aboutsummaryrefslogtreecommitdiffstats
path: root/src/mesa/drivers/dri/i965
Commit message (Collapse)AuthorAgeFilesLines
* i965/vs: Refactor min/max handling to share code.Kenneth Graunke2012-11-012-18/+21
| | | | | | | | v2: Properly use "conditionalmod" pre-Gen6, rather than the incorrectly copy-and-pasted "BRW_CONDITIONAL_G". Signed-off-by: Kenneth Graunke <[email protected]> Reviewed-by: Eric Anholt <[email protected]>
* i965/vs: Add support for emitting DPH opcodes.Kenneth Graunke2012-11-013-0/+6
| | | | | Signed-off-by: Kenneth Graunke <[email protected]> Reviewed-by: Eric Anholt <[email protected]>
* i965/vs: Only do INTEL_DEBUG=perf when there's a GLSL shader.Kenneth Graunke2012-11-011-3/+2
| | | | | | | | This will become necessary once we start supporting ARB programs and fixed function in this backend. Signed-off-by: Kenneth Graunke <[email protected]> Reviewed-by: Eric Anholt <[email protected]>
* i965/gen4: Fix assertion failures in depthstencil piglit tests.Eric Anholt2012-11-011-4/+5
| | | | | | Don't forget to set depth_mt even if !hiz_mt. Reviewed-by: Kenneth Graunke <[email protected]>
* i965: Add "alpha to coverage" to performance debug recompile messages.Kenneth Graunke2012-10-311-0/+1
| | | | | | | This was missing and got labeled "Something else". Signed-off-by: Kenneth Graunke <[email protected]> Reviewed-by: Eric Anholt <[email protected]>
* i965: Don't replicate data for zero-stride arrays when copying to VBOs.Kenneth Graunke2012-10-311-7/+6
| | | | | | | | | | | | | | When copy_array_to_vbo_array encountered an array with src_stride == 0 and dst_stride != 0, we would replicate out the single element to the whole size (max - min + 1). This is unnecessary: we can simply upload one copy and set the buffer's stride to 0. Decreases vertex upload overhead in an upcoming Steam for Linux title. Prior to this patch, copy_array_to_vbo_array appeared very high in the profile (Eric quoted 20%). After the patch, it disappeared completely. Reviewed-by: Eric Anholt <[email protected]> Signed-off-by: Kenneth Graunke <[email protected]>
* i965: Don't bother trying to extend the current vertex buffers.Kenneth Graunke2012-10-313-42/+1
| | | | | | | | | | | | | | | | | | | | This essentially reverts the following: commit c625aa19cb53ed27f91bfd16fea6ea727e9a5bbd Author: Chris Wilson <[email protected]> Date: Fri Feb 18 10:37:43 2011 +0000 intel: extend current vertex buffers While working on optimizing an upcoming Steam title, I broke this code. Eric expressed his doubts about this optimization, and noted that the original commit offered no performance data. I ran before and after benchmarks on Xonotic and Citybench, and found that this code made no difference. So, remove it to reduce complexity and make future work simpler. Reviewed-by: Eric Anholt <[email protected]>
* i965: Merge brw_prepare_query_begin() and brw_emit_query_begin().Eric Anholt2012-10-263-22/+7
| | | | | | | This is a leftover from when we had to split those two functions due to the separate BO validation step. Reviewed-by: Kenneth Graunke <[email protected]>
* i965: Rename misleading "active" field of brw->query.Eric Anholt2012-10-262-5/+5
| | | | | | | | | "Active" is an already-used term for the query being between glBeginQuery() and glEndQuery(), while this is tracking whether the start of the packet pair for emitting state has been inserted into the current batchbuffer. Reviewed-by: Kenneth Graunke <[email protected]>
* i965/vs: Preserve the type when copy propagating into an instruction.Kenneth Graunke2012-10-251-0/+1
| | | | | | | | | | | | | | | | | | | | | | Consider the following code, which reinterprets a register as a different type: mov(8) g6<1>F g1.4<0,4,1>.xF and(8) g5<1>.xUD g6<4,4,1>.xUD 0x7fffffffUD Copy propagation would notice that we can replace the use of g6 with g1.4 and eliminate the MOV. Unfortunately, it failed to preserve the UD type, incorrectly generating: and(8) g5<1>.xUD g6<4,4,1>.xF 0x7fffffffUD Found while debugging Ian's uncommitted ARB_vertex_program LOG opcode test with my new Mesa IR -> Vec4 IR translator. NOTE: This is a candidate for stable release branches. Signed-off-by: Kenneth Graunke <[email protected]> Reviewed-by: Eric Anholt <[email protected]>
