From 0ba9497e66a72e2f31e0e4d2b899d9bce686d698 Mon Sep 17 00:00:00 2001 From: Ian Romanick Date: Mon, 20 May 2019 17:25:01 -0700 Subject: [PATCH] intel/fs: Improve discard_if code generation MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit Previously we would blindly emit an sequence like: mov(1) f0.1<1>UW g1.14<0,1,0>UW ... cmp.l.f0(16) g7<1>F g5<8,8,1>F 0x41700000F /* 15F */ (+f0.1) cmp.z.f0.1(16) null<1>D g7<8,8,1>D 0D The first move sets the flags based on the initial execution mask. Later discard sequences contain a predicated compare that can only remove more SIMD channels. Often times the only user of the result from the first compare is the second compare. Instead, generate a sequence like mov(1) f0.1<1>UW g1.14<0,1,0>UW ... cmp.l.f0(16) g7<1>F g5<8,8,1>F 0x41700000F /* 15F */ (+f0.1) cmp.ge.f0.1(8) null<1>F g5<8,8,1>F 0x41700000F /* 15F */ If the results stored in g7 and f0.0 are not used, the comparison will be eliminated. This removes an instruction and potentially reduces register pressure. v2: Major re-write of the commit message (including fixing the assembly code). Suggested by Matt. All Gen8+ platforms had similar results. (Ice Lake shown) total instructions in shared programs: 17224434 -> 17198659 (-0.15%) instructions in affected programs: 2908125 -> 2882350 (-0.89%) helped: 18891 HURT: 5 helped stats (abs) min: 1 max: 12 x̄: 1.38 x̃: 1 helped stats (rel) min: 0.03% max: 25.00% x̄: 1.76% x̃: 1.02% HURT stats (abs) min: 9 max: 105 x̄: 51.40 x̃: 35 HURT stats (rel) min: 0.43% max: 4.92% x̄: 2.34% x̃: 1.56% 95% mean confidence interval for instructions value: -1.39 -1.34 95% mean confidence interval for instructions %-change: -1.79% -1.73% Instructions are helped. total cycles in shared programs: 361468458 -> 361170679 (-0.08%) cycles in affected programs: 38470116 -> 38172337 (-0.77%) helped: 16202 HURT: 1456 helped stats (abs) min: 1 max: 4473 x̄: 26.24 x̃: 18 helped stats (rel) min: <.01% max: 28.44% x̄: 2.90% x̃: 2.18% HURT stats (abs) min: 1 max: 5982 x̄: 87.51 x̃: 28 HURT stats (rel) min: <.01% max: 51.29% x̄: 5.48% x̃: 1.64% 95% mean confidence interval for cycles value: -18.24 -15.49 95% mean confidence interval for cycles %-change: -2.26% -2.14% Cycles are helped. total spills in shared programs: 12147 -> 12176 (0.24%) spills in affected programs: 175 -> 204 (16.57%) helped: 8 HURT: 5 total fills in shared programs: 25262 -> 25292 (0.12%) fills in affected programs: 269 -> 299 (11.15%) helped: 8 HURT: 5 Haswell total instructions in shared programs: 13530316 -> 13502647 (-0.20%) instructions in affected programs: 2507824 -> 2480155 (-1.10%) helped: 18859 HURT: 10 helped stats (abs) min: 1 max: 12 x̄: 1.48 x̃: 1 helped stats (rel) min: 0.03% max: 27.78% x̄: 2.38% x̃: 1.41% HURT stats (abs) min: 5 max: 39 x̄: 25.70 x̃: 31 HURT stats (rel) min: 0.22% max: 1.66% x̄: 1.09% x̃: 1.31% 95% mean confidence interval for instructions value: -1.49 -1.44 95% mean confidence interval for instructions %-change: -2.42% -2.34% Instructions are helped. total cycles in shared programs: 377865412 -> 377639034 (-0.06%) cycles in affected programs: 40169572 -> 39943194 (-0.56%) helped: 15550 HURT: 1938 helped stats (abs) min: 1 max: 2482 x̄: 25.67 x̃: 18 helped stats (rel) min: <.01% max: 37.77% x̄: 3.00% x̃: 2.25% HURT stats (abs) min: 1 max: 4862 x̄: 89.17 x̃: 35 HURT stats (rel) min: <.01% max: 67.67% x̄: 6.16% x̃: 2.75% 95% mean confidence