Instead of updating the clear color in anv before a resolve, just let
blorp handle that for us during fast clears.
v5: Update comment about HiZ clear color (Jordan).
Signed-off-by: Rafael Antognolli <rafael.antognolli@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
On Gen10+, instead of copying the clear color from the state buffer to
the surface state, just use the address of the state buffer in the
surface state directly. This way we can avoid the copy from state buffer
to surface state.
v4:
- Remove use_clear_address from anv code. (Jason)
- Use the helper to extract clear color from attachment (Jason)
Signed-off-by: Rafael Antognolli <rafael.antognolli@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Extract the code from color_attachment_compute_aux_usage, so we can
later reuse it to update the clear color state buffer.
Signed-off-by: Rafael Antognolli <rafael.antognolli@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
On Gen10, when emitting the surface state, use the value stored in the
clear color entry buffer by using a clear color address in the surface
state.
v4: Use the clear color offset from the clear_color_bo, when available.
Signed-off-by: Rafael Antognolli <rafael.antognolli@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
On Gen10, whenever we do a fast clear, blorp will update the clear color
state buffer for us, as long as we set the clear color address
correctly.
However, on a hiz clear, if the surface is already on the fast clear
state we skip the actual fast clear operation and, before gen10, only
updated the miptree. On gen10+ we need to update the clear value state
buffer too, since blorp will not be doing a fast clear and updating it
for us.
v4:
- do not use clear_value_size in the for loop
- Get the address of the clear color from the aux buffer or the
clear_color_bo, depending on which one is available.
- let core blorp update the clear color, but also update it when we
skip a fast clear depth.
v5: Better subject (Jordan).
v6: Remove outdated comment (Jason).
Signed-off-by: Rafael Antognolli <rafael.antognolli@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
In a follow up patch, we make use of clear_color_bo, which is in
mt->mcs_buf or mt->hiz_buf. To avoid duplicating more code that does the
same thing on both aux buffers, just use aux_buf already.
v5: Add aux_buf to brw_wm_surface_state too.
v6: Drop aux_surf and use aux_buf->surf instead (Jason).
Signed-off-by: Rafael Antognolli <rafael.antognolli@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Add an extra BO to store clear color when we receive the aux buffer from
the window system. Since we have no control over the aux buffer size in
this case, we need the new BO to store only the clear color.
v5:
- Better subject (Jordan).
- Drop alignment from brw_bo_alloc().
Signed-off-by: Rafael Antognolli <rafael.antognolli@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Similarly to vulkan where we store the clear value in the aux surface,
we can do the same in GL.
v2: Remove unneeded extra function.
v3: Use clear_value_state_size instead of clear_value_size.
v4:
- rename to clear_color_state_size
- store clear_color_bo and clear_color_offset in the aux buf struct
v5: Unreference clear color bo (Jordan)
Signed-off-by: Rafael Antognolli <rafael.antognolli@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
We always want to update the fast clear color during a fast clear on
i965. On anv, we are doing that before a resolve, but by adding support
to blorp, we can do a similar thing and update it during a fast clear
instead.
The goal is to remove some code from anv that does such update, and
centralize everything in blorp, hopefully removing a lot of code
duplication. It also allows us to have a similar behavior on gen < 9 and
gen >= 10.
v5: s/we/we are/ (Jordan)
Signed-off-by: Rafael Antognolli <rafael.antognolli@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
On gen10+, if surface->clear_color_addr is present, use it directly
intead of copying it to the surface state.
v4: Remove redundant #if clause for GEN <= 10 (Jason)
v5: Move flush after the reloc, and keep lower bits (Topi).
Signed-off-by: Rafael Antognolli <rafael.antognolli@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
gen10 can emit the clear color by setting it on a buffer somewhere, and
then adding only the address to the surface state.
This commit add support for that on isl_surf_fill_state, and if that is
requested, skip setting the clear value itself.
v2: Add assert to make sure we are at least on gen10.
Signed-off-by: Rafael Antognolli <rafael.antognolli@intel.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
The size of the clear color struct (expected by the hardware) is 8
dwords (isl_dev.ss.clear_value_state_size here). But we still need to
track the size of the clear color, used when memcopying it to/from the
state buffer. For that we keep isl_dev.ss.clear_value_size.
v4:
- Add struct to gen11 too (Jason, Jordan)
- Add field for Converted Clear Color to gen11 (Jason)
- Add clear_color_state_offset to differentiate from
clear_value_offset.
- Fix all the places where clear_value_size was used.
v5 (Jason):
- Split genxml changes to another commit.
- Remove unnecessary gen checks.
- Bring back missing offset increment to init_fast_clear_color().
v6 (Jason):
- On init_fast_clear_color, change:
addr.offset += 4 => sdi.Address.offset += i * 4
- Use GEN_GEN instead of GEN_VERSIONx10.
