1/*
2 * Copyright 2016 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 */
22
23/* To compile this assembly code:
24 * PROJECT=greenland ./sp3 cwsr_trap_handler_gfx9.asm -hex tmp.hex
25 */
26
27/* HW (GFX9) source code for CWSR trap handler */
28/* Version 18 + multiple trap handler */
29
30// this performance-optimal version was originally from Seven Xu at SRDC
31
32// Revison #18	 --...
33/* Rev History
34** #1. Branch from gc dv.   //gfxip/gfx9/main/src/test/suites/block/cs/sr/cs_trap_handler.sp3#1,#50, #51, #52-53(Skip, Already Fixed by PV), #54-56(merged),#57-58(mergerd, skiped-already fixed by PV)
35** #4. SR Memory Layout:
36**			 1. VGPR-SGPR-HWREG-{LDS}
37**			 2. tba_hi.bits.26 - reconfigured as the first wave in tg bits, for defer Save LDS for a threadgroup.. performance concern..
38** #5. Update: 1. Accurate g8sr_ts_save_d timestamp
39** #6. Update: 1. Fix s_barrier usage; 2. VGPR s/r using swizzle buffer?(NoNeed, already matched the swizzle pattern, more investigation)
40** #7. Update: 1. don't barrier if noLDS
41** #8. Branch: 1. Branch to ver#0, which is very similar to gc dv version
42**	       2. Fix SQ issue by s_sleep 2
43** #9. Update: 1. Fix scc restore failed issue, restore wave_status at last
44**	       2. optimize s_buffer save by burst 16sgprs...
45** #10. Update 1. Optimize restore sgpr by busrt 16 sgprs.
46** #11. Update 1. Add 2 more timestamp for debug version
47** #12. Update 1. Add VGPR SR using DWx4, some case improve and some case drop performance
48** #13. Integ  1. Always use MUBUF for PV trap shader...
49** #14. Update 1. s_buffer_store soft clause...
50** #15. Update 1. PERF - sclar write with glc:0/mtype0 to allow L2 combine. perf improvement a lot.
51** #16. Update 1. PRRF - UNROLL LDS_DMA got 2500cycle save in IP tree
52** #17. Update 1. FUNC - LDS_DMA has issues while ATC, replace with ds_read/buffer_store for save part[TODO restore part]
53**	       2. PERF - Save LDS before save VGPR to cover LDS save long latency...
54** #18. Update 1. FUNC - Implicitly estore STATUS.VCCZ, which is not writable by s_setreg_b32
55**	       2. FUNC - Handle non-CWSR traps
56*/
57
58var G8SR_WDMEM_HWREG_OFFSET = 0
59var G8SR_WDMEM_SGPR_OFFSET  = 128  // in bytes
60
61// Keep definition same as the app shader, These 2 time stamps are part of the app shader... Should before any Save and after restore.
62
63var G8SR_DEBUG_TIMESTAMP = 0
64var G8SR_DEBUG_TS_SAVE_D_OFFSET = 40*4	// ts_save_d timestamp offset relative to SGPR_SR_memory_offset
65var s_g8sr_ts_save_s	= s[34:35]   // save start
66var s_g8sr_ts_sq_save_msg  = s[36:37]	// The save shader send SAVEWAVE msg to spi
67var s_g8sr_ts_spi_wrexec   = s[38:39]	// the SPI write the sr address to SQ
68var s_g8sr_ts_save_d	= s[40:41]   // save end
69var s_g8sr_ts_restore_s = s[42:43]   // restore start
70var s_g8sr_ts_restore_d = s[44:45]   // restore end
71
72var G8SR_VGPR_SR_IN_DWX4 = 0
73var G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4 = 0x00100000	 // DWx4 stride is 4*4Bytes
74var G8SR_RESTORE_BUF_RSRC_WORD1_STRIDE_DWx4  = G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4
75
76
77/*************************************************************************/
78/*		    control on how to run the shader			 */
79/*************************************************************************/
80//any hack that needs to be made to run this code in EMU (either because various EMU code are not ready or no compute save & restore in EMU run)
81var EMU_RUN_HACK		    =	0
82var EMU_RUN_HACK_RESTORE_NORMAL	    =	0
83var EMU_RUN_HACK_SAVE_NORMAL_EXIT   =	0
84var EMU_RUN_HACK_SAVE_SINGLE_WAVE   =	0
85var EMU_RUN_HACK_SAVE_FIRST_TIME    =	0		    //for interrupted restore in which the first save is through EMU_RUN_HACK
86var SAVE_LDS			    =	1
87var WG_BASE_ADDR_LO		    =	0x9000a000
88var WG_BASE_ADDR_HI		    =	0x0
89var WAVE_SPACE			    =	0x5000		    //memory size that each wave occupies in workgroup state mem
90var CTX_SAVE_CONTROL		    =	0x0
91var CTX_RESTORE_CONTROL		    =	CTX_SAVE_CONTROL
92var SIM_RUN_HACK		    =	0		    //any hack that needs to be made to run this code in SIM (either because various RTL code are not ready or no compute save & restore in RTL run)
93var SGPR_SAVE_USE_SQC		    =	1		    //use SQC D$ to do the write
94var USE_MTBUF_INSTEAD_OF_MUBUF	    =	0		    //because TC EMU currently asserts on 0 of // overload DFMT field to carry 4 more bits of stride for MUBUF opcodes
95var SWIZZLE_EN			    =	0		    //whether we use swizzled buffer addressing
96var ACK_SQC_STORE		    =	1		    //workaround for suspected SQC store bug causing incorrect stores under concurrency
97
98/**************************************************************************/
99/*			variables					  */
100/**************************************************************************/
101var SQ_WAVE_STATUS_INST_ATC_SHIFT  = 23
102var SQ_WAVE_STATUS_INST_ATC_MASK   = 0x00800000
103var SQ_WAVE_STATUS_SPI_PRIO_SHIFT  = 1
104var SQ_WAVE_STATUS_SPI_PRIO_MASK   = 0x00000006
105var SQ_WAVE_STATUS_HALT_MASK       = 0x2000
106var SQ_WAVE_STATUS_PRE_SPI_PRIO_SHIFT   = 0
107var SQ_WAVE_STATUS_PRE_SPI_PRIO_SIZE    = 1
108var SQ_WAVE_STATUS_POST_SPI_PRIO_SHIFT  = 3
109var SQ_WAVE_STATUS_POST_SPI_PRIO_SIZE   = 29
110
111var SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT	= 12
112var SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE	= 9
113var SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT	= 8
114var SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE	= 6
115var SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SHIFT	= 24
116var SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SIZE	= 3			//FIXME	 sq.blk still has 4 bits at this time while SQ programming guide has 3 bits
117
118var SQ_WAVE_TRAPSTS_SAVECTX_MASK    =	0x400
119var SQ_WAVE_TRAPSTS_EXCE_MASK	    =	0x1FF			// Exception mask
120var SQ_WAVE_TRAPSTS_SAVECTX_SHIFT   =	10
121var SQ_WAVE_TRAPSTS_MEM_VIOL_MASK   =	0x100
122var SQ_WAVE_TRAPSTS_MEM_VIOL_SHIFT  =	8
123var SQ_WAVE_TRAPSTS_PRE_SAVECTX_MASK	=   0x3FF
124var SQ_WAVE_TRAPSTS_PRE_SAVECTX_SHIFT	=   0x0
125var SQ_WAVE_TRAPSTS_PRE_SAVECTX_SIZE	=   10
126var SQ_WAVE_TRAPSTS_POST_SAVECTX_MASK	=   0xFFFFF800
127var SQ_WAVE_TRAPSTS_POST_SAVECTX_SHIFT	=   11
128var SQ_WAVE_TRAPSTS_POST_SAVECTX_SIZE	=   21
129var SQ_WAVE_TRAPSTS_ILLEGAL_INST_MASK	=   0x800
130
131var SQ_WAVE_IB_STS_RCNT_SHIFT		=   16			//FIXME
132var SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT	=   15			//FIXME
133var SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK	= 0x1F8000
134var SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK_NEG	= 0x00007FFF	//FIXME
135
136var SQ_BUF_RSRC_WORD1_ATC_SHIFT	    =	24
137var SQ_BUF_RSRC_WORD3_MTYPE_SHIFT   =	27
138
139var TTMP11_SAVE_RCNT_FIRST_REPLAY_SHIFT	=   26			// bits [31:26] unused by SPI debug data
140var TTMP11_SAVE_RCNT_FIRST_REPLAY_MASK	=   0xFC000000
141
142/*	Save	    */
143var S_SAVE_BUF_RSRC_WORD1_STRIDE	=   0x00040000		//stride is 4 bytes
144var S_SAVE_BUF_RSRC_WORD3_MISC		=   0x00807FAC		//SQ_SEL_X/Y/Z/W, BUF_NUM_FORMAT_FLOAT, (0 for MUBUF stride[17:14] when ADD_TID_ENABLE and BUF_DATA_FORMAT_32 for MTBUF), ADD_TID_ENABLE
145
146var S_SAVE_SPI_INIT_ATC_MASK		=   0x08000000		//bit[27]: ATC bit
147var S_SAVE_SPI_INIT_ATC_SHIFT		=   27
