enhance test infra and testing; fixed 2 bugs caught by extra tests

rivosinc · Nov 6, 2024 · 8e0b4b6 · 8e0b4b6
1 parent deba355
commit 8e0b4b6
Show file tree

Hide file tree

Showing 20 changed files with 399 additions and 47 deletions.
diff --git a/include/rvvlm_powD.inc.h b/include/rvvlm_powD.inc.h
@@ -92,10 +92,12 @@
         vy = __riscv_vfmin_mu(current_cases, vy, vy, 0x1.0p53, vlen);          \
         vy = __riscv_vfmax_mu(current_cases, vy, vy, -0x1.0p53, vlen);         \
         VINT y_to_int = __riscv_vfcvt_x(current_cases, vy, vlen);              \
-        /* TODO: y_to_int_fp and y_is_int need to be used */                   \
         VFLOAT y_to_int_fp = __riscv_vfcvt_f(current_cases, y_to_int, vlen);   \
         VBOOL y_is_int = __riscv_vmfeq(current_cases, vy, y_to_int_fp, vlen);  \
         VINT sign_z = __riscv_vsll(y_to_int, 63, vlen);                        \
+        VINT zero;                                                             \
+        zero = __riscv_vxor(zero, zero, vlen);                                 \
+        sign_z = __riscv_vmerge(zero, sign_z, y_is_int, vlen);                 \
         /* the parity is used later on to manipulate sign, hence sll 63 bits   \
          */                                                                    \
                                                                                \

diff --git a/include/rvvlm_sinandcosD.inc.h b/include/rvvlm_sinandcosD.inc.h
@@ -67,6 +67,7 @@ void F_VER1(API) {
       VFLOAT S = __riscv_vfsub(r_hi, A, vlen);
       VFLOAT s = __riscv_vfsub(r_hi, S, vlen);
       s = __riscv_vfsub(s, A, vlen);
+      s = __riscv_vfsub(s, a, vlen);
       s = __riscv_vfnmsac(s, PIBY2_LO, n_flt, vlen);
       r = __riscv_vmerge(r, S, r_small, vlen);
       r_delta = __riscv_vmerge(r_delta, s, r_small, vlen);

diff --git a/include/rvvlm_sincosD.inc.h b/include/rvvlm_sincosD.inc.h
@@ -78,6 +78,7 @@ void F_VER1(API) {
       VFLOAT S = __riscv_vfsub(r_hi, A, vlen);
       VFLOAT s = __riscv_vfsub(r_hi, S, vlen);
       s = __riscv_vfsub(s, A, vlen);
+      s = __riscv_vfsub(s, a, vlen);
       s = __riscv_vfnmsac(s, PIBY2_LO, n_flt, vlen);
       r = __riscv_vmerge(r, S, r_small, vlen);
       r_delta = __riscv_vmerge(r_delta, s, r_small, vlen);

diff --git a/include/rvvlm_tanD.inc.h b/include/rvvlm_tanD.inc.h
@@ -66,6 +66,7 @@ void F_VER1(API) {
       VFLOAT S = __riscv_vfsub(r_hi, A, vlen);
       VFLOAT s = __riscv_vfsub(r_hi, S, vlen);
       s = __riscv_vfsub(s, A, vlen);
+      s = __riscv_vfsub(s, a, vlen);
       s = __riscv_vfnmsac(s, PIBY2_LO, n_flt, vlen);
       r = __riscv_vmerge(r, S, r_small, vlen);
       r_delta = __riscv_vmerge(r_delta, s, r_small, vlen);

