I initially implemented frexp() as an IR opcode with a lowering pass,
but since it returns a value and has an out-parameter, it would break
assumptions our optimization passes make about ir_expressions being pure
(i.e., having no side effects).
For example, if opt_tree_grafting encounters this code:
uniform float u;
void main()
{
int exp;
float f = frexp(u, out exp);
float g = float(exp)/256.0;
float h = float(exp) + 1.0;
gl_FragColor = vec4(f, g, h, g + h);
}
it may try to optimize it to this:
uniform float u;
void main()
{
int exp;
float g = float(exp)/256.0;
float h = float(exp) + 1.0;
gl_FragColor = vec4(frexp(u, out exp), g, h, g + h);
}
Some hardware has an instruction which performs frexp(), but we would
need some other compiler infrastructure to be able to generate it, such
as an intrinsics system that would allow backends to emit specific code
for particular bits of IR.
Reviewed-by: Paul Berry <stereotype441@gmail.com>
Note the parameter name change in the int version of ir_constant, to
avoid the conflict with the loop iterator.
v2: Make analogous change to builtin_builder::imm().
Reviewed-by: Paul Berry <stereotype441@gmail.com>
These data structures are used for debug output, so it wasn't hurting
anything that there were missing bits. But it's good to keep things
up to date.
This patch also adds static asserts so that the {brw,cache}_bits[]
arrays are the proper size, so that we don't forget to add to them in
the future. Unfortunately there's no convenient way to assert that
mesa_bits[] is the proper size.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
If the geometry shader refers to the built-in variable
gl_PrimitiveIDIn, we need to set a bit in 3DSTATE_GS to tell the
hardware to dispatch primitive ID to r1, and we need to leave room for
it when allocating registers.
Note: this feature doesn't yet work properly when software primitive
restart is in use (the primitive ID counter will incorrectly reset
with each primitive restart, since software primitive restart works by
performing multiple draw calls). I plan to address that in a future
patch series.
Fixes piglit test "spec/glsl-1.50/execution/geometry/primitive-id-in".
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
When we previously implemented primitive restart, we didn't add cases
to brw_primitive_restart.c's can_cut_index_handle_prims() for the
primitive types that are introduced with geometry shaders. It turns
out that all of the new primitive types are supported by hardware
primitive restart.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
As part of its support for geometry shaders, GL 3.2 introduces four
new primitive types: GL_LINES_ADJACENCY, GL_LINE_STRIP_ADJACENCY,
GL_TRIANGLES_ADJACENCY, and GL_TRIANGLE_STRIP_ADJACENCY.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Simply adjust wrap mode to clamp_to_edge. This is all that's needed for a
correct implementation for nearest filtering, and it's way better than
using repeat wrap for instance for linear filtering (though obviously this
doesn't actually do seamless filtering).
v2: fix s/t wrap not r/s...
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Specifying a miptree layout makes no sense for constant buffers.
This has no functional change since BRW_SURFACE_MIPMAPLAYOUT_BELOW is
just a #define for 0.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
Now we use gen7_upload_constant_state() for all three shader stages.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
This paves the way for using gen7_upload_constant_state for PS data.
The formula is copied from gen7_wm_state.c.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
This saves a bit of typing and shortens a few lines.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
GL 3.2 requires us to support 128 varying components for geometry
shader outputs and fragment shader inputs, and 64 varying components
otherwise. But there's no hardware limitation that restricts us to 64
varying components, and core Mesa doesn't currently allow different
stages to have different maximum values, so just go ahead and enable
128 varying components for all stages. This gets us better test
coverage anyway.
Even though we are only working on GL 3.2 support for gen7 right now,
gen6 also supports 128 varying components, so go ahead and switch it
on there too.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Previously we only ever did 1 URB write, since the maximum number of
varyings we support is small enough to fit in 1 URB write (when using
BRW_URB_SWIZZLE_NONE, which is what the pre-Gen7 GS always uses). But
we're about to increase the number of varying components we support
from 64 to 128.
