math: Add expm1 builtin function
Ported from the amd-builtins branch.
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Matt Arsenault <Matthew.Arsenault@amd.com>
CC: Tom Stellard <thomas.stellard@amd.com>
llvm-svn: 292334
diff --git a/libclc/generic/lib/math/tables.cl b/libclc/generic/lib/math/tables.cl
index c1dd85c..3e1e4db 100644
--- a/libclc/generic/lib/math/tables.cl
+++ b/libclc/generic/lib/math/tables.cl
@@ -608,6 +608,142 @@
(float2)(0x1.428000p+0f, 0x1.45f31ap-13f)
};
+DECLARE_TABLE(float, EXP_TBL, 65) = {
+ 0x1.000000p+0f,
+ 0x1.02c9a4p+0f,
+ 0x1.059b0ep+0f,
+ 0x1.087452p+0f,
+ 0x1.0b5586p+0f,
+ 0x1.0e3ec4p+0f,
+ 0x1.11301ep+0f,
+ 0x1.1429aap+0f,
+ 0x1.172b84p+0f,
+ 0x1.1a35bep+0f,
+ 0x1.1d4874p+0f,
+ 0x1.2063b8p+0f,
+ 0x1.2387a6p+0f,
+ 0x1.26b456p+0f,
+ 0x1.29e9e0p+0f,
+ 0x1.2d285ap+0f,
+ 0x1.306fe0p+0f,
+ 0x1.33c08cp+0f,
+ 0x1.371a74p+0f,
+ 0x1.3a7db4p+0f,
+ 0x1.3dea64p+0f,
+ 0x1.4160a2p+0f,
+ 0x1.44e086p+0f,
+ 0x1.486a2cp+0f,
+ 0x1.4bfdaep+0f,
+ 0x1.4f9b28p+0f,
+ 0x1.5342b6p+0f,
+ 0x1.56f474p+0f,
+ 0x1.5ab07ep+0f,
+ 0x1.5e76f2p+0f,
+ 0x1.6247ecp+0f,
+ 0x1.662388p+0f,
+ 0x1.6a09e6p+0f,
+ 0x1.6dfb24p+0f,
+ 0x1.71f75ep+0f,
+ 0x1.75feb6p+0f,
+ 0x1.7a1148p+0f,
+ 0x1.7e2f34p+0f,
+ 0x1.82589ap+0f,
+ 0x1.868d9ap+0f,
+ 0x1.8ace54p+0f,
+ 0x1.8f1aeap+0f,
+ 0x1.93737cp+0f,
+ 0x1.97d82ap+0f,
+ 0x1.9c4918p+0f,
+ 0x1.a0c668p+0f,
+ 0x1.a5503cp+0f,
+ 0x1.a9e6b6p+0f,
+ 0x1.ae89fap+0f,
+ 0x1.b33a2cp+0f,
+ 0x1.b7f770p+0f,
+ 0x1.bcc1eap+0f,
+ 0x1.c199bep+0f,
+ 0x1.c67f12p+0f,
+ 0x1.cb720ep+0f,
+ 0x1.d072d4p+0f,
+ 0x1.d5818ep+0f,
+ 0x1.da9e60p+0f,
+ 0x1.dfc974p+0f,
+ 0x1.e502eep+0f,
+ 0x1.ea4afap+0f,
+ 0x1.efa1bep+0f,
+ 0x1.f50766p+0f,
+ 0x1.fa7c18p+0f,
+ 0x1.000000p+1f,
+};
+
+DECLARE_TABLE(float2, EXP_TBL_EP, 65) = {
+ (float2) (0x1.000000p+0f, 0x0.000000p+0f),
+ (float2) (0x1.02c000p+0f, 0x1.347ceep-13f),
+ (float2) (0x1.058000p+0f, 0x1.b0d314p-12f),
+ (float2) (0x1.084000p+0f, 0x1.a28c3ap-11f),
+ (float2) (0x1.0b4000p+0f, 0x1.586cf8p-12f),
+ (float2) (0x1.0e0000p+0f, 0x1.f61968p-11f),
+ (float2) (0x1.110000p+0f, 0x1.80e808p-11f),
+ (float2) (0x1.140000p+0f, 0x1.4d5754p-11f),
+ (float2) (0x1.170000p+0f, 0x1.5c1e3ep-11f),
+ (float2) (0x1.1a0000p+0f, 0x1.adf5b6p-11f),
+ (float2) (0x1.1d4000p+0f, 0x1.0e62d0p-13f),
+ (float2) (0x1.204000p+0f, 0x1.1dc430p-11f),
+ (float2) (0x1.238000p+0f, 0x1.e9b9d4p-14f),
+ (float2) (0x1.268000p+0f, 0x1.a2b2f0p-11f),
+ (float2) (0x1.29c000p+0f, 0x1.4efa8ep-11f),
+ (float2) (0x1.2d0000p+0f, 0x1.42d372p-11f),
+ (float2) (0x1.304000p+0f, 0x1.7f0518p-11f),
+ (float2) (0x1.33c000p+0f, 0x1.164c82p-17f),
