2 * Copyright 2012 Nouveau Project
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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 AUTHORS OR COPYRIGHT HOLDERS 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.
22 * Authors: Christoph Bumiller
25 #include "nvc0/nvc0_context.h"
26 #include "nvc0/nve4_compute.h"
28 #include "codegen/nv50_ir_driver.h"
31 static void nve4_compute_dump_launch_desc(const struct nve4_cp_launch_desc
*);
32 static void gp100_compute_dump_launch_desc(const struct gp100_cp_launch_desc
*);
37 nve4_screen_compute_setup(struct nvc0_screen
*screen
,
38 struct nouveau_pushbuf
*push
)
40 struct nouveau_device
*dev
= screen
->base
.device
;
41 struct nouveau_object
*chan
= screen
->base
.channel
;
47 switch (dev
->chipset
& ~0xf) {
50 obj_class
= NVF0_COMPUTE_CLASS
; /* GK110 */
53 obj_class
= NVE4_COMPUTE_CLASS
; /* GK104 */
56 obj_class
= GM107_COMPUTE_CLASS
;
59 obj_class
= GM200_COMPUTE_CLASS
;
62 NOUVEAU_ERR("unsupported chipset: NV%02x\n", dev
->chipset
);
66 ret
= nouveau_object_new(chan
, 0xbeef00c0, obj_class
, NULL
, 0,
69 NOUVEAU_ERR("Failed to allocate compute object: %d\n", ret
);
73 BEGIN_NVC0(push
, SUBC_CP(NV01_SUBCHAN_OBJECT
), 1);
74 PUSH_DATA (push
, screen
->compute
->oclass
);
76 BEGIN_NVC0(push
, NVE4_CP(TEMP_ADDRESS_HIGH
), 2);
77 PUSH_DATAh(push
, screen
->tls
->offset
);
78 PUSH_DATA (push
, screen
->tls
->offset
);
79 /* No idea why there are 2. Divide size by 2 to be safe.
80 * Actually this might be per-MP TEMP size and looks like I'm only using
81 * 2 MPs instead of all 8.
83 BEGIN_NVC0(push
, NVE4_CP(MP_TEMP_SIZE_HIGH(0)), 3);
84 PUSH_DATAh(push
, screen
->tls
->size
/ screen
->mp_count
);
85 PUSH_DATA (push
, (screen
->tls
->size
/ screen
->mp_count
) & ~0x7fff);
86 PUSH_DATA (push
, 0xff);
87 BEGIN_NVC0(push
, NVE4_CP(MP_TEMP_SIZE_HIGH(1)), 3);
88 PUSH_DATAh(push
, screen
->tls
->size
/ screen
->mp_count
);
89 PUSH_DATA (push
, (screen
->tls
->size
/ screen
->mp_count
) & ~0x7fff);
90 PUSH_DATA (push
, 0xff);
92 /* Unified address space ? Who needs that ? Certainly not OpenCL.
94 * FATAL: Buffers with addresses inside [0x1000000, 0x3000000] will NOT be
95 * accessible. We cannot prevent that at the moment, so expect failure.