With 128 varyings, the most URB writes we'll have to do is 2, but it's
just as easy to write a general-purpose loop.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The "{VS,GS} URB Entry Allocation Size" fields of 3DSTATE_URB allow
values in the range 0-4, but they are U8-1 fields, so the range of
possible allocation sizes is 1-5. We were erroneously prohibiting a
size of 5.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Previously we only ever did 1 or 2 URB writes, since the maximum
number of varyings we support is small enough to fit in 2 URB writes.
But GL 3.2 requires the geometry shader to support 128 output varying
components, and this could require up to 3 URB writes.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Since the SF/SBE stage is only capable of performing arbitrary
reorderings of 16 varying slots, we can't arrange the fragment shader
inputs in an arbitrary order if there are more than 16 input varying
slots in use. We need to make sure that slots 16-31 match the
corresponding outputs of the previous pipeline stage.
The easiest way to accomplish this is to just make all varying slots
match up with the previous pipeline stage.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The for loop was rather silly. In addition to checking brw->gen < 6
on each loop iteration, it took pains to exclude bits from
fp->Base.InputsRead that don't correspond to fragment shader inputs.
But those bits would never have been set in the first place, since the
only bits that are ever set in fp->Base.InputsRead are fragment shader
inputs.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Now that the vertex shader output VUE map is determined solely by a
64-bit bitfield, we don't have to store it in its entirety in the
geometry shader program key; instead, we can just store the bitfield,
and let the geometry shader infer the VUE map at compile time.
This dramatically reduces the size of the geometry shader program key,
which we want to keep small since it gets recomputed whenever the
active program changes.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Previously, on Gen6+, we laid out the vertex (or geometry) shader VUE
map differently depending whether user clipping was active. If it was
active, we put the clip distances in slots 2 and 3 (where the clipper
expects them); if it was inactive, we assigned them in the order of
the gl_varying_slot enum.
This made for unnecessary recompiles, since turning clipping on/off
for a shader that used gl_ClipDistance might rearrange the varyings.
It also required extra bookkeeping, since it required the user
clipping flag to be provided to brw_compute_vue_map() as a parameter.
With this patch, we always put clip distances at in slots 2 and 3 if
they are written to. do_vs_prog() and do_gs_prog() are responsible
for ensuring that clip distances are written to when user clipping is
enabled (as do_vs_prog() previously did for gen4-5).
This makes the only input to brw_compute_vue_map() a bitfield of which
varyings the shader writes to, a fact that we'll take advantage of in
forthcoming patches.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Previously, if a fragment shader accessed gl_FragCoord or
gl_FrontFacing, we would assign them their own slots in the fragment
shader input attribute array, using up space that could be made
available to real varyings. This was not strictly necessary (since
these values are not true varyings, and are instead computed from
other data available in the FS payload). But we had to do it anyway
because the SF/SBE setup code assumed that every 1 bit in the
gl_program::InputsRead bitfield corresponded to a genuine varying
variable.
Now that the SF/SBE code consults brw_wm_prog_data and only sets up
the attributes that the fragment shader actually needs, we don't have
to do this anymore.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Previously, the SF/SBE setup code delivered varying inputs to the FS
in the order in which they appear in the gl_program::InputsRead
bitfield, since that's what the FS expects.
When we add support for more than 64 varying components, this will no
longer always be the case, because the Gen6+ SF/SBE stage is only
capable of performing arbitrary reorderings of 16 varying slots. So,
when there are more than 16 vec4's worth of varying inputs, the FS
will have to adjust the order its input varyings in order to partially
match the order of outputs from the geometry or vertex shader.
To allow extra flexibility in the ordering of FS varyings, this patch
causes the SF/SBE to deliver varying inputs to the FS in exactly the
order that the FS requests, by consulting brw_wm_prog_data::urb_setup
and brw_wm_prog_data::num_varying_inputs.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We always program the SF unit to start reading the vertex URB entry at
offset 1. In upcoming patches, we'll be adding FS code that relies on
this. So consistently use the constant BRW_SF_URB_ENTRY_READ_OFFSET
rather than hardcoding a 1.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Previously, we assumed that the number of varying inputs consumed by
the fragment shader was equal to the number of bits set in
gl_program::InputsRead. However, we'll soon be making two changes
that will cause that not to be true:
- We'll stop wasting varying input space for gl_FragCoord and
gl_FrontFacing, which aren't varyings.
