r600g: fix compile warnings in r600_cp_dma_copy_buffer on 32-bit gcc

r600g: fix up CP DMA for VM on cayman and TN

Need to add the virtual address.

Reviewed-by: Marek Olšák <maraeo@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

svga: use pipe_sampler_view_release() in svga_cleanup_tss_binding()

Fixes a crash when the Redway3D Turbine demo exits.  We've made this
change in other places in the past.  The root issue is texture objects
are being shared by multiple contexts and sampler views get shared too.
Sampler views have a context pointer and if that context gets deleted
we may try to reference that context when finally deleting the sampler
view.

pipe_sampler_view_release() avoids this problem because it takes
an explicit context.

Reviewed-by: Zack Rusin <zackr@vmware.com>

st/mesa: handle new GLSL IR enumerants in switch statements

To silence warnings about unhandled cases.

svga: add NULL pointer check in svga_create_sampler_state()

Note: This is a candidate for the 9.0 branch.

vbo: add a null pointer check to handle OOM instead of crashing

Note: This is a candidate for the 9.0 branch.

util: add new error checking code in vbuf helper

Check the return value of calls to u_upload_alloc() and
u_upload_data() and return early if needed.

Since we don't have a way to propagate errors all the way up to
Mesa through pipe_context::draw_vbo(), call debug_warn_once() so
the user might have some clue about OOM errors.

Note: This is a candidate for the 9.0 branch.

st/mesa: do proper error checking for u_upload_alloc() calls

We weren't properly checking the return value of these calls (and
calls to u_upload_data()) to detect OOM errors.

Note: This is a candidate for the 9.0 branch.

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>

util: add some defensive coding in u_upload_alloc()

Some callers of this function were checking the 'ptr' result to see if
the function failed.  But the correct way is to check the regular
return value for PIPE_ERROR_x.  Now we initialize all the returned
values at the top of the function in case we do hit an error (like OOM).

Callers are more likely to detect OOM conditions now.  But there
are some callers which don't do any error checking...

Note: This is a candidate for the 9.0 branch.

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>

glsl: use glsl_strtof() instead of glsl_strtod()

Since the result of those calls is always assigned to a float.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

glsl: add new glsl_strtof() function

Note, we could alternately implement this in terms of glsl_strtod()
with a (float) cast.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

softpipe: add casts to silence MSVC warnings

Reviewed-by: José Fonseca <jfonseca@vmware.com>

util: silence MSVC signed/unsigned comparison warnings

Reviewed-by: José Fonseca <jfonseca@vmware.com>

util: silence MSVC double->float conversion warnings

Reviewed-by: José Fonseca <jfonseca@vmware.com>

util: silence MSVC signed/unsigned warnings in debug_get_flags_option()

Reviewed-by: José Fonseca <jfonseca@vmware.com>

st/mesa: silence assorted MSVC warnings in DrawPixels code

Reviewed-by: José Fonseca <jfonseca@vmware.com>

swrast: silence a bunch of MSVC warnings

Reviewed-by: José Fonseca <jfonseca@vmware.com>

mesa: use GLbitfield64 when copying program inputs

Note: This is a candidate for the 9.0 branch.

Reviewed-by: José Fonseca <jfonseca@vmware.com>

mesa: add some casts to silence MSVC warnings

Reviewed-by: José Fonseca <jfonseca@vmware.com>

mesa: add casts in _mesa_GetTexParameterfv() to silence warnings

There are other similar int->float casts elsewhere in the function.

Reviewed-by: José Fonseca <jfonseca@vmware.com>

i965: Enable ARB_shading_language_packing

Reviewed-by: Chad Versace <chad.versace@linux.intel.com>
Reviewed-by: Paul Berry <stereotype441@gmail.com>

i965: Assert that the 4x8 pack/unpack operations have been lowered

Reviewed-by: Chad Versace <chad.versace@linux.intel.com>
Reviewed-by: Paul Berry <stereotype441@gmail.com>

i965: Lower the 4x8 pack/unpack operations

Reviewed-by: Chad Versace <chad.versace@linux.intel.com>
Reviewed-by: Paul Berry <stereotype441@gmail.com>

glsl: Add support for lowering 4x8 pack/unpack operations

Lower them to arithmetic and bit manipulation expressions.

