Merge pull request #7 from rivosinc/dev/PingTakPeterTang/hyperbolics

added sinh/cosh/tanh functions using mixed floating-point/fixed-point…

CMakeLists.txt

@@ -43,6 +43,12 @@ set(PROJECT_SOURCES
     ${CMAKE_CURRENT_SOURCE_DIR}/src/rvvlm_tanDI.c
     ${CMAKE_CURRENT_SOURCE_DIR}/src/rvvlm_tanpiD.c
     ${CMAKE_CURRENT_SOURCE_DIR}/src/rvvlm_tanpiDI.c
+    ${CMAKE_CURRENT_SOURCE_DIR}/src/rvvlm_coshD.c
+    ${CMAKE_CURRENT_SOURCE_DIR}/src/rvvlm_coshDI.c
+    ${CMAKE_CURRENT_SOURCE_DIR}/src/rvvlm_sinhD.c
+    ${CMAKE_CURRENT_SOURCE_DIR}/src/rvvlm_sinhDI.c
+    ${CMAKE_CURRENT_SOURCE_DIR}/src/rvvlm_tanhD.c
+    ${CMAKE_CURRENT_SOURCE_DIR}/src/rvvlm_tanhDI.c
 )
 
 add_library(vecm

include/rvvlm.h

@@ -262,6 +262,31 @@ union sui64_fp64 {
 #define RVVLM_TANPIDI_VSET_CONFIG "rvvlm_fp64m2.h"
 #define RVVLM_TANPIDI_MERGED rvvlm_tanpiI
 
+// FP64 cosh function configuration
+#define RVVLM_COSHD_VSET_CONFIG "rvvlm_fp64m2.h"
+#define RVVLM_COSHD_STD rvvlm_coshD_std
+#define RVVLM_COSHD_MIXED rvvlm_cosh
+
+#define RVVLM_COSHDI_VSET_CONFIG "rvvlm_fp64m2.h"
+#define RVVLM_COSHDI_STD rvvlm_coshI
+#define RVVLM_COSHDI_MIXED rvvlm_coshDI_mixed
+
+// FP64 sinh function configuration
+#define RVVLM_SINHD_VSET_CONFIG "rvvlm_fp64m2.h"
+#define RVVLM_SINHD_STD rvvlm_sinhD_std
+#define RVVLM_SINHD_MIXED rvvlm_sinh
+
+#define RVVLM_SINHDI_VSET_CONFIG "rvvlm_fp64m2.h"
+#define RVVLM_SINHDI_STD rvvlm_sinhI
+#define RVVLM_SINHDI_MIXED rvvlm_sinhDI_mixed
+
+// FP64 tanh function configuration
+#define RVVLM_TANHD_VSET_CONFIG "rvvlm_fp64m2.h"
+#define RVVLM_TANHD_STD rvvlm_tanh
+
+#define RVVLM_TANHDI_VSET_CONFIG "rvvlm_fp64m2.h"
+#define RVVLM_TANHDI_STD rvvlm_tanhI
+
 // Define the various tables for table-driven implementations
 extern int64_t expD_tbl64_fixedpt[64];
 extern int64_t logD_tbl128_fixedpt[128];
@@ -366,6 +391,26 @@ void RVVLM_TANPID_MERGED(size_t x_len, const double *x, double *y);
 void RVVLM_TANPIDI_MERGED(size_t x_len, const double *x, size_t stride_x,
                           double *y, size_t stride_y);
 
+void RVVLM_COSHD_STD(size_t x_len, const double *x, double *y);
+void RVVLM_COSHDI_STD(size_t x_len, const double *x, size_t stride_x, double *y,
+                      size_t stride_y);
+
+void RVVLM_COSHD_MIXED(size_t x_len, const double *x, double *y);
+void RVVLM_COSHDI_MIXED(size_t x_len, const double *x, size_t stride_x,
+                        double *y, size_t stride_y);
+
+void RVVLM_SINHD_STD(size_t x_len, const double *x, double *y);
+void RVVLM_SINHDI_STD(size_t x_len, const double *x, size_t stride_x, double *y,
+                      size_t stride_y);
+
+void RVVLM_SINHD_MIXED(size_t x_len, const double *x, double *y);
+void RVVLM_SINHDI_MIXED(size_t x_len, const double *x, size_t stride_x,
+                        double *y, size_t stride_y);
+
+void RVVLM_TANHD_STD(size_t x_len, const double *x, double *y);
+void RVVLM_TANHDI_STD(size_t x_len, const double *x, size_t stride_x, double *y,
+                      size_t stride_y);
+
 #ifdef __cplusplus
 }
 #endif

