qemu-devel.nongnu.org archive mirror
 help / color / mirror / Atom feed
* [Qemu-devel] [PATCH] target-ppc: Add quad precision muladd instructions
@ 2017-02-15  6:37 Bharata B Rao
  2017-02-15 22:13 ` Richard Henderson
  0 siblings, 1 reply; 6+ messages in thread
From: Bharata B Rao @ 2017-02-15  6:37 UTC (permalink / raw)
  To: qemu-devel; +Cc: qemu-ppc, david, rth, nikunj, Bharata B Rao

xsmaddqp:   VSX Scalar Multiply-Add Quad-Precision
xsmaddqpo:  VSX Scalar Multiply-Add Quad-Precision using round to Odd
xsnmaddqp:  VSX Scalar Negative Multiply-Add Quad-Precision
xsnmaddqpo: VSX Scalar Negative Multiply-Add Quad-Precision using round to Odd

xsmsubqp:   VSX Scalar Multiply-Subtract Quad-Precision
xsmsubqpo:  VSX Scalar Multiply-Subtract Quad-Precision using round to Odd
xsnmsubqp:  VSX Scalar Negative Multiply-Subtract Quad-Precision
xsnmsubqpo: VSX Scalar Negative Multiply-Subtract Quad-Precision
            using round to Odd

Signed-off-by: Bharata B Rao <bharata@linux.vnet.ibm.com>
---
 target/ppc/fpu_helper.c             | 69 +++++++++++++++++++++++++++++++++++++
 target/ppc/helper.h                 |  4 +++
 target/ppc/translate/vsx-impl.inc.c |  4 +++
 target/ppc/translate/vsx-ops.inc.c  |  4 +++
 4 files changed, 81 insertions(+)

diff --git a/target/ppc/fpu_helper.c b/target/ppc/fpu_helper.c
index 58aee64..201cafd 100644
--- a/target/ppc/fpu_helper.c
+++ b/target/ppc/fpu_helper.c
@@ -2425,6 +2425,75 @@ VSX_MADD(xvnmaddmsp, 4, float32, VsrW(i), NMADD_FLGS, 0, 0, 0)
 VSX_MADD(xvnmsubasp, 4, float32, VsrW(i), NMSUB_FLGS, 1, 0, 0)
 VSX_MADD(xvnmsubmsp, 4, float32, VsrW(i), NMSUB_FLGS, 0, 0, 0)
 
