summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorS. Gilles <sgilles@math.umd.edu>2018-07-31 18:36:05 -0400
committerS. Gilles <sgilles@math.umd.edu>2018-07-31 20:05:29 -0400
commita00a5db9878af9a52fe84cb1c0a7e9316924c7fa (patch)
treec4b49686d94eb0744f63bc353bba22ef8cbf33b2
parentf2e8e641bf8ee06fa5128c522b151a7f4d9d003b (diff)
downloadmc-a00a5db9878af9a52fe84cb1c0a7e9316924c7fa.tar.gz
Correct extra-precision division typo for atan.
-rw-r--r--lib/math/atan-impl.myr105
-rw-r--r--lib/math/test/atan-impl.myr78
2 files changed, 166 insertions, 17 deletions
diff --git a/lib/math/atan-impl.myr b/lib/math/atan-impl.myr
index 7591df6..78f235d 100644
--- a/lib/math/atan-impl.myr
+++ b/lib/math/atan-impl.myr
@@ -317,6 +317,18 @@ const C : (uint64, uint64, uint64, uint64, uint64, uint64, uint64)[257] = [
]
const atan32 = {x
+ /*
+ Irritating special rounding cases that would require
+ more extra accuracy than they're worth to correctly round
+ back from 64 to 32 bits.
+ */
+ var xb = std.flt32bits(x)
+ if xb == 0x3d8d6b23
+ -> std.flt32frombits(0x3d8d31c3)
+ elif xb == 0xbd8d6b23
+ -> std.flt32frombits(0xbd8d31c3)
+ ;;
+
var r = atan264((x : flt64), 1.0)
-> (r : flt32)
}
@@ -326,6 +338,16 @@ const atan64 = {x
}
const atan232 = {y, x
+ /* Handle the special cases of atan32 for consistency */
+ if x == 1.0
+ var yb = std.flt32bits(y)
+ if yb == 0x3d8d6b23
+ -> std.flt32frombits(0x3d8d31c3)
+ elif yb == 0xbd8d6b23
+ -> std.flt32frombits(0xbd8d31c3)
+ ;;
+ ;;
+
var r = atan264((y : flt64), (x : flt64))
-> (r : flt32)
}
@@ -335,36 +357,79 @@ const atan264 = {y, x
-> std.flt64nan()
;;
- var xabs = std.flt64frombits(std.flt64bits(x) & (~(0 : uint64)) >> 1)
- var yabs = std.flt64frombits(std.flt64bits(y) & (~(0 : uint64)) >> 1)
-
- var then_negate = false
- if (xabs == x) != (yabs == y)
- then_negate = true
- ;;
- x = xabs
- y = yabs
-
var xb = std.flt64bits(x)
- if xb == 0x0
- if then_negate
+ var yb = std.flt64bits(y)
+ var xpos, xe, ypos, ye
+ (xpos, xe, _) = std.flt64explode(x)
+ (ypos, ye, _) = std.flt64explode(y)
+ xpos = !xpos
+ ypos = !ypos
+
+ /* All the special cases */
+ if yb == 0x0000000000000000
+ if xpos
+ -> 0.0
+ else
+ -> std.flt64frombits(0x400921fb54442d18)
+ ;;
+ elif yb == 0x8000000000000000
+ if xpos
+ -> std.flt64frombits(0x8000000000000000)
+ else
+ -> std.flt64frombits(0xc00921fb54442d18)
+ ;;
+ elif xb == 0x0000000000000000 || xb == 0x8000000000000000
+ if ypos
-> std.flt64frombits(0x3ff921fb54442d18)
else
-> std.flt64frombits(0xbff921fb54442d18)
;;
;;
+ var xinf = (xe == 1024)
+ var yinf = (ye == 1024)
+
+ match (yinf, ypos, xinf, xpos)
+ | (true, true, true, false): -> std.flt64frombits(0x4002d97c7f3321d2)
+ | (true, false, true, false): -> std.flt64frombits(0xc002d97c7f3321d2)
+ | (true, true, true, true ): -> std.flt64frombits(0x3fe921fb54442d18)