Reviewed-by: Chad Versace <chad.versace@linux.intel.com>
Reviewed-by: Paul Berry <stereotype441@gmail.com>

glsl: Evaluate constant pack/unpack 4x8 expressions

That is, evaluate constant expressions for the following functions:
  packSnorm4x8, unpackSnorm4x8
  packUnorm4x8, unpackUnorm4x8

Reviewed-by: Chad Versace <chad.versace@linux.intel.com>
Reviewed-by: Paul Berry <stereotype441@gmail.com>

glsl: Extend ir_expression_operation for ARB_shading_language_packing

For each function {pack,unpack}{Snorm,Unorm}4x8, add a corresponding
opcode to enum ir_expression_operation. Validate the new opcodes in
ir_validate.cpp.

Reviewed-by: Chad Versace <chad.versace@linux.intel.com>
Reviewed-by: Paul Berry <stereotype441@gmail.com>

glsl: Add IR lisp for ARB_shading_language_packing

Reviewed-by: Chad Versace <chad.versace@linux.intel.com>
Reviewed-by: Paul Berry <stereotype441@gmail.com>

glsl: Add infrastructure for ARB_shading_language_packing

Reviewed-by: Chad Versace <chad.versace@linux.intel.com>
Reviewed-by: Paul Berry <stereotype441@gmail.com>

configure.ac: Don't set LLVM_LIBS when llvm is disabled

r600g: Don't build llvm_wrapper.cpp when we aren't using LLVM

We were using the NEED_RADEON_GALLIUM conditional to decide whether or not
to build llvm_wrapper.cpp, which is required for using the LLVM backend.
llvm_wrapper.cpp needs to be linked against the LLVM IPO libary
and this library is only added to LLVM_LIBS if either opencl or the
r600-llvm-compiler is enabled.

The NEED_RADEON_GALLIUM conditional is set to true when enabling the
radeonsi driver, so if the radeonsi and r600 drivers are enabled without
also enabling opencl or r600-llvm-compiler, llvm_wrapper.cpp will be
built, but the IPO library won't be added to LLVM_LIBS.  This was
causing unresolved symbol errors when buiding with this configuration.

https://bugs.freedesktop.org/show_bug.cgi?id=59831

Tested-by: Alex Deucher <alexander.deucher@amd.com>

i965: Pass in the glarray to get_surface_type.

Dereffing all the values in the two callers was just pointless, and
the function isn't inlined so there was actual code impact.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965: Remove nonsense comment.

vb.inputs_read has never been a thing, even in the initial import.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965: Remove NDEBUG undef that was snuck in.

If you want debug, set --enable-debug in your config flags.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965: reuse _mesa_sizeof_type for index buffer types.

The core Mesa code has just one more case than this (GL_BITMAP), so I
don't see any cause to special-case it.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965: Reuse precalculated ib_type_size value.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965: Drop debug check for knowing the size of a type.

This was added in b93684f5f311f89c965960ab42bfea71a397b180, but there's
no need for it -- get_size has to succeed, and it has an assert for us
in debug builds.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965: Stop worrying about alignment of vertex data.

For our current types, the required alignment is actually just 1 byte.
When we get doubles, we have to worry (those have to be aligned to the
natural size), but we don't have doubles yet and they'll just be a
special case.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965: Use the glarray _ElementSize that Mesa tracks for us.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

mesa: Print more informative debug for _mesa_do_init_remap_table().

This is the same logic from _mesa_map_function_array().

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Remove ir_variable::uniform_block

v2: A previous patch contained a spurious hunk that removed an
assignment to ir_variable::uniform_block.  That hunk was moved to this
patch.  Suggested by Carl Worth.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Allow dereferencing fields of an interface instance

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>

glsl: Allow elimination of uniform block members

glGetActiveUniform is not supposed to report block members that are not
active even if they are included in the layout of the block.  The block
layout is determined from the GLSL_TYPE_INTERFACE that defines the
block, so eliminating the ir_variables that correspond to the individual
fields is safe.

