If we have two (or more) smooth varyings like this:
nop t3; ldvary.rf0
fmul t5, t3, t0
fadd t6, t5, r5
nop t7; ldvary.rf0
fmul t9, t7, t0
fadd t10, t9, r5
nop t11; ldvary.rf0
fmul t13, t11, t0
fadd t14, t13, r5
We may be able to pipeline them like this:
nop ; nop ; ldvary.r4
nop ; fmul r0, r4, rf0 ; ldvary.r1
fadd rf13, r0, r5 ; fmul r2, r1, rf0 ; ldvary.r3
fadd rf12, r2, r5 ; fmul r4, r3, rf0 ; ldvary.r0
But in order to do this, we will need to manually tweak the
QPU scheduling.
This patch tracks information about ldvary sequences that are
good candidates for pipelining, and a follow-up patch will
use this information to pipeline them when we emit the QPU
code.
v2 (apinheiro):
- Rename the v3d_compile fields to avoid confusion with the qinst fields.
- Assert that a sequence's start instruction is not the same as the end.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9304>
If a new UBO load happens to read exactly at the offset right after the
previous UBO load (something that is fairly common, for example when
reading a matrix), we can skip the unifa write (with its 3 delay slots)
and just continue to call ldunifa to continue reading consecutive addresses.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9128>
Currently this is useful to clean up after DCEing leading ldunifa
instructions, but it can be expanded to handle more cases which
may allow to simplify the compiler code in places where we have
been trying to optimize manually for similar cases.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9128>
If we emit a new uniform and that uniform has already been emitted
in the same block we can just reuse that.
There is a balancing game here between reducing ldunif instructions
and not increasing register pressure too much though, so we put
a limit to how far back we are willing to look for a previous
definition of the uniform. Based on shader-db results, 20 instructions
produces best results.
total instructions in shared programs: 14928266 -> 14907432 (-0.14%)
instructions in affected programs: 6431841 -> 6411007 (-0.32%)
helped: 15270
HURT: 10772
Instructions are helped.
total uniforms in shared programs: 3944672 -> 3840276 (-2.65%)
uniforms in affected programs: 1827184 -> 1722788 (-5.71%)
helped: 30423
HURT: 845
Uniforms are helped.
total inst-and-stalls in shared programs: 14957813 -> 14936873 (-0.14%)
inst-and-stalls in affected programs: 6475349 -> 6454409 (-0.32%)
helped: 15287
HURT: 10852
Inst-and-stalls are helped.
v2 (Eric):
- consider ldunifrf too
- check that no other instruction writes to the register
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9077>
V3D 3.x has V3D_QPU_WADDR_TMU which in V3D 4.x is V3D_QPU_WADDR_UNIFA
(which isn't a TMU write address). This change passes a devinfo to
any functions that need to do these checks so we can account for the
target V3D version correctly.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8980>
The GL frontend can lower this weird GL feature away for us. This should
fix redeclaration of the gl_Color/SecondaryColor as centroid, since that
case had been missed in the !flat special case here.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8601>
We have been trying to avoid this by tracking fifo usages in the driver and
flushing all outstanding TMU sequences if we overflowed any of these, however,
this is actually not the most efficient strategy. Instead, we would like to
flush only enough operations to get things going again, which is better for
pipelining. Doing that in the driver would require some additional work, but
thankfully, it is not required, since this seems to be what the hardware does
automatically, so we can just remove overflow tracking for these two fifos
and enjoy the benefits.
This also further improves shader-db stats:
total instructions in shared programs: 8975062 -> 8955145 (-0.22%)
instructions in affected programs: 1637624 -> 1617707 (-1.22%)
helped: 4050
HURT: 2241
Instructions are helped.
total threads in shared programs: 236802 -> 237042 (0.10%)
threads in affected programs: 252 -> 492 (95.24%)
helped: 122
HURT: 2
Threads are helped.
total sfu-stalls in shared programs: 19901 -> 19592 (-1.55%)
sfu-stalls in affected programs: 4744 -> 4435 (-6.51%)
helped: 1248
HURT: 1051
Sfu-stalls are helped.
total inst-and-stalls in shared programs: 8994963 -> 8974737 (-0.22%)
inst-and-stalls in affected programs: 1636184 -> 1615958 (-1.24%)
helped: 4050
HURT: 2239
Inst-and-stalls are helped.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8825>
TMU pipelining can severely reduce our capacity to emit TMU spills,
causing us to fail to compile a shader we may otherwise be able to
compile. This is because pipelining extends the liveness of TMU
sequences by posponing the thread switch and LDTMU until a result
is needed, and we can't emit TMU spills while in the middle of a
TMU sequence.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8825>
This follows the same idea as for TMU general instructions of reusing
the existing infrastructure to first count required register writes and
flush outstanding TMU dependencies, and then emit the actual writes, which
requires that we split the code that decides about register writes to
a helper.