+/*
+ * Quadruple-precision version of multiply and add/subtract.
+ *
+ * This implementation is not 100% accurate as we truncate the
+ * intermediate result of multiplication and then add/subtract
+ * separately.
+ *
+ * TODO: When float128_muladd() becomes available, switch this
+ * implementation to use that instead of separate float128_mul()
+ * followed by float128_add().
+ */
+#define VSX_MADD_QP(op, maddflgs)                                             \
+void helper_##op(CPUPPCState *env, uint32_t opcode)                           \
+{                                                                             \
+    ppc_vsr_t xt_in, xa, xb, xt_out;                                          \
+                                                                              \
+    getVSR(rA(opcode) + 32, &xa, env);                                        \
+    getVSR(rB(opcode) + 32, &xb, env);                                        \
+    getVSR(rD(opcode) + 32, &xt_in, env);                                     \
+                                                                              \
+    xt_out = xt_in;                                                           \
+    helper_reset_fpstatus(env);                                               \
+    float_status tstat = env->fp_status;                                      \
+    if (unlikely(Rc(opcode) != 0)) {                                          \
+        tstat.float_rounding_mode = float_round_to_odd;                       \
+    }                                                                         \
+    set_float_exception_flags(0, &tstat);                                     \
+    xt_out.f128 = float128_mul(xa.f128, xt_in.f128, &tstat);                  \
+                                                                              \
+    if (maddflgs & float_muladd_negate_c) {                                   \
+        xb.VsrD(0) ^= 0x8000000000000000;                                     \
+    }                                                                         \
+    xt_out.f128 = float128_add(xt_out.f128, xb.f128, &tstat);                 \
+    env->fp_status.float_exception_flags |= tstat.float_exception_flags;      \
+                                                                              \
+    if (unlikely(tstat.float_exception_flags & float_flag_invalid)) {         \
+        if (float128_is_signaling_nan(xa.f128, &tstat) ||                     \
+            float128_is_signaling_nan(xt_in.f128, &tstat) ||                  \
+            float128_is_signaling_nan(xb.f128, &tstat)) {                     \
+            float_invalid_op_excp(env, POWERPC_EXCP_FP_VXSNAN, 1);            \
+            tstat.float_exception_flags &= ~float_flag_invalid;               \
+        }                                                                     \
+        if ((float128_is_infinity(xa.f128) && float128_is_zero(xt_in.f128)) ||\
+            (float128_is_zero(xa.f128) && float128_is_infinity(xt_in.f128))) {\
+            float_invalid_op_excp(env, POWERPC_EXCP_FP_VXIMZ, 1);             \
+            tstat.float_exception_flags &= ~float_flag_invalid;               \
+        }                                                                     \
+        if ((tstat.float_exception_flags & float_flag_invalid) &&             \
+            ((float128_is_infinity(xa.f128) ||                                \
+            float128_is_infinity(xt_in.f128)) &&                              \
+            float128_is_infinity(xb.f128))) {                                 \
+            float_invalid_op_excp(env, POWERPC_EXCP_FP_VXISI, 1);             \
+        }                                                                     \
+    }                                                                         \
+                                                                              \
+    helper_compute_fprf_float128(env, xt_out.f128);                           \
+    if ((maddflgs & float_muladd_negate_result) &&                            \
+        !float128_is_any_nan(xt_out.f128)) {                                  \
+        xt_out.VsrD(0) ^= 0x8000000000000000;                                 \
+    }                                                                         \
+    putVSR(rD(opcode) + 32, &xt_out, env);                                    \
+    float_check_status(env);                                                  \
+}
+
+VSX_MADD_QP(xsmaddqp, MADD_FLGS)
+VSX_MADD_QP(xsmsubqp, MSUB_FLGS)
+VSX_MADD_QP(xsnmaddqp, NMADD_FLGS)
+VSX_MADD_QP(xsnmsubqp, NMSUB_FLGS)
+
 /* VSX_SCALAR_CMP_DP - VSX scalar floating point compare double precision
  *   op    - instruction mnemonic
  *   cmp   - comparison operation
diff --git a/target/ppc/helper.h b/target/ppc/helper.h
index 6d77661..eade946 100644
--- a/target/ppc/helper.h
+++ b/target/ppc/helper.h
@@ -480,12 +480,16 @@ DEF_HELPER_2(xssqrtsp, void, env, i32)
 DEF_HELPER_2(xsrsqrtesp, void, env, i32)
 DEF_HELPER_2(xsmaddasp, void, env, i32)
 DEF_HELPER_2(xsmaddmsp, void, env, i32)
+DEF_HELPER_2(xsmaddqp, void, env, i32)
 DEF_HELPER_2(xsmsubasp, void, env, i32)
 DEF_HELPER_2(xsmsubmsp, void, env, i32)
+DEF_HELPER_2(xsmsubqp, void, env, i32)
 DEF_HELPER_2(xsnmaddasp, void, env, i32)
 DEF_HELPER_2(xsnmaddmsp, void, env, i32)
+DEF_HELPER_2(xsnmaddqp, void, env, i32)
 DEF_HELPER_2(xsnmsubasp, void, env, i32)
 DEF_HELPER_2(xsnmsubmsp, void, env, i32)
+DEF_HELPER_2(xsnmsubqp, void, env, i32)
 
