Time of first use cost of upload.

Time of first use cost of upload.

This test is to check whether there is a time-of-first-use
cost (e.g. driver doesn't upload to the texture until we
draw with it).

BUG=423481
Committed: https://crrev.com/f82e8f7c7ab63b38b715c664f3bb23a608b5d28b
Cr-Commit-Position: refs/heads/master@{#322131}

[sivag, 2015-03-23 11:27:31]

siva.gunturi@samsung.com changed reviewers:
+ dcastagna@chromium.org, piman@chromium.org

[sivag, 2015-03-23 11:28:25]

siva.gunturi@samsung.com changed reviewers:
- dcastagna@chromium.org

[sivag, 2015-03-23 11:29:43]

siva.gunturi@samsung.com changed reviewers:
+ dcastagna@chromium.org

[sivag, 2015-03-23 11:29:43]

ptal..

[Daniele Castagna, 2015-03-23 16:06:38]

On 2015/03/23 at 11:29:43, siva.gunturi wrote:
> ptal..

Hi!
Thank you for working on this.

Why did you decide to go with this approach?
Wouldn't it be easier to draw multiple times with the same texture and measure
the cost of the first draw versus the cost of the subsequent draws?

[sivag, 2015-03-23 16:33:57]

On 2015/03/23 16:06:38, Daniele Castagna wrote:
> On 2015/03/23 at 11:29:43, siva.gunturi wrote:
> > ptal..
> 
> Hi!
> Thank you for working on this.
> 
> Why did you decide to go with this approach?
> Wouldn't it be easier to draw multiple times with the same texture and measure
> the cost of the first draw versus the cost of the subsequent draws?

The test is added to check whether driver doesn't upload to the texture until we
draw with it,
First i uploaded to different textures and recorded their times.
Then recorded the draw readings separately. From gpu side this should tell us
whether upload
really happened when draw is called or , it didn't wait for draw to happen.

May be you are right, i just did the first draw for all textures here, we should
compare this
with subsequent draws to get the first draw cost.

Do you think this is valid only for first time use, or whenever we use the
texture to upload?

[Daniele Castagna, 2015-03-23 19:55:12]

On 2015/03/23 at 16:33:57, siva.gunturi wrote:
> On 2015/03/23 16:06:38, Daniele Castagna wrote:
> > On 2015/03/23 at 11:29:43, siva.gunturi wrote:
> > > ptal..
> > 
> > Hi!
> > Thank you for working on this.
> > 
> > Why did you decide to go with this approach?
> > Wouldn't it be easier to draw multiple times with the same texture and
measure
> > the cost of the first draw versus the cost of the subsequent draws?
> 
> The test is added to check whether driver doesn't upload to the texture until
we draw with it,
> First i uploaded to different textures and recorded their times.
> Then recorded the draw readings separately. From gpu side this should tell us
whether upload
> really happened when draw is called or , it didn't wait for draw to happen.
> 

Are you suggesting that the absence of a GPU cost in the upload phase implies
that there is going
to be a first time draw cost?
This seems like a really strong assumption. There might be no GPU cost in
uploading at all. There
might be delayed CPU cost the first time the texture is used.

IMO it'd be much better and easier just to measure the cost of drawing the first
time versus
drawing again with the same texture.

There might be other set-up costs for getting things ready to run the program
the first time
(e.g: first time a program is used). That's why in TextureUploadPerfTest.upload
we run a few
warm up uploads and draws.


If you want to test if there is a first time upload cost when drawing, I'd
suggest to change the
TextureUploadPerfTest.upload to make it draw twice using the same texture.



> May be you are right, i just did the first draw for all textures here, we
should compare this
> with subsequent draws to get the first draw cost.
> 
> Do you think this is valid only for first time use, or whenever we use the
texture to upload?

[sivag, 2015-03-24 17:07:06]

On 2015/03/23 19:55:12, Daniele Castagna wrote:
> On 2015/03/23 at 16:33:57, siva.gunturi wrote:
> > On 2015/03/23 16:06:38, Daniele Castagna wrote:
> > > On 2015/03/23 at 11:29:43, siva.gunturi wrote:
> > > > ptal..
> > > 
> > > Hi!
> > > Thank you for working on this.
> > > 
> > > Why did you decide to go with this approach?
> > > Wouldn't it be easier to draw multiple times with the same texture and
> measure
> > > the cost of the first draw versus the cost of the subsequent draws?
> > 
> > The test is added to check whether driver doesn't upload to the texture
until
> we draw with it,
> > First i uploaded to different textures and recorded their times.
> > Then recorded the draw readings separately. From gpu side this should tell
us
> whether upload
> > really happened when draw is called or , it didn't wait for draw to happen.
> > 
> 
> Are you suggesting that the absence of a GPU cost in the upload phase implies
> that there is going
> to be a first time draw cost?
> This seems like a really strong assumption. There might be no GPU cost in
> uploading at all. There
> might be delayed CPU cost the first time the texture is used.
> 
> IMO it'd be much better and easier just to measure the cost of drawing the
first
> time versus
> drawing again with the same texture.
> 
> There might be other set-up costs for getting things ready to run the program
> the first time
> (e.g: first time a program is used). That's why in
TextureUploadPerfTest.upload
> we run a few
> warm up uploads and draws.
> 
> 
> If you want to test if there is a first time upload cost when drawing, I'd
> suggest to change the
> TextureUploadPerfTest.upload to make it draw twice using the same texture.
> 
> 
> 
> > May be you are right, i just did the first draw for all textures here, we
> should compare this
> > with subsequent draws to get the first draw cost.
> > 
> > Do you think this is valid only for first time use, or whenever we use the
> texture to upload?

ptal...
I made the changes to TextureUploadPerfTest.upload by drawing twice and
recording first draw cost.

[Daniele Castagna, 2015-03-24 17:30:08]

dcastagna@chromium.org changed reviewers:
+ reveman@chromium.org
- piman@chromium.org

[Daniele Castagna, 2015-03-24 17:30:09]

Thank you!
LGTM

I'm adding reveman@ for owner approval.

https://codereview.chromium.org/1031463002/diff/40001/gpu/perftests/texture_u...
File gpu/perftests/texture_upload_perftest.cc (right):

https://codereview.chromium.org/1031463002/diff/40001/gpu/perftests/texture_u...
gpu/perftests/texture_upload_perftest.cc:381:
first_draw_timers.GetAsMeasurement("firstdrawarrayscost"));
nit: I'd remove the suffix "cost". Every measurement we're tracking is a cost.

[sivag, 2015-03-24 17:39:59]

Thanks daniele for the comments.
@reveman, ptal..

[reveman, 2015-03-24 17:51:38]

lgtm

[sivag, 2015-03-25 07:18:22]

The CQ bit was checked by siva.gunturi@samsung.com

[sivag, 2015-03-25 07:18:22]

The patchset sent to the CQ was uploaded after l-g-t-m from
reveman@chromium.org, dcastagna@chromium.org
Link to the patchset: https://codereview.chromium.org/1031463002/#ps60001
(title: "Remove the suffix "cost".")

[commit-bot: I haz the power, 2015-03-25 07:18:40]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1031463002/60001

[commit-bot: I haz the power, 2015-03-25 07:51:08]

Committed patchset #4 (id:60001)

[commit-bot: I haz the power, 2015-03-25 07:52:03]

Patchset 4 (id:??) landed as
https://crrev.com/f82e8f7c7ab63b38b715c664f3bb23a608b5d28b
Cr-Commit-Position: refs/heads/master@{#322131}