xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/sve/pow.c (revision f3087bef11543b42e0d69b708f367097a4118d24)
1 /*
2  * Double-precision SVE pow(x, y) function.
3  *
4  * Copyright (c) 2022-2025, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include "sv_math.h"
9 #include "test_sig.h"
10 #include "test_defs.h"
11 
12 /* This version share a similar algorithm as AOR scalar pow.
13 
14    The core computation consists in computing pow(x, y) as
15 
16      exp (y * log (x)).
17 
18    The algorithms for exp and log are very similar to scalar exp and log.
19    The log relies on table lookup for 3 variables and an order 8 polynomial.
20    It returns a high and a low contribution that are then passed to the exp,
21    to minimise the loss of accuracy in both routines.
22    The exp is based on 8-bit table lookup for scale and order-4 polynomial.
23    The SVE algorithm drops the tail in the exp computation at the price of
24    a lower accuracy, slightly above 1ULP.
25    The SVE algorithm also drops the special treatement of small (< 2^-65) and
26    large (> 2^63) finite values of |y|, as they only affect non-round to
27    nearest modes.
28 
29    Maximum measured error is 1.04 ULPs:
30    SV_NAME_D2 (pow) (0x1.3d2d45bc848acp+63, -0x1.a48a38b40cd43p-12)
31      got 0x1.f7116284221fcp-1
32     want 0x1.f7116284221fdp-1.  */
33 
34 /* Data is defined in v_pow_log_data.c.  */
35 #define N_LOG (1 << V_POW_LOG_TABLE_BITS)
36 #define Off 0x3fe6955500000000
37 
38 /* Data is defined in v_pow_exp_data.c.  */
39 #define N_EXP (1 << V_POW_EXP_TABLE_BITS)
40 #define SignBias (0x800 << V_POW_EXP_TABLE_BITS)
41 #define SmallExp 0x3c9 /* top12(0x1p-54).  */
42 #define BigExp 0x408   /* top12(512.).  */
43 #define ThresExp 0x03f /* BigExp - SmallExp.  */
44 #define HugeExp 0x409  /* top12(1024.).  */
45 
46 /* Constants associated with pow.  */
47 #define SmallBoundX 0x1p-126
48 #define SmallPowX 0x001 /* top12(0x1p-126).  */
49 #define BigPowX 0x7ff	/* top12(INFINITY).  */
50 #define ThresPowX 0x7fe /* BigPowX - SmallPowX.  */
51 #define SmallPowY 0x3be /* top12(0x1.e7b6p-65).  */
52 #define BigPowY 0x43e	/* top12(0x1.749p62).  */
53 #define ThresPowY 0x080 /* BigPowY - SmallPowY.  */
54 
55 static const struct data
56 {
57   double log_c0, log_c2, log_c4, log_c6, ln2_hi, ln2_lo;
58   double log_c1, log_c3, log_c5, off;
59   double n_over_ln2, exp_c2, ln2_over_n_hi, ln2_over_n_lo;
60   double exp_c0, exp_c1;
61 } data = {
62   .log_c0 = -0x1p-1,
63   .log_c1 = -0x1.555555555556p-1,
64   .log_c2 = 0x1.0000000000006p-1,
65   .log_c3 = 0x1.999999959554ep-1,
66   .log_c4 = -0x1.555555529a47ap-1,
67   .log_c5 = -0x1.2495b9b4845e9p0,
68   .log_c6 = 0x1.0002b8b263fc3p0,
69   .off = Off,
70   .exp_c0 = 0x1.fffffffffffd4p-2,
71   .exp_c1 = 0x1.5555571d6ef9p-3,
72   .exp_c2 = 0x1.5555576a5adcep-5,
73   .ln2_hi = 0x1.62e42fefa3800p-1,
74   .ln2_lo = 0x1.ef35793c76730p-45,
75   .n_over_ln2 = 0x1.71547652b82fep0 * N_EXP,
76   .ln2_over_n_hi = 0x1.62e42fefc0000p-9,
77   .ln2_over_n_lo = -0x1.c610ca86c3899p-45,
78 };
79 
80 /* Check if x is an integer.  */
81 static inline svbool_t
sv_isint(svbool_t pg,svfloat64_t x)82 sv_isint (svbool_t pg, svfloat64_t x)
83 {
84   return svcmpeq (pg, svrintz_z (pg, x), x);
85 }
86 
87 /* Check if x is real not integer valued.  */
88 static inline svbool_t
sv_isnotint(svbool_t pg,svfloat64_t x)89 sv_isnotint (svbool_t pg, svfloat64_t x)
90 {
91   return svcmpne (pg, svrintz_z (pg, x), x);
92 }
93 
94 /* Check if x is an odd integer.  */
95 static inline svbool_t
sv_isodd(svbool_t pg,svfloat64_t x)96 sv_isodd (svbool_t pg, svfloat64_t x)
97 {
98   svfloat64_t y = svmul_x (svptrue_b64 (), x, 0.5);
99   return sv_isnotint (pg, y);
