aboutsummaryrefslogtreecommitdiffstats
Commit message (Collapse)AuthorAgeFilesLines
* clover: Pass execution dimensions and offset to the kernel as implicit ↵Francisco Jerez2014-10-122-25/+70
| | | | | | arguments. Reviewed-by: Jan Vesely <[email protected]>
* clover: Add semantic information to module::argument for implicit parameter ↵Francisco Jerez2014-10-121-4/+12
| | | | | | passing. Reviewed-by: Jan Vesely <[email protected]>
* clover: Use unreachable() from util/macros.h instead of assert(0).Francisco Jerez2014-10-113-4/+4
| | | | Reviewed-by: Francisco Jerez <[email protected]>
* gallium: Add tokens for DragonFly BSD.Vinson Lee2014-10-101-0/+6
| | | | | Signed-off-by: Vinson Lee <[email protected]> Acked-by: Brian Paul <[email protected]>
* ilo: disassemble compacted instructionsChia-I Wu2014-10-114-2/+453
| | | | Signed-off-by: Chia-I Wu <[email protected]>
* glsl: improve accuracy of atan()Erik Faye-Lund2014-10-101-10/+55
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Our current atan()-approximation is pretty inaccurate at 1.0, so let's try to improve the situation by doing a direct approximation without going through atan. This new implementation uses an 11th degree polynomial to approximate atan in the [-1..1] range, and the following identitiy to reduce the entire range to [-1..1]: atan(x) = 0.5 * pi * sign(x) - atan(1.0 / x) This range-reduction idea is taken from the paper "Fast computation of Arctangent Functions for Embedded Applications: A Comparative Analysis" (Ukil et al. 2011). The polynomial that approximates atan(x) is: x * 0.9999793128310355 - x^3 * 0.3326756418091246 + x^5 * 0.1938924977115610 - x^7 * 0.1173503194786851 + x^9 * 0.0536813784310406 - x^11 * 0.0121323213173444 This polynomial was found with the following GNU Octave script: x = linspace(0, 1); y = atan(x); n = [1, 3, 5, 7, 9, 11]; format long; polyfitc(x, y, n) The polyfitc function is not built-in, but too long to include here. It can be downloaded from the following URL: http://www.mathworks.com/matlabcentral/fileexchange/47851-constraint-polynomial-fit/content/polyfitc.m This fixes the following piglit test: shaders/glsl-const-folding-01 Signed-off-by: Erik Faye-Lund <[email protected]> Reviewed-by: Ian Romanick <[email protected]>
* vc4: Use the fnv1 hash function instead of gallium util's crc32.Eric Anholt2014-10-101-2/+3
| | | | | Improves simulated norast performance on a little benchmark by 13.4012% +/- 2.08459% (n=13).
* vc4: Don't look up the compiled shaders unless state has changed.Eric Anholt2014-10-103-0/+28
| | | | | Improves simulated norast performance on a little benchmark by 38.0965% +/- 3.27534% (n=11).
* vc4: Actually clear the context's dirty flags.Eric Anholt2014-10-101-0/+1
| | | | | I was trying to skip state updates when !dirty, and suspiciously everything was always dirty.
* vc4: Optimize the other case of SEL_X_Y wih a 0 -> SEL_X_0(a).Eric Anholt2014-10-101-1/+23
| | | | Cleans up some output to be more obvious in a piglit test I'm looking at.
* mesa: fix error reported on gTexSubImage2D when level not validTapani Pälli2014-10-101-1/+1
| | | | | Signed-off-by: Tapani Pälli <[email protected]> Reviewed-by: Juha-Pekka Heikkila <[email protected]>
* i965: Fix register write checks.Kenneth Graunke2014-10-101-0/+2
| | | | | | | | | | | | | When mapping the buffer a second time, we need to use the new pointer, not the one from the previous mapping. Otherwise, we will most likely crash. Apparently, we've just been getting lucky and getting the same bo->virtual pointer in both cases. libdrm probably has a hand in that. Signed-off-by: Kenneth Graunke <[email protected]> Reviewed-by: Anuj Phogat <[email protected]> Cc: [email protected]
* vc4: Optimize out adds of 0.Eric Anholt2014-10-091-0/+26
|
* vc4: Optimize fmul(x, 0) and fmul(x, 1).Eric Anholt2014-10-091-0/+45
| | | | | This was being generated frequently by matrix multiplies of 2 and 3-channel vertex attributes (which have the 0 or 1 loaded in the shader).
* vc4: Factor out the turn-it-into-a-mov in opt_algebraic.Eric Anholt2014-10-091-10/+12
| | | | This will be used more in the next commits.