* i965/vs: Don't lose the MRF writemask when doing compute-to-MRF.Kenneth Graunke2012-10-251-0/+1
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Consider the following code sequence: mul(8) g4<1>F g1<0,4,1>.wzwwF g3<4,4,1>.wzwwF mov.sat(8) m1<1>.xyF g4<4,4,1>F mul(8) g4<1>F g1<0,4,1>.xxyxF g3<4,4,1>.xxyxF mov.sat(8) m1<1>.zwF g4<4,4,1>F The compute-to-MRF pass will discover the first mov.sat and attempt to replace it by rewriting earlier instructions. Everything works out, so it replaces scan_inst's destination file, reg, and reg_offset, resulting in: mul(8) m1<1>F g1<0,4,1>.wzwwF g3<4,4,1>.wzwwF mul(8) g4<1>F g1<0,4,1>.xxyxF g3<4,4,1>.xxyxF mov.sat(8) m1<1>.zwF g4<4,4,1>F Unfortunately, it loses the .xy writemask on the mov.sat's MRF destination. While this doesn't pose an immediate problem, it then proceeds to transform the second mov.sat, resulting in: mul(8) m1<1>F g1<0,4,1>.wzwwF g3<4,4,1>.wzwwF mul(8) m1<1>F g1<0,4,1>.xxyxF g3<4,4,1>.xxyxF Instead of writing both halves of the vector (like the original code), it overwrites the full vector both times, clobbering the desired .xy values. When encountering a MOV, the compute-to-MRF code scans for instructions which generate channels of the MOV source. It ensures that all necessary channels are available (possibly written by several instructions). In this case, *more* channels are available than necessary, so we want to take the subset that's actually used. Taking the bitwise and of both writemasks should accomplish that. This was discovered by analyzing an ARB_vertex_program test (glean/vertProg1/MUL test (with swizzle and masking)) with my new Mesa IR -> Vec4 IR translator code. However, it should be possible with GLSL programs as well. NOTE: This is a candidate for stable release branches. Signed-off-by: Kenneth Graunke <[email protected]> Reviewed-by: Eric Anholt <[email protected]>
* i965/vs: Fix debug dumping of VS push constants.Kenneth Graunke2012-10-251-1/+3
| | | | | | | | | | | | | | | | While copying the values into the batch space, we advance the param pointer. The debug code then tries to iterate over all the uploaded values, starting at param...which is now the end of the uploaded data, rather than the start. This patch saves a pointer to the start of push constant space before it gets altered and switches the debug code to use that. Tested by uncommenting the code and examining the output of glsl-vs-clamp-1.shader_test. Previously all values appeared to be zero. Signed-off-by: Kenneth Graunke <[email protected]> Reviewed-by: Eric Anholt <[email protected]>
* i965: Actually add support for GL_ANY_SAMPLES_PASSED from GL_ARB_oq2.Eric Anholt2012-10-221-0/+12
| | | | | | v2: Fix mangled sentence in the comment, and make the loop exit early. Reviewed-by: Ian Romanick <[email protected]> (v1)
* i965: Stop flushing the batch on timestamp queries, too.Eric Anholt2012-10-191-1/+0
| | | | | | | Given the usecase we have of trying to measure timestamps across individual draw calls, flushing will totally mess up what people are trying to measure. Reviewed-by: Kenneth Graunke <[email protected]>
* i965: Don't flush the batch immediately on EndQuery.Eric Anholt2012-10-191-5/+14
| | | | | | | | | | | | | | | | The theory I had when I wrote the code was that you wanted to minimize latency on your queries because the app was going to ask soon. Only, it turns out that everybody batches up their queries and asks for the results later (often after the next SwapBuffers!), so this was a pessimization. Until now, I had no workload where it mattered enough to benchmark. Recently I started playing some Minecraft, which uses tons of queries to decide whether to render chunks of the terrain. For that app, avoiding the flush in the query-generation loop improves performance 22.7% +/- 4.7% (n=3) on an apitrace capture of it (confirmed in game by watching the fps meter found by pressing F3, 15/16 -> 20/21 fps). Reviewed-by: Kenneth Graunke <[email protected]>
* i965/fs: Fix typo in refactor of brw_fs_reg_allocate.cpp.Eric Anholt2012-10-191-1/+1
| | | | | I'm amazed that my usual warnings check didn't catch this, and that this passed piglit.