148var S_SAVE_SPI_INIT_MTYPE_MASK		=   0x70000000		//bit[30:28]: Mtype
149var S_SAVE_SPI_INIT_MTYPE_SHIFT		=   28
150var S_SAVE_SPI_INIT_FIRST_WAVE_MASK	=   0x04000000		//bit[26]: FirstWaveInTG
151var S_SAVE_SPI_INIT_FIRST_WAVE_SHIFT	=   26
152
153var S_SAVE_PC_HI_RCNT_SHIFT		=   27			//FIXME	 check with Brian to ensure all fields other than PC[47:0] can be used
154var S_SAVE_PC_HI_RCNT_MASK		=   0xF8000000		//FIXME
155var S_SAVE_PC_HI_FIRST_REPLAY_SHIFT	=   26			//FIXME
156var S_SAVE_PC_HI_FIRST_REPLAY_MASK	=   0x04000000		//FIXME
157
158var s_save_spi_init_lo		    =	exec_lo
159var s_save_spi_init_hi		    =	exec_hi
160
161var s_save_pc_lo	    =	ttmp0		//{TTMP1, TTMP0} = {3'h0,pc_rewind[3:0], HT[0],trapID[7:0], PC[47:0]}
162var s_save_pc_hi	    =	ttmp1
163var s_save_exec_lo	    =	ttmp2
164var s_save_exec_hi	    =	ttmp3
165var s_save_tmp		    =	ttmp14
166var s_save_trapsts	    =	ttmp15		//not really used until the end of the SAVE routine
167var s_save_xnack_mask_lo    =	ttmp6
168var s_save_xnack_mask_hi    =	ttmp7
169var s_save_buf_rsrc0	    =	ttmp8
170var s_save_buf_rsrc1	    =	ttmp9
171var s_save_buf_rsrc2	    =	ttmp10
172var s_save_buf_rsrc3	    =	ttmp11
173var s_save_status	    =	ttmp12
174var s_save_mem_offset	    =	ttmp4
175var s_save_alloc_size	    =	s_save_trapsts		//conflict
176var s_save_m0		    =	ttmp5
177var s_save_ttmps_lo	    =	s_save_tmp		//no conflict
178var s_save_ttmps_hi	    =	s_save_trapsts		//no conflict
179
180/*	Restore	    */
181var S_RESTORE_BUF_RSRC_WORD1_STRIDE	    =	S_SAVE_BUF_RSRC_WORD1_STRIDE
182var S_RESTORE_BUF_RSRC_WORD3_MISC	    =	S_SAVE_BUF_RSRC_WORD3_MISC
183
184var S_RESTORE_SPI_INIT_ATC_MASK		    =	0x08000000	    //bit[27]: ATC bit
185var S_RESTORE_SPI_INIT_ATC_SHIFT	    =	27
186var S_RESTORE_SPI_INIT_MTYPE_MASK	    =	0x70000000	    //bit[30:28]: Mtype
187var S_RESTORE_SPI_INIT_MTYPE_SHIFT	    =	28
188var S_RESTORE_SPI_INIT_FIRST_WAVE_MASK	    =	0x04000000	    //bit[26]: FirstWaveInTG
189var S_RESTORE_SPI_INIT_FIRST_WAVE_SHIFT	    =	26
190
191var S_RESTORE_PC_HI_RCNT_SHIFT		    =	S_SAVE_PC_HI_RCNT_SHIFT
192var S_RESTORE_PC_HI_RCNT_MASK		    =	S_SAVE_PC_HI_RCNT_MASK
193var S_RESTORE_PC_HI_FIRST_REPLAY_SHIFT	    =	S_SAVE_PC_HI_FIRST_REPLAY_SHIFT
194var S_RESTORE_PC_HI_FIRST_REPLAY_MASK	    =	S_SAVE_PC_HI_FIRST_REPLAY_MASK
195
196var s_restore_spi_init_lo		    =	exec_lo
197var s_restore_spi_init_hi		    =	exec_hi
198
199var s_restore_mem_offset	=   ttmp12
200var s_restore_alloc_size	=   ttmp3
201var s_restore_tmp		=   ttmp2
202var s_restore_mem_offset_save	=   s_restore_tmp	//no conflict
203
204var s_restore_m0	    =	s_restore_alloc_size	//no conflict
205
206var s_restore_mode	    =	ttmp7
207
208var s_restore_pc_lo	    =	ttmp0
209var s_restore_pc_hi	    =	ttmp1
210var s_restore_exec_lo	    =	ttmp4
211var s_restore_exec_hi	    = 	ttmp5
212var s_restore_status	    =	ttmp14
213var s_restore_trapsts	    =	ttmp15
214var s_restore_xnack_mask_lo =	xnack_mask_lo
215var s_restore_xnack_mask_hi =	xnack_mask_hi
216var s_restore_buf_rsrc0	    =	ttmp8
217var s_restore_buf_rsrc1	    =	ttmp9
218var s_restore_buf_rsrc2	    =	ttmp10
219var s_restore_buf_rsrc3	    =	ttmp11
220var s_restore_ttmps_lo	    =	s_restore_tmp		//no conflict
221var s_restore_ttmps_hi	    =	s_restore_alloc_size	//no conflict
222
223/**************************************************************************/
224/*			trap handler entry points			  */
225/**************************************************************************/
226/* Shader Main*/
227
228shader main
229  asic(GFX9)
230  type(CS)
231
232
233    if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_RESTORE_NORMAL))		    //hack to use trap_id for determining save/restore
234	//FIXME VCCZ un-init assertion s_getreg_b32	s_save_status, hwreg(HW_REG_STATUS)	    //save STATUS since we will change SCC
235	s_and_b32 s_save_tmp, s_save_pc_hi, 0xffff0000		    //change SCC
236	s_cmp_eq_u32 s_save_tmp, 0x007e0000			    //Save: trap_id = 0x7e. Restore: trap_id = 0x7f.
237	s_cbranch_scc0 L_JUMP_TO_RESTORE			    //do not need to recover STATUS here  since we are going to RESTORE
238	//FIXME	 s_setreg_b32	hwreg(HW_REG_STATUS),	s_save_status	    //need to recover STATUS since we are going to SAVE
239	s_branch L_SKIP_RESTORE					    //NOT restore, SAVE actually
240    else
241	s_branch L_SKIP_RESTORE					    //NOT restore. might be a regular trap or save
242    end
243
244L_JUMP_TO_RESTORE:
245    s_branch L_RESTORE						    //restore
246
247L_SKIP_RESTORE:
248
249    s_getreg_b32    s_save_status, hwreg(HW_REG_STATUS)				    //save STATUS since we will change SCC
250    s_andn2_b32	    s_save_status, s_save_status, SQ_WAVE_STATUS_SPI_PRIO_MASK	    //check whether this is for save
251    s_getreg_b32    s_save_trapsts, hwreg(HW_REG_TRAPSTS)
252    s_and_b32       ttmp2, s_save_trapsts, SQ_WAVE_TRAPSTS_SAVECTX_MASK    //check whether this is for save
253    s_cbranch_scc1  L_SAVE					//this is the operation for save
254
255    // *********    Handle non-CWSR traps	*******************
256if (!EMU_RUN_HACK)
257    // Illegal instruction is a non-maskable exception which blocks context save.
258    // Halt the wavefront and return from the trap.
259    s_and_b32       ttmp2, s_save_trapsts, SQ_WAVE_TRAPSTS_ILLEGAL_INST_MASK
260    s_cbranch_scc1  L_HALT_WAVE
261
262    // If STATUS.MEM_VIOL is asserted then we cannot fetch from the TMA.
263    // Instead, halt the wavefront and return from the trap.
264    s_and_b32       ttmp2, s_save_trapsts, SQ_WAVE_TRAPSTS_MEM_VIOL_MASK
265    s_cbranch_scc0  L_FETCH_2ND_TRAP
266
267L_HALT_WAVE:
268    // If STATUS.HALT is set then this fault must come from SQC instruction fetch.
269    // We cannot prevent further faults. Spin wait until context saved.
270    s_and_b32       ttmp2, s_save_status, SQ_WAVE_STATUS_HALT_MASK
271    s_cbranch_scc0  L_NOT_ALREADY_HALTED
272
273L_WAIT_CTX_SAVE:
274    s_sleep         0x10
275    s_getreg_b32    ttmp2, hwreg(HW_REG_TRAPSTS)
276    s_and_b32       ttmp2, ttmp2, SQ_WAVE_TRAPSTS_SAVECTX_MASK
277    s_cbranch_scc0  L_WAIT_CTX_SAVE
278
279L_NOT_ALREADY_HALTED:
280    s_or_b32        s_save_status, s_save_status, SQ_WAVE_STATUS_HALT_MASK
281
282    // If the PC points to S_ENDPGM then context save will fail if STATUS.HALT is set.
283    // Rewind the PC to prevent this from occurring. The debugger compensates for this.
284    s_sub_u32       ttmp0, ttmp0, 0x8
285    s_subb_u32      ttmp1, ttmp1, 0x0
286
287L_FETCH_2ND_TRAP:
288    // Preserve and clear scalar XNACK state before issuing scalar reads.
289    // Save IB_STS.FIRST_REPLAY[15] and IB_STS.RCNT[20:16] into unused space ttmp11[31:26].
290    s_getreg_b32    ttmp2, hwreg(HW_REG_IB_STS)
291    s_and_b32       ttmp3, ttmp2, SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK
292    s_lshl_b32      ttmp3, ttmp3, (TTMP11_SAVE_RCNT_FIRST_REPLAY_SHIFT - SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT)
293    s_andn2_b32     ttmp11, ttmp11, TTMP11_SAVE_RCNT_FIRST_REPLAY_MASK
294    s_or_b32        ttmp11, ttmp11, ttmp3
295
296    s_andn2_b32     ttmp2, ttmp2, SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK
297    s_setreg_b32    hwreg(HW_REG_IB_STS), ttmp2
298
299    // Read second-level TBA/TMA from first-level TMA and jump if available.
