Move nv30, nv50 and nvc0 to nouveau.
[mesa.git] / src / gallium / drivers / nouveau / nvc0 / nvc0_screen.h
1 #ifndef __NVC0_SCREEN_H__
2 #define __NVC0_SCREEN_H__
3
4 #include "nouveau_screen.h"
5 #include "nouveau_mm.h"
6 #include "nouveau_fence.h"
7 #include "nouveau_heap.h"
8
9 #include "nv_object.xml.h"
10
11 #include "nvc0/nvc0_winsys.h"
12 #include "nvc0/nvc0_stateobj.h"
13
14 #define NVC0_TIC_MAX_ENTRIES 2048
15 #define NVC0_TSC_MAX_ENTRIES 2048
16
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
20
21 #define NVC0_MAX_SURFACE_SLOTS 16
22
23 struct nvc0_context;
24
25 struct nvc0_blitter;
26
27 struct nvc0_screen {
28 struct nouveau_screen base;
29
30 struct nvc0_context *cur_ctx;
31
32 int num_occlusion_queries_active;
33
34 struct nouveau_bo *text;
35 struct nouveau_bo *parm; /* for COMPUTE */
36 struct nouveau_bo *uniform_bo; /* for 3D */
37 struct nouveau_bo *tls;
38 struct nouveau_bo *txc; /* TIC (offset 0) and TSC (65536) */
39 struct nouveau_bo *poly_cache;
40
41 uint16_t mp_count;
42 uint16_t mp_count_compute; /* magic reg can make compute use fewer MPs */
43
44 struct nouveau_heap *text_heap;
45 struct nouveau_heap *lib_code; /* allocated from text_heap */
46
47 struct nvc0_blitter *blitter;
48
49 struct {
50 void **entries;
51 int next;
52 uint32_t lock[NVC0_TIC_MAX_ENTRIES / 32];
53 } tic;
54
55 struct {
56 void **entries;
57 int next;
58 uint32_t lock[NVC0_TSC_MAX_ENTRIES / 32];
59 } tsc;
60
61 struct {
62 struct nouveau_bo *bo;
63 uint32_t *map;
64 } fence;
65
66 struct {
67 struct nvc0_program *prog; /* compute state object to read MP counters */
68 struct pipe_query *mp_counter[8]; /* counter to query allocation */
69 uint8_t num_mp_pm_active[2];
70 boolean mp_counters_enabled;
71 } pm;
72
73 struct nouveau_mman *mm_VRAM_fe0;
74
75 struct nouveau_object *eng3d; /* sqrt(1/2)|kepler> + sqrt(1/2)|fermi> */
76 struct nouveau_object *eng2d;
77 struct nouveau_object *m2mf;
78 struct nouveau_object *compute;
79 };
80
81 static INLINE struct nvc0_screen *
82 nvc0_screen(struct pipe_screen *screen)
83 {
84 return (struct nvc0_screen *)screen;
85 }
86
87
88 /* Performance counter queries:
89 */
90 #define NVE4_PM_QUERY_COUNT 49
91 #define NVE4_PM_QUERY(i) (PIPE_QUERY_DRIVER_SPECIFIC + (i))
92 #define NVE4_PM_QUERY_LAST NVE4_PM_QUERY(NVE4_PM_QUERY_COUNT - 1)
93 #define NVE4_PM_QUERY_PROF_TRIGGER_0 0
94 #define NVE4_PM_QUERY_PROF_TRIGGER_1 1
95 #define NVE4_PM_QUERY_PROF_TRIGGER_2 2
96 #define NVE4_PM_QUERY_PROF_TRIGGER_3 3
