Bug 201689: Belt-and-suspenders detection of FMA.
Commit Message
On 10/14/2016 03:12 PM, H.J. Lu wrote:
> On Fri, Oct 14, 2016 at 11:09 AM, Carlos O'Donell <carlos@redhat.com> wrote:
>> In the Intel Architecture Instruction Set Extensions Programming reference
>> the recommended way to test for FMA in section '2.2.1 Detection of FMA'
>> is:
>>
>> "Application Software must identify that hardware supports AVX as explained
>> in ... after that it must also detect support for FMA..."
>>
>> We don't do that in glibc. We use osxsave to detect the use of xgetbv, and
>> after that we check for AVX and FMA orthogonally. It is conceivable that
>> you could have the AVX bit clear and the FMA bit in an undefined state.
>>
>> I have never seen a machine with the AVX bit clear and the FMA bit set, but
>> we should follow the intel specifications and adjust our check as the following
>> patch works.
>
> One can't write a program with FMA nor AVX2 without using AVX
> instructions. AVX2/FMA aren't usable if AVX isn't. We should do
>
> if (CPU_FEATURES_CPU_P (cpu_features, AVX))
> {
> Set AVX available
> if (CPU_FEATURES_CPU_P (cpu_features, AVX2))
> Set AVX2 available
> if (CPU_FEATURES_CPU_P (cpu_features, FMA))
> Set FMA available
> }
>
Agreed. I double checked the manual, here is a new patch.
Testing on x86_64 and i686.
OK to commit if the results are clean?
v2
- Move FMA and AVX2 check up into AVX check since they depend upon it.
2016-10-14 Carlos O'Donell <carlos@redhat.com>
[BZ #20689]
* sysdeps/x86/cpu-features.c: Only enable FMA and AVX2 if
AVX is usable.
---
Comments
On Mon, Oct 17, 2016 at 8:44 AM, Carlos O'Donell <carlos@redhat.com> wrote:
> On 10/14/2016 03:12 PM, H.J. Lu wrote:
>> On Fri, Oct 14, 2016 at 11:09 AM, Carlos O'Donell <carlos@redhat.com> wrote:
>>> In the Intel Architecture Instruction Set Extensions Programming reference
>>> the recommended way to test for FMA in section '2.2.1 Detection of FMA'
>>> is:
>>>
>>> "Application Software must identify that hardware supports AVX as explained
>>> in ... after that it must also detect support for FMA..."
>>>
>>> We don't do that in glibc. We use osxsave to detect the use of xgetbv, and
>>> after that we check for AVX and FMA orthogonally. It is conceivable that
>>> you could have the AVX bit clear and the FMA bit in an undefined state.
>>>
>>> I have never seen a machine with the AVX bit clear and the FMA bit set, but
>>> we should follow the intel specifications and adjust our check as the following
>>> patch works.
>>
>> One can't write a program with FMA nor AVX2 without using AVX
>> instructions. AVX2/FMA aren't usable if AVX isn't. We should do
>>
>> if (CPU_FEATURES_CPU_P (cpu_features, AVX))
>> {
>> Set AVX available
>> if (CPU_FEATURES_CPU_P (cpu_features, AVX2))
>> Set AVX2 available
>> if (CPU_FEATURES_CPU_P (cpu_features, FMA))
>> Set FMA available
>> }
>>
>
> Agreed. I double checked the manual, here is a new patch.
>
> Testing on x86_64 and i686.
>
> OK to commit if the results are clean?
>
> v2
> - Move FMA and AVX2 check up into AVX check since they depend upon it.
>
> 2016-10-14 Carlos O'Donell <carlos@redhat.com>
>
> [BZ #20689]
> * sysdeps/x86/cpu-features.c: Only enable FMA and AVX2 if
> AVX is usable.
>
> diff --git a/sysdeps/x86/cpu-features.c b/sysdeps/x86/cpu-features.c
> index 11b9af2..e228a76 100644
> --- a/sysdeps/x86/cpu-features.c
> +++ b/sysdeps/x86/cpu-features.c
> @@ -60,12 +60,20 @@ get_common_indeces (struct cpu_features *cpu_features,
> {
> /* Determine if AVX is usable. */
> if (CPU_FEATURES_CPU_P (cpu_features, AVX))
> - cpu_features->feature[index_arch_AVX_Usable]
> - |= bit_arch_AVX_Usable;
> - /* Determine if AVX2 is usable. */
> - if (CPU_FEATURES_CPU_P (cpu_features, AVX2))
> - cpu_features->feature[index_arch_AVX2_Usable]
> - |= bit_arch_AVX2_Usable;
> + {
> + cpu_features->feature[index_arch_AVX_Usable]
> + |= bit_arch_AVX_Usable;
> + /* The following features depend on AVX being usable. */
> + /* Determine if AVX2 is usable. */
> + if (CPU_FEATURES_CPU_P (cpu_features, AVX2))
> + cpu_features->feature[index_arch_AVX2_Usable]
> + |= bit_arch_AVX2_Usable;
> + /* Determine if FMA is usable. */
> + if (CPU_FEATURES_CPU_P (cpu_features, FMA))
> + cpu_features->feature[index_arch_FMA_Usable]
> + |= bit_arch_FMA_Usable;
> + }
> +
> /* Check if OPMASK state, upper 256-bit of ZMM0-ZMM15 and
> ZMM16-ZMM31 state are enabled. */
> if ((xcrlow & (bit_Opmask_state | bit_ZMM0_15_state
> @@ -83,10 +91,6 @@ get_common_indeces (struct cpu_features *cpu_features,
> |= bit_arch_AVX512DQ_Usable;
> }
> }
> - /* Determine if FMA is usable. */
> - if (CPU_FEATURES_CPU_P (cpu_features, FMA))
> - cpu_features->feature[index_arch_FMA_Usable]
> - |= bit_arch_FMA_Usable;
> }
> }
> }
> ---
>
This is OK.