100 }
101 
102 /* Returns 0 if not int, 1 if odd int, 2 if even int.  The argument is
103    the bit representation of a non-zero finite floating-point value.  */
104 static inline int
checkint(uint64_t iy)105 checkint (uint64_t iy)
106 {
107   int e = iy >> 52 & 0x7ff;
108   if (e < 0x3ff)
109     return 0;
110   if (e > 0x3ff + 52)
111     return 2;
112   if (iy & ((1ULL << (0x3ff + 52 - e)) - 1))
113     return 0;
114   if (iy & (1ULL << (0x3ff + 52 - e)))
115     return 1;
116   return 2;
117 }
118 
119 /* Top 12 bits (sign and exponent of each double float lane).  */
120 static inline svuint64_t
sv_top12(svfloat64_t x)121 sv_top12 (svfloat64_t x)
122 {
123   return svlsr_x (svptrue_b64 (), svreinterpret_u64 (x), 52);
124 }
125 
126 /* Returns 1 if input is the bit representation of 0, infinity or nan.  */
127 static inline int
zeroinfnan(uint64_t i)128 zeroinfnan (uint64_t i)
129 {
130   return 2 * i - 1 >= 2 * asuint64 (INFINITY) - 1;
131 }
132 
133 /* Returns 1 if input is the bit representation of 0, infinity or nan.  */
134 static inline svbool_t
sv_zeroinfnan(svbool_t pg,svuint64_t i)135 sv_zeroinfnan (svbool_t pg, svuint64_t i)
136 {
137   return svcmpge (pg, svsub_x (pg, svadd_x (pg, i, i), 1),
138 		  2 * asuint64 (INFINITY) - 1);
139 }
140 
141 /* Handle cases that may overflow or underflow when computing the result that
142    is scale*(1+TMP) without intermediate rounding.  The bit representation of
143    scale is in SBITS, however it has a computed exponent that may have
144    overflown into the sign bit so that needs to be adjusted before using it as
145    a double.  (int32_t)KI is the k used in the argument reduction and exponent
146    adjustment of scale, positive k here means the result may overflow and
147    negative k means the result may underflow.  */
148 static inline double
specialcase(double tmp,uint64_t sbits,uint64_t ki)149 specialcase (double tmp, uint64_t sbits, uint64_t ki)
150 {
151   double scale;
152   if ((ki & 0x80000000) == 0)
153     {
154       /* k > 0, the exponent of scale might have overflowed by <= 460.  */
155       sbits -= 1009ull << 52;
156       scale = asdouble (sbits);
157       return 0x1p1009 * (scale + scale * tmp);
158     }
159   /* k < 0, need special care in the subnormal range.  */
160   sbits += 1022ull << 52;
161   /* Note: sbits is signed scale.  */
162   scale = asdouble (sbits);
163   double y = scale + scale * tmp;
164   return 0x1p-1022 * y;
165 }
166 
167 /* Scalar fallback for special cases of SVE pow's exp.  */
168 static inline svfloat64_t
sv_call_specialcase(svfloat64_t x1,svuint64_t u1,svuint64_t u2,svfloat64_t y,svbool_t cmp)169 sv_call_specialcase (svfloat64_t x1, svuint64_t u1, svuint64_t u2,
170 		     svfloat64_t y, svbool_t cmp)
171 {
172   svbool_t p = svpfirst (cmp, svpfalse ());
173   while (svptest_any (cmp, p))
174     {
175       double sx1 = svclastb (p, 0, x1);
176       uint64_t su1 = svclastb (p, 0, u1);
177       uint64_t su2 = svclastb (p, 0, u2);
178       double elem = specialcase (sx1, su1, su2);
179       svfloat64_t y2 = sv_f64 (elem);
180       y = svsel (p, y2, y);
181       p = svpnext_b64 (cmp, p);
182     }
183   return y;
184 }
185 
186 /* Compute y+TAIL = log(x) where the rounded result is y and TAIL has about
187    additional 15 bits precision.  IX is the bit representation of x, but
188    normalized in the subnormal range using the sign bit for the exponent.  */
189 static inline svfloat64_t
sv_log_inline(svbool_t pg,svuint64_t ix,svfloat64_t * tail,const struct data * d)190 sv_log_inline (svbool_t pg, svuint64_t ix, svfloat64_t *tail,
191 	       const struct data *d)
192 {
193   /* x = 2^k z; where z is in range [Off,2*Off) and exact.