* vc4: Eliminate unused texture instructions.Eric Anholt2014-10-091-1/+21
|
* vc4: Dead code eliminate unused SF instructions.Eric Anholt2014-10-091-7/+26
|
* vc4: Prevent copy propagating out the MOVs from r4.Eric Anholt2014-10-091-1/+11
| | | | | | | | | Copy propagating these might result in reading the r4 after some other instruction has written r4. Just prevent all copy propagation of this for now. Fixes bad rendering with upcoming indirect register access support, where the copy propagation was consistently happening across another read.
* vc4: Split the coordinate shader to its own vc4_compiled_shader.Eric Anholt2014-10-093-89/+54
| | | | | | | | | | | Merging VS and CS into the same struct wasn't winning us anything except for not allocating a separate BO (but if we want to pack programs into BOs, we should pack not just those 2 programs together). What it was getting us was a bunch of code duplication about hash table lookups and propagating vc4_compile contents into a vc4_compiled_shader. I was about to make the situation worse with indirect uniform buffer access.
* vc4: Add #defines for the texture uniform fields.Eric Anholt2014-10-092-19/+113
| | | | | | I wanted to make another set of texture uploads for handling reladdr constants, and duplicating all the bitshifting looked like a terrible idea. In the process, this fixes a swap of the s/t texture wrap modes.
* vc4: Initialize undefined temporaries to 0.Eric Anholt2014-10-091-1/+6
| | | | | | | | | | Under the simulator, reading registers before writing them triggers an assertion failure. c->undef gets treated as r0, which will usually be written, but not if it's used in the first instruction. We should definitely not be aborting in this case, and return some sort of undefined value instead. Fixes glsl-user-varying-ff.
* i965: Skip uploading border color when unnecessary.Kenneth Graunke2014-10-091-2/+20
| | | | | | | | | | | | | | The border color is only needed when using the GL_CLAMP_TO_BORDER or (deprecated) GL_CLAMP wrap modes; all others ignore it, including the common GL_CLAMP_TO_EDGE and GL_REPEAT wrap modes. In those cases, we can skip uploading it entirely, saving a bit of space in the batchbuffer. Instead, we just point it at the start of the batch (offset 0); we have to program something, and that address is safe to read. Signed-off-by: Kenneth Graunke <[email protected]> Reviewed-by: Chris Forbes <[email protected]>
* i965: Use BDW_MOCS_PTE for renderbuffers.Kenneth Graunke2014-10-091-1/+1
| | | | | | | | | | | | | | | | | | | | | | | | | | | | Write-back caching cannot be used for buffers being scanned out by the display engine; surfaces used for scan-out must be write-through or uncached. I originally chose WT for render targets because it works in all cases. However, we really want to use write-back caching where possible, as it is more efficient. Most renderbuffers are not used for scanout - off-screen FBOs certainly are fine, and non-pageflipped backbuffers should be fine as well. So in most cases WB will work. However, we don't know what will be used for scan-out, so we instead simply use the PTE value specified by the kernel, as it knows these things. This matches our MOCS choice on Haswell. Fixes performance regressions since commit ee4484be3dc827cf15bcf109f5 in a microbenchmark (spotted by Eero Tamminen). Improves performance in GLBenchmark 2.7/EgyptHD by 7.44362% +/- 0.496939% (n=55) on a Broadwell GT2. Improves performance in a bunch of other microbenchmarks by ~15% or so. Signed-off-by: Kenneth Graunke <[email protected]> Reported-by: Eero Tamminen <[email protected]> Reviewed-by: Daniel Vetter <[email protected]> Reviewed-by: Kristian Høgsberg <[email protected]> Cc: [email protected]
* i965: Add a BRW_MOCS_PTE #define.Kenneth Graunke2014-10-091-3/+7
| | | | | | | | | | | | | | Like BDW_MOCS_WB and BDW_MOCS_WT, this specifies that we want to use all three caches (L3, LLC, and eLLC where available), but leaves the LLC caching mode up to the kernel's page table entry. This allows the kernel to pick WB/WT/UC based on whether it's using a buffer for scanout. Signed-off-by: Kenneth Graunke <[email protected]> Reviewed-by: Daniel Vetter <[email protected]> Reviewed-by: Kristian Høgsberg <[email protected]> Cc: [email protected]
* mesa: Make _mesa_print_arrays use stderr.Kenneth Graunke2014-10-091-3/+3
| | | | | | | | | These days, most driver debug output happens via stderr, not stdout. Some applications (such as Xephyr) also appear to close stdout which makes these messages go nowhere. Signed-off-by: Kenneth Graunke <[email protected]> Reviewed-by: Tapani Pälli <[email protected]>
* r600g,radeonsi: Always use GTT again for PIPE_USAGE_STREAM buffersMichel Dänzer2014-10-091-1/+3
| | | | | | | | | Putting those in VRAM can cause long pauses due to buffers being moved into / out of VRAM. Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=84662 Cc: [email protected] Reviewed-by: Alex Deucher <[email protected]>
* vc4: Optimize SF(ITOF(x)) -> SF(x).Eric Anholt2014-10-091-0/+16
| | | | | This is a common production of st_glsl_to_tgsi, because CMP takes a float argument.