diff --git a/test/include/near_NPiby2_tbl.h b/test/include/near_NPiby2_tbl.h
@@ -0,0 +1,71 @@
+// SPDX-FileCopyrightText: 2024 Rivos Inc.
+//
+// SPDX-License-Identifier: Apache-2.0
+
+#define NB_TEST_ARGS 63*3
+static const double dbl_near_NPiby2_tbl[NB_TEST_ARGS] = {
+   0x1.6c6cbc45dc8dep+9,0x1.6c6cbc45dc8dep+10,0x1.6c6cbc45dc8dep+11,
+   0x1.6c6cbc45dc8dep+12,0x1.635e3d74befcap+14,0x1.67e57cdd4dc54p+15,
+   0x1.65a1dd290660fp+16,0x1.bf9b3c6059d24p+17,0x1.39c6fd67805a7p+18,
+   0x1.39c6fd67805a7p+19,0x1.9eb7148f354d6p+20,0x1.9eb7148f354d6p+20,
+   0x1.d3ecce1f28274p+22,0x1.b951f1572eba5p+23,0x1.b951f1572eba5p+24,
+   0x1.b951f1572eba5p+25,0x1.b951f1572eba5p+26,0x1.b951f1572eba5p+27,
+   0x1.b951f1572eba5p+28,0x1.b951f1572eba5p+29,0x1.b951f1572eba5p+30,
+   0x1.b951f1572eba5p+31,0x1.5c9508c58aafap+32,0x1.5c9508c58aafap+32,
+   0x1.de5e5054e921bp+34,0x1.46546a5bd73ccp+35,0x1.46546a5bd73ccp+35,
+   0x1.46546a5bd73ccp+35,0x1.bb23eaa3db16dp+38,0x1.065c829d68730p+39,
+   0x1.065c829d68730p+39,0x1.065c829d68730p+39,0x1.065c829d68730p+39,
+   0x1.065c829d68730p+39,0x1.065c829d68730p+40,0x1.065c829d68730p+41,
+   0x1.f42d52c35675dp+46,0x1.7512069b7430dp+47,0x1.7512069b7430dp+48,
+   0x1.7512069b7430dp+49,0x1.44630cc2cad9dp+50,0x1.5cba89af1f855p+51,
+   0x1.5cba89af1f855p+52,0x1.56a4aa740a5a7p+53,0x1.59af9a1194efep+54,
+   0x1.59af9a1194efep+54,0x1.ae9608c734e12p+56,0x1.ae9608c734e12p+56,
+   0x1.c3cfa4749cdd7p+58,0x1.c3cfa4749cdd7p+59,0x1.c3cfa4749cdd7p+60,
+   0x1.c3cfa4749cdd7p+61,0x1.ed814f0220525p+62,0x1.ed814f0220525p+63,
+   0x1.a7f3bc5a7ed9ep+64,0x1.a7f3bc5a7ed9ep+64,0x1.a7f3bc5a7ed9ep+65,
+   0x1.782b7a20df6d4p+67,0x1.782b7a20df6d4p+67,0x1.782b7a20df6d4p+67,
+   0x1.782b7a20df6d4p+68,0x1.782b7a20df6d4p+69,0x1.782b7a20df6d4p+70,
+   0x1.782b7a20df6d4p+71,0x1.a4ffb379a019cp+74,0x1.308ab84507e83p+75,
+   0x1.99caa5236feeap+76,0x1.99caa5236feeap+76,0x1.99caa5236feeap+77,
+   0x1.99caa5236feeap+78,0x1.99caa5236feeap+79,0x1.e0664dbedfec5p+81,
+   0x1.e0664dbedfec5p+82,0x1.e0664dbedfec5p+83,0x1.e0664dbedfec5p+84,
+   0x1.f19e5d71b26bap+85,0x1.f19e5d71b26bap+85,0x1.f19e5d71b26bap+86,
+   0x1.f19e5d71b26bap+87,0x1.d4ab7990f194dp+89,0x1.66bd5424e5655p+90,
+   0x1.66bd5424e5655p+91,0x1.66bd5424e5655p+92,0x1.66bd5424e5655p+93,
+   0x1.66bd5424e5655p+94,0x1.c3dc1a597f202p+95,0x1.c3dc1a597f202p+95,
+   0x1.c3dc1a597f202p+96,0x1.e50fec1788957p+98,0x1.e50fec1788957p+99,
+   0x1.e50fec1788957p+100,0x1.e50fec1788957p+101,0x1.77d794c891d1cp+102,