97 BEGIN_NVC0(push
, NVE4_CP(LOCAL_BASE
), 1);
98 PUSH_DATA (push
, 0xff << 24);
99 BEGIN_NVC0(push
, NVE4_CP(SHARED_BASE
), 1);
100 PUSH_DATA (push
, 0xfe << 24);
102 BEGIN_NVC0(push
, NVE4_CP(CODE_ADDRESS_HIGH
), 2);
103 PUSH_DATAh(push
, screen
->text
->offset
);
104 PUSH_DATA (push
, screen
->text
->offset
);
106 BEGIN_NVC0(push
, SUBC_CP(0x0310), 1);
107 PUSH_DATA (push
, (obj_class
>= NVF0_COMPUTE_CLASS
) ? 0x400 : 0x300);
109 /* NOTE: these do not affect the state used by the 3D object */
110 BEGIN_NVC0(push
, NVE4_CP(TIC_ADDRESS_HIGH
), 3);
111 PUSH_DATAh(push
, screen
->txc
->offset
);
112 PUSH_DATA (push
, screen
->txc
->offset
);
113 PUSH_DATA (push
, NVC0_TIC_MAX_ENTRIES
- 1);
114 BEGIN_NVC0(push
, NVE4_CP(TSC_ADDRESS_HIGH
), 3);
115 PUSH_DATAh(push
, screen
->txc
->offset
+ 65536);
116 PUSH_DATA (push
, screen
->txc
->offset
+ 65536);
117 PUSH_DATA (push
, NVC0_TSC_MAX_ENTRIES
- 1);
119 if (obj_class
>= NVF0_COMPUTE_CLASS
) {
120 /* The blob calls GK110_COMPUTE.FIRMWARE[0x6], along with the args (0x1)
121 * passed with GK110_COMPUTE.GRAPH.SCRATCH[0x2]. This is currently
122 * disabled because our firmware doesn't support these commands and the
123 * GPU hangs if they are used. */
124 BEGIN_NIC0(push
, SUBC_CP(0x0248), 64);
125 for (i
= 63; i
>= 0; i
--)
126 PUSH_DATA(push
, 0x38000 | i
);
127 IMMED_NVC0(push
, SUBC_CP(NV50_GRAPH_SERIALIZE
), 0);
130 BEGIN_NVC0(push
, NVE4_CP(TEX_CB_INDEX
), 1);
131 PUSH_DATA (push
, 7); /* does not interfere with 3D */
133 /* Disabling this UNK command avoid a read fault when using texelFetch()
134 * from a compute shader for weird reasons.
135 if (obj_class == NVF0_COMPUTE_CLASS)
136 IMMED_NVC0(push, SUBC_CP(0x02c4), 1);
139 address
= screen
->uniform_bo
->offset
+ NVC0_CB_AUX_INFO(5);
141 /* MS sample coordinate offsets: these do not work with _ALT modes ! */
142 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH
), 2);
143 PUSH_DATAh(push
, address
+ NVC0_CB_AUX_MS_INFO
);
144 PUSH_DATA (push
, address
+ NVC0_CB_AUX_MS_INFO
);
145 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_LINE_LENGTH_IN
), 2);
146 PUSH_DATA (push
, 64);
148 BEGIN_1IC0(push
, NVE4_CP(UPLOAD_EXEC
), 17);
149 PUSH_DATA (push
, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR
| (0x20 << 1));
150 PUSH_DATA (push
, 0); /* 0 */
152 PUSH_DATA (push
, 1); /* 1 */
154 PUSH_DATA (push
, 0); /* 2 */
156 PUSH_DATA (push
, 1); /* 3 */
158 PUSH_DATA (push
, 2); /* 4 */
160 PUSH_DATA (push
, 3); /* 5 */
162 PUSH_DATA (push
, 2); /* 6 */
164 PUSH_DATA (push
, 3); /* 7 */
167 #ifdef NOUVEAU_NVE4_MP_TRAP_HANDLER
168 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH
), 2);
169 PUSH_DATAh(push
, screen
->parm
->offset
+ NVE4_CP_INPUT_TRAP_INFO_PTR
);
170 PUSH_DATA (push
, screen
->parm
->offset
+ NVE4_CP_INPUT_TRAP_INFO_PTR
);
171 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_LINE_LENGTH_IN
), 2);
172 PUSH_DATA (push
, 28);
174 BEGIN_1IC0(push
, NVE4_CP(UPLOAD_EXEC
), 8);
176 PUSH_DATA (push
, screen
->parm
->offset
+ NVE4_CP_PARAM_TRAP_INFO