Fixes gles3conform test
uniform_buffer_object_getuniformindices_for_for_nonexistent_or_not_active_uniform_names.

This also fixes the assertion failures (added in the previous commit) in
gles3conform uniform_buffer_object_index_of_not_active_block,
uniform_buffer_object_inherit_and_override_layouts, and
uniform_buffer_object_repeat_global_scope_layouts.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>

glsl: Calculate UBO data at link-time

Use the function added in the previous commit.

This temporarily causes gles3conform
uniform_buffer_object_index_of_not_active_block,
uniform_buffer_object_inherit_and_override_layouts, and
uniform_buffer_object_repeat_global_scope_layouts to assertion fail.
This is fixed in the next commit.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>

glsl: Add link_uniform_blocks to calculate all UBO data at link-time

Calculate all of the block member offsets, the IndexNames, and
everything else to do with every UBO.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>

glsl: Add a visitor to determine whether a uniform block is ever used

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>

glsl: Lower UBO references using link-time data instead of compile-time data

Pretty much all of the compile-time, per-compilation unit block data is
about to get the axe.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>

glsl: Add gl_uniform_buffer_variable::IndexName field

glGetUniformIndices requires that the block instance index not be
present in the name of queried uniforms.  However,
gl_uniform_buffer_variable::Name will include the instance index.  The
IndexName field is added to handle this difference.

Note that currently IndexName will always point to the same string as
Name.  This will change soon.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Carl Worth <cworth@cworth.org>

glsl: Make the align function available elsewhere in the linker

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Carl Worth <cworth@cworth.org>

glsl: Calculate link-time uniform block data without using compile-time block data

Pretty much all of the compile-time, per-compilation unit block data is
about to get the axe.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>

glsl: Assert that interfaces, like structures, are not seen as leaf types

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Carl Worth <cworth@cworth.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Add new uniform_field_visitor::process variant

This flavor takes a type and a base name.  It will be used to handle
cases where the block name (instead of the instance name) is used for an
interface block.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Carl Worth <cworth@cworth.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Recurse into uniform blocks just like uniform structures

v2: Inspite of the spell checker, spell recurse correctly.  Suggested by
Carl Worth.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Handle instance array declarations

v2: Add a comment and an assertion about the array size in the
non-instance name case.  Suggested by Paul Berry.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Track blocks in the symbol table using the glsl_type instead of the gl_uniform_block

Eventually the gl_uniform_block information won't be calculated until
linking.  Block names need to be checked for name clashes during
compiling, so we have to track it differently.

v2: Update the commit message.  Suggested by Carl Worth.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Add new uniform_field_visitor::visit_field variant

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Carl Worth <cworth@cworth.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Modify uniform_field_visitor::visit_field to take a row_major parameter

Not used yet, but the UBO layout visitor will use this.

v2: Remove a spruious hunk.  This is moved to the patch "glsl: Remove
ir_variable::uniform_block".  Suggested by Carl Worth.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Modify uniform_field_visitor::recursion to take a row_major parameter

Not used yet, but the UBO layout visitor will use this.

v2: Add some commentary as to why row_major is always set to false in
process.  Suggesed by Paul Berry.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Carl Worth <cworth@cworth.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Add a predicate to determine whether a variable is an interface block

For the first declaration below, there will be an ir_variable named
"instance" whose type and whose instance_type will be the same
glsl_type.  For the second declaration, there will be an ir_variable
named "f" whose type is float and whose instance_type is B2.

"instance" is an interface instance variable, but "f" is not.

uniform B1 {
    float f;
} instance;

uniform B2 {
    float f;
};

v2: Copy the comment message documentation into the code.  Suggested by
Paul Berry.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Carl Worth <cworth@cworth.org>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Require that indices into uniform block arrays be constants

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Carl Worth <cworth@cworth.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Add ir_variable::interface_type field

For variables that are in an interface block or are an instance of an
interface block, this is the GLSL_TYPE_INTERFACE type for that block.

