amd/amdkfd/cwsr_trap_handler_gfx8.asm

4  * Permission is hereby granted, free of charge, to any person obtaining a
9  * Software is furnished to do so, subject to the following conditions:
69 var S_SAVE_BUF_RSRC_WORD1_STRIDE        =   0x00040000          //stride is 4 bytes
165     s_andn2_b32     s_save_status, s_save_status, SQ_WAVE_STATUS_SPI_PRIO_MASK      //check whether this is for save
167     s_and_b32       s_save_trapsts, s_save_trapsts, SQ_WAVE_TRAPSTS_SAVECTX_MASK    //check whether this is for save
168     s_cbranch_scc1  L_SAVE                                      //this is the operation for save
172     /* read tba and tma for next level trap handler, ttmp4 is used as s_save_status */
182     s_and_b32       s_save_trapsts, s_save_trapsts, SQ_WAVE_TRAPSTS_EXCE_MASK // Check whether it is an exception
226     s_sleep 0x2                // sleep 1 (64clk) is not enough for 8 waves per SIMD, which will cause SQ hang, since the 7,8th wave could not get arbit to exec inst, while other waves are stuck into the sleep-loop and waiting for wrexec!=0
319     // SGPR is allocated in 16 SGPR granularity
332     s_cbranch_scc1  L_SAVE_SGPR_LOOP                                    //SGPR save is complete?
369     s_and_b32       s_save_alloc_size, s_save_alloc_size, 0xFFFFFFFF                //lds_size is zero?
372     s_barrier               //LDS is used? wait for other waves in the same TG
373     //s_and_b32     s_save_tmp, tba_hi, S_SAVE_SPI_INIT_FIRST_WAVE_MASK                //exec is still used here
374     s_and_b32       s_save_tmp, s_save_exec_hi, S_SAVE_SPI_INIT_FIRST_WAVE_MASK                //exec is still used here
431     s_lshl_b32      s_save_alloc_size, s_save_alloc_size, 2                         //Number of VGPRs = (vgpr_size + 1) * 4    (non-zero value)   //FIXME for GFX, zero is possible
458     s_cbranch_scc1  L_SAVE_VGPR_LOOP                                                //VGPR save is complete?
499     s_and_b32       s_restore_alloc_size, s_restore_alloc_size, 0xFFFFFFFF                  //lds_size is zero?
522     s_cbranch_scc1  L_RESTORE_LDS_LOOP                                                      //LDS restore is complete?
559     s_cbranch_scc1  L_RESTORE_VGPR_LOOP                                                     //VGPR restore (except v0) is complete?
606     s_cbranch_scc0  L_RESTORE_SGPR_LOOP             //SGPR restore (except s0) is complete?
634     s_waitcnt       lgkmcnt(0)                                                                                      //from now on, it is safe to restore STATUS and IB_STS
651     s_mov_b32       s_restore_tmp, 0x0                                                                              //IB_STS is zero
664     set_status_without_spi_prio(s_restore_status, s_restore_tmp) // SCC is included, which is changed by previous salu
669     s_rfe_restore_b64  s_restore_pc_lo, s_restore_m0            // s_restore_m0[0] is used to set STATUS.inst_atc
687         s_mov_b32 exec_lo, m0                   //assuming exec_lo is not needed anymore from this point on
720     // SQ LDS granularity is 64DW, while PGM_RSRC2.lds_size is in granularity 128DW
728     s_lshl_b32     s_vgpr_size_byte, s_vgpr_size_byte, (2+8) //Number of VGPRs = (vgpr_size + 1) * 4 * 64 * 4   (non-zero value)   //FIXME for GFX, zero is possible