+ | (true, false, true, true ): -> std.flt64frombits(0xbfe921fb54442d18)
+ | (true, true, false, _ ): -> std.flt64frombits(0x3ff921fb54442d18)
+ | (true, false, false, _ ): -> std.flt64frombits(0xbff921fb54442d18)
+ | (false, true, true, false): -> std.flt64frombits(0x400921fb54442d18)
+ | (false, false, true, false): -> std.flt64frombits(0xc00921fb54442d18)
+ | (false, true, true, true ): -> std.flt64frombits(0x0000000000000000)
+ | (false, false, true, true ): -> std.flt64frombits(0x8000000000000000)
+ | _:
+ ;;
+
+ /* Normal case. Here we just reduce y/x to [0, 1] */
+ var xabs = std.flt64frombits(std.flt64bits(x) & (~(0 : uint64)) >> 1)
+ var yabs = std.flt64frombits(std.flt64bits(y) & (~(0 : uint64)) >> 1)
+
+ var then_negate = !ypos
+ var then_subtract_from_pi = !xpos
+
+ x = xabs
+ y = yabs
+
+
var then_reverse = false
- if yabs > xabs
+ if y > x
then_reverse = true
std.swap(&x, &y)
;;
- /* Compute y/x = u + du; du is one Newton's iteration of division */
+ /* Compute y/x = u + du */
var u = y / x
- var t1, t2
- (t1, t2) = two_by_two(u, x)
- var du = y * ((y - t1) - t2)/x
+ var du
+ if u * x == y
+ du = 0.0
+ else
+ var t1, t2
+ (t1, t2) = two_by_two(u, x)
+ du = ((y - t1) - t2)/x
+ ;;
var ret = atan_reduced(u, du)
@@ -374,6 +439,12 @@ const atan264 = {y, x
var po2_lo = std.flt64frombits(pi_over_2[1])
ret = (po2_hi - ret) + po2_lo
;;
+ if then_subtract_from_pi
+ /* Compute pi - ret, or maybe -pi - ret */
+ var pi_hi = scale2(std.flt64frombits(pi_over_2[0]), 1)
+ var pi_lo = scale2(std.flt64frombits(pi_over_2[1]), 1)
+ ret = (pi_hi - ret) + pi_lo
+ ;;
if then_negate
ret = -1.0 * ret
diff --git a/lib/math/test/atan-impl.myr b/lib/math/test/atan-impl.myr
index c31d650..1c78639 100644
--- a/lib/math/test/atan-impl.myr
+++ b/lib/math/test/atan-impl.myr
@@ -47,6 +47,10 @@ const atan01 = {c
(0x4c120000, 0x3fc90fda),
(0x0c010000, 0x0c010000),
(0xc0070000, 0xbf9065b4),
+ (0x3d8d6b23, 0x3d8d31c3),
+ (0xbd8d6b23, 0xbd8d31c3),
+ (0xbf000000, 0xbeed6338),
+ (0xc1010000, 0xbfb94442),
][:]
for (x, y) : inputs
@@ -67,18 +71,92 @@ const atan01 = {c
}
const atan02 = {c
+ testr.fail(c, "WRITE THIS")
}
const atan03 = {c
+ var inputs : (uint32, uint32, uint32)[:] = [
+ (0x00000000, 0x80000000, 0x40490fdb), /* atan2(+0, -0) = +Pi */
+ (0x80000000, 0x80000000, 0xc0490fdb), /* atan2(-0, -0) = -Pi */
+ (0x00000000, 0x00000000, 0x00000000), /* atan2(+0, +0) = +0 */
+ (0x80000000, 0x00000000, 0x80000000), /* atan2(-0, +0) = -0 */
+ (0x00000000, 0xc5a33ab8, 0x40490fdb), /* atan2(+0, x < 0) = +Pi */
+ (0x00000000, 0x80000002, 0x40490fdb),
+ (0x00000000, 0xdddddddd, 0x40490fdb),
+ (0x80000000, 0xc5a33ab8, 0xc0490fdb), /* atan2(-0, x < 0) = -Pi */
+ (0x80000000, 0x80000002, 0xc0490fdb),
+ (0x80000000, 0xdddddddd, 0xc0490fdb),
+ (0x00000000, 0x35a33ab8, 0x00000000), /* atan2(+0, x > 0) = +0 */