Convert the ir_variable::is_in_uniform_block method added in the
previous commit to use this field instead of ir_variable::uniform_block.

v2: Fix the place-holder comment on ir_variable::interface_type.
Suggested by Paul Berry.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Carl Worth <cworth@cworth.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Add ir_variable::is_in_uniform_block predicate

The way a variable is tested for this property is about to change, and
this makes the code easier to modify.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Carl Worth <cworth@cworth.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Generate an interface type for uniform blocks

If the block has an instance name, add the instance name to the symbol
table instead of the individual fields.

Fixes the piglit test interface-name-access-without-interface-name.vert
for real.

v2: Update the comment before the assertion that interface block
definitions won't generate instructions.  Suggested by Paul Berry.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Paul Berry <stereotype441@gmail.com>

glsl: Add GLSL_TYPE_INTERFACE

Interfaces are structurally identical to structures from the compiler's
point of view.  They have some additional restrictions, and generally
GPUs use different instructions to access them.  Using a different base
type should make this a bit easier.

This commit also adds the glsl_type::interface_packing fields.  For
GLSL_TYPE_INTERFACE types, this will track the specified packing mode.
It is analogous to gl_uniform_buffer::_Packing.

v2: Add serveral missing GLSL_TYPE_INTERFACE cases in switch-statements.

v3: Add information about glsl_type::interface_packing.  Move row_major
checking in glsl_type::record_key_compare from this patch to the
previous patch.  Both suggested by Paul Berry.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Add row_major field to glsl_struct_field

For now, this will always be false.  In the near future, an "interface"
type will be added that shares a lot of infrastructure with structures.

v2: Move row_major checking in glsl_type::record_key_compare from the
next patch to this patch.  Suggested by Paul Berry.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Carl Worth <cworth@cworth.org>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Refactor out processing of structure fields

This will soon also be used for processing interface block fields.

v2: Add a comment explaining the interface of
ast_process_structure_or_interface_block.  Suggested by Paul Berry.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Carl Worth <cworth@cworth.org>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
Reviewed-by: Chad Versace <chad.versace@linux.intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Parse interface array size

The size is parsed and stored in the AST, but it is not used yet.
Processing of the array size is added in the patch "glsl: Handle
instance array declarations"

v2: Update the commit message (suggested by Carl Worth).  Add a comment
to ast_uniform_block::array_size (suggested by Paul Berry).

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Chad Versace <chad.versace@linux.intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Parse non-array uniform block instance names in GLSL ES 3.00.

In GLSL ES 3.00 (and GLSL 1.50), uniform blocks can have an associated
"instance name", which essentially namespaces the variables inside.

This patch adds basic parsing for this new feature, but doesn't yet hook
it up to actually do anything yet.

It does not support for arrays of interface blocks; a later commit will
take care of that.

This change temporarily regresses the piglit test
interface-name-access-without-interface-name.vert.  This shader failed
to compile before (the expected result), but it failed to compile for
the wrong reason.  This is not a real regression.

v2: Add some comments to ast_uniform_block::instance_name.  Suggested by
Paul Berry.

Reviewed-by: Carl Worth <cworth@cworth.org>
Reviewed-by: Chad Versace <chad.versace@linux.intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Refactor uniform block parser rules.

The existing code has a lot of duplication; the only difference between
the two cases is whether we merge in an additional layout qualifier.

Apparently creating a layout_qualifieropt rule that can be empty causes
a lot of conflicts and confusion.  However, refactoring out the guts of
the ast_uniform_block creation works fine.

Reviewed-by: Carl Worth <cworth@cworth.org>
Reviewed-by: Chad Versace <chad.versace@linux.intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

linker: Refactor intra-stage block compatabililty testing

Also slightly change the compatibility test.  Instead of comparing the
offsets of the block variables, compare the packing mode of the blocks.
Ideally we don't want to assign the offsets until a later stage of
linking.

