#
✕ nick:alyssa
#asahi-gpu · 2024-01-09 13:53
alyssaso only layer 0 is ever touched
#asahi-gpu · 2024-01-09 13:53
alyssai509vcb: in those copies, you have layers = 1 in the UAPI, and the end-of-tile ("Store pipeline") writes to a non-arrayed 2D PBE descriptor with a non-layered 2D image write
#asahi-gpu · 2024-01-09 13:51
alyssawe should probably make agxdecode smarter here but, effort
#asahi-gpu · 2024-01-09 13:50
alyssathe hardware decodes the layout (Linear) first
#asahi-gpu · 2024-01-09 13:50
alyssai509vcb: what you're seeing is that the Depth field overlaps the stride field
#asahi-gpu · 2024-01-09 03:17
alyssatrace?
#asahi-gpu · 2024-01-09 00:22
alyssa^^
#asahi-gpu · 2024-01-08 22:30
alyssa2 bugs, why not?
#asahi-gpu · 2024-01-06 21:36
alyssaNVK is the reference drivr
#asahi-gpu · 2024-01-06 01:23
alyssaread a bit more carefully
#asahi-gpu · 2024-01-05 18:57
alyssaagx_remap_varyings_vs control that
#asahi-gpu · 2024-01-05 18:56
alyssaassume it's after clip distance I guess
#asahi-gpu · 2024-01-05 18:56
alyssadoesn't remember the ordering off hand
#asahi-gpu · 2024-01-05 18:55
alyssabut viewport tgt + render tgt (== layer) work like clip distance as described there
#asahi-gpu · 2024-01-05 18:55
alyssaalso would appreciate a doc patch to https://docs.mesa3d.org/drivers/asahi.html#vertex-shader with what you've learned :)
#asahi-gpu · 2024-01-05 18:54
alyssasee the GL driver logic
#asahi-gpu · 2024-01-05 18:53
alyssa(there might be other bugs, but that's the only obvious one)
#asahi-gpu · 2024-01-05 18:52
alyssa`Output select` is missing the viewport target + render target bits to consume the fixed-function layer
#asahi-gpu · 2024-01-05 18:51
alyssathe VS is writing but it's not getting routed
#asahi-gpu · 2024-01-05 18:51
alyssaer, rather
#asahi-gpu · 2024-01-05 18:51
alyssai509vcb: the FS is reading the LAYER, but the VS isn't writing it
#asahi-gpu · 2024-01-05 17:31
alyssaknedlik: yep, branch agxv/main
#asahi-gpu · 2024-01-05 11:06
alyssanir_load_input(b, nir_imm_int(b, 0), .component = 0, .base = /* you get the idea */)
#asahi-gpu · 2024-01-05 11:05
alyssanir_intrinsic_instr_create is the "old" way, nowadays we prefer the much more concise helpers
#asahi-gpu · 2024-01-05 11:05
alyssai509vcb: https://gitlab.freedesktop.org/i509VCB/mesa/-/commit/02528b81be23d2b71ad8542f1e66c286e13df83d#78939f2a54b94a5f882175ecc0ac1f25b521df38_367_394
#asahi-gpu · 2024-01-05 11:04
alyssain which case, drop the u2f32 (unsigned to float) conversion
#asahi-gpu · 2024-01-05 11:04
alyssaunless it's meant to be an integer
#asahi-gpu · 2024-01-05 11:03
alyssanir_fadd
#asahi-gpu · 2024-01-05 11:03
alyssai509vcb: https://gitlab.freedesktop.org/i509VCB/mesa/-/commit/02528b81be23d2b71ad8542f1e66c286e13df83d#ccd82ce6d5431b9371915af7969698022e62444e_120_121
#asahi-gpu · 2024-01-04 23:11
alyssagrep nouveau/vulkan for load_layer_id to see how
#asahi-gpu · 2024-01-04 23:10
alyssas/the driver/agxv/
#asahi-gpu · 2024-01-04 23:10
alyssawhich we'll need to lower to load_input() in the driver
#asahi-gpu · 2024-01-04 23:10
alyssavk_meta_blit_resolve uses nir_load_layer_id
#asahi-gpu · 2024-01-04 23:10
alyssai509vcb: load the layer ID
#asahi-gpu · 2024-01-04 23:09
alyssano swizzling
#asahi-gpu · 2024-01-04 23:09
alyssai509vcb: vk copies are like memcpys
#asahi-gpu · 2024-01-04 20:18
alyssashould be compatible with e.g. rgba8_unorm?
#asahi-gpu · 2024-01-04 20:18
alyssai509vcb: what's wrong with bgra_snorm?
#asahi-gpu · 2024-01-04 19:58
alyssasigh
#asahi-gpu · 2024-01-04 19:56
alyssa..and growing
#asahi-gpu · 2024-01-04 19:56
alyssathe full thing is millions of tests
#asahi-gpu · 2024-01-04 19:55
alyssaj`ey: the vulkan cts is.... intense.
#asahi-gpu · 2024-01-04 18:42
alyssai509vcb: awesome!
#asahi-gpu · 2024-01-04 13:09
alyssathen in agx.agxv_CmdEndRendering.0012, we're downloading flots as floats, again canonicalizing/flushing
#asahi-gpu · 2024-01-04 13:08
alyssain e.g. agx.agxv_CmdEndRendering.0002 we're uploading floats as floats so the NaNs are getting canonicalized (and denorms flushed)
#asahi-gpu · 2024-01-04 13:07
alyssai509vcb: You need to do the same snorm/float casting logic for CopyBufferToImage and for CopyImageToBuffer, otherwise those will canonicalize NaNs/-1.0 too
#asahi-gpu · 2024-01-04 01:41
alyssahappy to take a look at the trace
#asahi-gpu · 2024-01-04 01:41
alyssathanks
#asahi-gpu · 2024-01-04 01:40
alyssaagxdecode?
#asahi-gpu · 2024-01-04 01:40
alyssa:S

plus ›