Re: Accessing AVX/AVX2 instruction in UEFI.

[Andrew Fish <afish@apple.com>]

Amit,

In regards to AVX/AVX2 performance how are you doing the measuring?

In EFI it is hard to measure wall clock time for things that take a long time. Basically there is no scheduler in EFI and no threads, but there are events. The events can preempt your App while it is running and the time spent in events would look to you like time spent in your App. 

Generally the time spent in events should be constant (hot plugging USB or other changes like that may have a noticeable impact). If the goal of the performance measurement is to make the system boot faster you care more about the delta, than the absolute time (so the event overhead does not matter). 

If you are just doing a computation that does not do any IO then you may be able to raise the TPL to prevent events from being part of your measurement. 

Thanks,

Andrew Fish

PS I assume your are measuring the RELEASE code since you are turning off optimization on the DEBUG code. 

> On May 4, 2017, at 5:22 AM, Amit kumar <akamit91@hotmail.com> wrote:
> 
> Here are the compiler flags
> [BuildOptions]
>  MSFT:DEBUG_*_*_CC_FLAGS = /Od /FAsc /GL-
>  MSFT:RELEASE_*_*_CC_FLAGS = /FAsc /D MDEPKG_NDEBUG
>  MSFT:RELEASE_*_*_DLINK_FLAGS = /BASE:0x10000  /ALIGN:4096 /FILEALIGN:4096
> 
> 
> ________________________________
> From: Amit kumar <akamit91@hotmail.com>
> Sent: Thursday, May 4, 2017 5:48:11 PM
> To: Andrew Fish
> Cc: Mike Kinney; edk2-devel@lists.01.org
> Subject: Re: [edk2] Accessing AVX/AVX2 instruction in UEFI.
> 
> 
> Yes am aligning the data at 32 byte boundary while allocating memory in both environments.
> 
> in windows using  _alligned_malloc(size,32);
> 
> in UEFI
> 
> Offset = (UINTN)src & 0xFF;
> 
> src = (CHAR8 *)((UINTN) src - Offset + 0x20);
> 
> 
> Thanks
> 
> Amit
> 
> ________________________________
> From: afish@apple.com <afish@apple.com> on behalf of Andrew Fish <afish@apple.com>
> Sent: Thursday, May 4, 2017 5:02:55 PM
> To: Amit kumar
> Cc: Mike Kinney; edk2-devel@lists.01.org
> Subject: Re: [edk2] Accessing AVX/AVX2 instruction in UEFI.
> 
> 
>> On May 4, 2017, at 4:13 AM, Amit kumar <akamit91@hotmail.com> wrote:
>> 
>> Hi,
>> 
>> 
>> Even after using AVX2 instruction my code shown no performance improvement in UEFI although there is substantial improvement when i run the similar code in windows .
>> 
>> Am i missing something ?
>> 
> 
> Is the data aligned the same in both environments?
> 
> Thanks,
> 
> Andrew Fish
> 
>> Using MSVC compiler and the codes written in ASM.
>> 
>> Thanks And Regards
>> 
>> Amit
>> 
>> ________________________________
>> From: edk2-devel <edk2-devel-bounces@lists.01.org> on behalf of Amit kumar <akamit91@hotmail.com>
>> Sent: Wednesday, May 3, 2017 11:18:39 AM
>> To: Kinney, Michael D; Andrew Fish
>> Cc: edk2-devel@lists.01.org
>> Subject: Re: [edk2] Accessing AVX/AVX2 instruction in UEFI.
>> 
>> Thank you Michael and Andrew
>> 
>> 
>> Regards
>> 
>> Amit
>> 
>> ________________________________
>> From: Kinney, Michael D <michael.d.kinney@intel.com>
>> Sent: Tuesday, May 2, 2017 10:33:45 PM
>> To: Andrew Fish; Amit kumar; Kinney, Michael D
>> Cc: edk2-devel@lists.01.org
>> Subject: RE: [edk2] Accessing AVX/AVX2 instruction in UEFI.
>> 
>> Amit,
>> 
>> The information from Andrew is correct.
>> 
>> The document that covers this topic is the
>> Intel(r) 64 and IA-32 Architectures Software Developer Manuals
>> 
>> https://software.intel.com/en-us/articles/intel-sdm
>> 
>> Volume 1, Section 13.5.3 describes the AVX State.  There are
>> More details about detecting and enabling different AVX features
>> in that document.
>> 
>> If the CPU supports AVX, then the basic assembly instructions
>> required to use AVX instructions are the following that sets
>> bits 0, 1, 2 of XCR0.
>> 
>>   mov     rcx, 0
>>   xgetbv
>>   or      rax, 0007h
>>   xsetbv
>> 