This is put in a new file called link_uniform_blocks.cpp.  Some new
functions related to uniform blocks are going to live in that file as
well.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Carl Worth <cworth@cworth.org>
Reviewed-by: Chad Versace <chad.versace@linux.intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

mesa: Track the packing mode of a UBO in gl_uniform_buffer

This allows the next patch to verify that two uniform blocks match
without first calculating the locations of the fields.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Chad Versace <chad.versace@linux.intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Replace most default cases in switches on GLSL type

This makes it easier to find switch-statements that need to be updated
after a new GLSL_TYPE_* is added because the compiler will generate a
warning.

Switch-statements that only had a small number of cases (e.g.,
everything in ir_constant_expression.cpp) were not modified.  I may
regret that decision when we eventually add support for doubles.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Carl Worth <cworth@cworth.org>
Reviewed-by: Chad Versace <chad.versace@linux.intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965: Correct gen6+ guardband calculation.

Too much attention was paid to the first paragraphs, and not enough to
the last little note that "oh, by the way, the rendered things
themselves still have to be clipped to just 8192 wide/high".

Fixes GTF's clip.c test with 4096 or higher width on ivb, where one of
the triangles got the upper half of its pixels dropped.

Tested-by: Ian Romanick <ian.d.romanick@intel.com>

i965: Use GL_RED for DEPTH_TEXTURE_MODE in ES 3.0 for unsized formats.

Khronos has apparently decided that depth textures with sized formats
(allowed with ARB_internalformat_query or ES 3.0) should be treated as
GL_RED, while unsized formats (an existing feature) should be treated
as GL_INTENSITY for compatibility with ES 2.0.

Ian is proposing changes to ARB_internalformat_query which will make
this actually legal and consistent.

A similar problem exists with GL 4.2, but we're going to ignore that
for the time being.

Tested on Ivybridge: no Piglit regressions; fixes 4 es3conform tests:
- depth_texture_fbo
- depth_texture_fbo_clear
- depth_texture_teximage
- depth_texture_texsubimage

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

i965: Bump maximum supported ES2 context version to 3.0

Since patch "i965: Validate requested GLES context version in
brwCreateContext", we have been able to create ES 3.0 contexts due to the
max version check.  So...bump the max version.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

i965/Gen6+: Enable ARB_ES3_compatibility extension

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

mesa/es3: Enable ES 3.0 API and shading language version

v2: Add ARB_internalformat_query to the list of required extensions.

v3: Add OES_depth_texture_cube_map to the list of required extensions.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>

scons: Add imports.c to builtin_compiler build.

Fixes build regression introduced by commit
eac030e38e3cdd4ed4534516e3d3a50c8a372719.

Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=59835

i965/fs/gen7: Fix fatal typo in unpackHalf2x16

s/src/src_w/

That little typo, which sneaked into v4 of the previous patch, generates
incorrect fs code.

Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

i965/fs/gen7: Emit code for GLSL 3.00 pack/unpack operations (v4)

v2: Remove lewd comment. [for idr]
v3: - Optimize away tmp register for packHalf2x16. [for anholt, paul]
    - Improve comments. [for anholt, paul]
    - Reduce near-duplicate code by removing vec4_visitor emit_pack/unpack
      methods. [for chadv]
v4: Factor our UD/W register conversion into helper function. [for anholt]

Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com> (v2)
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

i965/vs/gen7: Emit code for GLSL ES 3.00 pack/unpack operations (v3)

FIXME: This patch emits VS code that violates documented hardware
restrictions and then relies on undocumented behavior that results from
that violation.  This patch passes all tests, but should be fixed ASAP to
conform to the hardware documentation.

v2: Explain undocumented hardware behavior. Improve comments.
v3: Use ALU1 helper methods F32TO16() and F16TO32(). [for anholt]

Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com> (v1)
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

i965: Quote the PRM on a HorzStride subtlety

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

i965: Add opcodes for F32TO16 and F16TO32

The GLSL ES 3.00 operations packHalf2x16 and unpackHalf2x16 will emit
these opcodes.

