From 1583f49eaae0292eba1a04e67125bb4b92b33b0a Mon Sep 17 00:00:00 2001 From: Ian Romanick Date: Fri, 16 Feb 2018 17:33:13 -0800 Subject: [PATCH] i965/vec4: Allow CSE on subset VF constant loads MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit v2: Rewrite the code that generates the VF mask. Suggested by Ken. No changes on other platforms. Haswell, Ivy Bridge, and Sandy Bridge had similar results. (Haswell shown) total instructions in shared programs: 13059891 -> 13059884 (<.01%) instructions in affected programs: 431 -> 424 (-1.62%) helped: 7 HURT: 0 helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1 helped stats (rel) min: 1.19% max: 5.26% x̄: 2.05% x̃: 1.49% 95% mean confidence interval for instructions value: -1.00 -1.00 95% mean confidence interval for instructions %-change: -3.39% -0.71% Instructions are helped. total cycles in shared programs: 409260032 -> 409260018 (<.01%) cycles in affected programs: 4228 -> 4214 (-0.33%) helped: 7 HURT: 0 helped stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2 helped stats (rel) min: 0.28% max: 2.04% x̄: 0.54% x̃: 0.28% 95% mean confidence interval for cycles value: -2.00 -2.00 95% mean confidence interval for cycles %-change: -1.15% 0.07% Inconclusive result (%-change mean confidence interval includes 0). Signed-off-by: Ian Romanick Reviewed-by: Kenneth Graunke --- src/intel/compiler/brw_vec4_cse.cpp | 19 +++++++++++++++++++ 1 file changed, 19 insertions(+) diff --git a/src/intel/compiler/brw_vec4_cse.cpp b/src/intel/compiler/brw_vec4_cse.cpp index d9f08c96317..c9cf54c6f75 100644 --- a/src/intel/compiler/brw_vec4_cse.cpp +++ b/src/intel/compiler/brw_vec4_cse.cpp @@ -104,6 +104,25 @@ operands_match(const vec4_instruction *a, const vec4_instruction *b) return xs[0].equals(ys[0]) && ((xs[1].equals(ys[1]) && xs[2].equals(ys[2])) || (xs[2].equals(ys[1]) && xs[1].equals(ys[2]))); + } else if (a->opcode == BRW_OPCODE_MOV && + xs[0].file == IMM && + xs[0].type == BRW_REGISTER_TYPE_VF) { + src_reg tmp_x = xs[0]; + src_reg tmp_y = ys[0]; + + /* Smash out the values that are not part of the writemask. Otherwise + * the equals operator will fail due to mismatches in unused components. + */ + const unsigned ab_writemask = a->dst.writemask & b->dst.writemask; + const uint32_t mask = ((ab_writemask & WRITEMASK_X) ? 0x000000ff : 0) | + ((ab_writemask & WRITEMASK_Y) ? 0x0000ff00 : 0) | + ((ab_writemask & WRITEMASK_Z) ? 0x00ff0000 : 0) | + ((ab_writemask & WRITEMASK_W) ? 0xff000000 : 0); + + tmp_x.ud &= mask; + tmp_y.ud &= mask; + + return tmp_x.equals(tmp_y); } else if (!a->is_commutative()) { return xs[0].equals(ys[0]) && xs[1].equals(ys[1]) && xs[2].equals(ys[2]); } else { -- 2.30.2