300    // ttmp[2:5] and ttmp12 can be used (others hold SPI-initialized debug data)
301    // ttmp12 holds SQ_WAVE_STATUS
302    s_getreg_b32    ttmp14, hwreg(HW_REG_SQ_SHADER_TMA_LO)
303    s_getreg_b32    ttmp15, hwreg(HW_REG_SQ_SHADER_TMA_HI)
304    s_lshl_b64      [ttmp14, ttmp15], [ttmp14, ttmp15], 0x8
305    s_load_dwordx2  [ttmp2, ttmp3], [ttmp14, ttmp15], 0x0 glc:1 // second-level TBA
306    s_waitcnt       lgkmcnt(0)
307    s_load_dwordx2  [ttmp14, ttmp15], [ttmp14, ttmp15], 0x8 glc:1 // second-level TMA
308    s_waitcnt       lgkmcnt(0)
309    s_and_b64       [ttmp2, ttmp3], [ttmp2, ttmp3], [ttmp2, ttmp3]
310    s_cbranch_scc0  L_NO_NEXT_TRAP // second-level trap handler not been set
311    s_setpc_b64     [ttmp2, ttmp3] // jump to second-level trap handler
312
313L_NO_NEXT_TRAP:
314    s_getreg_b32    s_save_trapsts, hwreg(HW_REG_TRAPSTS)
315    s_and_b32	    s_save_trapsts, s_save_trapsts, SQ_WAVE_TRAPSTS_EXCE_MASK // Check whether it is an exception
316    s_cbranch_scc1  L_EXCP_CASE	  // Exception, jump back to the shader program directly.
317    s_add_u32	    ttmp0, ttmp0, 4   // S_TRAP case, add 4 to ttmp0
318    s_addc_u32	ttmp1, ttmp1, 0
319L_EXCP_CASE:
320    s_and_b32	ttmp1, ttmp1, 0xFFFF
321
322    // Restore SQ_WAVE_IB_STS.
323    s_lshr_b32      ttmp2, ttmp11, (TTMP11_SAVE_RCNT_FIRST_REPLAY_SHIFT - SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT)
324    s_and_b32       ttmp2, ttmp2, SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK
325    s_setreg_b32    hwreg(HW_REG_IB_STS), ttmp2
326
327    // Restore SQ_WAVE_STATUS.
328    s_and_b64       exec, exec, exec // Restore STATUS.EXECZ, not writable by s_setreg_b32
329    s_and_b64       vcc, vcc, vcc    // Restore STATUS.VCCZ, not writable by s_setreg_b32
330    set_status_without_spi_prio(s_save_status, ttmp2)
331
332    s_rfe_b64       [ttmp0, ttmp1]
333end
334    // *********	End handling of non-CWSR traps	 *******************
335
336/**************************************************************************/
337/*			save routine					  */
338/**************************************************************************/
339
340L_SAVE:
341
342if G8SR_DEBUG_TIMESTAMP
343	s_memrealtime	s_g8sr_ts_save_s
344	s_waitcnt lgkmcnt(0)	     //FIXME, will cause xnack??
345end
346
347    s_and_b32	    s_save_pc_hi, s_save_pc_hi, 0x0000ffff    //pc[47:32]
348
349    s_mov_b32	    s_save_tmp, 0							    //clear saveCtx bit
350    s_setreg_b32    hwreg(HW_REG_TRAPSTS, SQ_WAVE_TRAPSTS_SAVECTX_SHIFT, 1), s_save_tmp	    //clear saveCtx bit
351
352    s_getreg_b32    s_save_tmp, hwreg(HW_REG_IB_STS, SQ_WAVE_IB_STS_RCNT_SHIFT, SQ_WAVE_IB_STS_RCNT_SIZE)		    //save RCNT
353    s_lshl_b32	    s_save_tmp, s_save_tmp, S_SAVE_PC_HI_RCNT_SHIFT
354    s_or_b32	    s_save_pc_hi, s_save_pc_hi, s_save_tmp
355    s_getreg_b32    s_save_tmp, hwreg(HW_REG_IB_STS, SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT, SQ_WAVE_IB_STS_FIRST_REPLAY_SIZE)   //save FIRST_REPLAY
356    s_lshl_b32	    s_save_tmp, s_save_tmp, S_SAVE_PC_HI_FIRST_REPLAY_SHIFT
357    s_or_b32	    s_save_pc_hi, s_save_pc_hi, s_save_tmp
358    s_getreg_b32    s_save_tmp, hwreg(HW_REG_IB_STS)					    //clear RCNT and FIRST_REPLAY in IB_STS
359    s_and_b32	    s_save_tmp, s_save_tmp, SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK_NEG
360
361    s_setreg_b32    hwreg(HW_REG_IB_STS), s_save_tmp
362
363    /*	    inform SPI the readiness and wait for SPI's go signal */
364    s_mov_b32	    s_save_exec_lo, exec_lo						    //save EXEC and use EXEC for the go signal from SPI
365    s_mov_b32	    s_save_exec_hi, exec_hi
366    s_mov_b64	    exec,   0x0								    //clear EXEC to get ready to receive
367
368if G8SR_DEBUG_TIMESTAMP
369	s_memrealtime  s_g8sr_ts_sq_save_msg
370	s_waitcnt lgkmcnt(0)
371end
372
373    if (EMU_RUN_HACK)
374
375    else
376	s_sendmsg   sendmsg(MSG_SAVEWAVE)  //send SPI a message and wait for SPI's write to EXEC
377    end
378
379    // Set SPI_PRIO=2 to avoid starving instruction fetch in the waves we're waiting for.
380    s_or_b32 s_save_tmp, s_save_status, (2 << SQ_WAVE_STATUS_SPI_PRIO_SHIFT)
381    s_setreg_b32 hwreg(HW_REG_STATUS), s_save_tmp
382
383  L_SLEEP:
384    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
385
386    if (EMU_RUN_HACK)
387
388    else
389	s_cbranch_execz L_SLEEP
390    end
391
392if G8SR_DEBUG_TIMESTAMP
393	s_memrealtime  s_g8sr_ts_spi_wrexec
394	s_waitcnt lgkmcnt(0)
395end
396
397    if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_SAVE_SINGLE_WAVE))
398	//calculate wd_addr using absolute thread id
399	v_readlane_b32 s_save_tmp, v9, 0
400	s_lshr_b32 s_save_tmp, s_save_tmp, 6
401	s_mul_i32 s_save_tmp, s_save_tmp, WAVE_SPACE
402	s_add_i32 s_save_spi_init_lo, s_save_tmp, WG_BASE_ADDR_LO
403	s_mov_b32 s_save_spi_init_hi, WG_BASE_ADDR_HI
404	s_and_b32 s_save_spi_init_hi, s_save_spi_init_hi, CTX_SAVE_CONTROL
405    else
406    end
407    if ((EMU_RUN_HACK) && (EMU_RUN_HACK_SAVE_SINGLE_WAVE))
408	s_add_i32 s_save_spi_init_lo, s_save_tmp, WG_BASE_ADDR_LO
409	s_mov_b32 s_save_spi_init_hi, WG_BASE_ADDR_HI
410	s_and_b32 s_save_spi_init_hi, s_save_spi_init_hi, CTX_SAVE_CONTROL
411    else
412    end
413
414    // Save trap temporaries 4-11, 13 initialized by SPI debug dispatch logic
415    // ttmp SR memory offset : size(VGPR)+size(SGPR)+0x40
416    get_vgpr_size_bytes(s_save_ttmps_lo)
417    get_sgpr_size_bytes(s_save_ttmps_hi)
418    s_add_u32	    s_save_ttmps_lo, s_save_ttmps_lo, s_save_ttmps_hi
419    s_add_u32	    s_save_ttmps_lo, s_save_ttmps_lo, s_save_spi_init_lo
420    s_addc_u32	    s_save_ttmps_hi, s_save_spi_init_hi, 0x0
421    s_and_b32	    s_save_ttmps_hi, s_save_ttmps_hi, 0xFFFF
422    s_store_dwordx4 [ttmp4, ttmp5, ttmp6, ttmp7], [s_save_ttmps_lo, s_save_ttmps_hi], 0x50 glc:1
423    ack_sqc_store_workaround()
424    s_store_dwordx4 [ttmp8, ttmp9, ttmp10, ttmp11], [s_save_ttmps_lo, s_save_ttmps_hi], 0x60 glc:1
425    ack_sqc_store_workaround()
426    s_store_dword   ttmp13, [s_save_ttmps_lo, s_save_ttmps_hi], 0x74 glc:1
427    ack_sqc_store_workaround()
428
429    /*	    setup Resource Contants    */
430    s_mov_b32	    s_save_buf_rsrc0,	s_save_spi_init_lo							//base_addr_lo
431    s_and_b32	    s_save_buf_rsrc1,	s_save_spi_init_hi, 0x0000FFFF						//base_addr_hi
432    s_or_b32	    s_save_buf_rsrc1,	s_save_buf_rsrc1,  S_SAVE_BUF_RSRC_WORD1_STRIDE
433    s_mov_b32	    s_save_buf_rsrc2,	0									//NUM_RECORDS initial value = 0 (in bytes) although not neccessarily inited
434    s_mov_b32	    s_save_buf_rsrc3,	S_SAVE_BUF_RSRC_WORD3_MISC
435    s_and_b32	    s_save_tmp,		s_save_spi_init_hi, S_SAVE_SPI_INIT_ATC_MASK
436    s_lshr_b32	    s_save_tmp,		s_save_tmp, (S_SAVE_SPI_INIT_ATC_SHIFT-SQ_BUF_RSRC_WORD1_ATC_SHIFT)	    //get ATC bit into position
437    s_or_b32	    s_save_buf_rsrc3,	s_save_buf_rsrc3,  s_save_tmp						//or ATC
438    s_and_b32	    s_save_tmp,		s_save_spi_init_hi, S_SAVE_SPI_INIT_MTYPE_MASK
439    s_lshr_b32	    s_save_tmp,		s_save_tmp, (S_SAVE_SPI_INIT_MTYPE_SHIFT-SQ_BUF_RSRC_WORD3_MTYPE_SHIFT)	    //get MTYPE bits into position
440    s_or_b32	    s_save_buf_rsrc3,	s_save_buf_rsrc3,  s_save_tmp						//or MTYPE
441
442    //FIXME  right now s_save_m0/s_save_mem_offset use tma_lo/tma_hi  (might need to save them before using them?)
