Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]
Groups > linux.kernel > #1738466
| From | Greg Kroah-Hartman <gregkh@linuxfoundation.org> |
|---|---|
| Newsgroups | linux.kernel |
| Subject | [PATCH 4.9 24/77] MIPS: math-emu: <MADDF|MSUBF>.S: Fix accuracy (32-bit case) |
| Date | 2017-09-24 23:10 +0200 |
| Message-ID | <utmk1-5jg-17@gated-at.bofh.it> (permalink) |
| References | <utlQZ-4Ti-7@gated-at.bofh.it> |
| Organization | linux.* mail to news gateway |
4.9-stable review patch. If anyone has any objections, please let me know.
------------------
From: Douglas Leung <douglas.leung@imgtec.com>
commit b3b8e1eb27c523e32b6a8aa7ec8ac4754456af57 upstream.
Implement fused multiply-add with correct accuracy.
Fused multiply-add operation has better accuracy than respective
sequential execution of multiply and add operations applied on the
same inputs. This is because accuracy errors accumulate in latter
case.
This patch implements fused multiply-add with the same accuracy
as it is implemented in hardware, using 64-bit intermediate
calculations.
One test case example (raw bits) that this patch fixes:
MADDF.S fd,fs,ft:
fd = 0x22575225
fs = ft = 0x3727c5ac
Fixes: e24c3bec3e8e ("MIPS: math-emu: Add support for the MIPS R6 MADDF FPU instruction")
Fixes: 83d43305a1df ("MIPS: math-emu: Add support for the MIPS R6 MSUBF FPU instruction")
Signed-off-by: Douglas Leung <douglas.leung@imgtec.com>
Signed-off-by: Miodrag Dinic <miodrag.dinic@imgtec.com>
Signed-off-by: Goran Ferenc <goran.ferenc@imgtec.com>
Signed-off-by: Aleksandar Markovic <aleksandar.markovic@imgtec.com>
Cc: Douglas Leung <douglas.leung@imgtec.com>
Cc: Bo Hu <bohu@google.com>
Cc: James Hogan <james.hogan@imgtec.com>
Cc: Jin Qian <jinqian@google.com>
Cc: Paul Burton <paul.burton@imgtec.com>
Cc: Petar Jovanovic <petar.jovanovic@imgtec.com>
Cc: Raghu Gandham <raghu.gandham@imgtec.com>
Cc: linux-mips@linux-mips.org
Cc: linux-kernel@vger.kernel.org
Patchwork: https://patchwork.linux-mips.org/patch/16890/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
---
arch/mips/math-emu/ieee754sp.h | 4 +
arch/mips/math-emu/sp_maddf.c | 116 ++++++++++++++++-------------------------
2 files changed, 50 insertions(+), 70 deletions(-)
--- a/arch/mips/math-emu/ieee754sp.h
+++ b/arch/mips/math-emu/ieee754sp.h
@@ -45,6 +45,10 @@ static inline int ieee754sp_finite(union
return SPBEXP(x) != SP_EMAX + 1 + SP_EBIAS;
}
+/* 64 bit right shift with rounding */
+#define XSPSRS64(v, rs) \
+ (((rs) >= 64) ? ((v) != 0) : ((v) >> (rs)) | ((v) << (64-(rs)) != 0))
+
/* 3bit extended single precision sticky right shift */
#define XSPSRS(v, rs) \
((rs > (SP_FBITS+3))?1:((v) >> (rs)) | ((v) << (32-(rs)) != 0))
--- a/arch/mips/math-emu/sp_maddf.c
+++ b/arch/mips/math-emu/sp_maddf.c
@@ -21,14 +21,8 @@ static union ieee754sp _sp_maddf(union i
int re;
int rs;
unsigned rm;
- unsigned short lxm;
- unsigned short hxm;
- unsigned short lym;
- unsigned short hym;
- unsigned lrm;
- unsigned hrm;
- unsigned t;
- unsigned at;
+ uint64_t rm64;
+ uint64_t zm64;
int s;
COMPXSP;
@@ -170,108 +164,90 @@ static union ieee754sp _sp_maddf(union i
if (flags & MADDF_NEGATE_PRODUCT)
rs ^= 1;
- /* shunt to top of word */
- xm <<= 32 - (SP_FBITS + 1);
- ym <<= 32 - (SP_FBITS + 1);
+ /* Multiple 24 bit xm and ym to give 48 bit results */
+ rm64 = (uint64_t)xm * ym;
- /*
- * Multiply 32 bits xm, ym to give high 32 bits rm with stickness.