* vc4: Add some optimization of FADD(FSUB(0, x)).Eric Anholt2014-10-091-0/+31
| | | | | This is a common production of st_glsl_to_tgsi, which uses negate flags on source arguments to handle subtraction.
* vc4: Mostly fix offset calculation for NPOT mipmap levels.Eric Anholt2014-10-092-3/+23
| | | | | | | | | | | | | | The non-base NPOT levels are stored as POT-aligned images. We get that POT alignment by minifying the POT-aligned base level. This means that level strides are also POT aligned, so we have to tell the rendering mode config that our resource is larger than the actual requested area. Fixes the fbo-generatemipmap-formats NPOT cases. Regresses depthstencil-render-miplevels 273 * -- the texture presentation now works (where it was completely broken before), it looks like there's some overflow of image bounds happening at the lower miplevels.
* vc4: Move the mirrored kernel code to a kernel/ directory.Eric Anholt2014-10-0912-258/+383
| | | | Now this whole setup matches the kernel's file layout much more closely.
* vc4: Enable LIT lowering in TGSI instead of our own code.Eric Anholt2014-10-081-35/+1
| | | | This brings us the -128/128 clamping on the w component.
* vc4: Fix scalar math opcodes to replicate their result from the X channel.Eric Anholt2014-10-081-4/+16
| | | | | Thanks to robclark for pointing out that I was probably failing to do this when I reported a "bug" in his lowering code.
* ilo: fix rectlist on GEN7+Chia-I Wu2014-10-091-0/+3
| | | | | | It was broken by 343b014b57ecc5431477e090100e6a26edbda540. Signed-off-by: Chia-I Wu <[email protected]>
* vc4: Add support for two-sided color.Eric Anholt2014-10-082-18/+51
| | | | | | | | | | It's fairly easy, thanks to Rob Clark's lowering code. Fixes two-sided-lighting and 4 vertex-program-two-side testcases, while regressing 8 testcases that involve enabling two-sided color while only initializing one of the two colors in the VS. If you're enabling two sided color, it's of course expected that you really do set up both colors, so this is still an improvement (and when we set up a linker for TGSI, we'll hopefully fix those 8 fails).
* vc4: Enable POW lowering in TGSI instead of our own code.Eric Anholt2014-10-081-11/+1
|
* vc4: Enable DP lowering in TGSI instead of our own code.Eric Anholt2014-10-081-41/+3
|
* vc4: Start using tgsi_lowering for opcodes we haven't supported before.Eric Anholt2014-10-081-1/+15
|
* gallium: Rename freedreno parts of tgsi_lowering.[ch].Eric Anholt2014-10-083-31/+32
| | | | Acked-by: Rob Clark <[email protected]>
* gallium: Reformat tgsi_lowering.c for the normal style.Eric Anholt2014-10-082-1204/+1201
| | | | Acked-by: Rob Clark <[email protected]>
* gallium: Copy fd_lowering.[ch] to tgsi_lowering.[ch] for code sharing.Eric Anholt2014-10-082-0/+1662
| | | | | | | | Lots of drivers need to transform the weird instructions in TGSI into reasonable scalar ops, and this code can make those translations canonical. Acked-by: Rob Clark <[email protected]>
* vc4: Set unused raddr fields to QPU_R_NOP.Eric Anholt2014-10-081-16/+27
| | | | | | | The simulator assertion fails if you have a write to a reg and then a read (for example, in the NOP side of an instruction), even if the read isn't used for anything. By setting unused raddrs to NOP, we avoid the problem (since only the phsyical registers are tracked).
* vc4: Abstract out the field-merging logic for instructions.Eric Anholt2014-10-081-11/+17
| | | | I'm going to be doing the same logic for some more fields next.
