Fixes to get kwin_wayland starting with Lima #51

shadeslayer · 2018-04-25T12:18:33Z

No description provided.

Signed-off-by: Qiang Yu <[email protected]>

Singed-off-by: Qiang Yu <[email protected]>

Signed-off-by: Qiang Yu <[email protected]>

This is needed for lima gp compiler. Signed-off-by: Qiang Yu <[email protected]>

Signed-off-by: Qiang Yu <[email protected]>

Signed-off-by: Erico Nunes <[email protected]> Reviewed-by: Qiang Yu <[email protected]>

v2: (Qiang Yu) no need to record draw info in lima_context Signed-off-by: Erico Nunes <[email protected]> Signed-off-by: Qiang Yu <[email protected]>

Signed-off-by: Qiang Yu <[email protected]>

No other place except mesa need the libdrm_lima, so we can just move it into mesa so that no need to maintain the libdrm ABI backward compatibility and reduce some overhead. Advised-by: Rob Clark <[email protected]> Signed-off-by: Qiang Yu <[email protected]>

Signed-off-by: Qiang Yu <[email protected]>

Implement 'gpir_op_lt' alu opcode from nir 'nir_op_slt'. Tested on Mali400 on the Allwinner A20. Signed-off-by: Erico Nunes <[email protected]>

Implement 'gpir_op_select' alu opcode from nir 'nir_op_bcsel'. Tested on Mali400 on the Allwinner A20. Signed-off-by: Erico Nunes <[email protected]>

ir_binop_gequal needs to be converted to nir_op_sge when native integers are not supported in the driver. Otherwise it becomes no different than ir_binop_less after the conversion. Signed-off-by: Erico Nunes <[email protected]>

Implement 'gpir_op_ge' alu opcode from nir 'nir_op_sge'. Tested on Mali400 on the Allwinner A20. Signed-off-by: Erico Nunes <[email protected]>

Implement 'gpir_op_floor' alu opcode from nir 'nir_op_ffloor'. Tested on Mali400 on the Allwinner A20. Signed-off-by: Erico Nunes <[email protected]>

Implement 'gpir_op_sign' alu opcode from nir 'nir_op_fsign'. Tested on Mali400 on the Allwinner A20. Signed-off-by: Erico Nunes <[email protected]>

Signed-off-by: Koen Kooi <[email protected]>

Signed-off-by: Qiang Yu <[email protected]>

mali400 uses on-chip tile buffer to store depth and stencil values, so no separate buffer is needed if FBO doesn't have depth/stencil attachment We still need to figure out how to render to depth/stencil buffer. Signed-off-by: Vasily Khoruzhick <[email protected]>

Signed-off-by: Qiang Yu <[email protected]>

Signed-off-by: Vasily Khoruzhick <[email protected]>

There are no 'equal' or 'not-equal' opcodes in gp. eq (a == b) is lowered to and(a >= b, b >= a) ne (a != b) is lowered to or(a < b, b < a) Signed-off-by: Erico Nunes <[email protected]>

Support nir opcodes nir_op_fand and nir_op_for with gpir_op_min and gpir_op_max. Tested on Mali400 on the Allwinner A20. Signed-off-by: Erico Nunes <[email protected]>

Fixes yuq#49

anarsoul · 2018-04-26T01:10:45Z

src/gallium/drivers/lima/lima_texture.c

@@ -91,7 +91,7 @@ lima_update_tex_desc(struct lima_context *ctx, struct lima_sampler_state *sample

   /* TODO: - do we need to align width/height to 16?
            - does hardware support stride different from width? */
-   width = prsc->width0;
+   width = align(prsc->width0, 16);


That's probably not a right place to do it. We should align resource stride when it's allocated.

I guess the resource stride is already aligned when allocation, but the descriptor setup need a way to tell the width and stride difference. Texture UV space [0, 1] is mapped to pixel space [0, width], but texture is stored in memory in stride X height way. So we have to program in the descriptor both width and stride value. I can't see where current code reflect this, so we need descriptor changes for the width != stride case.

@yuq I don't think that hardware supports that.

robclark said that it's pretty common limitation of hardware, so we have to workaround it in the driver.

Again, let's dump to see how mali handle this case.

