2020-04-15 21:26:30

by 赵军奎

[permalink] [raw]
Subject: [PATCH] Optimized division operation to shift operation

On some processors, the / operate will call the compiler`s div lib,
which is low efficient, We can replace the / operation with shift,
so that we can replace the call of the division library with one
shift assembly instruction.

Signed-off-by: Bernard Zhao <[email protected]>
---
drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c | 4 ++--
drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c | 4 ++--
drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c | 4 ++--
3 files changed, 6 insertions(+), 6 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c
index b205039..66cd078 100644
--- a/drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c
+++ b/drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c
@@ -175,10 +175,10 @@ static int gmc_v6_0_mc_load_microcode(struct amdgpu_device *adev)
amdgpu_ucode_print_mc_hdr(&hdr->header);

adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
- regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+ regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
new_io_mc_regs = (const __le32 *)
(adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
- ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+ ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
new_fw_data = (const __le32 *)
(adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));

diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c
index 9da9596..ca26d63 100644
--- a/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c
+++ b/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c
@@ -193,10 +193,10 @@ static int gmc_v7_0_mc_load_microcode(struct amdgpu_device *adev)
amdgpu_ucode_print_mc_hdr(&hdr->header);

adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
- regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+ regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
io_mc_regs = (const __le32 *)
(adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
- ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+ ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
fw_data = (const __le32 *)
(adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));

diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c
index 27d83204..295039c 100644
--- a/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c
+++ b/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c
@@ -318,10 +318,10 @@ static int gmc_v8_0_tonga_mc_load_microcode(struct amdgpu_device *adev)
amdgpu_ucode_print_mc_hdr(&hdr->header);

adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
- regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+ regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
io_mc_regs = (const __le32 *)
(adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
- ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+ ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
fw_data = (const __le32 *)
(adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));

--
2.7.4


2020-04-15 22:23:26

by Christian König

[permalink] [raw]
Subject: Re: [PATCH] Optimized division operation to shift operation

Am 15.04.20 um 09:41 schrieb Jani Nikula:
> On Tue, 14 Apr 2020, Alex Deucher <[email protected]> wrote:
>> On Tue, Apr 14, 2020 at 9:05 AM Bernard Zhao <[email protected]> wrote:
>>> On some processors, the / operate will call the compiler`s div lib,
>>> which is low efficient, We can replace the / operation with shift,
>>> so that we can replace the call of the division library with one
>>> shift assembly instruction.
> This was applied already, and it's not in a driver I look after... but
> to me this feels like something that really should be
> justified. Using >> instead of / for multiples of 2 division mattered 20
> years ago, I'd be surprised if it still did on modern compilers.

I have similar worries, especially since we replace the "/ (4 * 2)" with
">> 3" it's making the code just a bit less readable.

And that the code runs exactly once while loading the driver and pushing
the firmware into the hardware. So performance is completely irrelevant
here.

Regards,
Christian.

>
> BR,
> Jani.
>
>
>>> Signed-off-by: Bernard Zhao <[email protected]>
>> Applied. thanks.
>>
>> Alex
>>
>>> ---
>>> drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c | 4 ++--
>>> drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c | 4 ++--
>>> drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c | 4 ++--
>>> 3 files changed, 6 insertions(+), 6 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c
>>> index b205039..66cd078 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c
>>> @@ -175,10 +175,10 @@ static int gmc_v6_0_mc_load_microcode(struct amdgpu_device *adev)
>>> amdgpu_ucode_print_mc_hdr(&hdr->header);
>>>
>>> adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
>>> - regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
>>> + regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
>>> new_io_mc_regs = (const __le32 *)
>>> (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
>>> - ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
>>> + ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
>>> new_fw_data = (const __le32 *)
>>> (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c
>>> index 9da9596..ca26d63 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c
>>> @@ -193,10 +193,10 @@ static int gmc_v7_0_mc_load_microcode(struct amdgpu_device *adev)
>>> amdgpu_ucode_print_mc_hdr(&hdr->header);
>>>
>>> adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
>>> - regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
>>> + regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
>>> io_mc_regs = (const __le32 *)
>>> (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
>>> - ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
>>> + ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
>>> fw_data = (const __le32 *)
>>> (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c
>>> index 27d83204..295039c 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c
>>> @@ -318,10 +318,10 @@ static int gmc_v8_0_tonga_mc_load_microcode(struct amdgpu_device *adev)
>>> amdgpu_ucode_print_mc_hdr(&hdr->header);
>>>
>>> adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
>>> - regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
>>> + regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
>>> io_mc_regs = (const __le32 *)
>>> (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
>>> - ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
>>> + ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
>>> fw_data = (const __le32 *)
>>> (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
>>>
>>> --
>>> 2.7.4
>>>
>>> _______________________________________________
>>> amd-gfx mailing list
>>> [email protected]
>>> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Cchristian.koenig%40amd.com%7C1e91f7edcfe0473b0d7008d7e11074a8%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637225333103893889&amp;sdata=VDJlEY2%2Bl1SSO8Fw1dYqqPFqQtyHpsxQ0Tm7iVOgJQY%3D&amp;reserved=0
>> _______________________________________________
>> dri-devel mailing list
>> [email protected]
>> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Fdri-devel&amp;data=02%7C01%7Cchristian.koenig%40amd.com%7C1e91f7edcfe0473b0d7008d7e11074a8%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637225333103893889&amp;sdata=EpqRRbCiksur%2BjMlVQplExuJsmw6UPODhyBOutOVukw%3D&amp;reserved=0

2020-04-15 22:48:01

by Daniel Vetter

[permalink] [raw]
Subject: Re: [PATCH] Optimized division operation to shift operation