* r600: Use DMA transfers in r600_copy_global_bufferNiels Ole Salscheider2014-10-072-17/+43
| | | | | | v2: Do not demote items that are already in the pool Signed-off-by: Niels Ole Salscheider <[email protected]>
* glsl: Optimize min/max expression treesIago Toral Quiroga2014-10-074-0/+478
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Original patch by Petri Latvala <[email protected]>: Add an optimization pass that drops min/max expression operands that can be proven to not contribute to the final result. The algorithm is similar to alpha-beta pruning on a minmax search, from the field of AI. This optimization pass can optimize min/max expressions where operands are min/max expressions. Such code can appear in shaders by itself, or as the result of clamp() or AMD_shader_trinary_minmax functions. This optimization pass improves the generated code for piglit's AMD_shader_trinary_minmax tests as follows: total instructions in shared programs: 75 -> 67 (-10.67%) instructions in affected programs: 60 -> 52 (-13.33%) GAINED: 0 LOST: 0 All tests (max3, min3, mid3) improved. A full shader-db run: total instructions in shared programs: 4293603 -> 4293575 (-0.00%) instructions in affected programs: 1188 -> 1160 (-2.36%) GAINED: 0 LOST: 0 Improvements happen in Guacamelee and Serious Sam 3. One shader from Dungeon Defenders is hurt by shader-db metrics (26 -> 28), because of dropping of a (constant float (0.00000)) operand, which was compiled to a saturate modifier. Version 2 by Iago Toral Quiroga <[email protected]>: Changes from review feedback: - Squashed various cosmetic changes sent by Matt Turner. - Make less_all_components return an enum rather than setting a class member. (Suggested by Mat Turner). Also, renamed it to compare_components. - Make less_all_components, smaller_constant and larger_constant static. (Suggested by Mat Turner) - Change mixmax_range to call its limits "low" and "high" instead of "range[0]" and "range[1]". (Suggested by Connor Abbot). - Use ir_builder swizzle helpers in swizzle_if_required(). (Suggested by Connor Abbot). - Make the logic more clearer by rearrenging the code and commenting. (Suggested by Connor Abbot). - Added comment to explain why we need to recurse twice. (Suggested by Connor Abbot). - If we cannot prune an expression, do not return early. Instead, attempt to prune its children. (Suggested by Connor Abbot). Other changes: - Instead of having a global "valid" visitor member, let the various functions that can determine this status return a boolean and check for its value to decide what to do in each case. This is more flexible and allows to recurse into children of parents that could not be prunned due to invalid ranges (so related to the last bullet in the review feedback). - Make sure we always check if a range is valid before working with it. Since any use of get_range, combine_range or range_intersection can invalidate a range we should check for this situation every time we use any of these functions. Version 3 by Iago Toral Quiroga <[email protected]>: Changes from review feedback: - Now we can make get_range, combine_range and range_intersection static too (suggested by Connor Abbot). - Do not return NULL when looking for the larger or greater constant into mixed vector constants. Instead, produce a new constant by doing a component-wise minmax. With this we can also remove of the validations when we call into these functions (suggested by Connor Abbot). - Add a comment explaining the meaning of the baserange argument in prune_expression (suggested by Connor Abbot). Other changes: - Eliminate minmax expressions operating on constant vectors with mixed values by resolving them. No piglit regressions observed with Version 3. Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=76861 Reviewed-by: Connor Abbott <[email protected]>
* glsl: do not emit error for non written varyings on OpenGL ESTapani Pälli2014-10-071-2/+16
| | | | | | | | | | | Patch fixes following test case from 'shaders-with-varyings' WebGL conformance suite: "vertex shader with unused varying and fragment shader with used varying must succeed" v2: emit still a warning if the condition happens (Ian) Signed-off-by: Tapani Pälli <[email protected]> Reviewed-by: Ian Romanick <[email protected]>
* radeonsi: Use dummy pixel shader if compilation of the real shader failedMichel Dänzer2014-10-073-7/+22
| | | | | | | Instead of crashing. Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=79155#c5 Reviewed-by: Marek Olšák <[email protected]>
* ilo: let shaders determine surface countsChia-I Wu2014-10-069-202/+267
| | | | | | | | When a shader needs N surfaces, we should upload N surfaces and not depend on how many are bound. This commit is larger than it should be because we did not export how many surfaces a surface uses before. Signed-off-by: Chia-I Wu <[email protected]>
* ilo: let shaders determine sampler countsChia-I Wu2014-10-0413-87/+98
| | | | | | | When a shader needs N samplers, we should upload N samplers and not depend on how many are bound. Signed-off-by: Chia-I Wu <[email protected]>
* tgsi: change tgsi_shader_info::properties to a one-dimensional arrayMarek Olšák2014-10-0413-24/+23
| | | | | | Reviewed-by: Roland Scheidegger <[email protected]> v2: fix svga too
* radeonsi: set number of userdata SGPRs of GS copy shader to 4Marek Olšák2014-10-043-10/+23
| | | | | | | It only needs the constant buffer with clip planes and read-write resources for the GS->VS ring and streamout. That's 2 pointers. Reviewed-by: Michel Dänzer <[email protected]>
generated by cgit v1.2.3 (git 2.39.1) at 2024-12-14 02:41:09 +0000