- For fragment shaders that have more than 16 varying inputs, we'll
adjust the layout of the inputs to account for the fact that the
SF/SBE pipeline stage can't reorder inputs beyond the first 16; if
there are GS outputs that the FS doens't use (or vice versa) this
may cause the number of FS varying inputs to change.
So, instead of trying to guess the number of FS inputs from
gl_program::InputsRead, simply read it from
brw_wm_prog_data:num_varying_inputs, which is guaranteed to be correct
since it's populated by fs_visitor::calculate_urb_setup().
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
On gen4-5, the FS stage reads varying inputs from URB entries that
were output by the SF thread, where each register stores the
interpolation setup for two components of a vec4, therefore the FS
urb_read_length is twice the number of FS input varyings. On gen6+,
varying inputs are directly deposited in the FS payload by the SF/SBE
fixed function logic, so urb_read_length is irrelevant.
However, in future patches, it will be nice to be able to consult
brw_wm_prog_data to determine how many varying inputs the FS expects
(rather than inferring it from gl_program::InputsRead). So instead of
storing urb_read_length, we simply store num_varying_inputs in
brw_wm_prog_data. On gen4-5, we multiply this by 2 to recover the URB
read length.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
At the moment, for Gen6+, the FS assumes that all varying inputs are
delivered to it in the order in which they appear in the
gl_program::InputsRead bitfield, and the SF/SBE setup code ensures
that they are delivered in this order.
When we add support for more than 64 varying components, this will no
longer always be possible, because the Gen6+ SF/SBE stage is only
capable of performing arbitrary reorderings of 16 varying slots.
To allow extra flexibility in the ordering of FS varyings, this patch
causes the FS to advertise exactly what ordering it expects.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
For each sampler type, this tests that:
- The base type is GLSL_TYPE_SAMPLER.
- The dimensionality is set correctly.
- The returned data type is correct.
- The sampler_array and sampler_shadow flags are set correctly.
- sampler_coordinate_components() returns the correct value.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <idr@freedesktop.org>
It seems a user app can get us into this state, I trigger the fail
running fbo-maxsize inside virgl, it fails to create the backing
storage for the texture object, but then segfaults here when it
should fail the completeness test.
Cc: "9.2" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Brian Paul <brianp@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Fix the return type and allow src and dst types for comparison
to be separate, this at least fixes the two test cases I've written.
v2: drop the u32->s32 change
Acked-by: Christoph Bumiller <christoph.bumiller@speed.at>
Signed-off-by: Dave Airlie <airlied@redhat.com>
When the old contents do not need to be preserved, it is faster to
create a new backing bo rather than stall.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
max_index may be 0xffffffff. The hardware does not need 1 + max_index
(although it does not hurt unless max_index wraps around to zero).
Signed-off-by: Rob Clark <robclark@freedesktop.org>
For mem->gmem we don't sample depth/stencil as it's native type. So we
need to setup the swizzle state for the sampler based on the format used
for sampling.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Needed by some games, like etuxracer and supertuxkart which use alpha
test rather than blending, to handle texture transparency.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
With a debug option to force DIRECT (mainly to make it easier for
capturing cmdstream dumps). Using INDIRECT for large shaders at least
makes a noticable reduction in CPU load, which helps for CPU limited
games.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Because of how the tiling works, we can't really flush at arbitrary
points very easily. So wraparound is handled by resetting to top of
ringbuffer. Previously this would stall until current rendering is
complete. Instead cycle through multiple ringbuffers to avoid a stall.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Emit markers by writing to scratch registers in order to "triangulate"
gpu lockup position from post-mortem register dump. By comparing
register values in post-mortem dump to command-stream, it is possible to
narrow down which DRAW_INDX caused the lockup.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Have a single helper that all draws come through.. mainly for a
convenient debug and instrumentation point.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