);
177 PUSH_DATAh(push
, screen
->parm
->offset
+ NVE4_CP_PARAM_TRAP_INFO
);
178 PUSH_DATA (push
, screen
->tls
->offset
);
179 PUSH_DATAh(push
, screen
->tls
->offset
);
180 PUSH_DATA (push
, screen
->tls
->size
/ 2); /* MP TEMP block size */
181 PUSH_DATA (push
, screen
->tls
->size
/ 2 / 64); /* warp TEMP block size */
182 PUSH_DATA (push
, 0); /* warp cfstack size */
185 BEGIN_NVC0(push
, NVE4_CP(FLUSH
), 1);
186 PUSH_DATA (push
, NVE4_COMPUTE_FLUSH_CB
);
192 gm107_compute_validate_surfaces(struct nvc0_context
*nvc0
,
193 struct pipe_image_view
*view
, int slot
)
195 struct nv04_resource
*res
= nv04_resource(view
->resource
);
196 struct nouveau_pushbuf
*push
= nvc0
->base
.pushbuf
;
197 struct nvc0_screen
*screen
= nvc0
->screen
;
198 struct nouveau_bo
*txc
= nvc0
->screen
->txc
;
199 struct nv50_tic_entry
*tic
;
203 tic
= nv50_tic_entry(nvc0
->images_tic
[s
][slot
]);
205 res
= nv04_resource(tic
->pipe
.texture
);
206 nvc0_update_tic(nvc0
, tic
, res
);
209 tic
->id
= nvc0_screen_tic_alloc(nvc0
->screen
, tic
);
211 /* upload the texture view */
212 PUSH_SPACE(push
, 16);
213 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH
), 2);
214 PUSH_DATAh(push
, txc
->offset
+ (tic
->id
* 32));
215 PUSH_DATA (push
, txc
->offset
+ (tic
->id
* 32));
216 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_LINE_LENGTH_IN
), 2);
217 PUSH_DATA (push
, 32);
219 BEGIN_1IC0(push
, NVE4_CP(UPLOAD_EXEC
), 9);
220 PUSH_DATA (push
, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR
| (0x20 << 1));
221 PUSH_DATAp(push
, &tic
->tic
[0], 8);
223 BEGIN_NIC0(push
, NVE4_CP(TIC_FLUSH
), 1);
224 PUSH_DATA (push
, (tic
->id
<< 4) | 1);
226 if (res
->status
& NOUVEAU_BUFFER_STATUS_GPU_WRITING
) {
227 BEGIN_NIC0(push
, NVE4_CP(TEX_CACHE_CTL
), 1);
228 PUSH_DATA (push
, (tic
->id
<< 4) | 1);
230 nvc0
->screen
->tic
.lock
[tic
->id
/ 32] |= 1 << (tic
->id
% 32);
232 res
->status
&= ~NOUVEAU_BUFFER_STATUS_GPU_WRITING
;
233 res
->status
|= NOUVEAU_BUFFER_STATUS_GPU_READING
;
235 BCTX_REFN(nvc0
->bufctx_cp
, CP_SUF
, res
, RD
);
237 address
= screen
->uniform_bo
->offset
+ NVC0_CB_AUX_INFO(s
);
239 /* upload the texture handle */
240 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH
), 2);
241 PUSH_DATAh(push
, address
+ NVC0_CB_AUX_TEX_INFO(slot
+ 32));
242 PUSH_DATA (push
, address
+ NVC0_CB_AUX_TEX_INFO(slot
+ 32));
243 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_LINE_LENGTH_IN
), 2);
245 PUSH_DATA (push
, 0x1);
246 BEGIN_1IC0(push
, NVE4_CP(UPLOAD_EXEC
), 2);
247 PUSH_DATA (push
, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR
| (0x20 << 1));
248 PUSH_DATA (push
, tic
->id
);
250 BEGIN_NVC0(push
, NVE4_CP(FLUSH
), 1);
251 PUSH_DATA (push
, NVE4_COMPUTE_FLUSH_CB
);
255 nve4_compute_validate_surfaces(struct nvc0_context
*nvc0
)
257 struct nouveau_pushbuf
*push
= nvc0
->base
.pushbuf
;
262 if (!nvc0
->images_dirty
[s
])
265 address
= nvc0
->screen
->uniform_bo
->offset
+ NVC0_CB_AUX_INFO(s
);
267 for (i
= 0; i
< NVC0_MAX_IMAGES
; ++i
) {
268 struct pipe_image_view