+   0x1.dae3a00451b5fp+103,0x1.dae3a00451b5fp+104,0x1.1c6cc29b3b153p+105,
+   0x1.d99e1681eada0p+106,0x1.d99e1681eada0p+106,0x1.d99e1681eada0p+106,
+   0x1.d99e1681eada0p+106,0x1.d99e1681eada0p+106,0x1.d99e1681eada0p+106,
+   0x1.f8d36a8915598p+112,0x1.ae04fdb542be8p+113,0x1.ae04fdb542be8p+113,
+   0x1.ae04fdb542be8p+113,0x1.ae04fdb542be8p+113,0x1.ae04fdb542be8p+114,
+   0x1.ed82892c91569p+118,0x1.47623428452f0p+119,0x1.47623428452f0p+119,
+   0x1.47623428452f0p+119,0x1.f7e7a76e07787p+122,0x1.f7e7a76e07787p+123,
+   0x1.e7426933f5760p+124,0x1.e7426933f5760p+124,0x1.e7426933f5760p+124,
+   0x1.e7426933f5760p+124,0x1.e7426933f5760p+124,0x1.e7426933f5760p+124,
+   0x1.f91a847a07755p+130,0x1.504cac51f1eafp+131,0x1.504cac51f1eafp+132,
+   0x1.504cac51f1eafp+133,0x1.504cac51f1eafp+134,0x1.504cac51f1eafp+135,
+   0x1.504cac51f1eafp+136,0x1.8a56c23d2dfe0p+137,0x1.8a56c23d2dfe0p+137,
+   0x1.b2e25ce13d6a7p+139,0x1.e42d3528e32a3p+140,0x1.fcd2a14cb60a1p+141,
+   0x1.fcd2a14cb60a1p+142,0x1.5ad5a62cb1cc9p+143,0x1.5ad5a62cb1cc9p+144,
+   0x1.5ad5a62cb1cc9p+145,0x1.5ad5a62cb1cc9p+146,0x1.f1cdf37f20ff1p+147,
+   0x1.a651ccd5e965dp+148,0x1.a651ccd5e965dp+149,0x1.b930d680374c2p+150,
+   0x1.b930d680374c2p+150,0x1.b930d680374c2p+151,0x1.b930d680374c2p+152,
+   0x1.641bcd12e3311p+154,0x1.205a1c297f6b9p+155,0x1.205a1c297f6b9p+156,
+   0x1.205a1c297f6b9p+157,0x1.17e1e60c52f2ep+158,0x1.17e1e60c52f2ep+158,
+   0x1.ee076da427656p+160,0x1.82f4a9d83d2c2p+161,0x1.82f4a9d83d2c2p+161,
+   0x1.682ff8e5429ddp+163,0x1.682ff8e5429ddp+164,0x1.682ff8e5429ddp+165,
+   0x1.682ff8e5429ddp+166,0x1.cca7515e2585cp+167,0x1.cca7515e2585cp+167,
+   0x1.bdf4d57f7f0c4p+169,0x1.c54e136ed2490p+170,0x1.c54e136ed2490p+170,
+   0x1.c54e136ed2490p+170,0x1.fc218660d7c2fp+173,0x1.fc218660d7c2fp+174,
+   0x1.d28d5c4c5e73bp+175,0x1.d28d5c4c5e73bp+176,0x1.d28d5c4c5e73bp+177,
+   0x1.2722f5d698acep+178,0x1.2722f5d698acep+178,0x1.2722f5d698acep+179,
+   0x1.76e1ee2edd343p+181,0x1.0539b48d14c55p+182,0x1.0539b48d14c55p+183,
+   0x1.0539b48d14c55p+184,0x1.0539b48d14c55p+185,0x1.6014d4a7e0086p+186,
+   0x1.b5441ee104c98p+187,0x1.b5441ee104c98p+187,0x1.f7fd817cb39b5p+189,
+   0x1.f7fd817cb39b5p+190,0x1.bb4c8e40cbe5dp+191,0x1.bb4c8e40cbe5dp+192,
+   0x1.3ecf49b990c61p+193,0x1.0090a775f3363p+194,0x1.1faff897c1fe2p+195,
+   0x1.1faff897c1fe2p+195,0x1.e7e44a78ac18cp+197,0x1.e7e44a78ac18cp+197
+};
+
diff --git a/test/include/test_infra.h b/test/include/test_infra.h
@@ -6,8 +6,53 @@
 