interval for cycles value: -14.42 -11.47 95% mean confidence interval for cycles %-change: -2.05% -1.91% Cycles are helped. total spills in shared programs: 26769 -> 26814 (0.17%) spills in affected programs: 826 -> 871 (5.45%) helped: 9 HURT: 10 total fills in shared programs: 38383 -> 38425 (0.11%) fills in affected programs: 834 -> 876 (5.04%) helped: 9 HURT: 10 LOST: 5 GAINED: 10 Ivy Bridge total instructions in shared programs: 12079250 -> 12044139 (-0.29%) instructions in affected programs: 2409680 -> 2374569 (-1.46%) helped: 16135 HURT: 0 helped stats (abs) min: 1 max: 23 x̄: 2.18 x̃: 2 helped stats (rel) min: 0.07% max: 37.50% x̄: 2.72% x̃: 1.68% 95% mean confidence interval for instructions value: -2.21 -2.14 95% mean confidence interval for instructions %-change: -2.76% -2.67% Instructions are helped. total cycles in shared programs: 180116747 -> 179900405 (-0.12%) cycles in affected programs: 25439823 -> 25223481 (-0.85%) helped: 13817 HURT: 1499 helped stats (abs) min: 1 max: 1886 x̄: 26.40 x̃: 18 helped stats (rel) min: <.01% max: 38.84% x̄: 2.57% x̃: 1.97% HURT stats (abs) min: 1 max: 3684 x̄: 98.99 x̃: 52 HURT stats (rel) min: <.01% max: 97.01% x̄: 6.37% x̃: 3.42% 95% mean confidence interval for cycles value: -15.68 -12.57 95% mean confidence interval for cycles %-change: -1.77% -1.63% Cycles are helped. LOST: 8 GAINED: 10 Sandy Bridge total instructions in shared programs: 10878990 -> 10863659 (-0.14%) instructions in affected programs: 1806702 -> 1791371 (-0.85%) helped: 13023 HURT: 0 helped stats (abs) min: 1 max: 5 x̄: 1.18 x̃: 1 helped stats (rel) min: 0.07% max: 13.79% x̄: 1.65% x̃: 1.10% 95% mean confidence interval for instructions value: -1.18 -1.17 95% mean confidence interval for instructions %-change: -1.68% -1.62% Instructions are helped. total cycles in shared programs: 154082878 -> 153862810 (-0.14%) cycles in affected programs: 20199374 -> 19979306 (-1.09%) helped: 12048 HURT: 510 helped stats (abs) min: 1 max: 323 x̄: 20.57 x̃: 18 helped stats (rel) min: 0.03% max: 17.78% x̄: 2.05% x̃: 1.52% HURT stats (abs) min: 1 max: 448 x̄: 54.39 x̃: 16 HURT stats (rel) min: 0.02% max: 37.98% x̄: 4.13% x̃: 1.17% 95% mean confidence interval for cycles value: -17.97 -17.08 95% mean confidence interval for cycles %-change: -1.84% -1.75% Cycles are helped. LOST: 1 GAINED: 0 Iron Lake total instructions in shared programs: 8155075 -> 8142729 (-0.15%) instructions in affected programs: 949495 -> 937149 (-1.30%) helped: 5810 HURT: 0 helped stats (abs) min: 1 max: 8 x̄: 2.12 x̃: 2 helped stats (rel) min: 0.10% max: 16.67% x̄: 2.53% x̃: 1.85% 95% mean confidence interval for instructions value: -2.14 -2.11 95% mean confidence interval for instructions %-change: -2.59% -2.48% Instructions are helped. total cycles in shared programs: 188584610 -> 188549632 (-0.02%) cycles in affected programs: 17274446 -> 17239468 (-0.20%) helped: 3881 HURT: 90 helped stats (abs) min: 2 max: 168 x̄: 9.08 x̃: 6 helped stats (rel) min: <.01% max: 23.53% x̄: 0.83% x̃: 0.30% HURT stats (abs) min: 2 max: 10 x̄: 2.80 x̃: 2 HURT stats (rel) min: <.01% max: 0.60% x̄: 0.10% x̃: 0.07% 95% mean confidence interval for cycles value: -9.35 -8.27 95% mean confidence interval for cycles %-change: -0.85% -0.77% Cycles are helped. GM45 total instructions in shared programs: 5019308 -> 5013119 (-0.12%) instructions in affected programs: 489028 -> 482839 (-1.27%) helped: 2912 HURT: 0 