@yuq I suspect it does extra copy in this case. Anyway, I'm running buildroot rootfs with no X11 nor wayland, so it's a bit tricky for me to implement new tests - Mali doesn't support EGL_MESA_platform_gbm so neither of demos from your gfx repo work.

OK, if it does extra copy, it'll be way much slower.

Maybe I can modify the gbm-surface-fbo for X11 and go for a dump next week after I get back from vacation.

EGL_KHR_platform_gbm is supported by recent malion e.g. rockchip, see https://github.com/xbmc/xbmc/pull/13746/files#diff-2b5179a78cb072ff496b13892efaa8b2R41 for a bit more info.

@koenkooi I still can't get it working. I replaced EGL_PLATFORM_GBM_MESA with EGL_PLATFORM_GBM_KHR and now it fails in eglCreatePlatformWindowSurfaceEXT() - it returns EGL_NO_SURFACE.

Also I had to replace /dev/dri/renderD128 with /dev/dri/card0

yuq · 2018-04-25T14:47:39Z

src/gallium/drivers/lima/lima_resource.c

@@ -310,6 +310,10 @@ lima_surface_destroy(struct pipe_context *pctx, struct pipe_surface *psurf)

         struct hash_entry *entry =
            _mesa_hash_table_search(ctx->plb_pp_stream, &key);
+
+        if (!entry)


How does this happen? lima_surface_create should always create the entry.

I'm unsure if I can provide a minimal test case for this, but I added this check due to a crash I experienced when trying to move Plasma's widget popups.

OK, but this case may lead to a design bug, if so, fix here is not the root cause.

BTW. how about the kwin wayland status? Seems you got something onscreen like the Plasma's widget with mesa-lima?

Ill spend some more time looking at this then.

Re plasmashell, that's correct Plasmashell renders fine (though the render width seems to be off). Launching apps seems to lead to corruption though. Something I hope to tackle next week.

By design, all surface is created by lima_surface_create and freed by lima_surface_destroy. And lima_surface_create should always create or reference count the entry. So if these two functions are called in balance, this case should not happen.

Nice to hear one desktop works (partially), it's a little surprise to me, thanks.

That's what I understood too, I was equally surprised that this was not the case.

And yeah, hopefully together we can get all of it working soon :D

yuq · 2018-04-26T02:01:04Z

src/gallium/drivers/lima/lima_texture.c

@@ -91,7 +91,7 @@ lima_update_tex_desc(struct lima_context *ctx, struct lima_sampler_state *sample

   /* TODO: - do we need to align width/height to 16?
            - does hardware support stride different from width? */
-   width = prsc->width0;
+   width = align(prsc->width0, 16);


I guess the resource stride is already aligned when allocation, but the descriptor setup need a way to tell the width and stride difference. Texture UV space [0, 1] is mapped to pixel space [0, width], but texture is stored in memory in stride X height way. So we have to program in the descriptor both width and stride value. I can't see where current code reflect this, so we need descriptor changes for the width != stride case.

shadeslayer · 2018-05-09T15:19:22Z

Hey
Any chance you got that dump to help align stride and width? :D

anarsoul · 2018-05-09T15:26:36Z

Dump made by yuq shows that Mali driver uses tiled format. We don't support it at the moment in Lima driver, so it'll require more work

…

On Wed, May 9, 2018, 08:19 Rohan Garg ***@***.***> wrote: Hey Any chance you got that dump to help align stride and width? :D — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#51 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAFAiSqZ7OIYOUR1ma0fCcJY-zUu9LEUks5twwj6gaJpZM4TjULO> .

shadeslayer · 2018-05-09T15:27:57Z

Oh where is the dump?

anarsoul · 2018-05-09T15:28:37Z

See issues list.