- Define the opcodes BRW_OPCODE_{F32TO16,F16TO32}.
- Add the opcodes to the brw_disasm table.
- Define convenience functions brw_{F32TO16,F16TO32}.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Acked-by: Paul Berry <stereotype441@gmail.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

i965: Lower the GLSL ES 3.00 pack/unpack operations (v2)

On gen < 7, we fully lower all operations to arithmetic and bitwise
operations.

On gen >= 7, we fully lower the Snorm2x16 and Unorm2x16 operations, and
partially lower the Half2x16 operations.

v2:
  - Comment that scalarization is needed only for SOA code [for idr].
  - Replace switch-statement with if-statement [for idr].
  - Remove misplaced hunk from previous patch [found by idr].

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Tuner <mattst88@gmail.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

glsl: Add lowering pass for GLSL ES 3.00 pack/unpack operations (v4)

Lower them to arithmetic and bit manipulation expressions.

v2: Rewrite using ir_builder [for idr].
v3: Comment typos. [for mattst88]
v4: Fix arithmetic error in comments.
    Factor out a shift instruction.
    Don't heap allocate factory.instructions.
    [for paul]

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com> (v2)
Reviewed-by: Matt Tuner <mattst88@gmail.com> (v3)
Reviewed-by: Paul Berry <stereotype441@gmail.com> (v4)
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

glsl: Fix type-deduction for and/or/xor expressions

In ir_expression's constructor, the cases for {bit,logic}_{and,or,xor}
failed to handle the case when both operands were vectors.

Note: This is a candidate for the stable branches.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

glsl: Reformat and/or/xor cases in ir_expression ctor

Replace tabs with spaces. According to docs/devinfo.html, Mesa's
indetation style is:
  indent -br -i3 -npcs --no-tabs infile.c -o outfile.c

This patch prevents whitespace weirdness in the next patch.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

glsl/ir_builder: Add helpers for making if-statements

Add two overloaded variants of
    ir_if *if_tree()

The new functions allow one to chain together if-trees within a single C++
expression that resembles a real if-statement.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

glsl/ir_builder: Add `enum writemask`

Using this enum improves the readibility of calls to assign(), whose third
argument is a writemask.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

glsl/ir_factory: Add helper method for making an ir_constant

Add method ir_factory::constant.  This little method constructs an
ir_constant using the factory's mem_ctx.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

glsl/ir_builder: Add more helpers for constructing expressions

Add the following functions, each of which construct the similarly named
ir expression:
    div, round_even, clamp

    equal, less, greater, lequal, gequal

    logic_not, logic_and, logic_or

    bit_not, bit_or, bit_and, lshift, rshift

    f2i, i2f, f2u, u2f, i2u, u2i

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

glsl/ir_factory: Initialize members to NULL in constructor

This eliminates unexpected behavior due to unitialized values.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

glsl: Evaluate constant GLSL ES 3.00 pack/unpack expressions (v3)

That is, evaluate constant expressions of the following functions:
  packSnorm2x16  unpackSnorm2x16
  packUnorm2x16  unpackUnorm2x16
  packHalf2x16   unpackHalf2x16

v2: Reuse _mesa_pack_float_to_half and its inverse to evaluate
    pack/unpackHalf2x16. [for idr]
v3: Whitespace fixes. [for mattst88]
    Don't cast neg floats directly to uint16; use an intermediate cast to
    int16. [for paul]

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com> (v2)
Reviewed-by: Paul Berry <stereotype441@gmail.com>
Reviewed-by: Matt Tuner <mattst88@gmail.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

mesa: Remove rounding bias in _mesa_float_to_half()

Not all float32 values can be exactly represented as a float16.
_mesa_float_to_half() rounded such intermediate float32 values to zero by
truncating unrepresentable bits in the mantissa.

This patch improves _mesa_float_to_half() by rounding intermediate float32
values to the nearest float16; when the float32 is exactly between two
float16 values we round to the one with an even mantissa. This behavior is
preferred over the old behavior because:
  - It has reduced bias relative to the old behavior.

  - It reproduces the behavior of real hardware: opcode F32TO16 in
    Intel's GPU ISA.