>> One additional item you need to be aware of is that UEFI firmware only
>> saves/Restores CPU registers required for the UEFI ABI calling convention
>> when a timer interrupt or exception is processed.
>> 
>> This means CPU state such as the YMM registers are not saved/restored
>> across an interrupt and may be modified if code in interrupt context
>> also uses YMM registers.
>> 
>> When you enable the use of extended registers, interrupts should be
>> saved/disabled and restored around the extended register usage.
>> 
>> You can use the following functions from MdePkg BaseLib to do this
>> 
>> /**
>> Disables CPU interrupts and returns the interrupt state prior to the disable
>> operation.
>> 
>> @retval TRUE  CPU interrupts were enabled on entry to this call.
>> @retval FALSE CPU interrupts were disabled on entry to this call.
>> 
>> **/
>> BOOLEAN
>> EFIAPI
>> SaveAndDisableInterrupts (
>> VOID
>> );
>> 
>> /**
>> Set the current CPU interrupt state.
>> 
>> Sets the current CPU interrupt state to the state specified by
>> InterruptState. If InterruptState is TRUE, then interrupts are enabled. If
>> InterruptState is FALSE, then interrupts are disabled. InterruptState is
>> returned.
>> 
>> @param  InterruptState  TRUE if interrupts should enabled. FALSE if
>>                         interrupts should be disabled.
>> 
>> @return InterruptState
>> 
>> **/
>> BOOLEAN
>> EFIAPI
>> SetInterruptState (
>> IN      BOOLEAN                   InterruptState
>> );
>> 
>> Algorithm:
>> ============
>> {
>> BOOLEAN  InterruptState;
>> 
>> InterruptState = SaveAndDisableInterrupts();
>> 
>> // Enable use of AVX/AVX2 instructions
>> 
>> // Use AVX/AVX2 instructions
>> 
>> SetInterruptState (InterruptState);
>> }
>> 
>> Best regards,
>> 
>> Mike
>> 
>>> -----Original Message-----
>>> From: edk2-devel [mailto:edk2-devel-bounces@lists.01.org] On Behalf Of Andrew Fish
>>> Sent: Tuesday, May 2, 2017 8:12 AM
>>> To: Amit kumar <akamit91@hotmail.com>
>>> Cc: edk2-devel@lists.01.org
>>> Subject: Re: [edk2] Accessing AVX/AVX2 instruction in UEFI.
>>> 
>>> 
>>>> On May 2, 2017, at 6:57 AM, Amit kumar <akamit91@hotmail.com> wrote:
>>>> 
>>>> Hi,
>>>> 
>>>> Am trying to optimize an application using AVX/AVX2, but my code hangs while trying
>>> to access YMM registers.
>>>> The instruction where my code hangs is :
>>>> 
>>>> 
>>>> vmovups ymm0, YMMWORD PTR [rax]
>>>> 
>>>> 
>>>> I have verified the cpuid in OS and it supports AVX and AVX2 instruction. Processor
>>> i7 6th gen.
>>>> Can somebody help me out here ? Is there a way to enable YMM registers ?
>>>> 
>>> 
>>> Amit,
>>> 
>>> I think these instructions will generate an illegal instruction fault until you enable
>>> AVX. You need to check the Cpu ID bits in your code, then write BIT18 of CR4. After
>>> that XGETBV/XSETBV instructions are enabled and you can or in the lower 2 bits of
>>> XCR0. This basic operation is in the Intel Docs, it is just hard to find. Usually the
>>> OS has done this for the programmer and all the code needs to do is check the CPU ID.
>>> 
>>> Thanks,
>>> 
>>> Andrew Fish
>>> 
>>>> 
>>>> Thanks And Regards
>>>> Amit Kumar
>>>> 
>>>> _______________________________________________
>>>> edk2-devel mailing list
>>>> edk2-devel@lists.01.org
>>>> https://lists.01.org/mailman/listinfo/edk2-devel
>>> 
>>> _______________________________________________
>>> edk2-devel mailing list
>>> edk2-devel@lists.01.org
>>> https://lists.01.org/mailman/listinfo/edk2-devel
>> _______________________________________________
>> edk2-devel mailing list
>> edk2-devel@lists.01.org
>> https://lists.01.org/mailman/listinfo/edk2-devel
>> _______________________________________________
>> edk2-devel mailing list
>> edk2-devel@lists.01.org
>> https://lists.01.org/mailman/listinfo/edk2-devel
> 
> _______________________________________________
> edk2-devel mailing list
> edk2-devel@lists.01.org
> https://lists.01.org/mailman/listinfo/edk2-devel