443    s_mov_b32	    s_save_m0,		m0								    //save M0
444
445    /*	    global mem offset		*/
446    s_mov_b32	    s_save_mem_offset,	0x0									//mem offset initial value = 0
447
448
449
450
451    /*	    save HW registers	*/
452    //////////////////////////////
453
454  L_SAVE_HWREG:
455	// HWREG SR memory offset : size(VGPR)+size(SGPR)
456       get_vgpr_size_bytes(s_save_mem_offset)
457       get_sgpr_size_bytes(s_save_tmp)
458       s_add_u32 s_save_mem_offset, s_save_mem_offset, s_save_tmp
459
460
461    s_mov_b32	    s_save_buf_rsrc2, 0x4				//NUM_RECORDS	in bytes
462    if (SWIZZLE_EN)
463	s_add_u32	s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
464    else
465	s_mov_b32	s_save_buf_rsrc2,  0x1000000				    //NUM_RECORDS in bytes
466    end
467
468
469    write_hwreg_to_mem(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset)			//M0
470
471    if ((EMU_RUN_HACK) && (EMU_RUN_HACK_SAVE_FIRST_TIME))
472	s_add_u32 s_save_pc_lo, s_save_pc_lo, 4		    //pc[31:0]+4
473	s_addc_u32 s_save_pc_hi, s_save_pc_hi, 0x0	    //carry bit over
474    end
475
476    write_hwreg_to_mem(s_save_pc_lo, s_save_buf_rsrc0, s_save_mem_offset)		    //PC
477    write_hwreg_to_mem(s_save_pc_hi, s_save_buf_rsrc0, s_save_mem_offset)
478    write_hwreg_to_mem(s_save_exec_lo, s_save_buf_rsrc0, s_save_mem_offset)		//EXEC
479    write_hwreg_to_mem(s_save_exec_hi, s_save_buf_rsrc0, s_save_mem_offset)
480    write_hwreg_to_mem(s_save_status, s_save_buf_rsrc0, s_save_mem_offset)		//STATUS
481
482    //s_save_trapsts conflicts with s_save_alloc_size
483    s_getreg_b32    s_save_trapsts, hwreg(HW_REG_TRAPSTS)
484    write_hwreg_to_mem(s_save_trapsts, s_save_buf_rsrc0, s_save_mem_offset)		//TRAPSTS
485
486    write_hwreg_to_mem(xnack_mask_lo, s_save_buf_rsrc0, s_save_mem_offset)	    //XNACK_MASK_LO
487    write_hwreg_to_mem(xnack_mask_hi, s_save_buf_rsrc0, s_save_mem_offset)	    //XNACK_MASK_HI
488
489    //use s_save_tmp would introduce conflict here between s_save_tmp and s_save_buf_rsrc2
490    s_getreg_b32    s_save_m0, hwreg(HW_REG_MODE)						    //MODE
491    write_hwreg_to_mem(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset)
492
493
494
495    /*	    the first wave in the threadgroup	 */
496    s_and_b32	    s_save_tmp, s_save_spi_init_hi, S_SAVE_SPI_INIT_FIRST_WAVE_MASK	// extract fisrt wave bit
497    s_mov_b32	     s_save_exec_hi, 0x0
498    s_or_b32	     s_save_exec_hi, s_save_tmp, s_save_exec_hi				 // save first wave bit to s_save_exec_hi.bits[26]
499
500
501    /*		save SGPRs	*/
502	// Save SGPR before LDS save, then the s0 to s4 can be used during LDS save...
503    //////////////////////////////
504
505    // SGPR SR memory offset : size(VGPR)
506    get_vgpr_size_bytes(s_save_mem_offset)
507    // TODO, change RSRC word to rearrange memory layout for SGPRS
508
509    s_getreg_b32    s_save_alloc_size, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SIZE)		//spgr_size
510    s_add_u32	    s_save_alloc_size, s_save_alloc_size, 1
511    s_lshl_b32	    s_save_alloc_size, s_save_alloc_size, 4			    //Number of SGPRs = (sgpr_size + 1) * 16   (non-zero value)
512
513    if (SGPR_SAVE_USE_SQC)
514	s_lshl_b32	s_save_buf_rsrc2,   s_save_alloc_size, 2		    //NUM_RECORDS in bytes
515    else
516	s_lshl_b32	s_save_buf_rsrc2,   s_save_alloc_size, 8		    //NUM_RECORDS in bytes (64 threads)
517    end
518
519    if (SWIZZLE_EN)
520	s_add_u32	s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
521    else
522	s_mov_b32	s_save_buf_rsrc2,  0x1000000				    //NUM_RECORDS in bytes
523    end
524
525
526    // backup s_save_buf_rsrc0,1 to s_save_pc_lo/hi, since write_16sgpr_to_mem function will change the rsrc0
527    //s_mov_b64 s_save_pc_lo, s_save_buf_rsrc0
528    s_mov_b64 s_save_xnack_mask_lo, s_save_buf_rsrc0
529    s_add_u32 s_save_buf_rsrc0, s_save_buf_rsrc0, s_save_mem_offset
530    s_addc_u32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0
531
532    s_mov_b32	    m0, 0x0			    //SGPR initial index value =0
533    s_nop	    0x0				    //Manually inserted wait states
534  L_SAVE_SGPR_LOOP:
535    // SGPR is allocated in 16 SGPR granularity
536    s_movrels_b64   s0, s0     //s0 = s[0+m0], s1 = s[1+m0]
537    s_movrels_b64   s2, s2     //s2 = s[2+m0], s3 = s[3+m0]
538    s_movrels_b64   s4, s4     //s4 = s[4+m0], s5 = s[5+m0]
539    s_movrels_b64   s6, s6     //s6 = s[6+m0], s7 = s[7+m0]
540    s_movrels_b64   s8, s8     //s8 = s[8+m0], s9 = s[9+m0]
541    s_movrels_b64   s10, s10   //s10 = s[10+m0], s11 = s[11+m0]
542    s_movrels_b64   s12, s12   //s12 = s[12+m0], s13 = s[13+m0]
543    s_movrels_b64   s14, s14   //s14 = s[14+m0], s15 = s[15+m0]
544
545    write_16sgpr_to_mem(s0, s_save_buf_rsrc0, s_save_mem_offset) //PV: the best performance should be using s_buffer_store_dwordx4
546    s_add_u32	    m0, m0, 16							    //next sgpr index
547    s_cmp_lt_u32    m0, s_save_alloc_size					    //scc = (m0 < s_save_alloc_size) ? 1 : 0
548    s_cbranch_scc1  L_SAVE_SGPR_LOOP					//SGPR save is complete?
549    // restore s_save_buf_rsrc0,1
550    //s_mov_b64 s_save_buf_rsrc0, s_save_pc_lo
551    s_mov_b64 s_save_buf_rsrc0, s_save_xnack_mask_lo
552
553
554
555
556    /*		save first 4 VGPR, then LDS save could use   */
557	// each wave will alloc 4 vgprs at least...
558    /////////////////////////////////////////////////////////////////////////////////////
559
560    s_mov_b32	    s_save_mem_offset, 0
561    s_mov_b32	    exec_lo, 0xFFFFFFFF						    //need every thread from now on
562    s_mov_b32	    exec_hi, 0xFFFFFFFF
563    s_mov_b32	    xnack_mask_lo, 0x0
564    s_mov_b32	    xnack_mask_hi, 0x0
565
566    if (SWIZZLE_EN)
567	s_add_u32	s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
568    else
569	s_mov_b32	s_save_buf_rsrc2,  0x1000000				    //NUM_RECORDS in bytes
570    end
571
572
573    // VGPR Allocated in 4-GPR granularity
574
575if G8SR_VGPR_SR_IN_DWX4
576	// the const stride for DWx4 is 4*4 bytes
577	s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF   // reset const stride to 0
578	s_or_b32  s_save_buf_rsrc1, s_save_buf_rsrc1, G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4  // const stride to 4*4 bytes
579
580	buffer_store_dwordx4 v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1
581
582	s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF   // reset const stride to 0
583	s_or_b32  s_save_buf_rsrc1, s_save_buf_rsrc1, S_SAVE_BUF_RSRC_WORD1_STRIDE  // reset const stride to 4 bytes
584else
585	buffer_store_dword v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1
586	buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1  offset:256
587	buffer_store_dword v2, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1  offset:256*2
588	buffer_store_dword v3, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1  offset:256*3
589end
590
591
592
593    /*		save LDS	*/
594    //////////////////////////////
595
596  L_SAVE_LDS:
597
598	// Change EXEC to all threads...
599    s_mov_b32	    exec_lo, 0xFFFFFFFF	  //need every thread from now on
600    s_mov_b32	    exec_hi, 0xFFFFFFFF
601
602    s_getreg_b32    s_save_alloc_size, hwreg(HW_REG_LDS_ALLOC,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE)		    //lds_size
603    s_and_b32	    s_save_alloc_size, s_save_alloc_size, 0xFFFFFFFF		    //lds_size is zero?