+ (float2) (0x1.370000p+0f, 0x1.a7373ap-12f),
+ (float2) (0x1.3a4000p+0f, 0x1.ed9a72p-11f),
+ (float2) (0x1.3dc000p+0f, 0x1.532608p-11f),
+ (float2) (0x1.414000p+0f, 0x1.0510fap-11f),
+ (float2) (0x1.44c000p+0f, 0x1.043030p-11f),
+ (float2) (0x1.484000p+0f, 0x1.515ae0p-11f),
+ (float2) (0x1.4bc000p+0f, 0x1.ed6a9ap-11f),
+ (float2) (0x1.4f8000p+0f, 0x1.b2769cp-12f),
+ (float2) (0x1.534000p+0f, 0x1.5ab4eap-15f),
+ (float2) (0x1.56c000p+0f, 0x1.a39b5ap-11f),
+ (float2) (0x1.5a8000p+0f, 0x1.83eea4p-11f),
+ (float2) (0x1.5e4000p+0f, 0x1.b78ad6p-11f),
+ (float2) (0x1.624000p+0f, 0x1.fac0e8p-14f),
+ (float2) (0x1.660000p+0f, 0x1.1c412ap-11f),
+ (float2) (0x1.6a0000p+0f, 0x1.3cccfep-13f),
+ (float2) (0x1.6dc000p+0f, 0x1.d91e32p-11f),
+ (float2) (0x1.71c000p+0f, 0x1.baf476p-11f),
+ (float2) (0x1.75c000p+0f, 0x1.f5ab20p-11f),
+ (float2) (0x1.7a0000p+0f, 0x1.1473eap-12f),
+ (float2) (0x1.7e0000p+0f, 0x1.799b66p-11f),
+ (float2) (0x1.824000p+0f, 0x1.89994cp-12f),
+ (float2) (0x1.868000p+0f, 0x1.b33688p-13f),
+ (float2) (0x1.8ac000p+0f, 0x1.ca8454p-13f),
+ (float2) (0x1.8f0000p+0f, 0x1.ae9914p-12f),
+ (float2) (0x1.934000p+0f, 0x1.9bd866p-11f),
+ (float2) (0x1.97c000p+0f, 0x1.829fdep-12f),
+ (float2) (0x1.9c4000p+0f, 0x1.230546p-13f),
+ (float2) (0x1.a0c000p+0f, 0x1.99ed76p-14f),
+ (float2) (0x1.a54000p+0f, 0x1.03b23ep-12f),
+ (float2) (0x1.a9c000p+0f, 0x1.35aabcp-11f),
+ (float2) (0x1.ae8000p+0f, 0x1.3f32b4p-13f),
+ (float2) (0x1.b30000p+0f, 0x1.d15c26p-11f),
+ (float2) (0x1.b7c000p+0f, 0x1.bb797cp-11f),
+ (float2) (0x1.bcc000p+0f, 0x1.e904bcp-16f),
+ (float2) (0x1.c18000p+0f, 0x1.9bdd84p-12f),
+ (float2) (0x1.c64000p+0f, 0x1.f8972ap-11f),
+ (float2) (0x1.cb4000p+0f, 0x1.906e76p-11f),
+ (float2) (0x1.d04000p+0f, 0x1.96a502p-11f),
+ (float2) (0x1.d58000p+0f, 0x1.8dcfbap-16f),
+ (float2) (0x1.da8000p+0f, 0x1.e603dap-12f),
+ (float2) (0x1.dfc000p+0f, 0x1.2e66f6p-13f),
+ (float2) (0x1.e50000p+0f, 0x1.773c58p-15f),
+ (float2) (0x1.ea4000p+0f, 0x1.5f4548p-13f),
+ (float2) (0x1.ef8000p+0f, 0x1.0df730p-11f),
+ (float2) (0x1.f50000p+0f, 0x1.d96db8p-14f),
+ (float2) (0x1.fa4000p+0f, 0x1.e0c0cep-11f),
+ (float2) (0x1.000000p+1f, 0x0.000000p+0f),
+};
+
TABLE_FUNCTION(float2, LOGE_TBL, loge_tbl);
TABLE_FUNCTION(float, LOG_INV_TBL, log_inv_tbl);
TABLE_FUNCTION(float2, LOG2_TBL, log2_tbl);
@@ -618,6 +754,8 @@
TABLE_FUNCTION(float2, SINHCOSH_TBL, sinhcosh_tbl);
TABLE_FUNCTION(float2, CBRT_TBL, cbrt_tbl);
+TABLE_FUNCTION(float, EXP_TBL, exp_tbl);
+TABLE_FUNCTION(float2, EXP_TBL_EP, exp_tbl_ep);
#ifdef cl_khr_fp64