1 #ifndef __NVC0_SCREEN_H__
2 #define __NVC0_SCREEN_H__
4 #include "nouveau_screen.h"
5 #include "nouveau_mm.h"
6 #include "nouveau_fence.h"
7 #include "nouveau_heap.h"
9 #include "nv_object.xml.h"
11 #include "nvc0/nvc0_winsys.h"
12 #include "nvc0/nvc0_stateobj.h"
14 #define NVC0_TIC_MAX_ENTRIES 2048
15 #define NVC0_TSC_MAX_ENTRIES 2048
17 /* doesn't count reserved slots (for auxiliary constants, immediates, etc.) */
18 #define NVC0_MAX_PIPE_CONSTBUFS 14
19 #define NVE4_MAX_PIPE_CONSTBUFS_COMPUTE 7
21 #define NVC0_MAX_SURFACE_SLOTS 16
23 #define NVC0_MAX_VIEWPORTS 16
31 struct nouveau_screen base
;
33 struct nvc0_context
*cur_ctx
;
35 int num_occlusion_queries_active
;
37 struct nouveau_bo
*text
;
38 struct nouveau_bo
*parm
; /* for COMPUTE */
39 struct nouveau_bo
*uniform_bo
; /* for 3D */
40 struct nouveau_bo
*tls
;
41 struct nouveau_bo
*txc
; /* TIC (offset 0) and TSC (65536) */
42 struct nouveau_bo
*poly_cache
;
45 uint16_t mp_count_compute
; /* magic reg can make compute use fewer MPs */
47 struct nouveau_heap
*text_heap
;
48 struct nouveau_heap
*lib_code
; /* allocated from text_heap */
50 struct nvc0_blitter
*blitter
;
55 uint32_t lock
[NVC0_TIC_MAX_ENTRIES
/ 32];
61 uint32_t lock
[NVC0_TSC_MAX_ENTRIES
/ 32];
65 struct nouveau_bo
*bo
;
70 struct nvc0_program
*prog
; /* compute state object to read MP counters */
71 struct pipe_query
*mp_counter
[8]; /* counter to query allocation */
72 uint8_t num_mp_pm_active
[2];
73 boolean mp_counters_enabled
;
76 struct nouveau_object
*eng3d
; /* sqrt(1/2)|kepler> + sqrt(1/2)|fermi> */
77 struct nouveau_object
*eng2d
;
78 struct nouveau_object
*m2mf
;
79 struct nouveau_object
*compute
;
80 struct nouveau_object
*nvsw
;
83 static INLINE
struct nvc0_screen
*
84 nvc0_screen(struct pipe_screen
*screen
)
86 return (struct nvc0_screen
*)screen
;
90 * Performance counters groups:
92 #define NVC0_QUERY_MP_COUNTER_GROUP 0
93 #define NVC0_QUERY_DRV_STAT_GROUP 1
95 /* Performance counter queries:
97 #define NVE4_PM_QUERY_COUNT 49
98 #define NVE4_PM_QUERY(i) (PIPE_QUERY_DRIVER_SPECIFIC + (i))
99 #define NVE4_PM_QUERY_LAST NVE4_PM_QUERY(NVE4_PM_QUERY_COUNT - 1)
100 #define NVE4_PM_QUERY_PROF_TRIGGER_0 0
101 #define NVE4_PM_QUERY_PROF_TRIGGER_1 1
102 #define NVE4_PM_QUERY_PROF_TRIGGER_2 2
103 #define NVE4_PM_QUERY_PROF_TRIGGER_3 3
104 #define NVE4_PM_QUERY_PROF_TRIGGER_4 4
105 #define NVE4_PM_QUERY_PROF_TRIGGER_5 5
106 #define NVE4_PM_QUERY_PROF_TRIGGER_6 6
107 #define NVE4_PM_QUERY_PROF_TRIGGER_7 7
108 #define NVE4_PM_QUERY_LAUNCHED_WARPS 8
109 #define NVE4_PM_QUERY_LAUNCHED_THREADS 9
110 #define NVE4_PM_QUERY_LAUNCHED_CTA 10