97 #define NVE4_PM_QUERY_PROF_TRIGGER_4 4
98 #define NVE4_PM_QUERY_PROF_TRIGGER_5 5
99 #define NVE4_PM_QUERY_PROF_TRIGGER_6 6
100 #define NVE4_PM_QUERY_PROF_TRIGGER_7 7
101 #define NVE4_PM_QUERY_LAUNCHED_WARPS 8
102 #define NVE4_PM_QUERY_LAUNCHED_THREADS 9
103 #define NVE4_PM_QUERY_LAUNCHED_CTA 10
104 #define NVE4_PM_QUERY_INST_ISSUED1 11
105 #define NVE4_PM_QUERY_INST_ISSUED2 12
106 #define NVE4_PM_QUERY_INST_EXECUTED 13
107 #define NVE4_PM_QUERY_LD_LOCAL 14
108 #define NVE4_PM_QUERY_ST_LOCAL 15
109 #define NVE4_PM_QUERY_LD_SHARED 16
110 #define NVE4_PM_QUERY_ST_SHARED 17
111 #define NVE4_PM_QUERY_L1_LOCAL_LOAD_HIT 18
112 #define NVE4_PM_QUERY_L1_LOCAL_LOAD_MISS 19
113 #define NVE4_PM_QUERY_L1_LOCAL_STORE_HIT 20
114 #define NVE4_PM_QUERY_L1_LOCAL_STORE_MISS 21
115 #define NVE4_PM_QUERY_GLD_REQUEST 22
116 #define NVE4_PM_QUERY_GST_REQUEST 23
117 #define NVE4_PM_QUERY_L1_GLOBAL_LOAD_HIT 24
118 #define NVE4_PM_QUERY_L1_GLOBAL_LOAD_MISS 25
119 #define NVE4_PM_QUERY_GLD_TRANSACTIONS_UNCACHED 26
120 #define NVE4_PM_QUERY_GST_TRANSACTIONS 27
121 #define NVE4_PM_QUERY_BRANCH 28
122 #define NVE4_PM_QUERY_BRANCH_DIVERGENT 29
123 #define NVE4_PM_QUERY_ACTIVE_WARPS 30
124 #define NVE4_PM_QUERY_ACTIVE_CYCLES 31
125 #define NVE4_PM_QUERY_INST_ISSUED 32
126 #define NVE4_PM_QUERY_ATOM_COUNT 33
127 #define NVE4_PM_QUERY_GRED_COUNT 34
128 #define NVE4_PM_QUERY_LD_SHARED_REPLAY 35
129 #define NVE4_PM_QUERY_ST_SHARED_REPLAY 36
130 #define NVE4_PM_QUERY_LD_LOCAL_TRANSACTIONS 37
131 #define NVE4_PM_QUERY_ST_LOCAL_TRANSACTIONS 38
132 #define NVE4_PM_QUERY_L1_LD_SHARED_TRANSACTIONS 39
133 #define NVE4_PM_QUERY_L1_ST_SHARED_TRANSACTIONS 40
134 #define NVE4_PM_QUERY_GLD_MEM_DIV_REPLAY 41
135 #define NVE4_PM_QUERY_GST_MEM_DIV_REPLAY 42
136 #define NVE4_PM_QUERY_METRIC_IPC 43
137 #define NVE4_PM_QUERY_METRIC_IPAC 44
138 #define NVE4_PM_QUERY_METRIC_IPEC 45
139 #define NVE4_PM_QUERY_METRIC_MP_OCCUPANCY 46
140 #define NVE4_PM_QUERY_METRIC_MP_EFFICIENCY 47
141 #define NVE4_PM_QUERY_METRIC_INST_REPLAY_OHEAD 48
142
143 /*
144 #define NVE4_PM_QUERY_GR_IDLE 50
145 #define NVE4_PM_QUERY_BSP_IDLE 51
146 #define NVE4_PM_QUERY_VP_IDLE 52
147 #define NVE4_PM_QUERY_PPP_IDLE 53
148 #define NVE4_PM_QUERY_CE0_IDLE 54
149 #define NVE4_PM_QUERY_CE1_IDLE 55
150 #define NVE4_PM_QUERY_CE2_IDLE 56
151 */
152 /* L2 queries (PCOUNTER) */
153 /*
154 #define NVE4_PM_QUERY_L2_SUBP_WRITE_L1_SECTOR_QUERIES 57
155 ...
156 */
157 /* TEX queries (PCOUNTER) */
158 /*
159 #define NVE4_PM_QUERY_TEX0_CACHE_SECTOR_QUERIES 58
160 ...