…

On Wed, May 9, 2018, 08:27 Rohan Garg ***@***.***> wrote: Oh where is the dump? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#51 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAFAiYR6wXKRXSabI6qgj-Dj-8rjkz1Fks5twwr-gaJpZM4TjULO> .

koenkooi · 2018-05-09T15:28:47Z

#53

shadeslayer · 2018-05-09T15:58:53Z

Aha, thanks :D

shadeslayer · 2018-06-05T01:32:43Z

Closing since we're working on a better approach with tiled textures.

yuq added 30 commits March 29, 2018 09:28

lima/gpir: adjust load/store/const child num

ecd0798

Signed-off-by: Qiang Yu <[email protected]>

lima/gpir: add instr insert load node

ba589f0

Signed-off-by: Qiang Yu <[email protected]>

lima/gpir: restructure the instr insert

907b26b

Singed-off-by: Qiang Yu <[email protected]>

lima/gpir: store/alu instr insert check free slot

4a41109

Signed-off-by: Qiang Yu <[email protected]>

lima/gpir: schedule fix move max place

bcd07a6

Signed-off-by: Qiang Yu <[email protected]>

lima/gpir: insert move node between load and store node

2253583

Signed-off-by: Qiang Yu <[email protected]>

lima/gpir: insert move node for store/load move to scheduler

e8502b5

Signed-off-by: Qiang Yu <[email protected]>

lima/gpir: reuse existing move node when schedule load node

f44fdba

Signed-off-by: Qiang Yu <[email protected]>

lima/gpir: lower neg node

7ce6ae0

Signed-off-by: Qiang Yu <[email protected]>

lima/gpir: remove copy node handle

7e0c5c6

Signed-off-by: Qiang Yu <[email protected]>

lima/gpir: remove alu node children_component

904dfcc

Signed-off-by: Qiang Yu <[email protected]>

nir: add load uniform lower to scalar

0c04ffc

This is needed for lima gp compiler. Signed-off-by: Qiang Yu <[email protected]>

lima/gpir: compiler support load uniform node

515fffe

Signed-off-by: Qiang Yu <[email protected]>

lima/gpir: not print already printed subtree in node print

c239d50

Signed-off-by: Qiang Yu <[email protected]>

lima/gpir: don't print const node after deleted

1b94f3d

Signed-off-by: Qiang Yu <[email protected]>

lima/gpir: fix add entry within a set loop

0ddbfcb

Signed-off-by: Qiang Yu <[email protected]>

lima/gpir: lower const load into uniform load

0f312f6

Signed-off-by: Qiang Yu <[email protected]>

lima: use user constant buffer

7247ac9

Signed-off-by: Qiang Yu <[email protected]>

lima: construct const buffer from st and shader const

3d1c400

Signed-off-by: Qiang Yu <[email protected]>

lima: add codegen and use it to replace the static vs

e45f774

Signed-off-by: Qiang Yu <[email protected]>

lima/gpir: seperate two store unit for store temp

8410aba

Signed-off-by: Qiang Yu <[email protected]>

lima/gpir: print codegen

e794966

Signed-off-by: Qiang Yu <[email protected]>

lima: replace using gpir_prog with vs state

68f9bdd

Signed-off-by: Qiang Yu <[email protected]>

lima: remove the stand alone compiler

9321e8d

Signed-off-by: Qiang Yu <[email protected]>

lima: fs use nir

62b3fe8

Signed-off-by: Qiang Yu <[email protected]>

lima: fix util_dynarray_init for 17.2

f75a8e3

Signed-off-by: Erico Nunes <[email protected]> Reviewed-by: Qiang Yu <[email protected]>

lima: remove index buffer update for 17.2

3b0ea68

v2: (Qiang Yu) no need to record draw info in lima_context Signed-off-by: Erico Nunes <[email protected]> Signed-off-by: Qiang Yu <[email protected]>

lima: screen/shader param switch use default

e78a744

Signed-off-by: Qiang Yu <[email protected]>

lima: use mesa hash table instead of libdrm one

578eb6d

Signed-off-by: Qiang Yu <[email protected]>

enunes and others added 18 commits April 12, 2018 06:28

lima/gpir: support lt alu opcode

f31a204

Implement 'gpir_op_lt' alu opcode from nir 'nir_op_slt'. Tested on Mali400 on the Allwinner A20. Signed-off-by: Erico Nunes <[email protected]>

lima/gpir: support select alu opcode

780a272

Implement 'gpir_op_select' alu opcode from nir 'nir_op_bcsel'. Tested on Mali400 on the Allwinner A20. Signed-off-by: Erico Nunes <[email protected]>

nir: fix ir_binop_gequal glsl_to_nir conversion

862f09d

ir_binop_gequal needs to be converted to nir_op_sge when native integers are not supported in the driver. Otherwise it becomes no different than ir_binop_less after the conversion. Signed-off-by: Erico Nunes <[email protected]>