* i965/vs: include format argument in debug printfTapani Pälli2012-10-191-1/+1
| | | | | | | | otherwise some compilers will throw error "error: format not a string literal and no format arguments" Signed-off-by: Tapani Pälli <[email protected]> Reviewed-by: Chad Versace <[email protected]>
* i965/fs: Statically allocate the reg_sets at context initialization.Eric Anholt2012-10-173-49/+61
| | | | | | | Now that we've replaced all the variable settings other than reg_width, it's easy to hang on to this (the expensive part of setting up the allocator). Reviewed-by: Kenneth Graunke <[email protected]>
* i965/fs: Allocate registers in the unused parts of the gen7 MRF hack range.Eric Anholt2012-10-172-1/+63
| | | | | | This should also reduce register pressure on gen7+, like the previous commit. Reviewed-by: Kenneth Graunke <[email protected]>
* i965/fs: Reduce the interference between payload regs and virtual GRFs.Eric Anholt2012-10-172-19/+152
| | | | | | | | | | | | | Improves performance of the Lightsmark penumbra shadows scene by 15.7% +/- 1.0% (n=15), by eliminating register spilling. (tested by smashing the list of scenes to have all other scenes have 0 duration -- includes additional rendering of scene description text that normally doesn't appear in that scene) v2: Allow allocation of all but g0/g1 of the payload. v3: Pull count_to_loop_end() out to a helper function. Reviewed-by: Kenneth Graunke <[email protected]> (v2, recommended v3)
* i965/fs: Expose the payload registers to the register allocator.Eric Anholt2012-10-171-7/+39
| | | | | | For now, nothing else can get allocated over them, but that will change. Reviewed-by: Kenneth Graunke <[email protected]>
* i965/fs: Remove extra allocation for classes[].Eric Anholt2012-10-171-1/+1
| | | | | | | This was to slot in the magic aligned pairs class, but it got moved to a descriptive name later. Reviewed-by: Kenneth Graunke <[email protected]>
* i965/fs: Make the register allocation class_sizes[] choice static.Eric Anholt2012-10-172-60/+42
| | | | | | | | Based on split_virtual_grfs(), we choose the same set every time, so set it in stone. This will help us avoid regenerating the somewhat expensive class/register set setup every compile. Reviewed-by: Kenneth Graunke <[email protected]>
* i965/vs: Improve live interval calculation.Eric Anholt2012-10-174-96/+388
| | | | | | | | | | This is derived from the FS visitor code for the same, but tracks each channel separately (otherwise, some typical fill-a-channel-at-a-time patterns would produce excessive live intervals across loops and cause spilling). Reviewed-by: Kenneth Graunke <[email protected]> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=48375 (crash -> failure, can turn into pass by forcing unrolling still)
* i965/vs: Fix the mlen of scratch read/write messages.Eric Anholt2012-10-171-2/+2
| | | | | | | | These messages always have m0 = g0 and m1 = offset, and write has m2 = data. Avoids regression in opt_compute_to_mrf() with a change to scratch writes to set up the data as an MRF write in the IR. Reviewed-by: Kenneth Graunke <[email protected]>
* i965: Make the cfg reusable from the VS.Eric Anholt2012-10-175-16/+16
| | | | Reviewed-by: Kenneth Graunke <[email protected]>
* i965: Share the predicate field between FS and VS.Eric Anholt2012-10-1713-33/+32
| | | | | | | Note that BRW_PREDICATE_NONE is 0 and BRW_PREDICATE_NORMAL is 1, so that's a lot like the true/false we had in the FS before. Reviewed-by: Kenneth Graunke <[email protected]>
* i965: Rename fs_cfg types to not mention fs.Eric Anholt2012-10-178-65/+65
| | | | | | | | fs_bblock_link -> bblock_link fs_bblock -> bblock_t (to avoid conflicting with all the fs_bblock *bblock) fs_cfg -> cfg_t (to avoid conflicting with all the fs_cfg *cfg) Reviewed-by: Kenneth Graunke <[email protected]>
* i965: Move brw_fs_cfg.* to brw_cfg.*.Eric Anholt2012-10-177-6/+6
