Fixed the way the values that span two Dwords are decoded.
Based on the start and end indices of the field, the Dwords
are fetched and decoded accordingly.
v2: rename dw to qw in gen_field_iterator_next
and remove extra white space (Anuj)
v3: change all instances of dw to qw (Anuj)
Earlier, 64-bit fields (such as most pointers on Gen8+)
weren't decoded correctly. gen_field_iterator_next seemed
to walk one DWord at a time, sets v.dw, and then passes it
to field(). So, even though field() takes a uint64_t, we're
passing it a uint32_t (which gets promoted, so the top 32
bits will always be zero). This seems pretty bogus... (Ken)
Signed-off-by: Sirisha Gandikota <Sirisha.Gandikota@intel.com>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
From the Vulkan spec:
Size is the number of bytes to fill, and must be either a multiple of 4,
or VK_WHOLE_SIZE to fill the range from offset to the end of the buffer.
If VK_WHOLE_SIZE is used and the remaining size of the buffer is not a
multiple of 4, then the nearest smaller multiple is used.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Make gen_device_info a mutable structure so we can update the fields that
can be refined by querying the kernel (like subslices and EU numbers).
This patch does not make any functional change, it just makes
gen_get_device_info() fill a structure rather than returning a const
pointer.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reproduces this commit :
commit 0fb85ac08d
Author: Kenneth Graunke <kenneth@whitecape.org>
Date: Mon Jun 6 21:37:34 2016 -0700
i965: Use the correct number of threads for compute shaders.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This reproduces this commit :
commit 2213ffdb4b
Author: Kenneth Graunke <kenneth@whitecape.org>
Date: Mon Jun 6 21:37:34 2016 -0700
i965: Allocate scratch space for the maximum number of compute threads.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Gen6 only has one additional restriction over Gen7+, so we just add it
to the existing gen7 function (which actually covers later gens too).
This should stop FINISHME spew when running GL on Sandybridge.
v2: Fix bytes per block vs. bits per block confusion (Jason) and
rename function to gen6_filter_tiling (Jason and Chad).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Without this bit set, the value in "L3 Atomic Disable" won't get applied by
the hardware so we won't properly get L3 atomic caching.
Fixes dEQP-VK.spirv_assembly.instruction.compute.opatomic.compex and 198 of
the dEQP-VK.image.atomic_operations.* tests on HSW
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
The Vulkan driver sets 3DSTATE_DRAWING_RECTANGLE once to MAX_INT x MAX_INT
at the GPU initialization time and never sets it again. The GL driver sets
it every time the framebuffer changes. Originally, blorp set it to the
size of the drawing area but meant we had to set it back in the Vulkan
driver. Instead, we can easily just do that in the GL driver's blorp_exec
implementation and not set it in blorp core.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Previously, we relied on a driver hook for 3DSTATE_MULTISAMPLE. However,
now that Vulkan and GL use the same sample positions, we can set up
3DSTATE_MULTISAMPLE directly in blorp and delete the driver hook.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Earlier, the loop pretends to loop over instructions from "start" to "end",
but the callers always pass 8192 for end, which is some huge bogus
value. The real loop termination condition is send-with-EOT or 0. (Ken)
Signed-off-by: Sirisha Gandikota <Sirisha.Gandikota@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Skylake adds new SENDS and SENDSC opcodes, which should be
handled in the send-with-EOT check. Make an is_send() helper
that checks if the opcode is SEND/SENDC/SENDS/SENDSC (Ken)
v2: Make is_send() much more crispier, Mix declaration and
code to make the code compact (Ken)
Signed-off-by: Sirisha Gandikota <Sirisha.Gandikota@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Remove the float/dword union and use the iter->p[f->start / 32]
directly as printf formatter %08x expects uint32_t (Ken)
v2: Make the cleanup much more crispier (Ken)
Signed-off-by: Sirisha Gandikota <Sirisha.Gandikota@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Since Vulkan doesn't allow single-slice 3D storage images, we need to just
set the base_array_layer and array_len to the full size of the 3-D LOD.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
This reverts commit 3943888c94. It turns out
that commit was pretty-much bogus since it breaks binding a 3-D texture as a
2-D storage image. The correct fix for the Vulkan CTS tests needs to be in
the Vulkan driver itself rather than ISL.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Everything that we were once using the blit2d framework for is now done
with blorp.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
This is a lot cleaner and easier to read than the old piles of if
statements.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Nanley Chery <nanley.g.chery@intel.com>
This allows us to #undef them later if we don't want them to persist
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
This allows us to #undef them later if we don't want them to persist
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
The original blorp_alloc_binding_table helper was supposed to return the
binding table offset and map along with the surface state maps. This isn't
quite what we want, however. What we really want is the binding table
offsets, surface state offsets, and surface state maps. In the GL driver,
the binding table map *is* an array of surface state offsets. However, in
Vulkan, this isn't quite true as the entries in the binding table are
surface state offsets combined with another binding table block offset.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Previously, we were relying on the fact that VALGRIND_MEMPOOL_FREE came
later on in the function to prevent "link->bo = bo" from causing an invalid
write. However, in the case where the size requested by the user is very
small (less than sizeof(struct anv_bo)), this isn't sufficient. Instead,
we should call VALGRIND_MEMPOOL_FREE early and then use VG_NOACCESS_WRITE.