 DEF_HELPER_2(xvadddp, void, env, i32)
 DEF_HELPER_2(xvsubdp, void, env, i32)
diff --git a/target/ppc/translate/vsx-impl.inc.c b/target/ppc/translate/vsx-impl.inc.c
index 7f12908..0a96e6b 100644
--- a/target/ppc/translate/vsx-impl.inc.c
+++ b/target/ppc/translate/vsx-impl.inc.c
@@ -853,12 +853,16 @@ GEN_VSX_HELPER_2(xssqrtsp, 0x16, 0x00, 0, PPC2_VSX207)
 GEN_VSX_HELPER_2(xsrsqrtesp, 0x14, 0x00, 0, PPC2_VSX207)
 GEN_VSX_HELPER_2(xsmaddasp, 0x04, 0x00, 0, PPC2_VSX207)
 GEN_VSX_HELPER_2(xsmaddmsp, 0x04, 0x01, 0, PPC2_VSX207)
+GEN_VSX_HELPER_2(xsmaddqp, 0x04, 0x0C, 0, PPC2_ISA300)
 GEN_VSX_HELPER_2(xsmsubasp, 0x04, 0x02, 0, PPC2_VSX207)
 GEN_VSX_HELPER_2(xsmsubmsp, 0x04, 0x03, 0, PPC2_VSX207)
+GEN_VSX_HELPER_2(xsmsubqp, 0x04, 0x0D, 0, PPC2_ISA300)
 GEN_VSX_HELPER_2(xsnmaddasp, 0x04, 0x10, 0, PPC2_VSX207)
 GEN_VSX_HELPER_2(xsnmaddmsp, 0x04, 0x11, 0, PPC2_VSX207)
+GEN_VSX_HELPER_2(xsnmaddqp, 0x04, 0x0E, 0, PPC2_ISA300)
 GEN_VSX_HELPER_2(xsnmsubasp, 0x04, 0x12, 0, PPC2_VSX207)
 GEN_VSX_HELPER_2(xsnmsubmsp, 0x04, 0x13, 0, PPC2_VSX207)
+GEN_VSX_HELPER_2(xsnmsubqp, 0x04, 0x0F, 0, PPC2_ISA300)
 GEN_VSX_HELPER_2(xscvsxdsp, 0x10, 0x13, 0, PPC2_VSX207)
 GEN_VSX_HELPER_2(xscvuxdsp, 0x10, 0x12, 0, PPC2_VSX207)
 GEN_VSX_HELPER_2(xststdcsp, 0x14, 0x12, 0, PPC2_ISA300)
diff --git a/target/ppc/translate/vsx-ops.inc.c b/target/ppc/translate/vsx-ops.inc.c
index 5030c4a..e770fab 100644
--- a/target/ppc/translate/vsx-ops.inc.c
+++ b/target/ppc/translate/vsx-ops.inc.c
@@ -237,12 +237,16 @@ GEN_XX2FORM(xssqrtsp,  0x16, 0x00, PPC2_VSX207),
 GEN_XX2FORM(xsrsqrtesp,  0x14, 0x00, PPC2_VSX207),
 GEN_XX3FORM(xsmaddasp, 0x04, 0x00, PPC2_VSX207),
 GEN_XX3FORM(xsmaddmsp, 0x04, 0x01, PPC2_VSX207),
+GEN_VSX_XFORM_300(xsmaddqp, 0x04, 0x0C, 0x0),
 GEN_XX3FORM(xsmsubasp, 0x04, 0x02, PPC2_VSX207),
 GEN_XX3FORM(xsmsubmsp, 0x04, 0x03, PPC2_VSX207),
+GEN_VSX_XFORM_300(xsmsubqp, 0x04, 0x0D, 0x0),
 GEN_XX3FORM(xsnmaddasp, 0x04, 0x10, PPC2_VSX207),
 GEN_XX3FORM(xsnmaddmsp, 0x04, 0x11, PPC2_VSX207),
+GEN_VSX_XFORM_300(xsnmaddqp, 0x04, 0x0E, 0x0),
 GEN_XX3FORM(xsnmsubasp, 0x04, 0x12, PPC2_VSX207),
 GEN_XX3FORM(xsnmsubmsp, 0x04, 0x13, PPC2_VSX207),
+GEN_VSX_XFORM_300(xsnmsubqp, 0x04, 0x0F, 0x0),
 GEN_XX2FORM(xscvsxdsp, 0x10, 0x13, PPC2_VSX207),
 GEN_XX2FORM(xscvuxdsp, 0x10, 0x12, PPC2_VSX207),
 
-- 
2.7.4

^ permalink raw reply related	[flat|nested] 6+ messages in thread

* Re: [Qemu-devel] [PATCH] target-ppc: Add quad precision muladd instructions
  2017-02-15  6:37 [Qemu-devel] [PATCH] target-ppc: Add quad precision muladd instructions Bharata B Rao
@ 2017-02-15 22:13 ` Richard Henderson
  2017-02-16  2:41   ` Bharata B Rao
  0 siblings, 1 reply; 6+ messages in thread
From: Richard Henderson @ 2017-02-15 22:13 UTC (permalink / raw)
  To: Bharata B Rao, qemu-devel; +Cc: qemu-ppc, nikunj, david

On 02/15/2017 05:37 PM, Bharata B Rao wrote:
> + *
> + * TODO: When float128_muladd() becomes available, switch this
> + * implementation to use that instead of separate float128_mul()
> + * followed by float128_add().

Let's just do that, rather than add something that can't pass tests.

You should be able to copy float64_muladd and, for the most part, s/128/256/ 
and s/64/128/.  Other of the magic numbers, like the implicit bit and the 
exponent bias, you get from float128_mul.


r~

^ permalink raw reply	[flat|nested] 6+ messages in thread

* Re: [Qemu-devel] [PATCH] target-ppc: Add quad precision muladd instructions
  2017-02-15 22:13 ` Richard Henderson
@ 2017-02-16  2:41   ` Bharata B Rao
  2017-02-16 20:53     ` Richard Henderson
  0 siblings, 1 reply; 6+ messages in thread
From: Bharata B Rao @ 2017-02-16  2:41 UTC (permalink / raw)
  To: Richard Henderson; +Cc: qemu-devel, qemu-ppc, nikunj, david

On Thu, Feb 16, 2017 at 09:13:31AM +1100, Richard Henderson wrote:
> On 02/15/2017 05:37 PM, Bharata B Rao wrote:
> > + *
> > + * TODO: When float128_muladd() becomes available, switch this
> > + * implementation to use that instead of separate float128_mul()
> > + * followed by float128_add().
> 
> Let's just do that, rather than add something that can't pass tests.
> 
> You should be able to copy float64_muladd and, for the most part, s/128/256/
> and s/64/128/.  Other of the magic numbers, like the implicit bit and the
> exponent bias, you get from float128_mul.