194      The range is split into N subintervals.
195      The ith subinterval contains z and c is near its center.  */
196   svuint64_t tmp = svsub_x (pg, ix, d->off);
197   svuint64_t i = svand_x (pg, svlsr_x (pg, tmp, 52 - V_POW_LOG_TABLE_BITS),
198 			  sv_u64 (N_LOG - 1));
199   svint64_t k = svasr_x (pg, svreinterpret_s64 (tmp), 52);
200   svuint64_t iz = svsub_x (pg, ix, svlsl_x (pg, svreinterpret_u64 (k), 52));
201   svfloat64_t z = svreinterpret_f64 (iz);
202   svfloat64_t kd = svcvt_f64_x (pg, k);
203 
204   /* log(x) = k*Ln2 + log(c) + log1p(z/c-1).  */
205   /* SVE lookup requires 3 separate lookup tables, as opposed to scalar version
206      that uses array of structures. We also do the lookup earlier in the code
207      to make sure it finishes as early as possible.  */
208   svfloat64_t invc = svld1_gather_index (pg, __v_pow_log_data.invc, i);
209   svfloat64_t logc = svld1_gather_index (pg, __v_pow_log_data.logc, i);
210   svfloat64_t logctail = svld1_gather_index (pg, __v_pow_log_data.logctail, i);
211 
212   /* Note: 1/c is j/N or j/N/2 where j is an integer in [N,2N) and
213      |z/c - 1| < 1/N, so r = z/c - 1 is exactly representible.  */
214   svfloat64_t r = svmad_x (pg, z, invc, -1.0);
215   /* k*Ln2 + log(c) + r.  */
216 
217   svfloat64_t ln2_hilo = svld1rq_f64 (svptrue_b64 (), &d->ln2_hi);
218   svfloat64_t t1 = svmla_lane_f64 (logc, kd, ln2_hilo, 0);
219   svfloat64_t t2 = svadd_x (pg, t1, r);
220   svfloat64_t lo1 = svmla_lane_f64 (logctail, kd, ln2_hilo, 1);
221   svfloat64_t lo2 = svadd_x (pg, svsub_x (pg, t1, t2), r);
222 
223   /* Evaluation is optimized assuming superscalar pipelined execution.  */
224 
225   svfloat64_t log_c02 = svld1rq_f64 (svptrue_b64 (), &d->log_c0);
226   svfloat64_t ar = svmul_lane_f64 (r, log_c02, 0);
227   svfloat64_t ar2 = svmul_x (svptrue_b64 (), r, ar);
228   svfloat64_t ar3 = svmul_x (svptrue_b64 (), r, ar2);
229   /* k*Ln2 + log(c) + r + A[0]*r*r.  */
230   svfloat64_t hi = svadd_x (pg, t2, ar2);
231   svfloat64_t lo3 = svmls_x (pg, ar2, ar, r);
232   svfloat64_t lo4 = svadd_x (pg, svsub_x (pg, t2, hi), ar2);
233   /* p = log1p(r) - r - A[0]*r*r.  */
234   /* p = (ar3 * (A[1] + r * A[2] + ar2 * (A[3] + r * A[4] + ar2 * (A[5] + r *
235      A[6])))).  */
236 
237   svfloat64_t log_c46 = svld1rq_f64 (svptrue_b64 (), &d->log_c4);
238   svfloat64_t a56 = svmla_lane_f64 (sv_f64 (d->log_c5), r, log_c46, 1);
239   svfloat64_t a34 = svmla_lane_f64 (sv_f64 (d->log_c3), r, log_c46, 0);
240   svfloat64_t a12 = svmla_lane_f64 (sv_f64 (d->log_c1), r, log_c02, 1);