We do, however, have to call VALGRIND_MEMPOOL_FREE after reading bo_in
because it may be stored in the bo itself.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
The time we want to restrict the Z range of a 3-D surface is when rendering
to it. For storage surfaces, we always want he full range. However, we
still need to set MinimumArrayElement and RenderTargetViewExtent to
sensible values so we'll just set them to the reasonable defaults we used
before we started respecting the base_array_layer and array_len.
This fixes a bunch of Vulkan CTS regressions caused by 48f195d7c6.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=97790
Reviewed-by: Chad Versace <chadversary@chromium.org>
Normally, using a non-linear tiling format helps improve cache locality by
ensuring that neighboring pixels are usually close-by in memory. For RGB
formats, this still sort-of holds, but it can also lead to rather terrible
memory access patterns where a single RGB pixel value crosses a tile
boundary and gets split into two pieces in different 4K pages. It also
makes for some rather awkward calculations because your tile size is no
longer an even multiple of surface element size. For these reasons, we
chose to simply never create tiled RGB images in the Vulkan driver.
The GL driver, however, is not so kind so we need to support it somehow. I
briefly toyed with a couple of different schemes but this is the best one I
could come up with. The fundamental problem is that a tile no longer
contains an integer number of surface elements. I briefly considered a
couple other options but found them wanting:
1) Using floats for the logical tile size. This leads to potential
rounding error problems.
2) When presented with a RGB format, just make the tile 3-times as wide.
This isn't so nice because now our tiles are no longer power-of-two
size. Also, it can force the row_pitch to be larger than needed which,
while not strictly a problem for ISL, causes incompatibility problems
with the way the GL driver chooses surface pitches.
The chosen method requires that you pay attention and not just assume that
your tile_info is in the units you think it is. However, it's nice because
it provides a nice "these are the units" declaration in isl_tile_info
itself. Previously, the tile_info wasn't usable as a stand-alone structure
because you had to also know the format. It also forces figuring out how
to deal with inconsistencies between tiling and format back to the caller
which is good because the two different consumers of isl_tile_info really
want to deal with it differently: Computation of the surface size wants
the fewest number of horizontal tiles possible while get_intratile_offset
is far more concerned with things aligning nicely.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Acked-by: Chad Versace <chadversary@chromium.org>
The restriction that Y-tiled surfaces must have valign == 4 only aplies to
render targets but we were applying it universally. This causes problems
if ISL_FORMAT_R32G32B32_FLOAT is used because it requires valign == 2; this
should be okay because you can't render to that format.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Chad Versace <chadversary@chromium.org>
When Ivy Bridge introduced array multisampling, someone made the decision
to do lots of stuff throughout the driver in terms of physical array layers
rather than logical array layers. In ISL, we use logical array layers most
of the time and it really makes no sense to use physical array layers in
the blorp API. Every time someone passes physical array layers into blorp
for an array multisampled surface, they're always divisible by the number
of samples and we divide right away.
Eventually, I'd like to rework most of the GL driver internals to use
logical array layers but that's going to be a big project and will probably
happen as part of the ISL conversion. For now, we'll do the conversion in
brw_blorp and let blorp just use the logical layers.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
The result of this calculation goes into an fma() in the shader and we
would like it to be as precise as possible. The division in particular
was a source of imprecision whenever dst1 - dst0 was not a power of two.
This prevents regressions in some of the new Vulkan CTS tests for blitting
using a filtering of NEAREST.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
While we're here, we also re-arrange the parameters to better match the
parameter order of blorp_blit.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