*view
= &nvc0
->images
[s
][i
];
270 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH
), 2);
271 PUSH_DATAh(push
, address
+ NVC0_CB_AUX_SU_INFO(i
));
272 PUSH_DATA (push
, address
+ NVC0_CB_AUX_SU_INFO(i
));
273 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_LINE_LENGTH_IN
), 2);
274 PUSH_DATA (push
, 16 * 4);
275 PUSH_DATA (push
, 0x1);
276 BEGIN_1IC0(push
, NVE4_CP(UPLOAD_EXEC
), 1 + 16);
277 PUSH_DATA (push
, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR
| (0x20 << 1));
279 if (view
->resource
) {
280 struct nv04_resource
*res
= nv04_resource(view
->resource
);
282 if (res
->base
.target
== PIPE_BUFFER
) {
283 if (view
->access
& PIPE_IMAGE_ACCESS_WRITE
)
284 nvc0_mark_image_range_valid(view
);
287 nve4_set_surface_info(push
, view
, nvc0
);
288 BCTX_REFN(nvc0
->bufctx_cp
, CP_SUF
, res
, RDWR
);
290 if (nvc0
->screen
->base
.class_3d
>= GM107_3D_CLASS
)
291 gm107_compute_validate_surfaces(nvc0
, view
, i
);
293 for (j
= 0; j
< 16; j
++)
299 /* Thankfully, textures with samplers follow the normal rules. */
301 nve4_compute_validate_samplers(struct nvc0_context
*nvc0
)
303 bool need_flush
= nve4_validate_tsc(nvc0
, 5);
305 BEGIN_NVC0(nvc0
->base
.pushbuf
, NVE4_CP(TSC_FLUSH
), 1);
306 PUSH_DATA (nvc0
->base
.pushbuf
, 0);
309 /* Invalidate all 3D samplers because they are aliased. */
310 for (int s
= 0; s
< 5; s
++)
311 nvc0
->samplers_dirty
[s
] = ~0;
312 nvc0
->dirty_3d
|= NVC0_NEW_3D_SAMPLERS
;
315 /* (Code duplicated at bottom for various non-convincing reasons.
316 * E.g. we might want to use the COMPUTE subchannel to upload TIC/TSC
317 * entries to avoid a subchannel switch.
318 * Same for texture cache flushes.
319 * Also, the bufctx differs, and more IFs in the 3D version looks ugly.)
321 static void nve4_compute_validate_textures(struct nvc0_context
*);
324 nve4_compute_set_tex_handles(struct nvc0_context
*nvc0
)
326 struct nouveau_pushbuf
*push
= nvc0
->base
.pushbuf
;
327 struct nvc0_screen
*screen
= nvc0
->screen
;
329 const unsigned s
= nvc0_shader_stage(PIPE_SHADER_COMPUTE
);
331 uint32_t dirty
= nvc0
->textures_dirty
[s
] | nvc0
->samplers_dirty
[s
];
336 n
= util_logbase2(dirty
) + 1 - i
;
339 address
= screen
->uniform_bo
->offset
+ NVC0_CB_AUX_INFO(s
);
341 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH
), 2);
342 PUSH_DATAh(push
, address
+ NVC0_CB_AUX_TEX_INFO(i
));
343 PUSH_DATA (push
, address
+ NVC0_CB_AUX_TEX_INFO(i
));
344 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_LINE_LENGTH_IN
), 2);
345 PUSH_DATA (push
, n
* 4);
346 PUSH_DATA (push
, 0x1);
347 BEGIN_1IC0(push
, NVE4_CP(UPLOAD_EXEC
), 1 + n
);
348 PUSH_DATA (push
, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR
| (0x20 << 1));
349 PUSH_DATAp(push
, &nvc0
->tex_handles
[s
][i
], n
);
351 BEGIN_NVC0(push
, NVE4_CP(FLUSH
), 1);
352 PUSH_DATA (push
, NVE4_COMPUTE_FLUSH_CB
);
354 nvc0
->textures_dirty
[s
] = 0;
355 nvc0
->samplers_dirty
[s
] = 0;
359 nve4_compute_validate_constbufs(struct nvc0_context
*nvc0
)
361 struct nouveau_pushbuf