 #include <stdio.h>
 
+#define DBL_NEG_INF (int64_t)0xFFF0000000000000
+#define DBL_NEG_GT1 (int64_t)0xBFF0000000000001
+#define DBL_NEG_ONE (int64_t)0xBFF0000000000000
+#define DBL_NEG_LT1 (int64_t)0xBFEFFFFFFFFFFFFF
+#define DBL_NEG_TINY (int64_t)0x800FFFFFFFFFFFFF
+#define DBL_NEG_ZERO (int64_t)0x8000000000000000
+#define DBL_POS_ZERO (int64_t)0x0000000000000000
+#define DBL_POS_TINY (int64_t)0x000FFFFFFFFFFFFF
+#define DBL_POS_LT1 (int64_t)0x3FEFFFFFFFFFFFFF
+#define DBL_POS_ONE (int64_t)0x3FF0000000000000
+#define DBL_POS_GT1 (int64_t)0x3FF0000000000001
+#define DBL_POS_INF (int64_t)0x7FF0000000000000
+#define DBL_QNAN (int64_t)0x7FF8000000000000
+#define DBL_SNAN (int64_t)0x7FF4000000000000
+
+#define NB_TV_FOR_TRIG 4
+#define TV_FOR_TRIG                                                            \
+  {DBL_QNAN,    DBL_QNAN, DBL_SNAN,    DBL_QNAN,                               \
+   DBL_NEG_INF, DBL_QNAN, DBL_POS_INF, DBL_QNAN}
+
+#define NB_TV_FOR_EXP 4
+#define TV_FOR_EXP                                                             \
+  {DBL_QNAN,    DBL_QNAN,     DBL_SNAN,    DBL_QNAN,                           \
+   DBL_NEG_INF, DBL_POS_ZERO, DBL_POS_INF, DBL_POS_INF}
+
+#define NB_TV_FOR_EXPM1 5
+#define TV_FOR_EXPM1                                                           \
+  {DBL_QNAN,    DBL_QNAN,    DBL_SNAN,    DBL_QNAN,     DBL_NEG_INF,           \
+   DBL_NEG_ONE, DBL_POS_INF, DBL_POS_INF, DBL_POS_ZERO, DBL_POS_ZERO}
+
+#define NB_TV_FOR_LOG 8
+#define TV_FOR_LOG                                                             \
+  {DBL_QNAN,     DBL_QNAN,    DBL_SNAN,     DBL_QNAN,                          \
+   DBL_NEG_INF,  DBL_QNAN,    DBL_NEG_GT1,  DBL_QNAN,                          \
+   DBL_NEG_TINY, DBL_QNAN,    DBL_NEG_ZERO, DBL_NEG_INF,                       \
+   DBL_POS_ZERO, DBL_NEG_INF, DBL_POS_INF,  DBL_POS_INF}
+
+#define NB_TV_FOR_LOG1P 7
+#define TV_FOR_LOG1P                                                           \
+  {DBL_QNAN,     DBL_QNAN,     DBL_SNAN,    DBL_QNAN,    DBL_NEG_INF,          \
+   DBL_QNAN,     DBL_NEG_GT1,  DBL_QNAN,    DBL_NEG_ONE, DBL_NEG_INF,          \
+   DBL_POS_ZERO, DBL_POS_ZERO, DBL_POS_INF, DBL_POS_INF}
+
 #define COMMENT(comment)                                                       \
-  { printf("\n=====\t" comment "\n"); }
+  {                                                                            \
+    printf("\n=====\t" comment "\n");                                          \
+  }
 
 // Most common interface: testing on 1 interval
 // for 1-in-1-out unit-stride function
@@ -20,8 +65,14 @@ void report_err_fp64(void (*test_func)(size_t, const double *, double *,
                      long double (*ref_func)(long double), int, double, double,
                      int, double = 1.0);
 
+void report_err_fp64(void (*test_func)(size_t, const double *, double *,
+                                       double *),
+                     long double (*ref_func)(long double), int, const double *,
+                     int, double = 1.0);
+
 void report_err_fp64(void (*test_func)(size_t, const double *, double *),
-                     long double (*ref_func)(long double), const double *, int);
+                     long double (*ref_func)(long double), const double *, int,
+                     double = 1.0);
 
 void report_mixederr_fp64(void (*test_func)(size_t, const double *, double *),
                           long double (*ref_func)(long double), double, double,
@@ -86,6 +137,13 @@ void show_special2_fp64(void (*test_func)(size_t, const double *,
                                           const double *, double *),
                         int, char *);
 
+void test_vectors_fp64(void (*test_func)(size_t, const double *, double *),
+                       int64_t *, int);
+
+void test_vectors2_fp64(void (*test_func)(size_t, const double *,
+                                          const double *, double *),
+                        int64_t *, int, int);
+
 void trig_2pi_reduction(long double, long double *, int64_t *);
 
 long double acospil(long double);

diff --git a/test/src/test_cos.cpp b/test/src/test_cos.cpp
@@ -9,12 +9,22 @@
 #include "test_infra.h"
 