| | | | Reviewed-by: Kenneth Graunke <[email protected]>
* i965: Make the FS and VS share a few visitor/instruction fields.Eric Anholt2012-10-175-24/+30
| | | | | | This will let us reuse brw_fs_cfg.cpp from brw_vec4_*. Reviewed-by: Kenneth Graunke <[email protected]>
* i965/vs: Trim the swizzle of the scratch write temporary.Eric Anholt2012-10-171-1/+16
| | | | | | | | | This fixes confusion by the upcoming live variable analysis which saw e.g. use of temp.w when only temp.xyz were initialized in the basic block, and concluded that temp.w must have come from outside of the block (even though it was never initialized anywhere). Reviewed-by: Kenneth Graunke <[email protected]>
* i965/vs: Do the temporary allocation in emit_scratch_write().Eric Anholt2012-10-173-21/+12
| | | | | | Both callers were doing basically the same thing, just written differently. Reviewed-by: Kenneth Graunke <[email protected]>
* i965/vs: Simplify emit_scratch_write() prototype.Eric Anholt2012-10-173-8/+6
| | | | | | | Both callers used (effectively) inst->dst as the argument, so just reference it. Reviewed-by: Kenneth Graunke <[email protected]>
* i965/vs: Add a little bit of IR-level debug ability.Eric Anholt2012-10-174-5/+96
| | | | | | | This is super basic, but it let me visualize a problem I had with opt_compute_to_mrf(). Reviewed-by: Kenneth Graunke <[email protected]>
* i965: Fix rendering to small mipmaps of depth/stencil buffers using a temp mt.Eric Anholt2012-10-164-121/+172
| | | | | | | Fixes 51 piglit tests (fbo-clear-formats, and most of the remaining failures in depthstencil). Reviewed-by: Kenneth Graunke <[email protected]>
* i965: Share the draw x/y offset masking code between main/blorp and all gens.Eric Anholt2012-10-166-108/+89
| | | | | | | | | | This code is twisty, and the comment before most of the blocks was actually giving me the opposite impression from its intention: We want to apply as much of our offset as possible through coarse tile-aligned adjustment, since we can do so independently per buffer, and apply the minimum we can through fine-grained drawing offset x/y, since it has to agree between all buffers. Reviewed-by: Kenneth Graunke <[email protected]>
* i965: Make a helper function for the renderbuffer temporary mt workaround.Eric Anholt2012-10-161-22/+1
| | | | | | | We now have a case of wanting to do that on gen6+ as well, so make this logic usable elsewhere. Reviewed-by: Kenneth Graunke <[email protected]>
* i965: Warn on a couple of workarounds in blending.Eric Anholt2012-10-161-2/+15
| | | | Reviewed-by: Kenneth Graunke <[email protected]>
* i965/fs: Fix segfault when using INTEL_DEBUG=perf with non-GLSL.Kenneth Graunke2012-10-161-1/+1
| | | | | | | | | | | | | | | Now that ARB programs and fixed function are routed through the new backend, shader might be NULL. Don't do INTEL_DEBUG=perf support in that case, since it relies on shader->compiled_once. Since INTEL_DEBUG=perf wasn't previously supported, this maintains the status quo. It might be nice to support it someday, however. This could be moved to brw_shader_program instead of brw_shader, but it appears even prog can be NULL in that case. Reviewed-by: Eric Anholt <[email protected]> Signed-off-by: Kenneth Graunke <[email protected]>
* mesa: Remove PROGRAM_WRITE_ONLY register type.Kenneth Graunke2012-10-161-1/+0
| | | | | | | More dead code. I'm not sure what it was for. Reviewed-by: Brian Paul <[email protected]> Reviewed-by: Eric Anholt <[email protected]>
* mesa: Remove support for named parameters.Kenneth Graunke2012-10-161-3/+1
| | | | | | | | | | | These were only part of NV_fragment_program, so we can kill them. The fact that PROGRAM_NAMED_PARAM appears in r200_vertprog.c is rather comedic, but also demonstrates that people just spam the various types of parameters everywhere because they're confusing. Reviewed-by: Brian Paul <[email protected]> Reviewed-by: Eric Anholt <[email protected]>