604    s_cbranch_scc0  L_SAVE_LDS_DONE									       //no lds used? jump to L_SAVE_DONE
605
606    s_barrier		    //LDS is used? wait for other waves in the same TG
607    s_and_b32	    s_save_tmp, s_save_exec_hi, S_SAVE_SPI_INIT_FIRST_WAVE_MASK		       //exec is still used here
608    s_cbranch_scc0  L_SAVE_LDS_DONE
609
610	// first wave do LDS save;
611
612    s_lshl_b32	    s_save_alloc_size, s_save_alloc_size, 6			    //LDS size in dwords = lds_size * 64dw
613    s_lshl_b32	    s_save_alloc_size, s_save_alloc_size, 2			    //LDS size in bytes
614    s_mov_b32	    s_save_buf_rsrc2,  s_save_alloc_size			    //NUM_RECORDS in bytes
615
616    // LDS at offset: size(VGPR)+SIZE(SGPR)+SIZE(HWREG)
617    //
618    get_vgpr_size_bytes(s_save_mem_offset)
619    get_sgpr_size_bytes(s_save_tmp)
620    s_add_u32  s_save_mem_offset, s_save_mem_offset, s_save_tmp
621    s_add_u32 s_save_mem_offset, s_save_mem_offset, get_hwreg_size_bytes()
622
623
624    if (SWIZZLE_EN)
625	s_add_u32	s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0	      //FIXME need to use swizzle to enable bounds checking?
626    else
627	s_mov_b32	s_save_buf_rsrc2,  0x1000000		      //NUM_RECORDS in bytes
628    end
629
630    s_mov_b32	    m0, 0x0						  //lds_offset initial value = 0
631
632
633var LDS_DMA_ENABLE = 0
634var UNROLL = 0
635if UNROLL==0 && LDS_DMA_ENABLE==1
636	s_mov_b32  s3, 256*2
637	s_nop 0
638	s_nop 0
639	s_nop 0
640  L_SAVE_LDS_LOOP:
641	//TODO: looks the 2 buffer_store/load clause for s/r will hurt performance.???
642    if (SAVE_LDS)     //SPI always alloc LDS space in 128DW granularity
643	    buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1		// first 64DW
644	    buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1 offset:256 // second 64DW
645    end
646
647    s_add_u32	    m0, m0, s3						//every buffer_store_lds does 256 bytes
648    s_add_u32	    s_save_mem_offset, s_save_mem_offset, s3				//mem offset increased by 256 bytes
649    s_cmp_lt_u32    m0, s_save_alloc_size						//scc=(m0 < s_save_alloc_size) ? 1 : 0
650    s_cbranch_scc1  L_SAVE_LDS_LOOP							//LDS save is complete?
651
652elsif LDS_DMA_ENABLE==1 && UNROLL==1 // UNROOL	, has ichace miss
653      // store from higest LDS address to lowest
654      s_mov_b32	 s3, 256*2
655      s_sub_u32	 m0, s_save_alloc_size, s3
656      s_add_u32 s_save_mem_offset, s_save_mem_offset, m0
657      s_lshr_b32 s_save_alloc_size, s_save_alloc_size, 9   // how many 128 trunks...
658      s_sub_u32 s_save_alloc_size, 128, s_save_alloc_size   // store from higheset addr to lowest
659      s_mul_i32 s_save_alloc_size, s_save_alloc_size, 6*4   // PC offset increment,  each LDS save block cost 6*4 Bytes instruction
660      s_add_u32 s_save_alloc_size, s_save_alloc_size, 3*4   //2is the below 2 inst...//s_addc and s_setpc
661      s_nop 0
662      s_nop 0
663      s_nop 0	//pad 3 dw to let LDS_DMA align with 64Bytes
664      s_getpc_b64 s[0:1]			      // reuse s[0:1], since s[0:1] already saved
665      s_add_u32	  s0, s0,s_save_alloc_size
666      s_addc_u32  s1, s1, 0
667      s_setpc_b64 s[0:1]
668
669
670       for var i =0; i< 128; i++
671	    // be careful to make here a 64Byte aligned address, which could improve performance...
672	    buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1 offset:0		// first 64DW
673	    buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1 offset:256		  // second 64DW
674
675	if i!=127
676	s_sub_u32  m0, m0, s3	   // use a sgpr to shrink 2DW-inst to 1DW inst to improve performance , i.e.  pack more LDS_DMA inst to one Cacheline
677	    s_sub_u32  s_save_mem_offset, s_save_mem_offset,  s3
678	    end
679       end
680
681else   // BUFFER_STORE
682      v_mbcnt_lo_u32_b32 v2, 0xffffffff, 0x0
683      v_mbcnt_hi_u32_b32 v3, 0xffffffff, v2	// tid
684      v_mul_i32_i24 v2, v3, 8	// tid*8
685      v_mov_b32 v3, 256*2
686      s_mov_b32 m0, 0x10000
687      s_mov_b32 s0, s_save_buf_rsrc3
688      s_and_b32 s_save_buf_rsrc3, s_save_buf_rsrc3, 0xFF7FFFFF	  // disable add_tid
689      s_or_b32 s_save_buf_rsrc3, s_save_buf_rsrc3, 0x58000   //DFMT
690
691L_SAVE_LDS_LOOP_VECTOR:
692      ds_read_b64 v[0:1], v2	//x =LDS[a], byte address
693      s_waitcnt lgkmcnt(0)
694      buffer_store_dwordx2  v[0:1], v2, s_save_buf_rsrc0, s_save_mem_offset offen:1  glc:1  slc:1
695//	s_waitcnt vmcnt(0)
696//	v_add_u32 v2, vcc[0:1], v2, v3
697      v_add_u32 v2, v2, v3
698      v_cmp_lt_u32 vcc[0:1], v2, s_save_alloc_size
699      s_cbranch_vccnz L_SAVE_LDS_LOOP_VECTOR
700
701      // restore rsrc3
702      s_mov_b32 s_save_buf_rsrc3, s0
703
704end
705
706L_SAVE_LDS_DONE:
707
708
709    /*		save VGPRs  - set the Rest VGPRs	*/
710    //////////////////////////////////////////////////////////////////////////////////////
711  L_SAVE_VGPR:
712    // VGPR SR memory offset: 0
713    // TODO rearrange the RSRC words to use swizzle for VGPR save...
714
715    s_mov_b32	    s_save_mem_offset, (0+256*4)				    // for the rest VGPRs
716    s_mov_b32	    exec_lo, 0xFFFFFFFF						    //need every thread from now on
717    s_mov_b32	    exec_hi, 0xFFFFFFFF
718
719    s_getreg_b32    s_save_alloc_size, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE)		    //vpgr_size
720    s_add_u32	    s_save_alloc_size, s_save_alloc_size, 1
721    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
722    s_lshl_b32	    s_save_buf_rsrc2,  s_save_alloc_size, 8			    //NUM_RECORDS in bytes (64 threads*4)
723    if (SWIZZLE_EN)
724	s_add_u32	s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
725    else
726	s_mov_b32	s_save_buf_rsrc2,  0x1000000				    //NUM_RECORDS in bytes
727    end
728
729
730    // VGPR Allocated in 4-GPR granularity
731
732if G8SR_VGPR_SR_IN_DWX4
733	// the const stride for DWx4 is 4*4 bytes
734	s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF   // reset const stride to 0
735	s_or_b32  s_save_buf_rsrc1, s_save_buf_rsrc1, G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4  // const stride to 4*4 bytes
736
737	s_mov_b32	  m0, 4	    // skip first 4 VGPRs
738	s_cmp_lt_u32	  m0, s_save_alloc_size
739	s_cbranch_scc0	  L_SAVE_VGPR_LOOP_END	    // no more vgprs
740
741	s_set_gpr_idx_on  m0, 0x1   // This will change M0
742	s_add_u32	  s_save_alloc_size, s_save_alloc_size, 0x1000	// because above inst change m0
743L_SAVE_VGPR_LOOP:
744	v_mov_b32	  v0, v0   // v0 = v[0+m0]
745	v_mov_b32	  v1, v1
746	v_mov_b32	  v2, v2
747	v_mov_b32	  v3, v3
748
749
750	buffer_store_dwordx4 v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1
751	s_add_u32	  m0, m0, 4
752	s_add_u32	  s_save_mem_offset, s_save_mem_offset, 256*4
753	s_cmp_lt_u32	  m0, s_save_alloc_size
754    s_cbranch_scc1  L_SAVE_VGPR_LOOP						    //VGPR save is complete?
755    s_set_gpr_idx_off
756L_SAVE_VGPR_LOOP_END:
757
758	s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF   // reset const stride to 0
759	s_or_b32  s_save_buf_rsrc1, s_save_buf_rsrc1, S_SAVE_BUF_RSRC_WORD1_STRIDE  // reset const stride to 4 bytes
760else
761    // VGPR store using dw burst
762    s_mov_b32	      m0, 0x4	//VGPR initial index value =0
763    s_cmp_lt_u32      m0, s_save_alloc_size
764    s_cbranch_scc0    L_SAVE_VGPR_END
765
766
767    s_set_gpr_idx_on	m0, 0x1 //M0[7:0] = M0[7:0] and M0[15:12] = 0x1
768    s_add_u32	    s_save_alloc_size, s_save_alloc_size, 0x1000		    //add 0x1000 since we compare m0 against it later
769
770  L_SAVE_VGPR_LOOP:
771    v_mov_b32	    v0, v0		//v0 = v[0+m0]
772    v_mov_b32	    v1, v1		//v0 = v[0+m0]
773    v_mov_b32	    v2, v2		//v0 = v[0+m0]
774    v_mov_b32	    v3, v3		//v0 = v[0+m0]
775
776    if(USE_MTBUF_INSTEAD_OF_MUBUF)
777	tbuffer_store_format_x v0, v0, s_save_buf_rsrc0, s_save_mem_offset format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1
778    else
779	buffer_store_dword v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1
780	buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1  offset:256
781	buffer_store_dword v2, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1  offset:256*2
782	buffer_store_dword v3, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1  offset:256*3
783    end
784
785    s_add_u32	    m0, m0, 4							    //next vgpr index
786    s_add_u32	    s_save_mem_offset, s_save_mem_offset, 256*4			    //every buffer_store_dword does 256 bytes
787    s_cmp_lt_u32    m0, s_save_alloc_size					    //scc = (m0 < s_save_alloc_size) ? 1 : 0
788    s_cbranch_scc1  L_SAVE_VGPR_LOOP						    //VGPR save is complete?