[jordan.l.justen@intel.com: isl_device_init changes]
Signed-off-by: Rafael Antognolli <rafael.antognolli@intel.com>
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
genxml does not support having two address fields with different names
but same position in the state struct. Both "Clear Color Address"
and "Clear Depth Address Low" mean the same thing, only for different
surface types.
To workaround this genxml limitation, rename "Clear Color Address"
to "Clear Value Address" and use it for both color and depth. Do the
same for the high bits.
TODO: add support for multiple addresses at the same position in the
xml.
v2: Combine high and low order bits into a single address field.
Signed-off-by: Rafael Antognolli <rafael.antognolli@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Some instructions contain fields that are either an address or a value
of some type based on the content of other fields, such as clear color
values vs address. That works fine if these fields are in the less
significant dword, the lower 32 bits of the address, because they get
OR'ed with the address. But if they are in the higher 32 bits, they get
discarded.
On Gen10 we have fields that share space with the higher 16 bits of the
address too. This commit makes sure those fields don't get discarded.
v5: Remove spurious whitespace (Jason).
Signed-off-by: Rafael Antognolli <rafael.antognolli@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
The lower bits seem to have extra fields in every platform but gen8
(even though we don't use them in gen9). So just go ahead and avoid
using them for the address.
v4: Use Jason's suggestion for comment explaining the change.
v5: Fix aux_address comment in anv_private.h (Jason)
Signed-off-by: Rafael Antognolli <rafael.antognolli@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Fix use of alloca() without #include <c99_alloca.h> in 1da345e5
vbo/vbo_context.c: In function '_vbo_draw_indirect':
vbo/vbo_context.c:284:34: error: implicit declaration of function 'alloca' [-Werror=implicit-function-declaration]
struct _mesa_prim *space = alloca(draw_count*sizeof(struct _mesa_prim));
^~~~~~
vbo/vbo_context.c:284:34: warning: initialization makes pointer from integer without a cast [-Wint-conversion]
Signed-off-by: Jon Turney <jon.turney@dronecode.org.uk>
Reviewed-by: Mathias Fröhlich <mathias.froehlich@web.de>
Disabled by default for now, it can be enabled with
RADV_PERFTEST=outoforder.
No CTS regressions on Polaris, and all Vulkan games I tested
look good as well.
Expect small performance improvements for applications where
out-of-order rasterization can be enabled by the driver.
Loosely based on RadeonSI.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
../src/intel/compiler/brw_reg.h: In function ‘bool brw_regs_negative_equal(const brw_reg*, const brw_reg*)’:
../src/intel/compiler/brw_reg.h:305:1: warning: control reaches end of non-void function [-Wreturn-type]
Introduced by 8f83eea71e ("i965: Add negative_equals methods").
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
From the SPIR-V spec, OpTypeImage:
"Depth is whether or not this image is a depth image. (Note that
whether or not depth comparisons are actually done is a property of
the sampling opcode, not of this type declaration.)"
The sampling opcodes that specify depth comparisons are
OpImageSample{Proj}Dref{Explicit,Implicit}Lod, so we should set
is_shadow only for these (we were using the deph property of the
image until now).
v2:
- Do the same for OpImageDrefGather.
- Set is_shadow to false if the sampling opcode is not one of these (Jason)
- Reuse an existing switch statement instead of adding a new one (Jason)
Fixes crashes in:
dEQP-VK.spirv_assembly.instruction.graphics.image_sampler.depth_property.*
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: mesa-stable@lists.freedesktop.org
Otherwise we may end up trying to coalesce in a case such as
ssa_1 = fadd r1, r2
r3.x = fneg(r2);
r3 = vec4(ssa_1, ssa_1.y, ...)
and that would cause us to move the writes to r3 from the vec to the
fadd which would re-order them with respect to the write from the fneg.
In order to solve this, we just don't coalesce if the destination of the
vec is not SSA. We could try to get clever and still coalesce if there
are no writes to the destination of the vec between the vec and the ALU
source. However, since registers only come from phi webs and indirects,
the chances of having a vec with a register destination that is actually
coalescable into its source is very slim.
Shader-db results on Haswell:
total instructions in shared programs: 13657906 -> 13659101 (<.01%)
instructions in affected programs: 149291 -> 150486 (0.80%)
helped: 0
HURT: 592
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105440
Fixes: 2458ea95c5 "nir/lower_vec_to_movs: Coalesce movs on-the-fly when possible"
Reported-by: Vadym Shovkoplias <vadym.shovkoplias@globallogic.com>
Tested-by: Vadym Shovkoplias <vadym.shovkoplias@globallogic.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