*push
= nvc0
->base
.pushbuf
;
364 while (nvc0
->constbuf_dirty
[s
]) {
365 int i
= ffs(nvc0
->constbuf_dirty
[s
]) - 1;
366 nvc0
->constbuf_dirty
[s
] &= ~(1 << i
);
368 if (nvc0
->constbuf
[s
][i
].user
) {
369 struct nouveau_bo
*bo
= nvc0
->screen
->uniform_bo
;
370 const unsigned base
= NVC0_CB_USR_INFO(s
);
371 const unsigned size
= nvc0
->constbuf
[s
][0].size
;
372 assert(i
== 0); /* we really only want OpenGL uniforms here */
373 assert(nvc0
->constbuf
[s
][0].u
.data
);
375 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH
), 2);
376 PUSH_DATAh(push
, bo
->offset
+ base
);
377 PUSH_DATA (push
, bo
->offset
+ base
);
378 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_LINE_LENGTH_IN
), 2);
379 PUSH_DATA (push
, size
);
380 PUSH_DATA (push
, 0x1);
381 BEGIN_1IC0(push
, NVE4_CP(UPLOAD_EXEC
), 1 + (size
/ 4));
382 PUSH_DATA (push
, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR
| (0x20 << 1));
383 PUSH_DATAp(push
, nvc0
->constbuf
[s
][0].u
.data
, size
/ 4);
386 struct nv04_resource
*res
=
387 nv04_resource(nvc0
->constbuf
[s
][i
].u
.buf
);
390 = nvc0
->screen
->uniform_bo
->offset
+ NVC0_CB_AUX_INFO(s
);
392 assert(i
> 0); /* we really only want uniform buffer objects */
394 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH
), 2);
395 PUSH_DATAh(push
, address
+ NVC0_CB_AUX_UBO_INFO(i
- 1));
396 PUSH_DATA (push
, address
+ NVC0_CB_AUX_UBO_INFO(i
- 1));
397 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_LINE_LENGTH_IN
), 2);
398 PUSH_DATA (push
, 4 * 4);
399 PUSH_DATA (push
, 0x1);
400 BEGIN_1IC0(push
, NVE4_CP(UPLOAD_EXEC
), 1 + 4);
401 PUSH_DATA (push
, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR
| (0x20 << 1));
403 PUSH_DATA (push
, res
->address
+ nvc0
->constbuf
[s
][i
].offset
);
404 PUSH_DATAh(push
, res
->address
+ nvc0
->constbuf
[s
][i
].offset
);
405 PUSH_DATA (push
, nvc0
->constbuf
[5][i
].size
);
407 BCTX_REFN(nvc0
->bufctx_cp
, CP_CB(i
), res
, RD
);
409 res
->cb_bindings
[s
] |= 1 << i
;
414 BEGIN_NVC0(push
, NVE4_CP(FLUSH
), 1);
415 PUSH_DATA (push
, NVE4_COMPUTE_FLUSH_CB
);
419 nve4_compute_validate_buffers(struct nvc0_context
*nvc0
)
421 struct nouveau_pushbuf
*push
= nvc0
->base
.pushbuf
;
426 address
= nvc0
->screen
->uniform_bo
->offset
+ NVC0_CB_AUX_INFO(s
);
428 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH
), 2);
429 PUSH_DATAh(push
, address
+ NVC0_CB_AUX_BUF_INFO(0));
430 PUSH_DATA (push
, address
+ NVC0_CB_AUX_BUF_INFO(0));
431 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_LINE_LENGTH_IN
), 2);
432 PUSH_DATA (push
, 4 * NVC0_MAX_BUFFERS
* 4);
433 PUSH_DATA (push
, 0x1);
434 BEGIN_1IC0(push
, NVE4_CP(UPLOAD_EXEC
), 1 + 4 * NVC0_MAX_BUFFERS
);
435 PUSH_DATA (push
, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR
| (0x20 << 1));
437 for (i
= 0; i
< NVC0_MAX_BUFFERS
; i
++) {
438 if (nvc0
->buffers
[s
][i
].buffer
) {
439 struct nv04_resource
*res
=
440 nv04_resource(nvc0
->buffers
[s
][i
].buffer
);
441 PUSH_DATA (push
, res
->address
+ nvc0
->buffers
[s
][i
].buffer_offset