helped stats (abs) min: 1 max: 8 x̄: 2.13 x̃: 2 helped stats (rel) min: 0.10% max: 16.67% x̄: 2.46% x̃: 1.81% 95% mean confidence interval for instructions value: -2.14 -2.11 95% mean confidence interval for instructions %-change: -2.54% -2.39% Instructions are helped. total cycles in shared programs: 129002592 -> 128977804 (-0.02%) cycles in affected programs: 12669152 -> 12644364 (-0.20%) helped: 2759 HURT: 37 helped stats (abs) min: 2 max: 168 x̄: 9.03 x̃: 4 helped stats (rel) min: <.01% max: 21.43% x̄: 0.75% x̃: 0.31% HURT stats (abs) min: 2 max: 10 x̄: 3.62 x̃: 4 HURT stats (rel) min: <.01% max: 0.41% x̄: 0.10% x̃: 0.04% 95% mean confidence interval for cycles value: -9.53 -8.20 95% mean confidence interval for cycles %-change: -0.79% -0.70% Cycles are helped. Reviewed-by: Caio Marcelo de Oliveira Filho Reviewed-by: Matt Turner --- src/intel/compiler/brw_fs_nir.cpp | 46 ++++++++++++++++++++++++++++++++++++--- 1 file changed, 43 insertions(+), 3 deletions(-) diff --git a/src/intel/compiler/brw_fs_nir.cpp b/src/intel/compiler/brw_fs_nir.cpp index b499542bb4a..e0818caae13 100644 --- a/src/intel/compiler/brw_fs_nir.cpp +++ b/src/intel/compiler/brw_fs_nir.cpp @@ -24,6 +24,7 @@ #include "compiler/glsl/ir.h" #include "brw_fs.h" #include "brw_nir.h" +#include "brw_eu.h" #include "nir_search_helpers.h" #include "util/u_math.h" #include "util/bitscan.h" @@ -3503,15 +3504,54 @@ fs_visitor::nir_emit_fs_intrinsic(const fs_builder &bld, * condition, we emit a CMP of g0 != g0, so all currently executing * channels will get turned off. */ - fs_inst *cmp; + fs_inst *cmp = NULL; if (instr->intrinsic == nir_intrinsic_discard_if) { - cmp = bld.CMP(bld.null_reg_f(), get_nir_src(instr->src[0]), - brw_imm_d(0), BRW_CONDITIONAL_Z); + nir_alu_instr *alu = nir_src_as_alu_instr(instr->src[0]); + + if (alu != NULL && + alu->op != nir_op_bcsel && + alu->op != nir_op_inot) { + /* Re-emit the instruction that generated the Boolean value, but + * do not store it. Since this instruction will be conditional, + * other instructions that want to use the real Boolean value may + * get garbage. This was a problem for piglit's fs-discard-exit-2 + * test. + * + * Ideally we'd detect that the instruction cannot have a + * conditional modifier before emitting the instructions. Alas, + * that is nigh impossible. Instead, we're going to assume the + * instruction (or last instruction) generated can have a + * conditional modifier. If it cannot, fallback to the old-style + * compare, and hope dead code elimination will clean up the + * extra instructions generated. + */ + nir_emit_alu(bld, alu, false); + + cmp = (fs_inst *) instructions.get_tail(); + if (cmp->conditional_mod == BRW_CONDITIONAL_NONE) { + if (cmp->can_do_cmod()) + cmp->conditional_mod = BRW_CONDITIONAL_Z; + else + cmp = NULL; + } else { + /* The old sequence that would have been generated is, + * basically, bool_result == false. This is equivalent to + * !bool_result, so negate the old modifier. + */ + cmp->conditional_mod = brw_negate_cmod(cmp->conditional_mod); + } + } + + if (cmp == NULL) { + cmp = bld.CMP(bld.null_reg_f(), get_nir_src(instr->src[0]), + brw_imm_d(0), BRW_CONDITIONAL_Z); + } } else { fs_reg some_reg = fs_reg(retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW)); cmp = bld.CMP(bld.null_reg_f(), some_reg, some_reg, BRW_CONDITIONAL_NZ); } + cmp->predicate = BRW_PREDICATE_NORMAL; cmp->flag_subreg = 1; -- 2.11.0