From: Sagar Ghuge Date: Wed, 15 Jan 2020 00:12:31 +0000 (-0800) Subject: intel/compiler: Clear accumulator register before EOT X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=a27542c5ddec8dd6a64a9c236cf6bea1db1b9e48;p=mesa.git intel/compiler: Clear accumulator register before EOT v2: (Francisco Jerez) - Drop vec4 changes. - Handle explicit acc0 operand and implicit one. - Make sure instruction is SIMD16, prediction is off and default mask control set to true. v3: (Francisco Jerez) - Clear accumulator only when it's written. - Use BRW_MASK_DISABLE instead of true. - Use correct width for brw_acc_reg(). - Fix last_inst_offset. v4: (Francisco Jerez) - Don't check for last instruction for accummulator write. Signed-off-by: Sagar Ghuge Reviewed-by: Francisco Jerez Reviewed-by: Matt Turner Tested-by: Marge Bot Part-of: --- diff --git a/src/intel/compiler/brw_fs_generator.cpp b/src/intel/compiler/brw_fs_generator.cpp index fc19a3283fc..7c8f7bde339 100644 --- a/src/intel/compiler/brw_fs_generator.cpp +++ b/src/intel/compiler/brw_fs_generator.cpp @@ -1711,6 +1711,7 @@ fs_generator::generate_code(const cfg_t *cfg, int dispatch_width, */ int spill_count = 0, fill_count = 0; int loop_count = 0, send_count = 0; + bool is_accum_used = false; struct disasm_info *disasm_info = disasm_initialize(devinfo, cfg); @@ -1741,6 +1742,23 @@ fs_generator::generate_code(const cfg_t *cfg, int dispatch_width, last_insn_offset = p->next_insn_offset; } + /* GEN:BUG:14010017096: + * + * Clear accumulator register before end of thread. + */ + if (inst->eot && is_accum_used && devinfo->gen >= 12) { + brw_set_default_exec_size(p, BRW_EXECUTE_16); + brw_set_default_mask_control(p, BRW_MASK_DISABLE); + brw_set_default_predicate_control(p, BRW_PREDICATE_NONE); + brw_MOV(p, brw_acc_reg(8), brw_imm_f(0.0f)); + last_insn_offset = p->next_insn_offset; + } + + if (!is_accum_used && !inst->eot) { + is_accum_used = inst->writes_accumulator_implicitly(devinfo) || + inst->dst.is_accumulator(); + } + if (unlikely(debug_flag)) disasm_annotate(disasm_info, inst, p->next_insn_offset);