- */
- lxm = xm & 0xffff;
- hxm = xm >> 16;
- lym = ym & 0xffff;
- hym = ym >> 16;
-
- lrm = lxm * lym; /* 16 * 16 => 32 */
- hrm = hxm * hym; /* 16 * 16 => 32 */
-
- t = lxm * hym; /* 16 * 16 => 32 */
- at = lrm + (t << 16);
- hrm += at < lrm;
- lrm = at;
- hrm = hrm + (t >> 16);
-
- t = hxm * lym; /* 16 * 16 => 32 */
- at = lrm + (t << 16);
- hrm += at < lrm;
- lrm = at;
- hrm = hrm + (t >> 16);
-
- rm = hrm | (lrm != 0);
+ /* Shunt to top of word */
+ rm64 = rm64 << 16;
- /*
- * Sticky shift down to normal rounding precision.
- */
- if ((int) rm < 0) {
- rm = (rm >> (32 - (SP_FBITS + 1 + 3))) |
- ((rm << (SP_FBITS + 1 + 3)) != 0);
+ /* Put explicit bit at bit 62 if necessary */
+ if ((int64_t) rm64 < 0) {
+ rm64 = rm64 >> 1;
re++;
- } else {
- rm = (rm >> (32 - (SP_FBITS + 1 + 3 + 1))) |
- ((rm << (SP_FBITS + 1 + 3 + 1)) != 0);
}
- assert(rm & (SP_HIDDEN_BIT << 3));
-
- if (zc == IEEE754_CLASS_ZERO)
- return ieee754sp_format(rs, re, rm);
- /* And now the addition */
+ assert(rm64 & (1 << 62));
- assert(zm & SP_HIDDEN_BIT);
+ if (zc == IEEE754_CLASS_ZERO) {
+ /*
+ * Move explicit bit from bit 62 to bit 26 since the
+ * ieee754sp_format code expects the mantissa to be
+ * 27 bits wide (24 + 3 rounding bits).
+ */
+ rm = XSPSRS64(rm64, (62 - 26));
+ return ieee754sp_format(rs, re, rm);
+ }
- /*
- * Provide guard,round and stick bit space.
- */
- zm <<= 3;
+ /* Move explicit bit from bit 23 to bit 62 */
+ zm64 = (uint64_t)zm << (62 - 23);
+ assert(zm64 & (1 << 62));
+ /* Make the exponents the same */
if (ze > re) {
/*
* Have to shift r fraction right to align.
*/
s = ze - re;
- rm = XSPSRS(rm, s);
+ rm64 = XSPSRS64(rm64, s);
re += s;
} else if (re > ze) {
/*
* Have to shift z fraction right to align.
*/
s = re - ze;
- zm = XSPSRS(zm, s);
+ zm64 = XSPSRS64(zm64, s);
ze += s;
}
assert(ze == re);
assert(ze <= SP_EMAX);
+ /* Do the addition */
if (zs == rs) {
/*
- * Generate 28 bit result of adding two 27 bit numbers
- * leaving result in zm, zs and ze.
+ * Generate 64 bit result by adding two 63 bit numbers
+ * leaving result in zm64, zs and ze.