Thanks.
On 10/17/2016 11:44 AM, Carlos O'Donell wrote:
> On 10/14/2016 03:12 PM, H.J. Lu wrote:
>> On Fri, Oct 14, 2016 at 11:09 AM, Carlos O'Donell <carlos@redhat.com> wrote:
>>> In the Intel Architecture Instruction Set Extensions Programming reference
>>> the recommended way to test for FMA in section '2.2.1 Detection of FMA'
>>> is:
>>>
>>> "Application Software must identify that hardware supports AVX as explained
>>> in ... after that it must also detect support for FMA..."
>>>
>>> We don't do that in glibc. We use osxsave to detect the use of xgetbv, and
>>> after that we check for AVX and FMA orthogonally. It is conceivable that
>>> you could have the AVX bit clear and the FMA bit in an undefined state.
>>>
>>> I have never seen a machine with the AVX bit clear and the FMA bit set, but
>>> we should follow the intel specifications and adjust our check as the following
>>> patch works.
>>
>> One can't write a program with FMA nor AVX2 without using AVX
>> instructions. AVX2/FMA aren't usable if AVX isn't. We should do
>>
>> if (CPU_FEATURES_CPU_P (cpu_features, AVX))
>> {
>> Set AVX available
>> if (CPU_FEATURES_CPU_P (cpu_features, AVX2))
>> Set AVX2 available
>> if (CPU_FEATURES_CPU_P (cpu_features, FMA))
>> Set FMA available
>> }
>>
>
> Agreed. I double checked the manual, here is a new patch.
>
> Testing on x86_64 and i686.
>
> OK to commit if the results are clean?
>
> v2
> - Move FMA and AVX2 check up into AVX check since they depend upon it.
>
> 2016-10-14 Carlos O'Donell <carlos@redhat.com>
>
> [BZ #20689]
> * sysdeps/x86/cpu-features.c: Only enable FMA and AVX2 if
> AVX is usable.
Committed. Thanks.
@@ -60,12 +60,20 @@ get_common_indeces (struct cpu_features *cpu_features,
{
/* Determine if AVX is usable. */
if (CPU_FEATURES_CPU_P (cpu_features, AVX))
- cpu_features->feature[index_arch_AVX_Usable]
- |= bit_arch_AVX_Usable;
- /* Determine if AVX2 is usable. */
- if (CPU_FEATURES_CPU_P (cpu_features, AVX2))
- cpu_features->feature[index_arch_AVX2_Usable]
- |= bit_arch_AVX2_Usable;
+ {
+ cpu_features->feature[index_arch_AVX_Usable]
+ |= bit_arch_AVX_Usable;
+ /* The following features depend on AVX being usable. */
+ /* Determine if AVX2 is usable. */
+ if (CPU_FEATURES_CPU_P (cpu_features, AVX2))
+ cpu_features->feature[index_arch_AVX2_Usable]
+ |= bit_arch_AVX2_Usable;
+ /* Determine if FMA is usable. */
+ if (CPU_FEATURES_CPU_P (cpu_features, FMA))
+ cpu_features->feature[index_arch_FMA_Usable]
+ |= bit_arch_FMA_Usable;
+ }
+
/* Check if OPMASK state, upper 256-bit of ZMM0-ZMM15 and
ZMM16-ZMM31 state are enabled. */
if ((xcrlow & (bit_Opmask_state | bit_ZMM0_15_state
@@ -83,10 +91,6 @@ get_common_indeces (struct cpu_features *cpu_features,
|= bit_arch_AVX512DQ_Usable;
}
}
- /* Determine if FMA is usable. */
- if (CPU_FEATURES_CPU_P (cpu_features, FMA))
- cpu_features->feature[index_arch_FMA_Usable]
- |= bit_arch_FMA_Usable;
}
}
}