241   svfloat64_t p = svmla_x (pg, a34, ar2, a56);
242   p = svmla_x (pg, a12, ar2, p);
243   p = svmul_x (svptrue_b64 (), ar3, p);
244   svfloat64_t lo = svadd_x (
245       pg, svadd_x (pg, svsub_x (pg, svadd_x (pg, lo1, lo2), lo3), lo4), p);
246   svfloat64_t y = svadd_x (pg, hi, lo);
247   *tail = svadd_x (pg, svsub_x (pg, hi, y), lo);
248   return y;
249 }
250 
251 static inline svfloat64_t
sv_exp_core(svbool_t pg,svfloat64_t x,svfloat64_t xtail,svuint64_t sign_bias,svfloat64_t * tmp,svuint64_t * sbits,svuint64_t * ki,const struct data * d)252 sv_exp_core (svbool_t pg, svfloat64_t x, svfloat64_t xtail,
253 	     svuint64_t sign_bias, svfloat64_t *tmp, svuint64_t *sbits,
254 	     svuint64_t *ki, const struct data *d)
255 {
256   /* exp(x) = 2^(k/N) * exp(r), with exp(r) in [2^(-1/2N),2^(1/2N)].  */
257   /* x = ln2/N*k + r, with int k and r in [-ln2/2N, ln2/2N].  */
258   svfloat64_t n_over_ln2_and_c2 = svld1rq_f64 (svptrue_b64 (), &d->n_over_ln2);
259   svfloat64_t z = svmul_lane_f64 (x, n_over_ln2_and_c2, 0);
260   /* z - kd is in [-1, 1] in non-nearest rounding modes.  */
261   svfloat64_t kd = svrinta_x (pg, z);
262   *ki = svreinterpret_u64 (svcvt_s64_x (pg, kd));
263 
264   svfloat64_t ln2_over_n_hilo
265       = svld1rq_f64 (svptrue_b64 (), &d->ln2_over_n_hi);
266   svfloat64_t r = x;
267   r = svmls_lane_f64 (r, kd, ln2_over_n_hilo, 0);
268   r = svmls_lane_f64 (r, kd, ln2_over_n_hilo, 1);
269   /* The code assumes 2^-200 < |xtail| < 2^-8/N.  */
270   r = svadd_x (pg, r, xtail);
271   /* 2^(k/N) ~= scale.  */
272   svuint64_t idx = svand_x (pg, *ki, N_EXP - 1);
273   svuint64_t top
274       = svlsl_x (pg, svadd_x (pg, *ki, sign_bias), 52 - V_POW_EXP_TABLE_BITS);
275   /* This is only a valid scale when -1023*N < k < 1024*N.  */
276   *sbits = svld1_gather_index (pg, __v_pow_exp_data.sbits, idx);
277   *sbits = svadd_x (pg, *sbits, top);
278   /* exp(x) = 2^(k/N) * exp(r) ~= scale + scale * (exp(r) - 1).  */
279   svfloat64_t r2 = svmul_x (svptrue_b64 (), r, r);
280   *tmp = svmla_lane_f64 (sv_f64 (d->exp_c1), r, n_over_ln2_and_c2, 1);
281   *tmp = svmla_x (pg, sv_f64 (d->exp_c0), r, *tmp);
282   *tmp = svmla_x (pg, r, r2, *tmp);
283   svfloat64_t scale = svreinterpret_f64 (*sbits);
284   /* Note: tmp == 0 or |tmp| > 2^-200 and scale > 2^-739, so there
285      is no spurious underflow here even without fma.  */
286   z = svmla_x (pg, scale, scale, *tmp);
287   return z;
288 }
289 
290 /* Computes sign*exp(x+xtail) where |xtail| < 2^-8/N and |xtail| <= |x|.