 TEST(cos, special) {
+  int nb_tv;
+  int64_t tv_in_out[2 * (NB_TV_FOR_TRIG)] = TV_FOR_TRIG;
+
+  nb_tv = NB_TV_FOR_TRIG;
+
+  test_vectors_fp64(rvvlm_cos, tv_in_out, nb_tv);
+}
+
+TEST(cos, near_NPiby2) {
+#include "near_NPiby2_tbl.h"
   unsigned long nb_tests;
-  double x_start, x_end;
 
   COMMENT("cos: current chosen algorithm; reduced argument in FP64 only")
 
-  show_special_fp64(rvvlm_cos, "Special Value handling of this function");
+  nb_tests = (NB_TEST_ARGS);
+  report_err_fp64(rvvlm_cos, cosl, dbl_near_NPiby2_tbl, nb_tests);
 }
 
 TEST(cos, small_args) {

diff --git a/test/src/test_exp.cpp b/test/src/test_exp.cpp
@@ -8,14 +8,21 @@
 #include "rvvlm.h"
 #include "test_infra.h"
 
+TEST(exp, special) {
+  int nb_tv;
+  int64_t tv_in_out[2 * (NB_TV_FOR_EXP)] = TV_FOR_EXP;
+
+  nb_tv = NB_TV_FOR_EXP;
+
+  test_vectors_fp64(rvvlm_exp, tv_in_out, nb_tv);
+}
+
 TEST(exp, small_args) {
   unsigned long nb_tests;
   double x_start, x_end;
 
   COMMENT("exp: current chosen algorithm; reduced argument in FP64 only")
 
-  show_special_fp64(rvvlm_exp, "Special Value handling of this function");
-
   x_start = -0.34;
   x_end = 0.34;
   nb_tests = 100000;

diff --git a/test/src/test_exp10.cpp b/test/src/test_exp10.cpp
@@ -8,14 +8,21 @@
 #include "rvvlm.h"
 #include "test_infra.h"
 
+TEST(exp10, special) {
+  int nb_tv;
+  int64_t tv_in_out[2 * (NB_TV_FOR_EXP)] = TV_FOR_EXP;
+
+  nb_tv = NB_TV_FOR_EXP;
+
+  test_vectors_fp64(rvvlm_exp10, tv_in_out, nb_tv);
+}
+
 TEST(exp10, small_args) {
   unsigned long nb_tests;
   double x_start, x_end;
 
   COMMENT("exp10: current chosen algorithm; reduced argument in FP64 only")
 
-  show_special_fp64(rvvlm_exp10, "Special Value handling of this function");
-
   x_start = -0.34;
   x_end = 0.34;
   nb_tests = 30000;

diff --git a/test/src/test_exp2.cpp b/test/src/test_exp2.cpp
@@ -8,14 +8,21 @@
 #include "rvvlm.h"
 #include "test_infra.h"
 
+TEST(exp2, special) {
+  int nb_tv;
+  int64_t tv_in_out[2 * (NB_TV_FOR_EXP)] = TV_FOR_EXP;
+
+  nb_tv = NB_TV_FOR_EXP;
+
+  test_vectors_fp64(rvvlm_exp2, tv_in_out, nb_tv);
+}
+
 TEST(exp2, small_args) {
   unsigned long nb_tests;
   double x_start, x_end;
 
   COMMENT("exp2: current chosen algorithm; reduced argument in FP64 only")
 
-  show_special_fp64(rvvlm_exp2, "Special Value handling of this function");
-
   x_start = -0.34;
   x_end = 0.34;
   nb_tests = 80000;

diff --git a/test/src/test_expm1.cpp b/test/src/test_expm1.cpp
@@ -8,14 +8,21 @@
 #include "rvvlm.h"
 #include "test_infra.h"
 
+TEST(expm1, special) {
+  int nb_tv;
+  int64_t tv_in_out[2 * (NB_TV_FOR_EXPM1)] = TV_FOR_EXPM1;
+
+  nb_tv = NB_TV_FOR_EXPM1;
+
+  test_vectors_fp64(rvvlm_expm1, tv_in_out, nb_tv);
+}
+
 TEST(expm1, small_args) {
   unsigned long nb_tests;
   double x_start, x_end;
 
   COMMENT("expm1: current chosen algorithm; reduced argument in FP64 only")
 
-  show_special_fp64(rvvlm_expm1, "Special Value handling of this function");
-
   x_start = -0.01;
   x_end = 0.01;
   nb_tests = 30000;