include/rvvlm_hyperbolicsD.h

@@ -0,0 +1,61 @@
+// SPDX-FileCopyrightText: 2023 Rivos Inc.
+//
+// SPDX-License-Identifier: Apache-2.0
+
+#if defined(COMPILE_FOR_SINH) || defined(COMPILE_FOR_TANH)
+#define GEN_EXCEPTIONS(special_args, vx, vlen)                                 \
+  __riscv_vfmadd((special_args), (vx), 0x1.0p-60, (vx), (vlen))
+#else
+#define GEN_EXCEPTIONS(special_args, vx, vlen)                                 \
+  __riscv_vfmadd((special_args), (vx), (vx), VFMV_VF(0x1.0p0, (vlen)), (vlen))
+#endif
+
+#if defined(COMPILE_FOR_SINH) || defined(COMPILE_FOR_COSH)
+#define EXCEPTION_HANDLING_HYPER(vx, expo_x, special_args, vy_special, vlen)   \
+  do {                                                                         \
+    VUINT vclass = __riscv_vfclass((vx), (vlen));                              \
+    VBOOL NaN_Inf;                                                             \
+    IDENTIFY(vclass, class_NaN | class_Inf, NaN_Inf, (vlen));                  \
+    VBOOL small_x = __riscv_vmsleu((expo_x), EXP_BIAS - MAN_LEN - 5, (vlen));  \
+    (special_args) = __riscv_vmor(NaN_Inf, small_x, (vlen));                   \
+    UINT nb_special_args = __riscv_vcpop((special_args), (vlen));              \
+    if (nb_special_args > 0) {                                                 \
+      (vy_special) = GEN_EXCEPTIONS((special_args), (vx), (vlen));             \
+      (vx) = __riscv_vfmerge((vx), fp_posZero, (special_args), (vlen));        \
+    }                                                                          \
+  } while (0)
+#else
+#define EXCEPTION_HANDLING_HYPER(vx, expo_x, special_args, vy_special, vlen)   \
+  do {                                                                         \
+    VUINT vclass = __riscv_vfclass((vx), (vlen));                              \
+    VBOOL NaN_Inf;                                                             \
+    IDENTIFY(vclass, class_NaN | class_Inf, NaN_Inf, (vlen));                  \
+    VBOOL small_x = __riscv_vmsleu((expo_x), EXP_BIAS - MAN_LEN - 5, (vlen));  \
+    (special_args) = __riscv_vmor(NaN_Inf, small_x, (vlen));                   \
+    UINT nb_special_args = __riscv_vcpop((special_args), (vlen));              \
+    if (nb_special_args > 0) {                                                 \
+      (vy_special) = GEN_EXCEPTIONS((special_args), (vx), (vlen));             \
+      VBOOL Inf;                                                               \
+      IDENTIFY(vclass, class_Inf, Inf, (vlen));                                \
+      VFLOAT one = VFMV_VF(fp_posOne, (vlen));                                 \
+      one = __riscv_vfsgnj(one, (vx), (vlen));                                 \
+      (vy_special) = __riscv_vmerge((vy_special), one, Inf, (vlen));           \
+      (vx) = __riscv_vfmerge((vx), fp_posZero, (special_args), (vlen));        \
+    }                                                                          \
+  } while (0)
+#endif
+
+#define LOG2_INV 0x1.71547652b82fep+0
+#define LOG2_HI 0x1.62e42fefa39efp-1
+#define LOG2_LO 0x1.abc9e3b39803fp-56
+
+#define ARGUMENT_REDUCTION(vx, n, r, r_delta, vlen)                            \
+  do {                                                                         \
+    VFLOAT n_flt = __riscv_vfmul((vx), LOG2_INV, (vlen));                      \
+    (n) = __riscv_vfcvt_x(n_flt, (vlen));                                      \
+    n_flt = __riscv_vfcvt_f((n), (vlen));                                      \
+    (r_delta) = __riscv_vfnmsac((vx), LOG2_HI, n_flt, (vlen));                 \
+    (r) = __riscv_vfnmsac((r_delta), LOG2_LO, n_flt, (vlen));                  \
+    (r_delta) = __riscv_vfsub((r_delta), (r), (vlen));                         \
+    (r_delta) = __riscv_vfnmsac((r_delta), LOG2_LO, n_flt, (vlen));            \
+  } while (0)