We also need to start using a component mask instead of the number
of components that we need to read with a particular TMU operation.
v2: update tmu_writes for V3D_QPU_WADDR_TMUOFF
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8825>
This creates the basic infrastructure to implement TMU pipelining and
applies it to general TMU. Follow-up patches will expand this
to texture and image/load store operations.
TMU pipelining means that we don't immediately end TMU sequences,
and instead, we postpone the thread switch and LDTMU (for loads)
or TMUWT (for stores) until we really need to do them.
For loads, we may need to flush them if another instruction reads
the result of a load operation. We can detect this because in that
case ntq_get_src() will not find the definition for that ssa/reg
(since we have not emitted the LDTMU instructions for it yet), so
when that happens, we flush all pending TMU operations and then
try again to find the definition for the source.
We also need to flush pending TMU operations when we reach the end
of a control flow block, to prevent the case where we emit a TMU
operation in a block, but then we read the result in another block
possibly under control flow.
It is also required to flush across barriers and discards to honor
their semantics.
Since this change doesn't implement pipelining for texture and
image load/store, we also need to flush outstanding TMU operations
if we ever have to emit one of these. This will be corrected with
follow-up patches.
Finally, the TMU has 3 fifos where it can queue TMU operations.
These fifos have limited capacity, depending on the number of threads
used to compile the shader, so we also need to ensure that we
don't have too many outstanding TMU requests and flush pending
TMU operations if a new TMU operation would overflow any of these
fifos. While overflowing the Input and Config fifos only leads
to stalls (which we want to avoid anyway), overflowing the Output
fifo is incorrect and would end up with a broken shader. This means
that we need to know how many TMU register writes are required
to emit a TMU operation and use that information to decide if we need
to flush pending TMU operations before we emit any register
writes for the new TMU operation.
v2: fix TMU flushing for NIR registers reads (jasuarez)
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8825>
Similarly to if statements, uniform loops are now emitted without
predication, using simple branches for breaks and continues. The
uniformity of the loop is determined by running the
nir_divergence_analysis pass.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7726>
So far the v3d compiler has them combined, as for OpenGL both are the
same. This change is intended to fit the v3d compiler better with
Vulkan, where they are separate concepts.
Note that NIR has them separate for a long time, both on nir_variable
and on some NIR lowerings.
v2: (from Iago feedback)
* Use key->num_tex/sampler_used to iterate through the array
* Fill up num_samplers_used on v3d, assert that is the same that
num_tex_used if possible.
v3: (Iago)
* Assert num_tex/samplers_used is smaller that tex/sampler array size.
v4: Update assert mentioned on v3 to use <= instead of < (detected by CI)
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
squash! broadcom/compiler: separate texture/sampler info from v3d_key
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7545>
So far the support for R/B swapping in vertex attributes were for the
generic attributes.
But there are cases like glSecondaryColorPointer() supporting BGRA
formats that require the R/B swapping to be also allowed in the
non-generic vertex attributes (in this case, in the COLOR1 attribute).
v2:
- Don't split line (Iago)
Signed-off-by: Juan A. Suarez Romero <jasuarez@igalia.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7196>
It is not used right now, so keeping it adds some noise/confusion.
So far configuring Z test are done through the CFG_BITS. See
v3dX(emit_state) at v3dx_emit.c for v3d, and pack_cfg_bits at
v3dv_pipeline.c for v3dv. There flags like z_updates_enable and others
are filled up.
That key field seems like a leftover coming from using vc4 as
reference, as that driver defines and uses a field with name name.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7421>
Some shaders that need to spill hundreds of registers can take very long times
to compile as each allocation attempt spills a single register and restarts
the allocation process. We can significantly cut down these times if we allow
the compiler to spill in batches, which should be possible if we are spilling
uniforms, which is in fact the kind of spills that we do first because they
have lower cost than TMU spills.
Doing this could cause us to slightly over spill in some cases (depending on
the chosen batch size) leading to slightly worse performance, so we only
enable this behavior after we have started to spill over a certain threshold,
at which point we assume that performance won't be good and we want to
favor compilation speed instead.
v2:
- Keep it simple and just try to spill a fixed amount of registers in a
batch instead of trying to compute this dynamically based on accumulated
spills and current register pressure. (Eric).
v3:
- Check if the node is valid before doing anything with it.