+ (0x00000000, 0x00000002, 0x00000000),
+ (0x00000000, 0x4ddddddd, 0x00000000),
+ (0x80000000, 0x35a33ab8, 0x80000000), /* atan2(-0, x > 0) = -0 */
+ (0x80000000, 0x00000002, 0x80000000),
+ (0x80000000, 0x4ddddddd, 0x80000000),
+ (0xdddddddd, 0x00000000, 0xbfc90fdb), /* atan2(y < 0, 0) = -Pi/2 */
+ (0xc5a33ab8, 0x00000000, 0xbfc90fdb),
+ (0x80000002, 0x00000000, 0xbfc90fdb),
+ (0x4ddddddd, 0x00000000, 0x3fc90fdb), /* atan2(y > 0, 0) = +Pi/2 */
+ (0x35a33ab8, 0x00000000, 0x3fc90fdb),
+ (0x00000002, 0x00000000, 0x3fc90fdb),
+ (0x7f800000, 0xff800000, 0x4016cbe4), /* atan2(+Inf, -Inf) = +3*Pi/4 */
+ (0xff800000, 0xff800000, 0xc016cbe4), /* atan2(-Inf, -Inf) = -3*Pi/4 */
+ (0x7f800000, 0x7f800000, 0x3f490fdb), /* atan2(+Inf, +Inf) = +Pi/4 */
+ (0xff800000, 0x7f800000, 0xbf490fdb), /* atan2(-Inf, +Inf) = -Pi/4 */
+ (0x7f800000, 0x4ddddddd, 0x3fc90fdb), /* atan2(+Inf, finite) = +Pi/2 */
+ (0x7f800000, 0x00000001, 0x3fc90fdb),
+ (0x7f800000, 0x80000004, 0x3fc90fdb),
+ (0x7f800000, 0xfedcba87, 0x3fc90fdb),
+ (0xff800000, 0x4ddddddd, 0xbfc90fdb), /* atan2(-Inf, finite) = -Pi/2 */
+ (0xff800000, 0x00000001, 0xbfc90fdb),
+ (0xff800000, 0x80000004, 0xbfc90fdb),
+ (0xff800000, 0xfedcba87, 0xbfc90fdb),
+ (0x6a520b4c, 0xff800000, 0x40490fdb), /* atan2(finite > 0, -Inf) = +Pi */
+ (0x35a33ab8, 0xff800000, 0x40490fdb),
+ (0x55a33ab8, 0xff800000, 0x40490fdb),
+ (0xea520b4c, 0xff800000, 0xc0490fdb), /* atan2(finite < 0, -Inf) = -Pi */
+ (0x95a33ab8, 0xff800000, 0xc0490fdb),
+ (0xc5a33ab8, 0xff800000, 0xc0490fdb),
+ (0x6a520b4c, 0x7f800000, 0x00000000), /* atan2(finite > 0, +Inf) = +0 */
+ (0x35a33ab8, 0x7f800000, 0x00000000),
+ (0x55a33ab8, 0x7f800000, 0x00000000),
+ (0xea520b4c, 0x7f800000, 0x80000000), /* atan2(finite < 0, +Inf) = -0 */
+ (0x95a33ab8, 0x7f800000, 0x80000000),
+ (0xc5a33ab8, 0x7f800000, 0x80000000),
+ (0x1aae129e, 0xde263fa8, 0x40490fdb), /* misc */
+ (0xb76e98b6, 0xdbeb6637, 0xc0490fdb),
+ (0x7112fd5b, 0x7509b252, 0x3b88a34d),
+ (0xe53215fe, 0xcd0f08fc, 0xbfc90fdb),
+ (0xcd47c963, 0x85268f36, 0xbfc90fdb),
+ (0xfacd1adc, 0x79fd5d79, 0xbfa2b8c8),
+ (0xfa3f79f2, 0xf5f06269, 0xbfc96033),
+ (0xddc7b749, 0x5f3d9db0, 0xbe060c09),
+ (0x63c8ee47, 0x792ac38f, 0x2a169cbe),
+ (0xe3c24a4f, 0xe0f9b02f, 0xbfcba1c1),
+ (0xe1f9385d, 0xe317764d, 0xc03c145d)
+ ][:]
+
+ for (y, x, z_exp) : inputs
+ var xf : flt32 = std.flt32frombits(x)
+ var yf : flt32 = std.flt32frombits(y)
+ var zf_act : flt32 = math.atan2(yf, xf)
+ var z_act : uint32 = std.flt32bits(zf_act)
+
+ testr.check(c, same32(z_act, z_exp),
+ "atan(0x{b=16,w=8,p=0}, 0x{b=16,w=8,p=0}) = 0x{b=16,w=8,p=0}, should be 0x{b=16,w=8,p=0}",
+ y, x, z_act, z_exp)
+ ;;
}
const atan04 = {c
+ testr.fail(c, "WRITE THIS")
}
const atan05 = {c
+ testr.fail(c, "WRITE THIS")
}
const atan06 = {c
+ testr.fail(c, "WRITE THIS")
}
const atan07 = {c