Instead of removing every instruction in add_insts_from_block(), just
move the instruction to its scheduled location. This is a step towards
doing both bottom-up and top-down scheduling without conflicts.
Note that this patch changes cycle counts for programs because it begins
including control flow instructions in the estimates.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
I think when this code was written, basic blocks were always ended by a
control flow instruction or an end-of-thread message. That's no longer
the case, and removing this restriction actually helps things:
instructions in affected programs: 7267 -> 7244 (-0.32%)
helped: 4
total cycles in shared programs: 66559580 -> 66431900 (-0.19%)
cycles in affected programs: 28310152 -> 28182472 (-0.45%)
helped: 9577
HURT: 879
GAINED: 2
The addition of the is_control_flow() checks is not a functional change,
since the add_insts_from_block() does not put them in the list of
instructions to schedule. I plan to change this in a later patch.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Missing this causes an assertion failure in the scheduler with the next
patch.
Additionally, this gives cmod propagation enough information to optimize
code better.
total instructions in shared programs: 7112991 -> 7112852 (-0.00%)
instructions in affected programs: 25704 -> 25565 (-0.54%)
helped: 139
total cycles in shared programs: 64812898 -> 64810674 (-0.00%)
cycles in affected programs: 127224 -> 125000 (-1.75%)
helped: 139
Acked-by: Francisco Jerez <currojerez@riseup.net>
This reverts commit d0e1d6b7e2.
The change in the vec4 code is a mistake -- there's never an
FS_OPCODE_FB_WRITE in vec4 code.
The change in the fs code had the (harmless) effect of not recognizing
an FB_WRITE as a scheduling barrier even if it was marked EOT --
harmless because the scheduler marked the last instruction of a block as
a barrier, something I'm changing in the following patches.
This will be reimplemented later in the series.
All of these were simply code for "architecture register file" (and in
the case of destinations, "not the null register").
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
These printed the cycle count the last basic block (sched.time is set
per basic block!). We have accurate, full program, data printed
elsewhere.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Failed to update state tracker with new buffer interface.
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This enables in shader defined transform feedback mode even if the
only place xfb_stride is defined is on the global out.
We don't worry about xfb_buffer since Issue 22 c) in the spec says:
"If the shader has an "xfb_buffer" qualifier identifying a buffer,
but doesn't declare "xfb_offset" on anything associated with it,
what happens?
...
variables not qualified with "xfb_offset" are not captured, which
makes the associated "xfb_buffer" qualifier irrelevant."
Reviewed-by: Dave Airlie <airlied@redhat.com>
When we move to the next buffer we need to reset the stream
so that we don't generate an error message about streams not
matching.
Reviewed-by: Dave Airlie <airlied@redhat.com>
This moves the check until after we have done the stride
calculation and applies it to the xfb_* qualifiers.
Reviewed-by: Dave Airlie <airlied@redhat.com>
From the ARB_enhanced_layous spec:
"It is a compile-time or link-time error to have any *xfb_offset*
that overflows *xfb_stride*, whether stated on declarations before
or after the *xfb_stride*, or in different compilation units.
...
When no *xfb_stride* is specified for a buffer, the stride of a
buffer will be the smallest needed to hold the variable placed at
the highest offset, including any required padding."
Reviewed-by: Dave Airlie <airlied@redhat.com>
Here we use the built-in validation in
ast_layout_expression::process_qualifier_constant() to check for mismatching
global out strides on buffers in a single shader.
From the ARB_enhanced_layouts spec:
"While *xfb_stride* can be declared multiple times for the same buffer,
it is a compile-time or link-time error to have different values
specified for the stride for the same buffer."
For intrastage validation a new helper link_xfb_stride_layout_qualifiers()
is created. We also take this opportunity to make sure stride is at least
a multiple of 4, we will validate doubles at a later stage.
From the ARB_enhanced_layouts spec:
"If the buffer is capturing any double-typed outputs, the stride must
be a multiple of 8, otherwise it must be a multiple of 4, or a
compile-time or link-time error results."
Finally we update store_tfeedback_info() to apply the strides to
LinkedTransformFeedback and update the buffers bitmask to mark any global
buffers with a stride as active. For example a shader with:
layout (xfb_buffer = 0, xfb_offset = 0) out vec4 gs_fs;
layout (xfb_buffer = 1, xfb_stride = 64) out;
Is expected to have a buffer bound to both 0 and 1.
From the ARB_enhanced_layouts spec:
"A binding point requires a bound buffer object if and only if its
associated stride in the program object used for transform feedback
primitive capture is non-zero."
Reviewed-by: Dave Airlie <airlied@redhat.com>
This will be used in a following patch to implement interface
query support for TRANSFORM_FEEDBACK_BUFFER.
Reviewed-by: Dave Airlie <airlied@redhat.com>
This allows us to print the correct binding point when not all
buffers declared in the shader are bound.
For example if we use a single buffer:
layout(xfb_buffer=2, offset=0) out vec4 v;
We now print '2' when the buffer is not bound rather than '0'.
Reviewed-by: Dave Airlie <airlied@redhat.com>
This adds the initial infrastructure for enabling transform feedback
mode via in shader qualifiers and adds initial buffer support.
Reviewed-by: Dave Airlie <airlied@redhat.com>
This function checks for any xfb_* qualifiers which will enable
transform feedback mode and cause any API defined xfb varyings
to be ignored.