789    s_set_gpr_idx_off
790end
791
792L_SAVE_VGPR_END:
793
794
795
796
797
798
799    /*	   S_PGM_END_SAVED  */				    //FIXME  graphics ONLY
800    if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_SAVE_NORMAL_EXIT))
801	s_and_b32 s_save_pc_hi, s_save_pc_hi, 0x0000ffff    //pc[47:32]
802	s_add_u32 s_save_pc_lo, s_save_pc_lo, 4		    //pc[31:0]+4
803	s_addc_u32 s_save_pc_hi, s_save_pc_hi, 0x0	    //carry bit over
804	s_rfe_b64 s_save_pc_lo				    //Return to the main shader program
805    else
806    end
807
808// Save Done timestamp
809if G8SR_DEBUG_TIMESTAMP
810	s_memrealtime	s_g8sr_ts_save_d
811	// SGPR SR memory offset : size(VGPR)
812	get_vgpr_size_bytes(s_save_mem_offset)
813	s_add_u32 s_save_mem_offset, s_save_mem_offset, G8SR_DEBUG_TS_SAVE_D_OFFSET
814	s_waitcnt lgkmcnt(0)	     //FIXME, will cause xnack??
815	// Need reset rsrc2??
816	s_mov_b32 m0, s_save_mem_offset
817	s_mov_b32 s_save_buf_rsrc2,  0x1000000					//NUM_RECORDS in bytes
818	s_buffer_store_dwordx2 s_g8sr_ts_save_d, s_save_buf_rsrc0, m0	    glc:1
819end
820
821
822    s_branch	L_END_PGM
823
824
825
826/**************************************************************************/
827/*			restore routine					  */
828/**************************************************************************/
829
830L_RESTORE:
831    /*	    Setup Resource Contants    */
832    if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_RESTORE_NORMAL))
833	//calculate wd_addr using absolute thread id
834	v_readlane_b32 s_restore_tmp, v9, 0
835	s_lshr_b32 s_restore_tmp, s_restore_tmp, 6
836	s_mul_i32 s_restore_tmp, s_restore_tmp, WAVE_SPACE
837	s_add_i32 s_restore_spi_init_lo, s_restore_tmp, WG_BASE_ADDR_LO
838	s_mov_b32 s_restore_spi_init_hi, WG_BASE_ADDR_HI
839	s_and_b32 s_restore_spi_init_hi, s_restore_spi_init_hi, CTX_RESTORE_CONTROL
840    else
841    end
842
843if G8SR_DEBUG_TIMESTAMP
844	s_memrealtime	s_g8sr_ts_restore_s
845	s_waitcnt lgkmcnt(0)	     //FIXME, will cause xnack??
846	// tma_lo/hi are sgpr 110, 111, which will not used for 112 SGPR allocated case...
847	s_mov_b32 s_restore_pc_lo, s_g8sr_ts_restore_s[0]
848	s_mov_b32 s_restore_pc_hi, s_g8sr_ts_restore_s[1]   //backup ts to ttmp0/1, sicne exec will be finally restored..
849end
850
851
852
853    s_mov_b32	    s_restore_buf_rsrc0,    s_restore_spi_init_lo							    //base_addr_lo
854    s_and_b32	    s_restore_buf_rsrc1,    s_restore_spi_init_hi, 0x0000FFFF						    //base_addr_hi
855    s_or_b32	    s_restore_buf_rsrc1,    s_restore_buf_rsrc1,  S_RESTORE_BUF_RSRC_WORD1_STRIDE
856    s_mov_b32	    s_restore_buf_rsrc2,    0										    //NUM_RECORDS initial value = 0 (in bytes)
857    s_mov_b32	    s_restore_buf_rsrc3,    S_RESTORE_BUF_RSRC_WORD3_MISC
858    s_and_b32	    s_restore_tmp,	    s_restore_spi_init_hi, S_RESTORE_SPI_INIT_ATC_MASK
859    s_lshr_b32	    s_restore_tmp,	    s_restore_tmp, (S_RESTORE_SPI_INIT_ATC_SHIFT-SQ_BUF_RSRC_WORD1_ATC_SHIFT)	    //get ATC bit into position
860    s_or_b32	    s_restore_buf_rsrc3,    s_restore_buf_rsrc3,  s_restore_tmp						    //or ATC
861    s_and_b32	    s_restore_tmp,	    s_restore_spi_init_hi, S_RESTORE_SPI_INIT_MTYPE_MASK
862    s_lshr_b32	    s_restore_tmp,	    s_restore_tmp, (S_RESTORE_SPI_INIT_MTYPE_SHIFT-SQ_BUF_RSRC_WORD3_MTYPE_SHIFT)   //get MTYPE bits into position
863    s_or_b32	    s_restore_buf_rsrc3,    s_restore_buf_rsrc3,  s_restore_tmp						    //or MTYPE
864
865    /*	    global mem offset		*/
866//  s_mov_b32	    s_restore_mem_offset, 0x0				    //mem offset initial value = 0
867
868    /*	    the first wave in the threadgroup	 */
869    s_and_b32	    s_restore_tmp, s_restore_spi_init_hi, S_RESTORE_SPI_INIT_FIRST_WAVE_MASK
870    s_cbranch_scc0  L_RESTORE_VGPR
871
872    /*		restore LDS	*/
873    //////////////////////////////
874  L_RESTORE_LDS:
875
876    s_mov_b32	    exec_lo, 0xFFFFFFFF							    //need every thread from now on   //be consistent with SAVE although can be moved ahead
877    s_mov_b32	    exec_hi, 0xFFFFFFFF
878
879    s_getreg_b32    s_restore_alloc_size, hwreg(HW_REG_LDS_ALLOC,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE)		//lds_size
880    s_and_b32	    s_restore_alloc_size, s_restore_alloc_size, 0xFFFFFFFF		    //lds_size is zero?
881    s_cbranch_scc0  L_RESTORE_VGPR							    //no lds used? jump to L_RESTORE_VGPR
882    s_lshl_b32	    s_restore_alloc_size, s_restore_alloc_size, 6			    //LDS size in dwords = lds_size * 64dw
883    s_lshl_b32	    s_restore_alloc_size, s_restore_alloc_size, 2			    //LDS size in bytes
884    s_mov_b32	    s_restore_buf_rsrc2,    s_restore_alloc_size			    //NUM_RECORDS in bytes
885
886    // LDS at offset: size(VGPR)+SIZE(SGPR)+SIZE(HWREG)
887    //
888    get_vgpr_size_bytes(s_restore_mem_offset)
889    get_sgpr_size_bytes(s_restore_tmp)
890    s_add_u32  s_restore_mem_offset, s_restore_mem_offset, s_restore_tmp
891    s_add_u32  s_restore_mem_offset, s_restore_mem_offset, get_hwreg_size_bytes()	     //FIXME, Check if offset overflow???
892
893
894    if (SWIZZLE_EN)
895	s_add_u32	s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
896    else
897	s_mov_b32	s_restore_buf_rsrc2,  0x1000000					    //NUM_RECORDS in bytes
898    end
899    s_mov_b32	    m0, 0x0								    //lds_offset initial value = 0
900
901  L_RESTORE_LDS_LOOP:
902    if (SAVE_LDS)
903	buffer_load_dword   v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset lds:1		       // first 64DW
904	buffer_load_dword   v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset lds:1 offset:256	       // second 64DW
905    end
906    s_add_u32	    m0, m0, 256*2						// 128 DW
907    s_add_u32	    s_restore_mem_offset, s_restore_mem_offset, 256*2		//mem offset increased by 128DW
908    s_cmp_lt_u32    m0, s_restore_alloc_size					//scc=(m0 < s_restore_alloc_size) ? 1 : 0
909    s_cbranch_scc1  L_RESTORE_LDS_LOOP							    //LDS restore is complete?
