This basically reverts commit
2acc719374.
With the previous change, we're not batchbuffer limited any
longer. So we actually start seeing a performance difference
between X and Y tiling. X tiling is funny because it is
faster for screen-aligned quads but slower in games. So let's
use Y tiling which is 10% faster overall.
Now that we don't throttle at every batchbuffer, we can shrink
the size of batchbuffers to achieve early flushing. This gives
a significant speed boost in a lot of games (on the order of
20%).
It should be TGSI_TYPE_UNSIGNED, not TGSI_TYPE_FLOAT.
Fixed also gallivm not_emit_cpu() to use uint build context.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Acked-by: Roland Scheidegger <sroland@vmware.com>
Move the body of tgsi_opcode_infer_dst_type() to a new helper function,
tgsi_opcode_infer_type(), and call the helper function from
tgsi_opcode_infer_dst_type(). The diff looks complicated simply because the
code is moved around.
A following commit will make tgsi_opcode_infer_src_type() call
tgsi_opcode_infer_type().
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Acked-by: Roland Scheidegger <sroland@vmware.com>
Reorder opcodes by their assigned numbers. This makes it easier to see the
differences between tgsi_opcode_infer_src_type() and
tgsi_opcode_infer_dst_type().
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Acked-by: Roland Scheidegger <sroland@vmware.com>
Make use of tgsi_util_get_texture_coord_dim() to replace the big switch table.
There is a subtle difference with this change. When TXP is used with an array
texture, the layer is now also projected. This behavior matches the TGSI doc.
Since GLSL does not allow TXP on an array texture, I am not sure which
behavior is correct or preferred.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Acked-by: Roland Scheidegger <sroland@vmware.com>
This util function returns the dimension of the texture coordinates for a
texture target, and the location of the shadow reference value.
For example, when the texture target is TGSI_TEXTURE_SHADOW2D, the dimension
of the texture coordinates is 2, and the location of the ref value is 2
(that is, the Z channel).
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Acked-by: Roland Scheidegger <sroland@vmware.com>
Fixes a regression in firefox's unaccelerated compositing path for WebGL
with the introduction of Y tiling.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=64213
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We accidentally "fixed" the piglit test for this when introducing Y
tiling, since this path stopped being executed. In reenabling this path
for Y tiling, we ended up regressing it again, so just fix it.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=59439
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Adds the remaining integer opcodes, and some opcodes are moved to more
appropriate places, along with getting rid of the (already nearly empty)
ps_2_x section. Though the CAP bits for some of these are still a bit in
the air so the documentation isn't quite as watertight as is desirable.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
We can replace CNDxx with MOV (and possibly eliminate after
propagation) in following cases:
If src1 is equal to src2 in CNDxx instruction then the result doesn't
depend on condition and we can replace the instruction with
"MOV dst, src1".
If src0 is const then we can evaluate the condition at compile time and
also replace it with MOV.
Signed-off-by: Vadim Girlin <vadimgirlin@gmail.com>
Use the same limit for kcache constants in alu group on r6xx as on other
chips (two const pairs). Relaxing this will require additional checks to
make sure that all 4 consts in the group come from 2 kcache sets (clause
limit), probably without noticeable improvements of shader performance.
Signed-off-by: Vadim Girlin <vadimgirlin@gmail.com>
Everyone was doing effectively the same thing, except for some funky code
reuse in Intel, and swrast mistakenly recomputing _BaseFormat instead of
using the texture's _BaseFormat. swrast's sRGB handling is left in place,
though it should be done by using _mesa_get_render_format() at render time
instead (as-is, it will miss updates to GL_FRAMEBUFFER_SRGB).
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We're looking for the logical width of our level, which is what
image->Width2/Height2 is. The previous code relied on MSAA textures being
only level 0.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
There was some comment about trying to avoid marking resolves in
updownsample, but if the downsample is never actually rendered to, then
the required resolve tracked in the downsample will never be executed, so
who cares?
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
v2: Only lower bitfieldInsert to BFM+BFI (and don't lower
bitfieldExtract at all) since three-source instructions are now
usable in the vertex shader.
v3: Lower bitfield_insert in the same pass with everything else, since
it doesn't produce any instructions to be lowered (the other two
lowering passes that were in a previous iteration of this series
emitted subtractions which needed to be lowered).
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz> [v2]
v2: Rebase on LRP addition.
Use fix_3src_operand() when emitting BFE and BFI2.
Add BFE and BFI2 to is_3src_inst check in
brw_vec4_copy_propagation.cpp.
Subtract result of FBH from 31 (unless an error) to convert
MSB counts to LSB counts
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
Don't bother scalarizing ir_binop_bfm, since its results are
identical for all channels.
v2: Subtract result of FBH from 31 (unless an error) to convert
MSB counts to LSB counts.
v3: Use op0->clone() in ir_triop_bfi to prevent (var_ref
channel_expressions) from appearing multiple times in the IR.
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz> [v2]
Specifically
bfe - for bitfieldExtract()
bfi1 and bfi2 - for bitfieldInsert()
bfrev - for bitfieldReverse()
cbit - for bitCount()
fbh - for findMSB()
fbl - for findLSB()
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
Also update asserts to allow BFE and BFI2, which take (unsigned)
doubleword arguments.
v2: Allow BRW_REGISTER_TYPE_UD for src1 and src2 as well.
Assert that src2.type (instead of src0.type) matches dest.type since
it's the primary argument and src0 and src1 might correctly have
different types.
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz> [v1]
i965/Gen7+ and Radeon/Evergreen+ have bfm/bfi instructions to implement
bitfieldInsert() from ARB_gpu_shader5.
v2: Add ir_binop_bfm and ir_triop_bfi to st_glsl_to_tgsi.cpp.
Remove spurious temporary assignment and dereference.
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
v2: Order bits from LSB end (31 - count) for ir_unop_find_msb.
v3: Add ir_triop_bitfield_extract as an exception to the op[0]->type ==
op[1]->type assertion in ir_constant_expression.cpp.
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz> [v2]
This library is very small, so there is not much to gain from building
it as a shared library. Also, when linking statically with LLVM, a
shared libradeonllvm exports LLVM symbols and creates problems when
used with other shared objects that also link statically to LLVM.
Reviewed-by: Mathias.Froehlich@web.de
The LLVM C API is considered stable and should never change, so it
is much more desirable to use than the LLVM C++ API, which is constantly in
flux.
v2:
- Split target initialization and lookup into separate functions
Reviewed-by: Mathias.Froehlich@web.de
This does not solve all of the problems with using LLVM in a
multithreaded enivronment, but it should help in some cases.
Reviewed-by: Mathias.Froehlich@web.de