111 #define NVE4_PM_QUERY_INST_ISSUED1 11
112 #define NVE4_PM_QUERY_INST_ISSUED2 12
113 #define NVE4_PM_QUERY_INST_EXECUTED 13
114 #define NVE4_PM_QUERY_LD_LOCAL 14
115 #define NVE4_PM_QUERY_ST_LOCAL 15
116 #define NVE4_PM_QUERY_LD_SHARED 16
117 #define NVE4_PM_QUERY_ST_SHARED 17
118 #define NVE4_PM_QUERY_L1_LOCAL_LOAD_HIT 18
119 #define NVE4_PM_QUERY_L1_LOCAL_LOAD_MISS 19
120 #define NVE4_PM_QUERY_L1_LOCAL_STORE_HIT 20
121 #define NVE4_PM_QUERY_L1_LOCAL_STORE_MISS 21
122 #define NVE4_PM_QUERY_GLD_REQUEST 22
123 #define NVE4_PM_QUERY_GST_REQUEST 23
124 #define NVE4_PM_QUERY_L1_GLOBAL_LOAD_HIT 24
125 #define NVE4_PM_QUERY_L1_GLOBAL_LOAD_MISS 25
126 #define NVE4_PM_QUERY_GLD_TRANSACTIONS_UNCACHED 26
127 #define NVE4_PM_QUERY_GST_TRANSACTIONS 27
128 #define NVE4_PM_QUERY_BRANCH 28
129 #define NVE4_PM_QUERY_BRANCH_DIVERGENT 29
130 #define NVE4_PM_QUERY_ACTIVE_WARPS 30
131 #define NVE4_PM_QUERY_ACTIVE_CYCLES 31
132 #define NVE4_PM_QUERY_INST_ISSUED 32
133 #define NVE4_PM_QUERY_ATOM_COUNT 33
134 #define NVE4_PM_QUERY_GRED_COUNT 34
135 #define NVE4_PM_QUERY_LD_SHARED_REPLAY 35
136 #define NVE4_PM_QUERY_ST_SHARED_REPLAY 36
137 #define NVE4_PM_QUERY_LD_LOCAL_TRANSACTIONS 37
138 #define NVE4_PM_QUERY_ST_LOCAL_TRANSACTIONS 38
139 #define NVE4_PM_QUERY_L1_LD_SHARED_TRANSACTIONS 39
140 #define NVE4_PM_QUERY_L1_ST_SHARED_TRANSACTIONS 40
141 #define NVE4_PM_QUERY_GLD_MEM_DIV_REPLAY 41
142 #define NVE4_PM_QUERY_GST_MEM_DIV_REPLAY 42
143 #define NVE4_PM_QUERY_METRIC_IPC 43
144 #define NVE4_PM_QUERY_METRIC_IPAC 44
145 #define NVE4_PM_QUERY_METRIC_IPEC 45
146 #define NVE4_PM_QUERY_METRIC_MP_OCCUPANCY 46
147 #define NVE4_PM_QUERY_METRIC_MP_EFFICIENCY 47
148 #define NVE4_PM_QUERY_METRIC_INST_REPLAY_OHEAD 48
151 #define NVE4_PM_QUERY_GR_IDLE 50
152 #define NVE4_PM_QUERY_BSP_IDLE 51
153 #define NVE4_PM_QUERY_VP_IDLE 52
154 #define NVE4_PM_QUERY_PPP_IDLE 53
155 #define NVE4_PM_QUERY_CE0_IDLE 54
156 #define NVE4_PM_QUERY_CE1_IDLE 55
157 #define NVE4_PM_QUERY_CE2_IDLE 56
159 /* L2 queries (PCOUNTER) */
161 #define NVE4_PM_QUERY_L2_SUBP_WRITE_L1_SECTOR_QUERIES 57
164 /* TEX queries (PCOUNTER) */
166 #define NVE4_PM_QUERY_TEX0_CACHE_SECTOR_QUERIES 58
170 #define NVC0_PM_QUERY_COUNT 31
171 #define NVC0_PM_QUERY(i) (PIPE_QUERY_DRIVER_SPECIFIC + 2048 + (i))
172 #define NVC0_PM_QUERY_LAST NVC0_PM_QUERY(NVC0_PM_QUERY_COUNT - 1)
173 #define NVC0_PM_QUERY_INST_EXECUTED 0
174 #define NVC0_PM_QUERY_BRANCH 1
175 #define NVC0_PM_QUERY_BRANCH_DIVERGENT 2
176 #define NVC0_PM_QUERY_ACTIVE_WARPS 3
177 #define NVC0_PM_QUERY_ACTIVE_CYCLES 4
178 #define NVC0_PM_QUERY_LAUNCHED_WARPS 5
179 #define NVC0_PM_QUERY_LAUNCHED_THREADS 6