  - By reproducing the behavior of the GPU (at least on Intel hardware),
    compile-time evaluation of constant packHalf2x16 GLSL expressions will
    result in the same value as if the expression were executed on the GPU.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

mesa,glsl: Move round_to_even() from glsl to mesa/main (v2)

Move round_to_even's definition to mesa/main so that _mesa_float_to_half()
can use it in order to eliminate rounding bias.

In additon to moving the fuction definition, prefix its name with "_mesa",
just as all other functions in mesa/main are prefixed.

v2: Fix Android build.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

glsl/standalone_scaffolding: Add stub for _mesa_warning()

A subsequent patch will add mesa/main/imports.c as a dependency to the
compiler, which in turn requires that _mesa_warning() be defined.

The real definition of _mesa_warning() is in mesa/main/errors.c, but to
pull that file into the standalone scaffolding would require transitively
pulling in the dispatch tables.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Acked-by: Paul Berry <stereotype441@gmail.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

glsl: Extend ir_expression_operation for GLSL 3.00 pack/unpack functions (v2)

For each function {pack,unpack}{Snorm,Unorm,Half}2x16, add a corresponding
opcode to enum ir_expression_operation.  Validate the new opcodes in
ir_validate.cpp.

Also, add opcodes for scalarized variants of the Half2x16 functions.  (The
code generator for the i965 fragment shader requires that all vector
operations be scalarized.  A lowering pass, to be added later, will
scalarize the Half2x16 functions).

v2: Fix assertion message in ir_to_mesa [for idr].

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Tuner <mattst88@gmail.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

glsl: Add IR lisp for GLSL ES 3.00 pack/unpack functions

For each of the following functions, add a declaration to
builtins/profiles/300es.glsl and create new file
builtins/ir/${funcname}.ir:

  packSnorm2x16  unpackSnorm2x16
  packUnorm2x16  unpackUnorm2x16
  packHalf2x16   unpackHalf2x16

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Tuner <mattst88@gmail.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

glsl: Fix typo in comment

s/num_operands()/get_num_operands()/

Discovered because Eclipse failed to resolve the false reference.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

i965/disasm: Fix horizontal stride of dest registers

The bug: The printed horizontal stride was the numerical value of the
  BRW_HORIZONTAL_$N enum.
The fix: Translate the enum before printing.

Note: This is a candidate for the stable releases.
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

intel: Fix glCopyTexSubImage on buffers whose width >= 32kbytes

When possible, glCopyTexSubImage calls are performed using the
hardware blitter.  However, according to the Ivy Bridge PRM, Vol1
Part4, section 1.2.1.2 (Graphics Data Size Limitations):

    The BLT engine is capable of transferring very large quantities of
    graphics data. Any graphics data read from and written to the
    destination is permitted to represent a number of pixels that
    occupies up to 65,536 scan lines and up to 32,768 bytes per scan
    line at the destination. The maximum number of pixels that may be
    represented per scan line’s worth of graphics data depends on the
    color depth.

With an RGBA32F color buffer (which has 16 bytes per pixel) this
imposes a maximum width of 2048 pixels.  Other pixel formats have
accordingly larger limits.

To make matters worse, if the pitch of the buffer is 32k or greater,
intel_copy_texsubimage's call to intelEmitCopyBlit will overflow
intelEmitCopyBlit's src_pitch and dst_pitch parameters (which are
16-bit signed integers).

We can conveniently avoid both problems by avoiding use of the blitter
when the miptree's pitch is >= 32k.

Fixes gles3conform "framebuffer_blit_functionality_magnifying_blit"
tests when the buffer width is equal to 8192.

Note: this is very similar to the recent patch "intel: Fix ReadPixels
on buffers whose width >= 32kbytes" except that it applies to
glCopyTexSubImage instead of glReadPixels.  In a future patch it would
be nice to refactor the code so that (a) overflow is avoided, and (b)
intelEmitCopyBlit is responsible for checking whether the blitter can
handle the width, so that all callers of intelEmitCopyBlit work
properly, rather than just these two.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>