910
911
912    /*		restore VGPRs	    */
913    //////////////////////////////
914  L_RESTORE_VGPR:
915	// VGPR SR memory offset : 0
916    s_mov_b32	    s_restore_mem_offset, 0x0
917    s_mov_b32	    exec_lo, 0xFFFFFFFF							    //need every thread from now on   //be consistent with SAVE although can be moved ahead
918    s_mov_b32	    exec_hi, 0xFFFFFFFF
919
920    s_getreg_b32    s_restore_alloc_size, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE)	//vpgr_size
921    s_add_u32	    s_restore_alloc_size, s_restore_alloc_size, 1
922    s_lshl_b32	    s_restore_alloc_size, s_restore_alloc_size, 2			    //Number of VGPRs = (vgpr_size + 1) * 4    (non-zero value)
923    s_lshl_b32	    s_restore_buf_rsrc2,  s_restore_alloc_size, 8			    //NUM_RECORDS in bytes (64 threads*4)
924    if (SWIZZLE_EN)
925	s_add_u32	s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
926    else
927	s_mov_b32	s_restore_buf_rsrc2,  0x1000000					    //NUM_RECORDS in bytes
928    end
929
930if G8SR_VGPR_SR_IN_DWX4
931     get_vgpr_size_bytes(s_restore_mem_offset)
932     s_sub_u32	       s_restore_mem_offset, s_restore_mem_offset, 256*4
933
934     // the const stride for DWx4 is 4*4 bytes
935     s_and_b32 s_restore_buf_rsrc1, s_restore_buf_rsrc1, 0x0000FFFF   // reset const stride to 0
936     s_or_b32  s_restore_buf_rsrc1, s_restore_buf_rsrc1, G8SR_RESTORE_BUF_RSRC_WORD1_STRIDE_DWx4  // const stride to 4*4 bytes
937
938     s_mov_b32	       m0, s_restore_alloc_size
939     s_set_gpr_idx_on  m0, 0x8	  // Note.. This will change m0
940
941L_RESTORE_VGPR_LOOP:
942     buffer_load_dwordx4 v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1
943     s_waitcnt vmcnt(0)
944     s_sub_u32	       m0, m0, 4
945     v_mov_b32	       v0, v0	// v[0+m0] = v0
946     v_mov_b32	       v1, v1
947     v_mov_b32	       v2, v2
948     v_mov_b32	       v3, v3
949     s_sub_u32	       s_restore_mem_offset, s_restore_mem_offset, 256*4
950     s_cmp_eq_u32      m0, 0x8000
951     s_cbranch_scc0    L_RESTORE_VGPR_LOOP
952     s_set_gpr_idx_off
953
954     s_and_b32 s_restore_buf_rsrc1, s_restore_buf_rsrc1, 0x0000FFFF   // reset const stride to 0
955     s_or_b32  s_restore_buf_rsrc1, s_restore_buf_rsrc1, S_RESTORE_BUF_RSRC_WORD1_STRIDE  // const stride to 4*4 bytes
956
957else
958    // VGPR load using dw burst
959    s_mov_b32	    s_restore_mem_offset_save, s_restore_mem_offset	// restore start with v1, v0 will be the last
960    s_add_u32	    s_restore_mem_offset, s_restore_mem_offset, 256*4
961    s_mov_b32	    m0, 4				//VGPR initial index value = 1
962    s_set_gpr_idx_on  m0, 0x8			    //M0[7:0] = M0[7:0] and M0[15:12] = 0x8
963    s_add_u32	    s_restore_alloc_size, s_restore_alloc_size, 0x8000			    //add 0x8000 since we compare m0 against it later
964
965  L_RESTORE_VGPR_LOOP:
966    if(USE_MTBUF_INSTEAD_OF_MUBUF)
967	tbuffer_load_format_x v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1
968    else
969	buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1
970	buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:256
971	buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:256*2
972	buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:256*3
973    end
974    s_waitcnt	    vmcnt(0)								    //ensure data ready
975    v_mov_b32	    v0, v0								    //v[0+m0] = v0
976    v_mov_b32	    v1, v1
977    v_mov_b32	    v2, v2
978    v_mov_b32	    v3, v3
979    s_add_u32	    m0, m0, 4								    //next vgpr index
980    s_add_u32	    s_restore_mem_offset, s_restore_mem_offset, 256*4				//every buffer_load_dword does 256 bytes
981    s_cmp_lt_u32    m0, s_restore_alloc_size						    //scc = (m0 < s_restore_alloc_size) ? 1 : 0
982    s_cbranch_scc1  L_RESTORE_VGPR_LOOP							    //VGPR restore (except v0) is complete?
983    s_set_gpr_idx_off
984											    /* VGPR restore on v0 */
985    if(USE_MTBUF_INSTEAD_OF_MUBUF)
986	tbuffer_load_format_x v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1
987    else
988	buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save    slc:1 glc:1
989	buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save    slc:1 glc:1 offset:256
990	buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save    slc:1 glc:1 offset:256*2
991	buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save    slc:1 glc:1 offset:256*3
992    end
993
994end
995
996    /*		restore SGPRs	    */
997    //////////////////////////////
998
999    // SGPR SR memory offset : size(VGPR)
1000    get_vgpr_size_bytes(s_restore_mem_offset)
1001    get_sgpr_size_bytes(s_restore_tmp)
1002    s_add_u32 s_restore_mem_offset, s_restore_mem_offset, s_restore_tmp
1003    s_sub_u32 s_restore_mem_offset, s_restore_mem_offset, 16*4	   // restore SGPR from S[n] to S[0], by 16 sgprs group
1004    // TODO, change RSRC word to rearrange memory layout for SGPRS
1005
1006    s_getreg_b32    s_restore_alloc_size, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SIZE)		    //spgr_size
1007    s_add_u32	    s_restore_alloc_size, s_restore_alloc_size, 1
1008    s_lshl_b32	    s_restore_alloc_size, s_restore_alloc_size, 4			    //Number of SGPRs = (sgpr_size + 1) * 16   (non-zero value)
1009
1010    if (SGPR_SAVE_USE_SQC)
1011	s_lshl_b32	s_restore_buf_rsrc2,	s_restore_alloc_size, 2			    //NUM_RECORDS in bytes
1012    else
1013	s_lshl_b32	s_restore_buf_rsrc2,	s_restore_alloc_size, 8			    //NUM_RECORDS in bytes (64 threads)
1014    end
1015    if (SWIZZLE_EN)
1016	s_add_u32	s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
1017    else
1018	s_mov_b32	s_restore_buf_rsrc2,  0x1000000					    //NUM_RECORDS in bytes
1019    end
1020
1021    s_mov_b32 m0, s_restore_alloc_size
1022
1023 L_RESTORE_SGPR_LOOP:
1024    read_16sgpr_from_mem(s0, s_restore_buf_rsrc0, s_restore_mem_offset)	 //PV: further performance improvement can be made
1025    s_waitcnt	    lgkmcnt(0)								    //ensure data ready
1026
1027    s_sub_u32 m0, m0, 16    // Restore from S[n] to S[0]
1028    s_nop 0 // hazard SALU M0=> S_MOVREL
1029
1030    s_movreld_b64   s0, s0	//s[0+m0] = s0
1031    s_movreld_b64   s2, s2
1032    s_movreld_b64   s4, s4
1033    s_movreld_b64   s6, s6
1034    s_movreld_b64   s8, s8
1035    s_movreld_b64   s10, s10
1036    s_movreld_b64   s12, s12
1037    s_movreld_b64   s14, s14
1038
1039    s_cmp_eq_u32    m0, 0		//scc = (m0 < s_restore_alloc_size) ? 1 : 0
1040    s_cbranch_scc0  L_RESTORE_SGPR_LOOP		    //SGPR restore (except s0) is complete?
1041
1042    /*	    restore HW registers    */
1043    //////////////////////////////
1044  L_RESTORE_HWREG:
1045
1046
1047if G8SR_DEBUG_TIMESTAMP
1048      s_mov_b32 s_g8sr_ts_restore_s[0], s_restore_pc_lo
1049      s_mov_b32 s_g8sr_ts_restore_s[1], s_restore_pc_hi
1050end
1051
1052    // HWREG SR memory offset : size(VGPR)+size(SGPR)
1053    get_vgpr_size_bytes(s_restore_mem_offset)
1054    get_sgpr_size_bytes(s_restore_tmp)
1055    s_add_u32 s_restore_mem_offset, s_restore_mem_offset, s_restore_tmp
1056
1057
1058    s_mov_b32	    s_restore_buf_rsrc2, 0x4						    //NUM_RECORDS   in bytes
1059    if (SWIZZLE_EN)
1060	s_add_u32	s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
1061    else
1062	s_mov_b32	s_restore_buf_rsrc2,  0x1000000					    //NUM_RECORDS in bytes
1063    end
1064
1065    read_hwreg_from_mem(s_restore_m0, s_restore_buf_rsrc0, s_restore_mem_offset)		    //M0
1066    read_hwreg_from_mem(s_restore_pc_lo, s_restore_buf_rsrc0, s_restore_mem_offset)		//PC
1067    read_hwreg_from_mem(s_restore_pc_hi, s_restore_buf_rsrc0, s_restore_mem_offset)
1068    read_hwreg_from_mem(s_restore_exec_lo, s_restore_buf_rsrc0, s_restore_mem_offset)		    //EXEC
1069    read_hwreg_from_mem(s_restore_exec_hi, s_restore_buf_rsrc0, s_restore_mem_offset)
1070    read_hwreg_from_mem(s_restore_status, s_restore_buf_rsrc0, s_restore_mem_offset)		    //STATUS
1071    read_hwreg_from_mem(s_restore_trapsts, s_restore_buf_rsrc0, s_restore_mem_offset)		    //TRAPSTS
1072    read_hwreg_from_mem(xnack_mask_lo, s_restore_buf_rsrc0, s_restore_mem_offset)		    //XNACK_MASK_LO
1073    read_hwreg_from_mem(xnack_mask_hi, s_restore_buf_rsrc0, s_restore_mem_offset)		    //XNACK_MASK_HI
1074    read_hwreg_from_mem(s_restore_mode, s_restore_buf_rsrc0, s_restore_mem_offset)		//MODE