180 #define NVC0_PM_QUERY_LD_SHARED 7
181 #define NVC0_PM_QUERY_ST_SHARED 8
182 #define NVC0_PM_QUERY_LD_LOCAL 9
183 #define NVC0_PM_QUERY_ST_LOCAL 10
184 #define NVC0_PM_QUERY_GRED_COUNT 11
185 #define NVC0_PM_QUERY_ATOM_COUNT 12
186 #define NVC0_PM_QUERY_GLD_REQUEST 13
187 #define NVC0_PM_QUERY_GST_REQUEST 14
188 #define NVC0_PM_QUERY_INST_ISSUED1_0 15
189 #define NVC0_PM_QUERY_INST_ISSUED1_1 16
190 #define NVC0_PM_QUERY_INST_ISSUED2_0 17
191 #define NVC0_PM_QUERY_INST_ISSUED2_1 18
192 #define NVC0_PM_QUERY_TH_INST_EXECUTED_0 19
193 #define NVC0_PM_QUERY_TH_INST_EXECUTED_1 20
194 #define NVC0_PM_QUERY_TH_INST_EXECUTED_2 21
195 #define NVC0_PM_QUERY_TH_INST_EXECUTED_3 22
196 #define NVC0_PM_QUERY_PROF_TRIGGER_0 23
197 #define NVC0_PM_QUERY_PROF_TRIGGER_1 24
198 #define NVC0_PM_QUERY_PROF_TRIGGER_2 25
199 #define NVC0_PM_QUERY_PROF_TRIGGER_3 26
200 #define NVC0_PM_QUERY_PROF_TRIGGER_4 27
201 #define NVC0_PM_QUERY_PROF_TRIGGER_5 28
202 #define NVC0_PM_QUERY_PROF_TRIGGER_6 29
203 #define NVC0_PM_QUERY_PROF_TRIGGER_7 30
205 /* Driver statistics queries:
207 #ifdef NOUVEAU_ENABLE_DRIVER_STATISTICS
209 #define NVC0_QUERY_DRV_STAT(i) (PIPE_QUERY_DRIVER_SPECIFIC + 1024 + (i))
210 #define NVC0_QUERY_DRV_STAT_COUNT 29
211 #define NVC0_QUERY_DRV_STAT_LAST NVC0_QUERY_DRV_STAT(NVC0_QUERY_DRV_STAT_COUNT - 1)
212 #define NVC0_QUERY_DRV_STAT_TEX_OBJECT_CURRENT_COUNT 0
213 #define NVC0_QUERY_DRV_STAT_TEX_OBJECT_CURRENT_BYTES 1
214 #define NVC0_QUERY_DRV_STAT_BUF_OBJECT_CURRENT_COUNT 2
215 #define NVC0_QUERY_DRV_STAT_BUF_OBJECT_CURRENT_BYTES_VID 3
216 #define NVC0_QUERY_DRV_STAT_BUF_OBJECT_CURRENT_BYTES_SYS 4
217 #define NVC0_QUERY_DRV_STAT_TEX_TRANSFERS_READ 5
218 #define NVC0_QUERY_DRV_STAT_TEX_TRANSFERS_WRITE 6
219 #define NVC0_QUERY_DRV_STAT_TEX_COPY_COUNT 7
220 #define NVC0_QUERY_DRV_STAT_TEX_BLIT_COUNT 8
221 #define NVC0_QUERY_DRV_STAT_TEX_CACHE_FLUSH_COUNT 9
222 #define NVC0_QUERY_DRV_STAT_BUF_TRANSFERS_READ 10
223 #define NVC0_QUERY_DRV_STAT_BUF_TRANSFERS_WRITE 11
224 #define NVC0_QUERY_DRV_STAT_BUF_READ_BYTES_STAGING_VID 12
225 #define NVC0_QUERY_DRV_STAT_BUF_WRITE_BYTES_DIRECT 13
226 #define NVC0_QUERY_DRV_STAT_BUF_WRITE_BYTES_STAGING_VID 14
227 #define NVC0_QUERY_DRV_STAT_BUF_WRITE_BYTES_STAGING_SYS 15
228 #define NVC0_QUERY_DRV_STAT_BUF_COPY_BYTES 16
229 #define NVC0_QUERY_DRV_STAT_BUF_NON_KERNEL_FENCE_SYNC_COUNT 17
230 #define NVC0_QUERY_DRV_STAT_ANY_NON_KERNEL_FENCE_SYNC_COUNT 18
231 #define NVC0_QUERY_DRV_STAT_QUERY_SYNC_COUNT 19
232 #define NVC0_QUERY_DRV_STAT_GPU_SERIALIZE_COUNT 20
233 #define NVC0_QUERY_DRV_STAT_DRAW_CALLS_ARRAY 21
234 #define NVC0_QUERY_DRV_STAT_DRAW_CALLS_INDEXED 22
