diff options
author | Rob Clark <[email protected]> | 2020-05-27 13:50:05 -0700 |
---|---|---|
committer | Marge Bot <[email protected]> | 2020-05-29 00:38:28 +0000 |
commit | 6f391262003e2d58395dd17d2cf1e1a6807f7a0a (patch) | |
tree | 63dd56418d0a6a6db466b2668aa48658a1cc4823 /src/freedreno | |
parent | a3947f9d247619043ac9a2c17f746d2fbfb0e5ac (diff) |
freedreno/a6xx: document LRZ flag buffer
Doesn't seem to be a big win, although I could still be missing
something in my implementation. But might as well add the
documentation.
Signed-off-by: Rob Clark <[email protected]>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5217>
Diffstat (limited to 'src/freedreno')
-rw-r--r-- | src/freedreno/registers/a6xx.xml | 33 |
1 files changed, 32 insertions, 1 deletions
diff --git a/src/freedreno/registers/a6xx.xml b/src/freedreno/registers/a6xx.xml index a718b24149f..72c0c384fca 100644 --- a/src/freedreno/registers/a6xx.xml +++ b/src/freedreno/registers/a6xx.xml @@ -1973,7 +1973,7 @@ to upconvert to 32b float internally? <bitfield name="LRZ_WRITE" pos="1" type="boolean"/> <doc>update MAX instead of MIN value, ie. GL_GREATER/GL_GEQUAL</doc> <bitfield name="GREATER" pos="2" type="boolean"/> - <bitfield name="UNK3" pos="3" type="boolean"/> + <bitfield name="FC_ENABLE" pos="3" type="boolean"/> <!-- set when depth-test + depth-write enabled --> <bitfield name="Z_TEST_ENABLE" pos="4" type="boolean"/> </reg32> @@ -1988,6 +1988,37 @@ to upconvert to 32b float internally? <bitfield name="PITCH" low="0" high="10" shr="5" type="uint"/> <bitfield name="ARRAY_PITCH" low="11" high="21" shr="5" type="uint"/> <!-- ??? --> </reg32> + + <!-- + The LRZ "fast clear" buffer is initialized to zero's by blob, and + read/written when GRAS_LRZ_CNTL.FC_ENABLE (b3) is set. It appears + to store 1b/block. It appears that '0' means block has original + depth clear value, and '1' means that the corresponding block in + LRZ has been modified. Ignoring alignment/padding, the size is + given by the formula: + + // calculate LRZ size from depth size: + if (nr_samples == 4) { + width *= 2; + height *= 2; + } else if (nr_samples == 2) { + height *= 2; + } + + lrz_width = div_round_up(width, 8); + lrz_heigh = div_round_up(height, 8); + + // calculate # of blocks: + nblocksx = div_round_up(lrz_width, 16); + nblocksy = div_round_up(lrz_height, 4); + + // fast-clear buffer is 1bit/block: + fc_sz = div_round_up(nblocksx * nblocksy, 8); + + In practice the blob seems to switch off FC_ENABLE once the size + increases beyond 1 page. Not sure if that is an actual limit or + not. + --> <reg32 offset="0x8106" name="GRAS_LRZ_FAST_CLEAR_BUFFER_BASE_LO"/> <reg32 offset="0x8107" name="GRAS_LRZ_FAST_CLEAR_BUFFER_BASE_HI"/> <reg32 offset="0x8106" name="GRAS_LRZ_FAST_CLEAR_BUFFER_BASE" type="waddress"/> |