* mesa: Remove the EmitNVTempInitialization shader compiler option.Kenneth Graunke2012-10-161-1/+0
| | | | | | | Nobody uses it anymore. Reviewed-by: Brian Paul <[email protected]> Reviewed-by: Eric Anholt <[email protected]>
* intel: Remove NV_vertex_program support.Eric Anholt2012-10-152-7/+0
| | | | | | | | | | | | We were holding on to this code because we were aware that NWN 1 had some support for vertex programs -- no other linux programs I've come across would use it (since other software also has ARB_vp or GLSL support). Only, it turns out that NWN doesn't even give us any vertex programs. Given that we have known issues where the extension has never been fully supported, just give up on it. Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=46795 Reviewed-by: Brian Paul <[email protected]>
* i965/vp: Remove more code for unused opcodes.Eric Anholt2012-10-151-55/+1
| | | | | | | These don't appear in ARB_vp or NV_vp and I missed that fact on the first pass of removing dead opcodes. Reviewed-by: Brian Paul <[email protected]>
* i965/vs: Add support for splitting virtual GRFs.Eric Anholt2012-10-103-0/+62
| | | | | | | | | | This should improve our ability to register allocate without spilling. Unfortuantely, due to the live variable analysis being ignorant of loops, we still have register allocation failures on some programs. v2: Add more context to the comment explaining the function. Reviewed-by: Kenneth Graunke <[email protected]> (v1)
* i965/vs: Try again when we've successfully spilled a reg.Eric Anholt2012-10-103-10/+16
| | | | | | | | Before, we'd spill one reg, then continue on without actually register allocating, then assertion fail when we tried to use a vgrf number as a register number. Reviewed-by: Kenneth Graunke <[email protected]>
* i965/vs: Implement register spilling.Kenneth Graunke2012-10-103-1/+144
| | | | | | | | | | | | | | To validate this code, I ran piglit -t vs quick.tests with the "go spill everything" debugging code enabled. There was only one regression: glsl-vs-unroll-explosion simply ran out of registers. This should be fine in the real world, since no one actually spills every single register. NOTE: This is a candidate for the 9.0 branch. Even if it proves to have bugs, it's likely better than simply failing to compile. Signed-off-by: Kenneth Graunke <[email protected]> Reviewed-by: Eric Anholt <[email protected]>
* i965/vs: Fix unit mismatch in scratch base_offset parameter.Kenneth Graunke2012-10-103-4/+8
| | | | | | | | | | | | | | | | | | | | | move_grf_array_access_to_scratch() calculates scratch buffer offsets in bytes. However, emit_scratch_read/write() expects the base_offset parameter to be measured in OWords. As a result, a shader using a scratch read/write offset greater than zero (in practice, a shader containing more than one variable in scratch) would use too large an offset, frequently exceeding the available scratch space. This patch corrects the mismatch by removing spurious conversion from OWords to bytes in move_grf_array_access_to_scratch(). This is based on a patch by Paul Berry. NOTE: This is a candidate for stable release branches. Signed-off-by: Kenneth Graunke <[email protected]> Reviewed-by: Eric Anholt <[email protected]>
* i965: Delete some dead code from brw_eu_emit.c.Kenneth Graunke2012-10-092-120/+0
| | | | Presumably some of this was used by the old fragment shader backend.
* i965/fs: Improve performance of copy/constant propagation.Eric Anholt2012-10-082-18/+30
| | | | | | | | Use a simple chaining hash table for the ACP. This is not really very good, because we still do a full walk of the tree per destination write, but it still reduces fp-long-alu runtime from 5.3 to 3.9s. Reviewed-by: Kenneth Graunke <[email protected]>
generated by cgit v1.2.3 (git 2.39.1) at 2024-12-02 11:41:32 +0000