lima/gpir: support ge alu opcode

7bfc8ff

Implement 'gpir_op_ge' alu opcode from nir 'nir_op_sge'. Tested on Mali400 on the Allwinner A20. Signed-off-by: Erico Nunes <[email protected]>

lima/gpir: support floor alu opcode

f5d4ad3

Implement 'gpir_op_floor' alu opcode from nir 'nir_op_ffloor'. Tested on Mali400 on the Allwinner A20. Signed-off-by: Erico Nunes <[email protected]>

lima/gpir: support sign alu opcode

1e3fd4f

Implement 'gpir_op_sign' alu opcode from nir 'nir_op_fsign'. Tested on Mali400 on the Allwinner A20. Signed-off-by: Erico Nunes <[email protected]>

lima: replace "unsupport" in error messages with "unsupported"

c4d4ca0

Signed-off-by: Koen Kooi <[email protected]>

lima nir: alias nir_op_fmov to ppir_op_mov

bc3710f

Signed-off-by: Koen Kooi <[email protected]>

lima: use DLBU for mali450 PP jobs

1c7700f

Signed-off-by: Qiang Yu <[email protected]>

lima: use LIMA_CTX_PLB_BLK_SIZE in lima_update_plb

376b3c8

Signed-off-by: Qiang Yu <[email protected]>

lima: set DLBU reg according to LIMA_CTX_PLB_BLK_SIZE

13bf02f

Signed-off-by: Qiang Yu <[email protected]>

lima/ppir: implement nir_op_fmin

2cfcf46

Signed-off-by: Vasily Khoruzhick <[email protected]>

lima/gpir: implement eq and ne lowering functions

4eaa09d

There are no 'equal' or 'not-equal' opcodes in gp. eq (a == b) is lowered to and(a >= b, b >= a) ne (a != b) is lowered to or(a < b, b < a) Signed-off-by: Erico Nunes <[email protected]>

lima/gpir: support and and or operations

2dc0e24

Support nir opcodes nir_op_fand and nir_op_for with gpir_op_min and gpir_op_max. Tested on Mali400 on the Allwinner A20. Signed-off-by: Erico Nunes <[email protected]>

Do not update textures when there are no textures

709e3b2

Fixes yuq#49

Align width to 16 in order to fix screen corruption

a684154

Don't try to access null pointers

3229620

anarsoul reviewed Apr 26, 2018

View reviewed changes

yuq requested changes Apr 26, 2018

View reviewed changes

yuq force-pushed the lima-18.0 branch 2 times, most recently from ee7c7ee to ac95540 Compare May 8, 2018 01:49

shadeslayer closed this Jun 5, 2018

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fixes to get kwin_wayland starting with Lima #51

Fixes to get kwin_wayland starting with Lima #51

shadeslayer commented Apr 25, 2018

anarsoul Apr 26, 2018

yuq Apr 26, 2018

anarsoul Apr 26, 2018

anarsoul Apr 26, 2018

yuq Apr 26, 2018

anarsoul Apr 26, 2018

yuq Apr 26, 2018

koenkooi Apr 26, 2018

anarsoul May 1, 2018

anarsoul May 1, 2018

yuq Apr 25, 2018

shadeslayer Apr 26, 2018

yuq Apr 26, 2018

shadeslayer Apr 26, 2018 •

edited

Loading

yuq Apr 26, 2018

shadeslayer Apr 26, 2018

yuq Apr 26, 2018

shadeslayer commented May 9, 2018

anarsoul commented May 9, 2018 via email

shadeslayer commented May 9, 2018

anarsoul commented May 9, 2018 via email

koenkooi commented May 9, 2018

shadeslayer commented May 9, 2018

shadeslayer commented Jun 5, 2018

Fixes to get kwin_wayland starting with Lima #51

Fixes to get kwin_wayland starting with Lima #51

Conversation

shadeslayer commented Apr 25, 2018

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

shadeslayer Apr 26, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

shadeslayer commented May 9, 2018

anarsoul commented May 9, 2018 via email

shadeslayer commented May 9, 2018

anarsoul commented May 9, 2018 via email

koenkooi commented May 9, 2018

shadeslayer commented May 9, 2018

shadeslayer commented Jun 5, 2018

shadeslayer Apr 26, 2018 •

edited

Loading