291    The sign_bias argument is SignBias or 0 and sets the sign to -1 or 1.  */
292 static inline svfloat64_t
sv_exp_inline(svbool_t pg,svfloat64_t x,svfloat64_t xtail,svuint64_t sign_bias,const struct data * d)293 sv_exp_inline (svbool_t pg, svfloat64_t x, svfloat64_t xtail,
294 	       svuint64_t sign_bias, const struct data *d)
295 {
296   /* 3 types of special cases: tiny (uflow and spurious uflow), huge (oflow)
297      and other cases of large values of x (scale * (1 + TMP) oflow).  */
298   svuint64_t abstop = svand_x (pg, sv_top12 (x), 0x7ff);
299   /* |x| is large (|x| >= 512) or tiny (|x| <= 0x1p-54).  */
300   svbool_t uoflow = svcmpge (pg, svsub_x (pg, abstop, SmallExp), ThresExp);
301 
302   svfloat64_t tmp;
303   svuint64_t sbits, ki;
304   if (unlikely (svptest_any (pg, uoflow)))
305     {
306       svfloat64_t z
307 	  = sv_exp_core (pg, x, xtail, sign_bias, &tmp, &sbits, &ki, d);
308 
309       /* |x| is tiny (|x| <= 0x1p-54).  */
310       svbool_t uflow
311 	  = svcmpge (pg, svsub_x (pg, abstop, SmallExp), 0x80000000);
312       uflow = svand_z (pg, uoflow, uflow);
313       /* |x| is huge (|x| >= 1024).  */
314       svbool_t oflow = svcmpge (pg, abstop, HugeExp);
315       oflow = svand_z (pg, uoflow, svbic_z (pg, oflow, uflow));
316 
317       /* For large |x| values (512 < |x| < 1024) scale * (1 + TMP) can overflow
318     or underflow.  */
319       svbool_t special = svbic_z (pg, uoflow, svorr_z (pg, uflow, oflow));
320 
321       /* Update result with special and large cases.  */
322       z = sv_call_specialcase (tmp, sbits, ki, z, special);
323 
324       /* Handle underflow and overflow.  */
325       svbool_t x_is_neg = svcmplt (pg, x, 0);
326       svuint64_t sign_mask
327 	  = svlsl_x (pg, sign_bias, 52 - V_POW_EXP_TABLE_BITS);
328       svfloat64_t res_uoflow
329 	  = svsel (x_is_neg, sv_f64 (0.0), sv_f64 (INFINITY));
330       res_uoflow = svreinterpret_f64 (
331 	  svorr_x (pg, svreinterpret_u64 (res_uoflow), sign_mask));
332       /* Avoid spurious underflow for tiny x.  */
333       svfloat64_t res_spurious_uflow
334 	  = svreinterpret_f64 (svorr_x (pg, sign_mask, 0x3ff0000000000000));
335 
336       z = svsel (oflow, res_uoflow, z);
337       z = svsel (uflow, res_spurious_uflow, z);
338       return z;
339     }
340 
341   return sv_exp_core (pg, x, xtail, sign_bias, &tmp, &sbits, &ki, d);
342 }
343 
344 static inline double
pow_sc(double x,double y)345 pow_sc (double x, double y)
346 {
347   uint64_t ix = asuint64 (x);
348   uint64_t iy = asuint64 (y);
349   /* Special cases: |x| or |y| is 0, inf or nan.  */
350   if (unlikely (zeroinfnan (iy)))
351     {
352       if (2 * iy == 0)
353 	return issignaling_inline (x) ? x + y : 1.0;
354       if (ix == asuint64 (1.0))
355 	return issignaling_inline (y) ? x + y : 1.0;
356       if (2 * ix > 2 * asuint64 (INFINITY) || 2 * iy > 2 * asuint64 (INFINITY))
357 	return x + y;