235 #define NVC0_QUERY_DRV_STAT_DRAW_CALLS_FALLBACK_COUNT 23
236 #define NVC0_QUERY_DRV_STAT_USER_BUFFER_UPLOAD_BYTES 24
237 #define NVC0_QUERY_DRV_STAT_CONSTBUF_UPLOAD_COUNT 25
238 #define NVC0_QUERY_DRV_STAT_CONSTBUF_UPLOAD_BYTES 26
239 #define NVC0_QUERY_DRV_STAT_PUSHBUF_COUNT 27
240 #define NVC0_QUERY_DRV_STAT_RESOURCE_VALIDATE_COUNT 28
244 #define NVC0_QUERY_DRV_STAT_COUNT 0
248 int nvc0_screen_get_driver_query_info(struct pipe_screen
*, unsigned,
249 struct pipe_driver_query_info
*);
251 int nvc0_screen_get_driver_query_group_info(struct pipe_screen
*, unsigned,
252 struct pipe_driver_query_group_info
*);
254 boolean
nvc0_blitter_create(struct nvc0_screen
*);
255 void nvc0_blitter_destroy(struct nvc0_screen
*);
257 void nvc0_screen_make_buffers_resident(struct nvc0_screen
*);
259 int nvc0_screen_tic_alloc(struct nvc0_screen
*, void *);
260 int nvc0_screen_tsc_alloc(struct nvc0_screen
*, void *);
262 int nve4_screen_compute_setup(struct nvc0_screen
*, struct nouveau_pushbuf
*);
263 int nvc0_screen_compute_setup(struct nvc0_screen
*, struct nouveau_pushbuf
*);
265 boolean
nvc0_screen_resize_tls_area(struct nvc0_screen
*, uint32_t lpos
,
266 uint32_t lneg
, uint32_t cstack
);
269 nvc0_resource_fence(struct nv04_resource
*res
, uint32_t flags
)
271 struct nvc0_screen
*screen
= nvc0_screen(res
->base
.screen
);
274 nouveau_fence_ref(screen
->base
.fence
.current
, &res
->fence
);
275 if (flags
& NOUVEAU_BO_WR
)
276 nouveau_fence_ref(screen
->base
.fence
.current
, &res
->fence_wr
);
281 nvc0_resource_validate(struct nv04_resource
*res
, uint32_t flags
)
283 if (likely(res
->bo
)) {
284 if (flags
& NOUVEAU_BO_WR
)
285 res
->status
|= NOUVEAU_BUFFER_STATUS_GPU_WRITING
|
286 NOUVEAU_BUFFER_STATUS_DIRTY
;
287 if (flags
& NOUVEAU_BO_RD
)
288 res
->status
|= NOUVEAU_BUFFER_STATUS_GPU_READING
;
290 nvc0_resource_fence(res
, flags
);
301 extern const struct nvc0_format nvc0_format_table
[];
304 nvc0_screen_tic_unlock(struct nvc0_screen
*screen
, struct nv50_tic_entry
*tic
)
307 screen
->tic
.lock
[tic
->id
/ 32] &= ~(1 << (tic
->id
% 32));
311 nvc0_screen_tsc_unlock(struct nvc0_screen
*screen
, struct nv50_tsc_entry
*tsc
)
314 screen
->tsc
.lock
[tsc
->id
/ 32] &= ~(1 << (tsc
->id
% 32));
318 nvc0_screen_tic_free(struct nvc0_screen
*screen
, struct nv50_tic_entry
*tic
)
321 screen
->tic
.entries
[tic
->id
] = NULL
;
322 screen
->tic
.lock
[tic
->id
/ 32] &= ~(1 << (tic
->id
% 32));
327 nvc0_screen_tsc_free(struct nvc0_screen
*screen
, struct nv50_tsc_entry
*tsc
)
330 screen
->tsc
.entries
[tsc
->id
] = NULL
;
331 screen
->tsc
.lock
[tsc
->id
/ 32] &= ~(1 << (tsc
->id
% 32));