1075
1076    s_waitcnt	    lgkmcnt(0)											    //from now on, it is safe to restore STATUS and IB_STS
1077
1078    //for normal save & restore, the saved PC points to the next inst to execute, no adjustment needs to be made, otherwise:
1079    if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_RESTORE_NORMAL))
1080	s_add_u32 s_restore_pc_lo, s_restore_pc_lo, 8		 //pc[31:0]+8	  //two back-to-back s_trap are used (first for save and second for restore)
1081	s_addc_u32  s_restore_pc_hi, s_restore_pc_hi, 0x0	 //carry bit over
1082    end
1083    if ((EMU_RUN_HACK) && (EMU_RUN_HACK_RESTORE_NORMAL))
1084	s_add_u32 s_restore_pc_lo, s_restore_pc_lo, 4		 //pc[31:0]+4	  // save is hack through s_trap but restore is normal
1085	s_addc_u32  s_restore_pc_hi, s_restore_pc_hi, 0x0	 //carry bit over
1086    end
1087
1088    s_mov_b32	    m0,		s_restore_m0
1089    s_mov_b32	    exec_lo,	s_restore_exec_lo
1090    s_mov_b32	    exec_hi,	s_restore_exec_hi
1091
1092    s_and_b32	    s_restore_m0, SQ_WAVE_TRAPSTS_PRE_SAVECTX_MASK, s_restore_trapsts
1093    s_setreg_b32    hwreg(HW_REG_TRAPSTS, SQ_WAVE_TRAPSTS_PRE_SAVECTX_SHIFT, SQ_WAVE_TRAPSTS_PRE_SAVECTX_SIZE), s_restore_m0
1094    s_and_b32	    s_restore_m0, SQ_WAVE_TRAPSTS_POST_SAVECTX_MASK, s_restore_trapsts
1095    s_lshr_b32	    s_restore_m0, s_restore_m0, SQ_WAVE_TRAPSTS_POST_SAVECTX_SHIFT
1096    s_setreg_b32    hwreg(HW_REG_TRAPSTS, SQ_WAVE_TRAPSTS_POST_SAVECTX_SHIFT, SQ_WAVE_TRAPSTS_POST_SAVECTX_SIZE), s_restore_m0
1097    //s_setreg_b32  hwreg(HW_REG_TRAPSTS),  s_restore_trapsts	   //don't overwrite SAVECTX bit as it may be set through external SAVECTX during restore
1098    s_setreg_b32    hwreg(HW_REG_MODE),	    s_restore_mode
1099
1100    // Restore trap temporaries 4-11, 13 initialized by SPI debug dispatch logic
1101    // ttmp SR memory offset : size(VGPR)+size(SGPR)+0x40
1102    get_vgpr_size_bytes(s_restore_ttmps_lo)
1103    get_sgpr_size_bytes(s_restore_ttmps_hi)
1104    s_add_u32	    s_restore_ttmps_lo, s_restore_ttmps_lo, s_restore_ttmps_hi
1105    s_add_u32	    s_restore_ttmps_lo, s_restore_ttmps_lo, s_restore_buf_rsrc0
1106    s_addc_u32	    s_restore_ttmps_hi, s_restore_buf_rsrc1, 0x0
1107    s_and_b32	    s_restore_ttmps_hi, s_restore_ttmps_hi, 0xFFFF
1108    s_load_dwordx4  [ttmp4, ttmp5, ttmp6, ttmp7], [s_restore_ttmps_lo, s_restore_ttmps_hi], 0x50 glc:1
1109    s_load_dwordx4  [ttmp8, ttmp9, ttmp10, ttmp11], [s_restore_ttmps_lo, s_restore_ttmps_hi], 0x60 glc:1
1110    s_load_dword    ttmp13, [s_restore_ttmps_lo, s_restore_ttmps_hi], 0x74 glc:1
1111    s_waitcnt	    lgkmcnt(0)
1112
1113    //reuse s_restore_m0 as a temp register
1114    s_and_b32	    s_restore_m0, s_restore_pc_hi, S_SAVE_PC_HI_RCNT_MASK
1115    s_lshr_b32	    s_restore_m0, s_restore_m0, S_SAVE_PC_HI_RCNT_SHIFT
1116    s_lshl_b32	    s_restore_m0, s_restore_m0, SQ_WAVE_IB_STS_RCNT_SHIFT
1117    s_mov_b32	    s_restore_tmp, 0x0										    //IB_STS is zero
1118    s_or_b32	    s_restore_tmp, s_restore_tmp, s_restore_m0
1119    s_and_b32	    s_restore_m0, s_restore_pc_hi, S_SAVE_PC_HI_FIRST_REPLAY_MASK
1120    s_lshr_b32	    s_restore_m0, s_restore_m0, S_SAVE_PC_HI_FIRST_REPLAY_SHIFT
1121    s_lshl_b32	    s_restore_m0, s_restore_m0, SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT
1122    s_or_b32	    s_restore_tmp, s_restore_tmp, s_restore_m0
1123    s_and_b32	    s_restore_m0, s_restore_status, SQ_WAVE_STATUS_INST_ATC_MASK
1124    s_lshr_b32	    s_restore_m0, s_restore_m0, SQ_WAVE_STATUS_INST_ATC_SHIFT
1125    s_setreg_b32    hwreg(HW_REG_IB_STS),   s_restore_tmp
1126
1127    s_and_b32 s_restore_pc_hi, s_restore_pc_hi, 0x0000ffff	//pc[47:32]	   //Do it here in order not to affect STATUS
1128    s_and_b64	 exec, exec, exec  // Restore STATUS.EXECZ, not writable by s_setreg_b32
1129    s_and_b64	 vcc, vcc, vcc	// Restore STATUS.VCCZ, not writable by s_setreg_b32
1130    set_status_without_spi_prio(s_restore_status, s_restore_tmp) // SCC is included, which is changed by previous salu
1131
1132    s_barrier							//barrier to ensure the readiness of LDS before access attempts from any other wave in the same TG //FIXME not performance-optimal at this time
1133
1134if G8SR_DEBUG_TIMESTAMP
1135    s_memrealtime s_g8sr_ts_restore_d
1136    s_waitcnt lgkmcnt(0)
1137end
1138
1139//  s_rfe_b64 s_restore_pc_lo					//Return to the main shader program and resume execution
1140    s_rfe_restore_b64  s_restore_pc_lo, s_restore_m0		// s_restore_m0[0] is used to set STATUS.inst_atc
1141
1142
1143/**************************************************************************/
1144/*			the END						  */
1145/**************************************************************************/
1146L_END_PGM:
1147    s_endpgm
1148
1149end
1150
1151
1152/**************************************************************************/
1153/*			the helper functions				  */
1154/**************************************************************************/
1155
1156//Only for save hwreg to mem
1157function write_hwreg_to_mem(s, s_rsrc, s_mem_offset)
1158	s_mov_b32 exec_lo, m0			//assuming exec_lo is not needed anymore from this point on
1159	s_mov_b32 m0, s_mem_offset
1160	s_buffer_store_dword s, s_rsrc, m0	glc:1
1161	ack_sqc_store_workaround()
1162	s_add_u32	s_mem_offset, s_mem_offset, 4
1163	s_mov_b32   m0, exec_lo
1164end
1165
1166
1167// HWREG are saved before SGPRs, so all HWREG could be use.
1168function write_16sgpr_to_mem(s, s_rsrc, s_mem_offset)
1169
1170	s_buffer_store_dwordx4 s[0], s_rsrc, 0	glc:1
1171	ack_sqc_store_workaround()
1172	s_buffer_store_dwordx4 s[4], s_rsrc, 16	 glc:1
1173	ack_sqc_store_workaround()
1174	s_buffer_store_dwordx4 s[8], s_rsrc, 32	 glc:1
1175	ack_sqc_store_workaround()
1176	s_buffer_store_dwordx4 s[12], s_rsrc, 48 glc:1
1177	ack_sqc_store_workaround()
1178	s_add_u32	s_rsrc[0], s_rsrc[0], 4*16
1179	s_addc_u32	s_rsrc[1], s_rsrc[1], 0x0	      // +scc
1180end
1181
1182
1183function read_hwreg_from_mem(s, s_rsrc, s_mem_offset)
1184    s_buffer_load_dword s, s_rsrc, s_mem_offset	    glc:1
1185    s_add_u32	    s_mem_offset, s_mem_offset, 4
1186end
1187
1188function read_16sgpr_from_mem(s, s_rsrc, s_mem_offset)
1189    s_buffer_load_dwordx16 s, s_rsrc, s_mem_offset	glc:1
1190    s_sub_u32	    s_mem_offset, s_mem_offset, 4*16
1191end
1192
1193
1194
1195function get_lds_size_bytes(s_lds_size_byte)
1196    // SQ LDS granularity is 64DW, while PGM_RSRC2.lds_size is in granularity 128DW
1197    s_getreg_b32   s_lds_size_byte, hwreg(HW_REG_LDS_ALLOC, SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT, SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE)		// lds_size
1198    s_lshl_b32	   s_lds_size_byte, s_lds_size_byte, 8			    //LDS size in dwords = lds_size * 64 *4Bytes    // granularity 64DW
1199end
1200
1201function get_vgpr_size_bytes(s_vgpr_size_byte)
1202    s_getreg_b32   s_vgpr_size_byte, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE)	 //vpgr_size
1203    s_add_u32	   s_vgpr_size_byte, s_vgpr_size_byte, 1
1204    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
1205end
1206
1207function get_sgpr_size_bytes(s_sgpr_size_byte)
1208    s_getreg_b32   s_sgpr_size_byte, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SIZE)	 //spgr_size
1209    s_add_u32	   s_sgpr_size_byte, s_sgpr_size_byte, 1
1210    s_lshl_b32	   s_sgpr_size_byte, s_sgpr_size_byte, 6 //Number of SGPRs = (sgpr_size + 1) * 16 *4   (non-zero value)
1211end
1212
1213function get_hwreg_size_bytes
1214    return 128 //HWREG size 128 bytes
1215end
1216
1217function ack_sqc_store_workaround
1218    if ACK_SQC_STORE
1219        s_waitcnt lgkmcnt(0)
1220    end
1221end
1222
1223function set_status_without_spi_prio(status, tmp)
1224    // Do not restore STATUS.SPI_PRIO since scheduler may have raised it.
1225    s_lshr_b32      tmp, status, SQ_WAVE_STATUS_POST_SPI_PRIO_SHIFT
1226    s_setreg_b32    hwreg(HW_REG_STATUS, SQ_WAVE_STATUS_POST_SPI_PRIO_SHIFT, SQ_WAVE_STATUS_POST_SPI_PRIO_SIZE), tmp
1227    s_nop           0x2 // avoid S_SETREG => S_SETREG hazard
1228    s_setreg_b32    hwreg(HW_REG_STATUS, SQ_WAVE_STATUS_PRE_SPI_PRIO_SHIFT, SQ_WAVE_STATUS_PRE_SPI_PRIO_SIZE), status
1229end
1230