I started like that but got lost somewhere down that path...

It needs at least the following new functions to be implemented:

propagateFloat128MulAddNaN
shortShift256Left
shift256RightJamming
add256
sub256

It all looked doable, but the magic numbers used around the code that
does eventual multiplication looked difficult to understand and I couldn't
deduce that from float128_mul. For some reason float128_mul implements
multipliction via multiplication and addition (mul128To256 & add128). There
is no equivalent to this in float64_muladd.

Let me make another attempt at this.

Regards,
Bharata.

^ permalink raw reply	[flat|nested] 6+ messages in thread

* Re: [Qemu-devel] [PATCH] target-ppc: Add quad precision muladd instructions
  2017-02-16  2:41   ` Bharata B Rao
@ 2017-02-16 20:53     ` Richard Henderson
  0 siblings, 0 replies; 6+ messages in thread
From: Richard Henderson @ 2017-02-16 20:53 UTC (permalink / raw)
  To: bharata; +Cc: qemu-devel, qemu-ppc, nikunj, david

On 02/16/2017 01:41 PM, Bharata B Rao wrote:
> For some reason float128_mul implements
> multipliction via multiplication and addition (mul128To256 & add128). There
> is no equivalent to this in float64_muladd.

Yes, I don't understand that myself.  It really appears to be a bug.


r~

^ permalink raw reply	[flat|nested] 6+ messages in thread

* Re: [Qemu-devel] [PATCH] target-ppc: Add quad precision muladd instructions
@ 2020-09-22  8:04 David Hildenbrand
  2020-09-22  9:22 ` Bharata B Rao
  0 siblings, 1 reply; 6+ messages in thread
From: David Hildenbrand @ 2020-09-22  8:04 UTC (permalink / raw)
  To: qemu-devel@nongnu.org, qemu-s390x, Bharata B Rao,
	Alex Bennée
  Cc: David Gibson, qemu-ppc, nikunj, Richard Henderson

Hi guys,

I just stumbled over

https://lore.kernel.org/qemu-devel/1487140636-19955-1-git-send-email-bharata@linux.vnet.ibm.com/

while looking for the state of float128_muladd(), as I need that for
s390x as well.

@Bharata, did you manage to implement a prototype? Looking into
float128_mul() makes my head spin.

-- 
Thanks,

David / dhildenb



^ permalink raw reply	[flat|nested] 6+ messages in thread

* Re: [Qemu-devel] [PATCH] target-ppc: Add quad precision muladd instructions
  2020-09-22  8:04 David Hildenbrand
@ 2020-09-22  9:22 ` Bharata B Rao
  0 siblings, 0 replies; 6+ messages in thread
From: Bharata B Rao @ 2020-09-22  9:22 UTC (permalink / raw)
  To: David Hildenbrand
  Cc: qemu-devel@nongnu.org, qemu-s390x, nikunj.dadhania, qemu-ppc,
	Bharata B Rao, David Gibson, Alex Bennée, Richard Henderson

On Tue, Sep 22, 2020 at 10:04:53AM +0200, David Hildenbrand wrote:
> Hi guys,
> 
> I just stumbled over
> 
> https://lore.kernel.org/qemu-devel/1487140636-19955-1-git-send-email-bharata@linux.vnet.ibm.com/
> 
> while looking for the state of float128_muladd(), as I need that for
> s390x as well.
> 
> @Bharata, did you manage to implement a prototype? Looking into
> float128_mul() makes my head spin.

I did try but couldn't figure out a way to implement some of the
dependent functions. Hence we still don't have quad-precision
multiply-add/sub instructions in ppc.

Regards,
Bharata.


^ permalink raw reply	[flat|nested] 6+ messages in thread

end of thread, other threads:[~2020-09-22  9:25 UTC | newest]

Thread overview: 6+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2017-02-15  6:37 [Qemu-devel] [PATCH] target-ppc: Add quad precision muladd instructions Bharata B Rao
2017-02-15 22:13 ` Richard Henderson
2017-02-16  2:41   ` Bharata B Rao
2017-02-16 20:53     ` Richard Henderson
  -- strict thread matches above, loose matches on Subject: below --
2020-09-22  8:04 David Hildenbrand
2020-09-22  9:22 ` Bharata B Rao

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).