- Drop the environment variable to select batch size and just fix it to 20.
With this we can take this CTS test from 35 minutes down to about 3 minutes:
dEQP-VK.ssbo.layout.random.all_shared_buffer.5
Reviewed-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
For example, regarding gl_SampleID, the GLSL spec states:
"Any static use of this variable in a fragment shader causes the
entire shader to be evaluated per-sample."
So we need to track if the fragment shader does anything that implicitly
enables per-sample shading in the compiler for the driver to
auto-enable sample rate shading if needed.
v2:
- Instead of tracking reads of gl_SampleID, check SYSTEM_BIT_SAMPLE_ID
and SYSTEM_BIT_SAMPLE_POS as well as the sample layout qualifier like
other drivers are doing to activate this behavior (Eric).
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com> (v1)
Reviewed-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
We will need this in Vulkan to support vertex format
VK_FORMAT_B8G8R8A8_UNORM. The hardware doesn't allow to swizzle
vertex attribute components, so we need to do it in the shader.
v2:
- Use nir_intrinsic_io_semantics() to retrieve the location instead
of looping through the shader input variables (Eric).
- Assert that we only have one component (Eric).
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com> (v1)
Reviewed-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
Vulkan lowers gl_InstanceIndex to load_base_instance +
load_instance_id, so we need to implement loading the base instance in
the compiler.
The base instance is set by the BASE_VERTEX_BASE_INSTANCE command
right before the instanced draw call and it is included in the VPM
payload together with the InstanceID and VertexID if this is requested
by the shader record.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
We would need on OpenGL to update values for all the textures used. On
OpenGL that value can be always took from the context or the nir
shader, but there are cases on Vulkan that it is not the case, or
would force up to recompute it.
Acked-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
To ask to debug a registr allocation failure
(V3D_DEBUG_REGISTER_ALLOCATION seemed too long to me).
When a fallback register allocation algorithm was added, if the
register allocation fails, it only dumpg the current vir with the
register pressure info with the failed fallback. But if we want do
debug the problem, we would be interested on both.
Additionally, it was strange that we got the full vir dump with the
failure even if no debug option was set.
Additionally we add shaderdb like stats for those failures, to make
easier to compare one and the other.
v2: keep a small warning message in case both register allocation
algorithms fails (Neil)
Reviewed-by: Neil Roberts <nroberts@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6999>
v3d_compile is now split out into a helper function that gets called a
second time if compilation fails the first time with the result
reporting the register allocation failed. The second time it is run with
the fallback scheduler to try and increase the chances of successfully
allocating the registers.
v2: Add a performance debug message when using the fallback scheduler.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5953>
Instead of just having a bool status for the failure, there is now an
enum so that the compilation can report a more detailed status.
Currently this is only used to report whether the failure was due to
failed register allocation. The “failed” bool doesn’t seem to actually
have been used anywhere so this doesn’t really change a lot.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5953>
When line smoothing is enabled, the driver now increases the width of
the line so that it can add some semi-transparent pixels to either side
of the line. A lowering pass is added which modifies the alpha component
of every write to fragment output 0 so that if the fragment is outside
the width of the line then the alpha is reduced. It additionally
discards fragments that are completely invisible. It might seem bad to
use discard on a tiled renderer but the assumption is that any bad
effects from using discard will also happen anyway because of enabling
alpha blending.
v2: Disable the line smoothing pass entirely when the framebuffer
contains an integer colour output or one with no alpha channel.
Calculate the coverage once upfront and store in a global variable
instead of calculating each time an output write is modified. Also
do the conditional discard once upfront.
v3: Don’t check whether the output buffer has an alpha channel. Only
look at output 0. Use aa_line_width intrinsic instead of calculating
the real line width in the shader. Clamp the coverage as part of the
global variable, not per output write.
Reviewed-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5624>
When geometry shaders write a value to gl_Layer that doesn't correspond to
an existing layer in the target framebuffer the rendering behavior is
undefined according to the spec, however, there are CTS tests that trigger
this scenario on purpose, probably to ensure that nothing terrible happens.