It also counts the number of varyings that have a xfb_offset
qualifier and finally it calls the create_xfb_varying_names()
helper to generate the names of varyings to be caputured.
Reviewed-by: Dave Airlie <airlied@redhat.com>
This will be used to get a count of the number of varying name
strings we are required to generate for use with the query api.
Reviewed-by: Dave Airlie <airlied@redhat.com>
When we have an interface block like:
layout (xfb_buffer = 0, xfb_offset = 0) out Block {
vec4 var1;
layout (xfb_stride = 32) vec4 var2;
vec4 var3;
};
We take into account the stride of var2 when calculating the offset
for var3.
Reviewed-by: Dave Airlie <airlied@redhat.com>
From the ARB_enhanced_layouts spec:
"The *xfb_stride* qualifier specifies how many bytes are consumed
by each captured vertex. It applies to the transform feedback
buffer for that declaration, whether it is inherited or explicitly
declared. It can be applied to variables, blocks, block members,
or just the qualifier out. If the buffer is capturing any
double-typed outputs, the stride must be a multiple of 8, otherwise
it must be a multiple of 4, or a compile-time or link-time error
results.
...
The resulting stride (implicit or explicit) must be less than or
equal to the implementation-dependent constant
gl_MaxTransformFeedbackInterleavedComponents."
Reviewed-by: Dave Airlie <airlied@redhat.com>
Also copies the qualifier values to GLSL IR.
From the ARB_enhanced_layouts spec:
"The *xfb_buffer* qualifier can be applied to the qualifier out,
to output variables, to output blocks, and to output block
members. Shaders in the transform feedback capturing mode have
an initial global default of
layout(xfb_buffer = 0) out;
This default can be changed by declaring a different buffer with
xfb_buffer on the interface qualifier out. This is the only way
the global default can be changed. When a variable or output block
is declared without an xfb_buffer qualifier, it inherits the global
default buffer. When a variable or output block is declared with an
xfb_buffer qualifier, it has that declared buffer. All members of a
block inherit the block's buffer. A member is allowed to declare
an xfb_buffer, but it must match the buffer inherited from its
block, or a compile-time error results.
The *xfb_buffer* qualifier follows the same conventions, behavior,
defaults, and inheritance rules as the qualifier stream, and the
examples for stream apply here as well. This includes a block's
inheritance of the current global default buffer, a block member's
inheritance of the block's buffer, and the requirement that any
*xfb_buffer* declared on a block member must match the buffer
inherited from the block.
...
It is a compile-time error to specify an *xfb_buffer* that is
greater than the implementation-dependent constant
gl_MaxTransformFeedbackBuffers."
Reviewed-by: Dave Airlie <airlied@redhat.com>
Since any of the xfb_* qualifiers trigger the shader to be in
transform feedback mode we need an extra field to track if
the xfb_buffer on interface members was set explicitly since
xfb_buffer will always have a default value.
Reviewed-by: Dave Airlie <airlied@redhat.com>
These will be used to hold qualifier values for interface and
struct members.
Support is added to the struct/interface constructors to copy these
fields upon creation.
We also update record_compare() to ensure we don't reuse a glsl_type
with the wrong xfb_* qualifier values.
Reviewed-by: Dave Airlie <airlied@redhat.com>
This adds validation for all qualifiers as allowed by the
table in Section 4.4 (Layout Qualifiers) of the GLSL 4.5 spec.
Reviewed-by: Dave Airlie <airlied@redhat.com>
We reuse the existing offset field for holding the xfb_offset
expression but create a new flag as to avoid hitting the rules
for the offset qualifier for UBOs.
xfb_buffer qualifiers require extra processing when merging as
they can be applied to global out defaults. We just apply the
same rules as we do for the stream qualifier as the spec says:
"The *xfb_buffer* qualifier follows the same conventions,
behavior, defaults, and inheritance rules as the qualifier
stream, and the examples for stream apply here as well."
For xfb_stride we push everything into a global out field for
later processing as xfb_stride applies to the entire buffer.
We still need to have a separate field to store per variable
strides because they can still effect implicit offsets
e.g. when applied to block members with implicit offsets.
Reviewed-by: Dave Airlie <airlied@redhat.com>
Firstly this updates the named interface lowering pass to store the
interface without the arrays removed.
Note we need to remove the arrays in the interface/varying matching
code to not regress things but in future this should be fixed
futher as it would seem we currently successfully match interface
blocks with differnt array sizes.
Since we now know if the interface was an array we can reduce the
IR flags from_named_ifc_block_array and from_named_ifc_block_nonarray
to just from_named_ifc_block.
Next rather than having a different code path for named interface
blocks in program_resource_visitor we just make use of the one used
by UBOs this allows us to now handle arrays of arrays correctly.
Finally we add a new param to the recursion function
named_ifc_member this is because we only want to process a single
member at a time. Note that this is also the glsl_struct_field
from the original ifc type before lowering rather than the type
from the lowered variable. This fixes a bug in Mesa where we would
generate the names like WithInstArray[0].g[0][0] when it should be
WithInstArray[0].g[0] for the following interface.
out WithInstArray {
float g[3];
} instArray[2];
Reviewed-by: Dave Airlie <airlied@redhat.com>
It seems expected that both lhs and rhs could be of type error_type
in this code however the TCS case wasn't expecting it.
Fixes segfault in an enhanced layouts GL CTS test.
Reviewed-by: Dave Airlie <airlied@redhat.com>