161 */
162
163 #define NVC0_PM_QUERY_COUNT 31
164 #define NVC0_PM_QUERY(i) (PIPE_QUERY_DRIVER_SPECIFIC + 2048 + (i))
165 #define NVC0_PM_QUERY_LAST NVC0_PM_QUERY(NVC0_PM_QUERY_COUNT - 1)
166 #define NVC0_PM_QUERY_INST_EXECUTED 0
167 #define NVC0_PM_QUERY_BRANCH 1
168 #define NVC0_PM_QUERY_BRANCH_DIVERGENT 2
169 #define NVC0_PM_QUERY_ACTIVE_WARPS 3
170 #define NVC0_PM_QUERY_ACTIVE_CYCLES 4
171 #define NVC0_PM_QUERY_LAUNCHED_WARPS 5
172 #define NVC0_PM_QUERY_LAUNCHED_THREADS 6
173 #define NVC0_PM_QUERY_LD_SHARED 7
174 #define NVC0_PM_QUERY_ST_SHARED 8
175 #define NVC0_PM_QUERY_LD_LOCAL 9
176 #define NVC0_PM_QUERY_ST_LOCAL 10
177 #define NVC0_PM_QUERY_GRED_COUNT 11
178 #define NVC0_PM_QUERY_ATOM_COUNT 12
179 #define NVC0_PM_QUERY_GLD_REQUEST 13
180 #define NVC0_PM_QUERY_GST_REQUEST 14
181 #define NVC0_PM_QUERY_INST_ISSUED1_0 15
182 #define NVC0_PM_QUERY_INST_ISSUED1_1 16
183 #define NVC0_PM_QUERY_INST_ISSUED2_0 17
184 #define NVC0_PM_QUERY_INST_ISSUED2_1 18
185 #define NVC0_PM_QUERY_TH_INST_EXECUTED_0 19
186 #define NVC0_PM_QUERY_TH_INST_EXECUTED_1 20
187 #define NVC0_PM_QUERY_TH_INST_EXECUTED_2 21
188 #define NVC0_PM_QUERY_TH_INST_EXECUTED_3 22
189 #define NVC0_PM_QUERY_PROF_TRIGGER_0 23
190 #define NVC0_PM_QUERY_PROF_TRIGGER_1 24
191 #define NVC0_PM_QUERY_PROF_TRIGGER_2 25
192 #define NVC0_PM_QUERY_PROF_TRIGGER_3 26
193 #define NVC0_PM_QUERY_PROF_TRIGGER_4 27
194 #define NVC0_PM_QUERY_PROF_TRIGGER_5 28
195 #define NVC0_PM_QUERY_PROF_TRIGGER_6 29
196 #define NVC0_PM_QUERY_PROF_TRIGGER_7 30
197
198 /* Driver statistics queries:
199 */
200 #ifdef NOUVEAU_ENABLE_DRIVER_STATISTICS
201
202 #define NVC0_QUERY_DRV_STAT(i) (PIPE_QUERY_DRIVER_SPECIFIC + 1024 + (i))
203 #define NVC0_QUERY_DRV_STAT_COUNT 29
204 #define NVC0_QUERY_DRV_STAT_LAST NVC0_QUERY_DRV_STAT(NVC0_QUERY_DRV_STAT_COUNT - 1)
205 #define NVC0_QUERY_DRV_STAT_TEX_OBJECT_CURRENT_COUNT 0
206 #define NVC0_QUERY_DRV_STAT_TEX_OBJECT_CURRENT_BYTES 1
207 #define NVC0_QUERY_DRV_STAT_BUF_OBJECT_CURRENT_COUNT 2
208 #define NVC0_QUERY_DRV_STAT_BUF_OBJECT_CURRENT_BYTES_VID 3
209 #define NVC0_QUERY_DRV_STAT_BUF_OBJECT_CURRENT_BYTES_SYS 4
210 #define NVC0_QUERY_DRV_STAT_TEX_TRANSFERS_READ 5
211 #define NVC0_QUERY_DRV_STAT_TEX_TRANSFERS_WRITE 6
212 #define NVC0_QUERY_DRV_STAT_TEX_COPY_COUNT 7
213 #define NVC0_QUERY_DRV_STAT_TEX_BLIT_COUNT 8
214 #define NVC0_QUERY_DRV_STAT_TEX_CACHE_FLUSH_COUNT 9
215 #define NVC0_QUERY_DRV_STAT_BUF_TRANSFERS_READ 10
216 #define NVC0_QUERY_DRV_STAT_BUF_TRANSFERS_WRITE 11
217 #define NVC0_QUERY_DRV_STAT_BUF_READ_BYTES_STAGING_VID 12
218 #define NVC0_QUERY_DRV_STAT_BUF_WRITE_BYTES_DIRECT 13
219 #define NVC0_QUERY_DRV_STAT_BUF_WRITE_BYTES_STAGING_VID 14
220 #define NVC0_QUERY_DRV_STAT_BUF_WRITE_BYTES_STAGING_SYS 15
221 #define NVC0_QUERY_DRV_STAT_BUF_COPY_BYTES 16
222 #define NVC0_QUERY_DRV_STAT_BUF_NON_KERNEL_FENCE_SYNC_COUNT 17
223 #define NVC0_QUERY_DRV_STAT_ANY_NON_KERNEL_FENCE_SYNC_COUNT 18
224 #define NVC0_QUERY_DRV_STAT_QUERY_SYNC_COUNT 19
225 #define NVC0_QUERY_DRV_STAT_GPU_SERIALIZE_COUNT 20
226 #define NVC0_QUERY_DRV_STAT_DRAW_CALLS_ARRAY 21
