For each block, we compute an instruction dependency graph, which allows us to maintain a worklist of instructions able to be scheduled (no dependencies). We traverse this graph backwards, emitting instructions backwards and reversing the block at the end, to allow conditionals and writeout to be expressed naturally and to enable liveness tracking in the future.
At the moment, we do not schedule for register pressure; we just aim to minimize bundles (register pressure is capped here by a distance metric). Scheduling for register pressure will be the next step as soon as this series is finalized; at that point, we'll augment mir_choose_instruction and mir_choose_bundle with pressure-aware heuristics. Overall, I'm quite happy with how this came out. Full stats below. Highlight is a 10% reduction of cycle count for glmark shaders. total instructions in shared programs: 3500 -> 3472 (-0.80%) instructions in affected programs: 685 -> 657 (-4.09%) helped: 15 HURT: 1 helped stats (abs) min: 1 max: 5 x̄: 1.93 x̃: 1 helped stats (rel) min: 1.49% max: 12.20% x̄: 4.65% x̃: 3.45% HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 HURT stats (rel) min: 4.00% max: 4.00% x̄: 4.00% x̃: 4.00% 95% mean confidence interval for instructions value: -2.61 -0.89 95% mean confidence interval for instructions %-change: -6.03% -2.18% Instructions are helped. total bundles in shared programs: 2020 -> 1804 (-10.69%) bundles in affected programs: 1343 -> 1127 (-16.08%) helped: 45 HURT: 0 helped stats (abs) min: 1 max: 48 x̄: 4.80 x̃: 2 helped stats (rel) min: 5.56% max: 34.78% x̄: 13.50% x̃: 12.50% 95% mean confidence interval for bundles value: -7.15 -2.45 95% mean confidence interval for bundles %-change: -15.98% -11.03% Bundles are helped. total quadwords in shared programs: 3336 -> 3130 (-6.18%) quadwords in affected programs: 2162 -> 1956 (-9.53%) helped: 51 HURT: 0 helped stats (abs) min: 1 max: 17 x̄: 4.04 x̃: 2 helped stats (rel) min: 1.72% max: 22.92% x̄: 10.94% x̃: 11.11% 95% mean confidence interval for quadwords value: -5.32 -2.76 95% mean confidence interval for quadwords %-change: -12.47% -9.42% Quadwords are helped. total registers in shared programs: 343 -> 327 (-4.66%) registers in affected programs: 123 -> 107 (-13.01%) helped: 22 HURT: 5 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 12.50% max: 33.33% x̄: 25.61% x̃: 25.00% HURT stats (abs) min: 1 max: 2 x̄: 1.20 x̃: 1 HURT stats (rel) min: 8.33% max: 100.00% x̄: 37.00% x̃: 20.00% 95% mean confidence interval for registers value: -0.94 -0.24 95% mean confidence interval for registers %-change: -25.65% -2.38% Registers are helped. total threads in shared programs: 451 -> 459 (1.77%) threads in affected programs: 8 -> 16 (100.00%) helped: 4 HURT: 0 helped stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2 helped stats (rel) min: 100.00% max: 100.00% x̄: 100.00% x̃: 100.00% 95% mean confidence interval for threads value: 2.00 2.00 95% mean confidence interval for threads %-change: 100.00% 100.00% total spills in shared programs: 3 -> 3 (0.00%) spills in affected programs: 0 -> 0 helped: 0 HURT: 0 total fills in shared programs: 5 -> 6 (20.00%) fills in affected programs: 5 -> 6 (20.00%) helped: 0 HURT: 1 Alyssa Rosenzweig (30): pan/midgard: Add missing parans in SWIZZLE definition pan/midgard: Fix component count handling for ldst pan/midgard: Squeeze indices before scheduling pan/midgard: Add flatten_mir helper pan/midgard: Calculate dependency graph pan/midgard: Initialize worklist pan/midgard: Add mir_choose_instruction stub pan/midgard: Add mir_update_worklist helper pan/midgard: Add mir_choose_bundle helper pan/midgard: Add mir_schedule_texture/ldst/alu helpers pan/midgard: Remove csel constant unit force pan/midgard: Add constant intersection filters pan/midgard: Add predicate->exclude pan/midgard: Implement predicate->unit pan/midgard: Add helpers for scheduling conditionals pan/midgard: Extend csel_swizzle to branches pan/midgard: Implement load/store pairing pan/midgard: Add mir_choose_alu helper pan/midgard: Add distance metric to choose_instruction pan/midgard: Use new scheduler pan/midgard: Don't double check SCALAR units pan/midgard: Extend choose_instruction for scalar units pan/midgard: Schedule to smul/sadd pan/midgard: Only one conditional per bundle allowed pan/midgard: Allow 6 instructions per bundle pan/midgard: Allow writeout to see into the future pan/midgard: Tightly pack 32-bit constants pan/midgard: Add mir_flip helper pan/midgard: Add csel invert optimization pan/midgard: Allow scheduling conditions with constants src/panfrost/midgard/compiler.h | 27 +- src/panfrost/midgard/helpers.h | 17 +- src/panfrost/midgard/midgard_compile.c | 109 +- src/panfrost/midgard/midgard_opt_invert.c | 38 +- src/panfrost/midgard/midgard_schedule.c | 1303 ++++++++++++++------- src/panfrost/midgard/mir.c | 85 +- 6 files changed, 965 insertions(+), 614 deletions(-) -- 2.23.0 _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