For V3D, this situation is problematic because the binner uses the layer
index to select the offset to write into the tile state data, and we only
allocate tile state for MAX2(num_layers, 1), so we want to make sure we
don't produce values that would lead to out of bounds writes. The simulator
has an assert to catch this, although we haven't observed issues in actual
hardware it is probably best to play safe.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Geometry shaders can output many vertices and thus have higher VPM memory
pressure as a result. It is possible that too wide geometry shader dispatches
exceed the maximum available VPM output allocated, in which case we need
to reduce the dispatch width until we can fit the VPM memory requirements.
Supported dispatch widths for geometry shaders are 16, 8, 4, 1.
There is a limit in the number of VPM output sectors that can be used by a
geometry shader that we can meet by lowering the dispatch width at compile
time, however, at draw time we need to revisit this number and, together with
other elements that can contribute to total VPM memory requirements, decide
on a configuration that can fit the program into the available VPM memory.
Ideally, we also want to aim for not using more than half of the available
memory so we that we can run a pair of bin and render programs in parallel.
v2: fixed language in comment and typo in commit log. (Alejandro)
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Most of the relevant work happens in the v3d_nir_lower_io. Since
geometry shaders can write any number of output vertices, this pass
injects a few variables into the shader code to keep track of things
like the number of vertices emitted or the offsets into the VPM
of the current vertex output, etc. This is also where we handle
EmitVertex() and EmitPrimitive() intrinsics.
The geometry shader VPM output layout has a specific structure
with a 32-bit general header, then another 32-bit header slot for
each output vertex, and finally the actual vertex data.
When vertex shaders are paired with geometry shaders we also need
to consider the following:
- Only geometry shaders emit fixed function outputs.
- The coordinate shader used for the vertex stage during binning must
not drop varyings other than those used by transform feedback, since
these may be read by the binning GS.
v2:
- Use MAX3 instead of a chain of MAX2 (Alejandro).
- Make all loop variables unsigned in ntq_setup_gs_inputs (Alejandro)
- Update comment in IO owering so it includes the GS stage (Alejandro)
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Until now this made sense because we always paired vertex shaders
with fragment shaders, but as soon as we implement geometry and
tessellation shaders that will no longer be the case, so rename
this to (num_)used_outputs.
v2: Use 'used_outputs' instead of ns_outputs, which is more explicit (Eric).
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
That we set for any TMU write on spills and general tmu. It is then
used as part of v3d_emit_gl_shader_state later.
v2: add a new flag instead at v3d_compiler instead of dirty the flag
at v3dx if there is any spill (change suggested by Eric, added by
Alejandro)
v3: set this for anything that is not a load and do it also in
v3d40_vir_emit_image_load_store (Eric)
Reviewed-by: Eric Anholt <eric@anholt.net>
SFU operations have a latency of 2 cicles, so if their results
are used in the following cycle to a SFU instruction, the GPU
stalls for an extra cycle until the result is available.
This adds the number of stalls to the shader-db debug mode and
sum of instruction + stalls to evaluate optimizations to schedule
instructions that avoid generating sfu-stalls.
v2: Rename v3d_qpu_generates_sfu_stalls to v3d_qpu_instr_is_sfu (Eric)
Reviewed-by: Eric Anholt <eric@anholt.net>
Among other things, this avoid the need of loading 1/-1 constants (so
one less operation).
The removed comment suggest the option of adding support on NIR for
inc/dec. Intel just uses an auxiliar method to get which hw operation
is needed, so no lowering is needed. And at the same time, being so
small, seems unreasonable to try to add a general one on NIR
itself. It is more easy to just adapt the method here (that is what
the patch does right now).
It is worth to note that we are not getting any change on shader-db
stats because all those methods are used on the usual shader-db set
with shaders needing GLSL > 4.2. In general there aren't too many GLSL
ES 3.1 tests.
As an alternative, we captured the GLES3/GLSL31/GLS32 used on
vk-gl-cts, even if that is not a real life usage of shaders. With
those we get the following:
total instructions in shared programs: 1217022 -> 1217013 (<.01%)
instructions in affected programs: 117 -> 108 (-7.69%)
helped: 6
HURT: 0
helped stats (abs) min: 1 max: 2 x̄: 1.50 x̃: 1
helped stats (rel) min: 3.57% max: 10.00% x̄: 8.09% x̃: 9.09%
95% mean confidence interval for instructions value: -2.07 -0.93
95% mean confidence interval for instructions %-change: -10.54% -5.64%
Instructions are helped.
Note that the shaders helped are really low because most of the
vk-gl-cts tests using AtomicInc/Dec/Add are mostly used on compute
shaders. Although right now there is a branch around with CS support,
the usual is doing the stats against master.
Reviewed-by: Eric Anholt <eric@anholt.net>