227 #define NVC0_QUERY_DRV_STAT_DRAW_CALLS_INDEXED 22
228 #define NVC0_QUERY_DRV_STAT_DRAW_CALLS_FALLBACK_COUNT 23
229 #define NVC0_QUERY_DRV_STAT_USER_BUFFER_UPLOAD_BYTES 24
230 #define NVC0_QUERY_DRV_STAT_CONSTBUF_UPLOAD_COUNT 25
231 #define NVC0_QUERY_DRV_STAT_CONSTBUF_UPLOAD_BYTES 26
232 #define NVC0_QUERY_DRV_STAT_PUSHBUF_COUNT 27
233 #define NVC0_QUERY_DRV_STAT_RESOURCE_VALIDATE_COUNT 28
234
235 #else
236
237 #define NVC0_QUERY_DRV_STAT_COUNT 0
238
239 #endif
240
241 int nvc0_screen_get_driver_query_info(struct pipe_screen *, unsigned,
242 struct pipe_driver_query_info *);
243
244 boolean nvc0_blitter_create(struct nvc0_screen *);
245 void nvc0_blitter_destroy(struct nvc0_screen *);
246
247 void nvc0_screen_make_buffers_resident(struct nvc0_screen *);
248
249 int nvc0_screen_tic_alloc(struct nvc0_screen *, void *);
250 int nvc0_screen_tsc_alloc(struct nvc0_screen *, void *);
251
252 int nve4_screen_compute_setup(struct nvc0_screen *, struct nouveau_pushbuf *);
253 int nvc0_screen_compute_setup(struct nvc0_screen *, struct nouveau_pushbuf *);
254
255 boolean nvc0_screen_resize_tls_area(struct nvc0_screen *, uint32_t lpos,
256 uint32_t lneg, uint32_t cstack);
257
258 static INLINE void
259 nvc0_resource_fence(struct nv04_resource *res, uint32_t flags)
260 {
261 struct nvc0_screen *screen = nvc0_screen(res->base.screen);
262
263 if (res->mm) {
264 nouveau_fence_ref(screen->base.fence.current, &res->fence);
265 if (flags & NOUVEAU_BO_WR)
266 nouveau_fence_ref(screen->base.fence.current, &res->fence_wr);
267 }
268 }
269
270 static INLINE void
271 nvc0_resource_validate(struct nv04_resource *res, uint32_t flags)
272 {
273 if (likely(res->bo)) {
274 if (flags & NOUVEAU_BO_WR)
275 res->status |= NOUVEAU_BUFFER_STATUS_GPU_WRITING |
276 NOUVEAU_BUFFER_STATUS_DIRTY;
277 if (flags & NOUVEAU_BO_RD)
278 res->status |= NOUVEAU_BUFFER_STATUS_GPU_READING;
279
280 nvc0_resource_fence(res, flags);
281 }
282 }
283
284 struct nvc0_format {
285 uint32_t rt;
286 uint32_t tic;
287 uint32_t vtx;
288 uint32_t usage;
289 };
290
291 extern const struct nvc0_format nvc0_format_table[];
292
293 static INLINE void
294 nvc0_screen_tic_unlock(struct nvc0_screen *screen, struct nv50_tic_entry *tic)
295 {
296 if (tic->id >= 0)
297 screen->tic.lock[tic->id / 32] &= ~(1 << (tic->id % 32));
298 }
299
300 static INLINE void
301 nvc0_screen_tsc_unlock(struct nvc0_screen *screen, struct nv50_tsc_entry *tsc)
302 {
303 if (tsc->id >= 0)
304 screen->tsc.lock[tsc->id / 32] &= ~(1 << (tsc->id % 32));
305 }
306
307 static INLINE void
308 nvc0_screen_tic_free(struct nvc0_screen *screen, struct nv50_tic_entry *tic)
309 {
310 if (tic->id >= 0) {
311 screen->tic.entries[tic->id] = NULL;
312 screen->tic.lock[tic->id / 32] &= ~(1 << (tic->id % 32));
313 }
314 }
315
316 static INLINE void
317 nvc0_screen_tsc_free(struct nvc0_screen *screen, struct nv50_tsc_entry *tsc)
318 {
319 if (tsc->id >= 0) {
320 screen->tsc.entries[tsc->id] = NULL;
321 screen->tsc.lock[tsc->id / 32] &= ~(1 << (tsc->id % 32));
322 }
323 }
324
325 #endif