*/
- zm = zm + rm;
-
- if (zm >> (SP_FBITS + 1 + 3)) { /* carry out */
- zm = XSPSRS1(zm);
+ zm64 = zm64 + rm64;
+ if ((int64_t)zm64 < 0) { /* carry out */
+ zm64 = XSPSRS1(zm64);
ze++;
}
} else {
- if (zm >= rm) {
- zm = zm - rm;
+ if (zm64 >= rm64) {
+ zm64 = zm64 - rm64;
} else {
- zm = rm - zm;
+ zm64 = rm64 - zm64;
zs = rs;
}
- if (zm == 0)
+ if (zm64 == 0)
return ieee754sp_zero(ieee754_csr.rm == FPU_CSR_RD);
/*
- * Normalize in extended single precision
+ * Put explicit bit at bit 62 if necessary.
*/
- while ((zm >> (SP_MBITS + 3)) == 0) {
- zm <<= 1;
+ while ((zm64 >> 62) == 0) {
+ zm64 <<= 1;
ze--;
}
-
}
+
+ /*
+ * Move explicit bit from bit 62 to bit 26 since the
+ * ieee754sp_format code expects the mantissa to be
+ * 27 bits wide (24 + 3 rounding bits).
+ */
+ zm = XSPSRS64(zm64, (62 - 26));
+
return ieee754sp_format(zs, ze, zm);
}
Back to linux.kernel | Previous | Next — Previous in thread | Next in thread | Find similar | Unroll thread
[PATCH 4.9 00/77] 4.9.52-stable review Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 22:40 +0200
[PATCH 4.9 60/77] PCI: pciehp: Report power fault only once until we clear it Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 22:40 +0200
[PATCH 4.9 30/77] powerpc: Fix DAR reporting when alignment handler faults Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 22:40 +0200
[PATCH 4.9 55/77] ftrace: Fix memleak when unregistering dynamic ops when tracing disabled Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 22:40 +0200
[PATCH 4.9 42/77] scsi: zfcp: trace high part of "new" 64 bit SCSI LUN Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 22:40 +0200
[PATCH 4.9 54/77] ftrace: Fix selftest goto location on error Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 22:40 +0200
[PATCH 4.9 75/77] mac80211_hwsim: Use proper TX power Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 22:40 +0200
[PATCH 4.9 21/77] MIPS: math-emu: <MADDF|MSUBF>.<D|S>: Fix some cases of infinite inputs Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 22:40 +0200
[PATCH 4.9 27/77] [PATCH - RESEND] crypto: AF_ALG - remove SGL terminator indicator when chaining Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 22:40 +0200
[PATCH 4.9 04/77] orangefs: Dont clear SGID when inheriting ACLs Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 22:40 +0200
[PATCH 4.9 06/77] drm/sun4i: Implement drm_driver lastclose to restore fbdev console Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 22:40 +0200
[PATCH 4.9 44/77] scsi: megaraid_sas: Check valid aen class range to avoid kernel panic Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 22:40 +0200
[PATCH 4.9 63/77] s390/mm: fix race on mm->context.flush_mm Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 22:50 +0200
[PATCH 4.9 65/77] media: uvcvideo: Prevent heap overflow when accessing mapped controls Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 22:50 +0200
[PATCH 4.9 67/77] bcache: initialize dirty stripes in flash_dev_run() Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 22:50 +0200
[PATCH 4.9 61/77] net/netfilter/nf_conntrack_core: Fix net_conntrack_lock() Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 22:50 +0200
[PATCH 4.9 56/77] tracing: Add barrier to trace_printk() buffer nesting modification Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:10 +0200
[PATCH 4.9 68/77] bcache: Fix leak of bdev reference Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:10 +0200
[PATCH 4.9 24/77] MIPS: math-emu: <MADDF|MSUBF>.S: Fix accuracy (32-bit case) Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:10 +0200
[PATCH 4.9 74/77] mac80211: fix VLAN handling with TXQs Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:10 +0200
[PATCH 4.9 59/77] PCI: shpchp: Enable bridge bus mastering if MSI is enabled Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:10 +0200
[PATCH 4.9 58/77] ARC: Re-enable MMU upon Machine Check exception Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:10 +0200
[PATCH 4.9 26/77] crypto: ccp - Fix XTS-AES-128 support on v5 CCPs Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:10 +0200
[PATCH 4.9 66/77] PM / devfreq: Fix memory leak when fail to register device Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:10 +0200
[PATCH 4.9 64/77] media: v4l2-compat-ioctl32: Fix timespec conversion Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:10 +0200
[PATCH 4.9 69/77] bcache: do not subtract sectors_to_gc for bypassed IO Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:10 +0200
