) 2023, Arm Limited.
+ * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
+ */
+
+#include "mathlib.h"
+#include "sv_math.h"
+#include "pl_sig.h"
+#include "pl_test.h"
+#include "poly_sve_f32.h"
+
+static const struct data
+{
+  float poly[6];
+  float range_val;
+} data = {
+  /* Taylor series coefficents for sin(pi * x).  */
+  .poly = { 0x1.921fb6p1f, -0x1.4abbcep2f, 0x1.466bc6p1f, -0x1.32d2ccp-1f,
+	    0x1.50783p-4f, -0x1.e30750p-8f },
+  .range_val = 0x1p31f,
+};
+
+/* A fast SVE implementation of cospif.
+   Maximum error: 2.60 ULP:
+   _ZGVsMxv_cospif(+/-0x1.cae664p-4) got 0x1.e09c9ep-1
+				    want 0x1.e09c98p-1.  */
+svfloat32_t SV_NAME_F1 (cospi) (svfloat32_t x, const svbool_t pg)
+{
+  const struct data *d = ptr_barrier (&data);
+
+  /* Using cospi(x) = sinpi(0.5 - x)
+     range reduction and offset into sinpi range -1/2 .. 1/2
+     r = 0.5 - |x - rint(x)|.  */
+  svfloat32_t n = svrinta_x (pg, x);
+  svfloat32_t r = svsub_x (pg, x, n);
+  r = svsub_x (pg, sv_f32 (0.5f), svabs_x (pg, r));
+
+  /* Result should be negated based on if n is odd or not.
+     If ax >= 2^31, the result will always be positive.  */
+  svbool_t cmp = svaclt (pg, x, d->range_val);
+  svuint32_t intn = svreinterpret_u32 (svcvt_s32_x (pg, n));
+  svuint32_t sign = svlsl_z (cmp, intn, 31);
+
+  /* y = sin(r).  */
+  svfloat32_t r2 = svmul_x (pg, r, r);
+  svfloat32_t y = sv_horner_5_f32_x (pg, r2, d->poly);
+  y = svmul_x (pg, y, r);
+
+  return svreinterpret_f32 (sveor_x (pg, svreinterpret_u32 (y), sign));
+}
+
+PL_SIG (SV, F, 1, cospi, -0.9, 0.9)
+PL_TEST_ULP (SV_NAME_F1 (cospi), 2.08)
+PL_TEST_INTERVAL (SV_NAME_F1 (cospi), 0, 0x1p-31, 5000)
+PL_TEST_INTERVAL (SV_NAME_F1 (cospi), -0, -0x1p-31, 5000)
+PL_TEST_INTERVAL (SV_NAME_F1 (cospi), 0x1p-31, 0.5, 10000)
+PL_TEST_INTERVAL (SV_NAME_F1 (cospi), -0x1p-31, -0.5, 10000)
+PL_TEST_INTERVAL (SV_NAME_F1 (cospi), 0.5, 0x1p31f, 10000)
+PL_TEST_INTERVAL (SV_NAME_F1 (cospi), -0.5, -0x1p31f, 10000)
+PL_TEST_INTERVAL (SV_NAME_F1 (cospi), 0x1p31f, inf, 10000)
+PL_TEST_INTERVAL (SV_NAME_F1 (cospi), -0x1p31f, -inf, 10000)
+PL_TEST_INTERVAL (SV_NAME_F1 (cospi), inf, inf, 1)
+PL_TEST_INTERVAL (SV_NAME_F1 (cospi), -inf, -inf, 1)
\ No newline at end of file
diff --git a/pl/math/sv_sinpif_2u5.c b/pl/math/sv_sinpif_2u5.c
new file mode 100644
index 0000000..52c815b
--- /dev/null
+++ b/pl/math/sv_sinpif_2u5.c
@@ -0,0 +1,59 @@
+/*
+ * Single-precision SVE sinpi(x) function.
+ *
+ * Copyright (c) 2023, Arm Limited.
+ * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
+ */
+
+#include "mathlib.h"
+#include "sv_math.h"
+#include "pl_sig.h"
+#include "pl_test.h"
+#include "poly_sve_f32.h"
+
+static const struct data
+{
+  float poly[6];
+} data = {
+  /* Taylor series coefficents for sin(pi * x).  */
+  .poly = { 0x1.921fb6p1f, -0x1.4abbcep2f, 0x1.466bc6p1f, -0x1.32d2ccp-1f,
+	    0x1.50783p-4f, -0x1.e30750p-8f },
+};
+
+/* A fast SVE implementation of sinpif.
+   Maximum error 2.48 ULP:
+   _ZGVsMxv_sinpif(0x1.d062b6p-2) got 0x1.fa8c06p-1
+				 want 0x1.fa8c02p-1.  */
+svfloat32_t SV_NAME_F1 (sinpi) (svfloat32_t x, const svbool_t pg)
+{
+  const struct data *d = ptr_barrier (&data);
+
+  /* range reduction into -1/2 .. 1/2
+     with n = rint(x) and r = r - n.  */
+  svfloat32_t n = svrinta_x (pg, x);
+  svfloat32_t r = svsub_x (pg, x, n);
+
+  /* Result should be negated based on if n is odd or not.  */
+  svuint32_t intn = svreinterpret_u32 (svcvt_s32_x (pg, n));
+  svuint32_t sign = svlsl_z (pg, intn, 31);
+
+  /* y = sin(r).  */
+  svfloat32_t r2 = svmul_x (pg, r, r);
+  svfloat32_t y = sv_horner_5_f32_x (pg, r2, d->poly);
+  y = svmul_x (pg, y, r);
+
+  return svreinterpret_f32 (sveor_x (pg, svreinterpret_u32 (y), sign));
+}
+
+PL_SIG (SV, F, 1, sinpi, -0.9, 0.9)
+PL_TEST_ULP (SV_NAME_F1 (sinpi), 1.99)
+PL_TEST_INTERVAL (SV_NAME_F1 (sinpi), 0, 0x1p-31, 5000)
+PL_TEST_INTERVAL (SV_NAME_F1 (sinpi), -0, -0x1p-31, 5000)
+PL_TEST_INTERVAL (SV_NAME_F1 (sinpi), 0x1p-31, 0.5, 10000)
+PL_TEST_INTERVAL (SV_NAME_F1 (sinpi), -0x1p-31, -0.5, 10000)
+PL_TEST_INTERVAL (SV_NAME_F1 (sinpi), 0.5, 0x1p22f, 10000)
+PL_TEST_INTERVAL (SV_NAME_F1 (sinpi), -0.5, -0x1p22f, 10000)
+PL_TEST_INTERVAL (SV_NAME_F1 (sinpi), 0x1p22f, inf, 10000)
+PL_TEST_INTERVAL (SV_NAME_F1 (sinpi), -0x1p22f, -inf, 10000)
+PL_TEST_INTERVAL (SV_NAME_F1 (sinpi), inf, inf, 1)
+PL_TEST_INTERVAL (SV_NAME_F1 (sinpi), -inf, -inf, 1)
\ No newline at end of file