On Wed, Apr 15, 2020 at 9:57 AM Christian König
<[email protected]> wrote:
>
> Am 15.04.20 um 09:41 schrieb Jani Nikula:
> > On Tue, 14 Apr 2020, Alex Deucher <[email protected]> wrote:
> >> On Tue, Apr 14, 2020 at 9:05 AM Bernard Zhao <[email protected]> wrote:
> >>> On some processors, the / operate will call the compiler`s div lib,
> >>> which is low efficient, We can replace the / operation with shift,
> >>> so that we can replace the call of the division library with one
> >>> shift assembly instruction.
> > This was applied already, and it's not in a driver I look after... but
> > to me this feels like something that really should be
> > justified. Using >> instead of / for multiples of 2 division mattered 20
> > years ago, I'd be surprised if it still did on modern compilers.
>
> I have similar worries, especially since we replace the "/ (4 * 2)" with
> ">> 3" it's making the code just a bit less readable.
>
> And that the code runs exactly once while loading the driver and pushing
> the firmware into the hardware. So performance is completely irrelevant
> here.

Yeah, and even in general I'd really want to see proof that such bad
compilers exist first. Doing a peephole pass for stuff like this is
like undergrad compiler course stuff. Also I'd trust the compiler's
call on which is faster much more than humans doing a manual peephole
pass.

Or am I just massively biased from all the layman following of what
kind of our gl/vk/compute compiler people are doing on a daily basis?
-Daniel

>
> Regards,
> Christian.
>
> >
> > BR,
> > Jani.
> >
> >
> >>> Signed-off-by: Bernard Zhao <[email protected]>
> >> Applied. thanks.
> >>
> >> Alex
> >>
> >>> ---
> >>> drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c | 4 ++--
> >>> drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c | 4 ++--
> >>> drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c | 4 ++--
> >>> 3 files changed, 6 insertions(+), 6 deletions(-)
> >>>
> >>> diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c
> >>> index b205039..66cd078 100644
> >>> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c
> >>> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c
> >>> @@ -175,10 +175,10 @@ static int gmc_v6_0_mc_load_microcode(struct amdgpu_device *adev)
> >>> amdgpu_ucode_print_mc_hdr(&hdr->header);
> >>>
> >>> adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
> >>> - regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
> >>> + regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
> >>> new_io_mc_regs = (const __le32 *)
> >>> (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
> >>> - ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
> >>> + ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
> >>> new_fw_data = (const __le32 *)
> >>> (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
> >>>
> >>> diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c
> >>> index 9da9596..ca26d63 100644
> >>> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c
> >>> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c
> >>> @@ -193,10 +193,10 @@ static int gmc_v7_0_mc_load_microcode(struct amdgpu_device *adev)
> >>> amdgpu_ucode_print_mc_hdr(&hdr->header);
> >>>
> >>> adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
> >>> - regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
> >>> + regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
> >>> io_mc_regs = (const __le32 *)
> >>> (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
> >>> - ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
> >>> + ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
> >>> fw_data = (const __le32 *)
> >>> (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
> >>>
> >>> diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c
> >>> index 27d83204..295039c 100644
> >>> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c
> >>> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c
> >>> @@ -318,10 +318,10 @@ static int gmc_v8_0_tonga_mc_load_microcode(struct amdgpu_device *adev)
> >>> amdgpu_ucode_print_mc_hdr(&hdr->header);
> >>>
> >>> adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
> >>> - regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
> >>> + regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
> >>> io_mc_regs = (const __le32 *)
> >>> (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
> >>> - ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
> >>> + ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
> >>> fw_data = (const __le32 *)
> >>> (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
> >>>
> >>> --
> >>> 2.7.4
> >>>
> >>> _______________________________________________
> >>> amd-gfx mailing list
> >>> [email protected]
> >>> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Cchristian.koenig%40amd.com%7C1e91f7edcfe0473b0d7008d7e11074a8%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637225333103893889&amp;sdata=VDJlEY2%2Bl1SSO8Fw1dYqqPFqQtyHpsxQ0Tm7iVOgJQY%3D&amp;reserved=0
> >> _______________________________________________
> >> dri-devel mailing list
> >> [email protected]
> >> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Fdri-devel&amp;data=02%7C01%7Cchristian.koenig%40amd.com%7C1e91f7edcfe0473b0d7008d7e11074a8%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637225333103893889&amp;sdata=EpqRRbCiksur%2BjMlVQplExuJsmw6UPODhyBOutOVukw%3D&amp;reserved=0
>
> _______________________________________________
> dri-devel mailing list
> [email protected]
> https://lists.freedesktop.org/mailman/listinfo/dri-devel



--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

2020-04-15 22:50:58

by David Laight

[permalink] [raw]
Subject: RE: [PATCH] Optimized division operation to shift operation

From: Christian König
> Sent: 15 April 2020 08:57
> Am 15.04.20 um 09:41 schrieb Jani Nikula:
> > On Tue, 14 Apr 2020, Alex Deucher <[email protected]> wrote:
> >> On Tue, Apr 14, 2020 at 9:05 AM Bernard Zhao <[email protected]> wrote:
> >>> On some processors, the / operate will call the compiler`s div lib,
> >>> which is low efficient, We can replace the / operation with shift,
> >>> so that we can replace the call of the division library with one
> >>> shift assembly instruction.
> > This was applied already, and it's not in a driver I look after... but
> > to me this feels like something that really should be
> > justified. Using >> instead of / for multiples of 2 division mattered 20
> > years ago, I'd be surprised if it still did on modern compilers.
>
> I have similar worries, especially since we replace the "/ (4 * 2)" with
> ">> 3" it's making the code just a bit less readable.
>
> And that the code runs exactly once while loading the driver and pushing
> the firmware into the hardware. So performance is completely irrelevant
> here.

Force the division to be unsigned and the compiler will use a shift.

David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)