[PATCH 4.9 73/77] bcache: fix bch_hprint crash and improve output Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:10 +0200
[PATCH 4.9 62/77] s390/mm: fix local TLB flushing vs. detach of an mm address space Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:10 +0200
[PATCH 4.9 57/77] tracing: Apply trace_clock changes to instance max buffer Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:10 +0200
[PATCH 4.9 18/77] MIPS: math-emu: MINA.<D|S>: Fix some cases of infinity and zero inputs Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:10 +0200
[PATCH 4.9 76/77] mac80211: flush hw_roc_start work before cancelling the ROC Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:10 +0200
[PATCH 4.9 39/77] scsi: zfcp: fix missing trace records for early returns in TMF eh handlers Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 40/77] scsi: zfcp: fix payload with full FCP_RSP IU in SCSI trace records Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 41/77] scsi: zfcp: trace HBA FSF response by default on dismiss or timedout late response Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 34/77] skd: Submit requests to firmware before triggering the doorbell Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 33/77] skd: Avoid that module unloading triggers a use-after-free Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 50/77] scsi: sg: factor out sg_fill_request_table() Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 11/77] pinctrl/amd: save pin registers over suspend/resume Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 48/77] scsi: sg: use standard lists for sg_requests Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 03/77] mm: prevent double decrease of nr_reserved_highatomic Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 35/77] scsi: zfcp: fix queuecommand for scsi_eh commands when DIX enabled Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 47/77] scsi: sg: remove save_scat_len Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 17/77] MIPS: math-emu: <MAXA|MINA>.<D|S>: Fix cases of both infinite inputs Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 45/77] scsi: megaraid_sas: Return pended IOCTLs with cmd_status MFI_STAT_WRONG_STATE in case adapter is dead Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 31/77] block: Relax a check in blk_start_queue() Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 53/77] scsi: qla2xxx: Fix an integer overflow in sysfs code Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 10/77] tty: fix __tty_insert_flip_char regression Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 29/77] ext4: fix quota inconsistency during orphan cleanup for read-only mounts Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 43/77] scsi: megaraid_sas: set minimum value of resetwaittime to be 1 secs Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 12/77] Input: i8042 - add Gigabyte P57 to the keyboard reset table Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 16/77] MIPS: math-emu: <MAXA|MINA>.<D|S>: Fix cases of input values with opposite signs Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 15/77] MIPS: math-emu: <MAX|MIN>.<D|S>: Fix cases of both inputs negative Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 07/77] IB/addr: Fix setting source address in addr6_resolve() Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 09/77] tty: improve tty_insert_flip_char() slow path Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 19/77] MIPS: math-emu: Handle zero accumulator case in MADDF and MSUBF separately Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 51/77] scsi: sg: fixup infoleak when using SG_GET_REQUEST_TABLE Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 28/77] ext4: fix incorrect quotaoff if the quota feature is enabled Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 05/77] IB/{qib, hfi1}: Avoid flow control testing for RDMA write operation Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
[PATCH 4.9 20/77] MIPS: math-emu: <MADDF|MSUBF>.<D|S>: Fix NaN propagation Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-24 23:20 +0200
Re: [PATCH 4.9 00/77] 4.9.52-stable review Guenter Roeck <linux@roeck-us.net> - 2017-09-25 03:10 +0200
Re: [PATCH 4.9 00/77] 4.9.52-stable review Tom Gall <tom.gall@linaro.org> - 2017-09-25 06:40 +0200
Re: [PATCH 4.9 00/77] 4.9.52-stable review Greg Kroah-Hartman <gregkh@linuxfoundation.org> - 2017-09-25 08:30 +0200
Re: [PATCH 4.9 00/77] 4.9.52-stable review Shuah Khan <shuahkh@osg.samsung.com> - 2017-09-26 01:20 +0200
csiph-web