358       if (2 * ix == 2 * asuint64 (1.0))
359 	return 1.0;
360       if ((2 * ix < 2 * asuint64 (1.0)) == !(iy >> 63))
361 	return 0.0; /* |x|<1 && y==inf or |x|>1 && y==-inf.  */
362       return y * y;
363     }
364   if (unlikely (zeroinfnan (ix)))
365     {
366       double_t x2 = x * x;
367       if (ix >> 63 && checkint (iy) == 1)
368 	x2 = -x2;
369       return (iy >> 63) ? 1 / x2 : x2;
370     }
371   return x;
372 }
373 
SV_NAME_D2(pow)374 svfloat64_t SV_NAME_D2 (pow) (svfloat64_t x, svfloat64_t y, const svbool_t pg)
375 {
376   const struct data *d = ptr_barrier (&data);
377 
378   /* This preamble handles special case conditions used in the final scalar
379      fallbacks. It also updates ix and sign_bias, that are used in the core
380      computation too, i.e., exp( y * log (x) ).  */
381   svuint64_t vix0 = svreinterpret_u64 (x);
382   svuint64_t viy0 = svreinterpret_u64 (y);
383 
384   /* Negative x cases.  */
385   svbool_t xisneg = svcmplt (pg, x, 0);
386 
387   /* Set sign_bias and ix depending on sign of x and nature of y.  */
388   svbool_t yint_or_xpos = pg;
389   svuint64_t sign_bias = sv_u64 (0);
390   svuint64_t vix = vix0;
391   if (unlikely (svptest_any (pg, xisneg)))
392     {
393       /* Determine nature of y.  */
394       yint_or_xpos = sv_isint (xisneg, y);
395       svbool_t yisodd_xisneg = sv_isodd (xisneg, y);
396       /* ix set to abs(ix) if y is integer.  */
397       vix = svand_m (yint_or_xpos, vix0, 0x7fffffffffffffff);
398       /* Set to SignBias if x is negative and y is odd.  */
399       sign_bias = svsel (yisodd_xisneg, sv_u64 (SignBias), sv_u64 (0));
400     }
401 
402   /* Small cases of x: |x| < 0x1p-126.  */
403   svbool_t xsmall = svaclt (yint_or_xpos, x, SmallBoundX);
404   if (unlikely (svptest_any (yint_or_xpos, xsmall)))
405     {
406       /* Normalize subnormal x so exponent becomes negative.  */
407       svuint64_t vtopx = svlsr_x (svptrue_b64 (), vix, 52);
408       svbool_t topx_is_null = svcmpeq (xsmall, vtopx, 0);
409 
410       svuint64_t vix_norm = svreinterpret_u64 (svmul_m (xsmall, x, 0x1p52));
411       vix_norm = svand_m (xsmall, vix_norm, 0x7fffffffffffffff);
412       vix_norm = svsub_m (xsmall, vix_norm, 52ULL << 52);
413       vix = svsel (topx_is_null, vix_norm, vix);
414     }
415 
416   /* y_hi = log(ix, &y_lo).  */
417   svfloat64_t vlo;
418   svfloat64_t vhi = sv_log_inline (yint_or_xpos, vix, &vlo, d);
419 
420   /* z = exp(y_hi, y_lo, sign_bias).  */
421   svfloat64_t vehi = svmul_x (svptrue_b64 (), y, vhi);
422   svfloat64_t vemi = svmls_x (yint_or_xpos, vehi, y, vhi);
423   svfloat64_t velo = svnmls_x (yint_or_xpos, vemi, y, vlo);
424   svfloat64_t vz = sv_exp_inline (yint_or_xpos, vehi, velo, sign_bias, d);
425 
426   /* Cases of finite y and finite negative x.  */
427   vz = svsel (yint_or_xpos, vz, sv_f64 (__builtin_nan ("")));
428 
