Apart from the fact that the radeon.h/r600_states.h editing is a nightmare, this
wasn't so bad.
passes piglit user-clip test now also trivial tests.
Signed-off-by: Dave Airlie <airlied@redhat.com>
check_os_katmai_support checks that the operating system running on a
SSE-capable processor supports SSE. This is necessary for unpatched
2.2.x and earlier kernels. 2.4.x and later kernels support SSE.
check_os_katmai_support will disable SSE capabilities for 32-bit x86
operating systems for which there is no code path. Currently, this
function handles Linux, Windows, and several BSDs. Mac OS, Cygwin, and
Solaris are several operating systems with no code paths.
Rather than add code for the unhandled operating systems, remove this
function altogether. This will fix SSE detection on all recent 32-bit
x86 operating systems. This completely breaks functionality on unpatched
2.2.x and earlier kernels, although there are likely no Gallium3D users
on such operating systems.
Changes in v2:
- Change function name
Currently draw_llvm refuses to create itself on non-SSE2 CPUs due to
an alleged LLVM bug.
However, this is implemented improperly, because other parts of draw
still attempt to access draw->llvm, resulting in segfaults.
Instead, put the check in debug_get_option_draw_use_llvm, check that
before calling draw_llvm_create, and then check whether draw->llvm is
non-null everywhere else.
NOTE: Win64 is untested, and is thus currently disabled.
If you have such a system, please enable it and report whether it works.
To enable it, change src/gallium/auxiliary/translate/translate.c
Changes in v5:
- On Win64, preserve %xmm6 and %xmm7 as required by the ABI
- Use _WIN64 instead of WIN64
Changes in v4:
- Use x86_target() and x86_target_caps()
- Enable translate_sse in x86-64, but not in Win64
Changes in v3:
- Win64 support (untested)
- Use u_cpu_detect.h constants instead of #ifs
Changes in v2:
- Minimize #ifs
- Give a name to magic number CHANNELS_0001
- Add support for CPUs without SSE (only memcpy and swizzles, like non SSE2)
- Fixed comments
translate_sse is currently very limited to the point of
being useless in essentially all cases.
In particular, it only support some float32 and unorm8
formats and doesn't work on x86-64.
This commit rewrites it to support:
1. Dumb memory copy for any pair of identical formats
2. All formats that are swizzles of each other
3. Converting 32/64-bit floats and all 8/16/32-bit integers to 32-bit float
4. Converting unorm8/snorm8 to snorm16 and uscaled8/sscaled8 to sscaled16
5. Support for x86-64 (doesn't take advantage of it in any way though)
This new translate can even be useful to translate index buffers for
cards that lack 8-bit index support.
It passes the testsuite I wrote, but note that this is a major change, and more
testing would be great.
Changes in v5:
- Add sse2_movdqa
Changes in v4:
- Use _WIN64 instead of WIN64
Changes in v3:
- Add target and target caps functions, so that they could be different in
principle from the current CPU and they don't need #ifs to check
Changes in v2:
- Win64 support (untested)
- Use u_cpu_detect.h constants instead of #ifs
This commit adds minimal x86-64 support: only movs between registers
are supported for r8-r15, and x64_rexw() must be used to ask for 64-bit
operations.
It also adds several new instructions for the new translate_sse code.
movdqa
Currently translate_sse puts two trivial wrappers in the translate vtable.
These slow it down and enlarge the source code for no gain, except perhaps
the ability to set a breakpoint there, so remove them.
Breakpoints can be set on the caller of the translate functions, with no
loss of functionality.
Changes in v3:
- If we can do a copy, don't try to get an emit func, as that can assert(0)
Changes in v2:
- Add comment regarding copy_size
When used in GPU drivers, translate can be used to simultaneously
perform a gather operation, and convert away from unsupported formats.
In this use case, input and output formats will often be identical: clearly
it would make sense to use a memcpy in this case.
Instead, translate will insist to convert to and from 32-bit floating point
numbers.
This is not only extremely expensive, but it also loses precision for
32/64-bit integers and 64-bit floating point numbers.
This patch changes translate_generic to just use memcpy if the formats are
identical, non-blocked, and with an integral number of bytes per pixel (note
that all sensible vertex formats are like this).
Triangle strip alternates the front/back orientation of its triangles.
max_vertices was made even so that varray never splitted a triangle
strip at the wrong positions.
It did not work with triangle strips with adjacencies. And it is no
longer relevant with vsplit.
vcache decomposes primitives while vsplit splits primitives. Splitting
is generally easier to do and is faster. More importantly, vcache
depends on flatshade_first to decompose. The outputs may have incorrect
vertex order which is significant to GS.
vsplit is based on varray. It sets the split flags when a primitive is
splitted. It also has support for indexed primitives.
For indexed primitives, unlike vcache, vsplit splits the primitives
instead of decomposes them.
A primitive may be splitted in frontends. The splitted primitives
should convey certain flag bits so that the decomposer can correctly
decide the stipple or edge flags.
This commit adds flags to draw_prim_info and updates the decomposer to
honor the flags. Frontends and middle ends will be updated later.
This changes r300_destroy_context() so it can be called
on a partially-initialized context, and uses it when
r300_create_context() hits a fatal error.
This makes sure r300_create_context() doesn't leak memory
or neglect to call r300_update_num_contexts() when it fails.
Signed-off-by: Marek Olšák <maraeo@gmail.com>
This fixes a potential bug if (has_hyperz) is false
(it would still init the atom as if has_hyperz were true).
Signed-off-by: Marek Olšák <maraeo@gmail.com>
For 16 and 64 pixel levels, calculate a mask which is linear in x and
y (ie not in the swizzle layout).
When iterating over full and partial masks, figure out position by
manipulating the bit number set in the mask, rather than relying on
postion arrays.
Similarly, calculate the lower-level c values from dcdx, dcdy and the
position rather than relying on the step array.