429   /* Special cases of x or y: zero, inf and nan.  */
430   svbool_t xspecial = sv_zeroinfnan (svptrue_b64 (), vix0);
431   svbool_t yspecial = sv_zeroinfnan (svptrue_b64 (), viy0);
432   svbool_t special = svorr_z (svptrue_b64 (), xspecial, yspecial);
433 
434   /* Cases of zero/inf/nan x or y.  */
435   if (unlikely (svptest_any (svptrue_b64 (), special)))
436     vz = sv_call2_f64 (pow_sc, x, y, vz, special);
437 
438   return vz;
439 }
440 
441 TEST_SIG (SV, D, 2, pow)
442 TEST_ULP (SV_NAME_D2 (pow), 0.55)
443 TEST_DISABLE_FENV (SV_NAME_D2 (pow))
444 /* Wide intervals spanning the whole domain but shared between x and y.  */
445 #define SV_POW_INTERVAL2(xlo, xhi, ylo, yhi, n)                               \
446   TEST_INTERVAL2 (SV_NAME_D2 (pow), xlo, xhi, ylo, yhi, n)                    \
447   TEST_INTERVAL2 (SV_NAME_D2 (pow), xlo, xhi, -ylo, -yhi, n)                  \
448   TEST_INTERVAL2 (SV_NAME_D2 (pow), -xlo, -xhi, ylo, yhi, n)                  \
449   TEST_INTERVAL2 (SV_NAME_D2 (pow), -xlo, -xhi, -ylo, -yhi, n)
450 #define EXPAND(str) str##000000000
451 #define SHL52(str) EXPAND (str)
452 SV_POW_INTERVAL2 (0, SHL52 (SmallPowX), 0, inf, 40000)
453 SV_POW_INTERVAL2 (SHL52 (SmallPowX), SHL52 (BigPowX), 0, inf, 40000)
454 SV_POW_INTERVAL2 (SHL52 (BigPowX), inf, 0, inf, 40000)
455 SV_POW_INTERVAL2 (0, inf, 0, SHL52 (SmallPowY), 40000)
456 SV_POW_INTERVAL2 (0, inf, SHL52 (SmallPowY), SHL52 (BigPowY), 40000)
457 SV_POW_INTERVAL2 (0, inf, SHL52 (BigPowY), inf, 40000)
458 SV_POW_INTERVAL2 (0, inf, 0, inf, 1000)
459 /* x~1 or y~1.  */
460 SV_POW_INTERVAL2 (0x1p-1, 0x1p1, 0x1p-10, 0x1p10, 10000)
461 SV_POW_INTERVAL2 (0x1.ep-1, 0x1.1p0, 0x1p8, 0x1p16, 10000)
462 SV_POW_INTERVAL2 (0x1p-500, 0x1p500, 0x1p-1, 0x1p1, 10000)
463 /* around estimated argmaxs of ULP error.  */
464 SV_POW_INTERVAL2 (0x1p-300, 0x1p-200, 0x1p-20, 0x1p-10, 10000)
465 SV_POW_INTERVAL2 (0x1p50, 0x1p100, 0x1p-20, 0x1p-10, 10000)
466 /* x is negative, y is odd or even integer, or y is real not integer.  */
467 TEST_INTERVAL2 (SV_NAME_D2 (pow), -0.0, -10.0, 3.0, 3.0, 10000)
468 TEST_INTERVAL2 (SV_NAME_D2 (pow), -0.0, -10.0, 4.0, 4.0, 10000)
469 TEST_INTERVAL2 (SV_NAME_D2 (pow), -0.0, -10.0, 0.0, 10.0, 10000)
470 TEST_INTERVAL2 (SV_NAME_D2 (pow), 0.0, 10.0, -0.0, -10.0, 10000)
471 /* |x| is inf, y is odd or even integer, or y is real not integer.  */
472 SV_POW_INTERVAL2 (inf, inf, 0.5, 0.5, 1)
473 SV_POW_INTERVAL2 (inf, inf, 1.0, 1.0, 1)
474 SV_POW_INTERVAL2 (inf, inf, 2.0, 2.0, 1)
475 SV_POW_INTERVAL2 (inf, inf, 3.0, 3.0, 1)
476 /* 0.0^y.  */
477 SV_POW_INTERVAL2 (0.0, 0.0, 0.0, 0x1p120, 1000)
478 /* 1.0^y.  */
479 TEST_INTERVAL2 (SV_NAME_D2 (pow), 1.0, 1.0, 0.0, 0x1p-50, 1000)
480 TEST_INTERVAL2 (SV_NAME_D2 (pow), 1.0, 1.0, 0x1p-50, 1.0, 1000)
481 TEST_INTERVAL2 (SV_NAME_D2 (pow), 1.0, 1.0, 1.0, 0x1p100, 1000)
482 TEST_INTERVAL2 (SV_NAME_D2 (pow), 1.0, 1.0, -1.0, -0x1p120, 1000)
483 CLOSE_SVE_ATTR
484