);
442 PUSH_DATAh(push
, res
->address
+ nvc0
->buffers
[s
][i
].buffer_offset
);
443 PUSH_DATA (push
, nvc0
->buffers
[s
][i
].buffer_size
);
445 BCTX_REFN(nvc0
->bufctx_cp
, CP_BUF
, res
, RDWR
);
446 util_range_add(&res
->valid_buffer_range
,
447 nvc0
->buffers
[s
][i
].buffer_offset
,
448 nvc0
->buffers
[s
][i
].buffer_offset
+
449 nvc0
->buffers
[s
][i
].buffer_size
);
459 static struct nvc0_state_validate
460 validate_list_cp
[] = {
461 { nvc0_compprog_validate
, NVC0_NEW_CP_PROGRAM
},
462 { nve4_compute_validate_textures
, NVC0_NEW_CP_TEXTURES
},
463 { nve4_compute_validate_samplers
, NVC0_NEW_CP_SAMPLERS
},
464 { nve4_compute_set_tex_handles
, NVC0_NEW_CP_TEXTURES
|
465 NVC0_NEW_CP_SAMPLERS
},
466 { nve4_compute_validate_surfaces
, NVC0_NEW_CP_SURFACES
},
467 { nvc0_compute_validate_globals
, NVC0_NEW_CP_GLOBALS
},
468 { nve4_compute_validate_buffers
, NVC0_NEW_CP_BUFFERS
},
469 { nve4_compute_validate_constbufs
, NVC0_NEW_CP_CONSTBUF
},
473 nve4_state_validate_cp(struct nvc0_context
*nvc0
, uint32_t mask
)
477 ret
= nvc0_state_validate(nvc0
, mask
, validate_list_cp
,
478 ARRAY_SIZE(validate_list_cp
), &nvc0
->dirty_cp
,
481 if (unlikely(nvc0
->state
.flushed
))
482 nvc0_bufctx_fence(nvc0
, nvc0
->bufctx_cp
, true);
487 nve4_compute_upload_input(struct nvc0_context
*nvc0
,
488 const struct pipe_grid_info
*info
)
490 struct nvc0_screen
*screen
= nvc0
->screen
;
491 struct nouveau_pushbuf
*push
= nvc0
->base
.pushbuf
;
492 struct nvc0_program
*cp
= nvc0
->compprog
;
495 address
= screen
->uniform_bo
->offset
+ NVC0_CB_AUX_INFO(5);
498 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH
), 2);
499 PUSH_DATAh(push
, screen
->uniform_bo
->offset
+ NVC0_CB_USR_INFO(5));
500 PUSH_DATA (push
, screen
->uniform_bo
->offset
+ NVC0_CB_USR_INFO(5));
501 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_LINE_LENGTH_IN
), 2);
502 PUSH_DATA (push
, cp
->parm_size
);
503 PUSH_DATA (push
, 0x1);
504 BEGIN_1IC0(push
, NVE4_CP(UPLOAD_EXEC
), 1 + (cp
->parm_size
/ 4));
505 PUSH_DATA (push
, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR
| (0x20 << 1));
506 PUSH_DATAp(push
, info
->input
, cp
->parm_size
/ 4);
508 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH
), 2);
509 PUSH_DATAh(push
, address
+ NVC0_CB_AUX_GRID_INFO(0));
510 PUSH_DATA (push
, address
+ NVC0_CB_AUX_GRID_INFO(0));
511 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_LINE_LENGTH_IN
), 2);
512 PUSH_DATA (push
, 8 * 4);
513 PUSH_DATA (push
, 0x1);
515 if (unlikely(info
->indirect
)) {
516 struct nv04_resource
*res
= nv04_resource(info
->indirect
);
517 uint32_t offset
= res
->offset
+ info
->indirect_offset
;
519 nouveau_pushbuf_space(push
, 32, 0, 1);
520 PUSH_REFN(push
, res
->bo
, NOUVEAU_BO_RD
| res
->domain
);
522 BEGIN_1IC0(push
, NVE4_CP(UPLOAD_EXEC
), 1 + 8);
523 PUSH_DATA (push
, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR
| (0x20 << 1));
524 PUSH_DATAp(push
, info
->block
, 3);
525 nouveau_pushbuf_data(push
, res
->bo
, offset
,
526 NVC0_IB_ENTRY_1_NO_PREFETCH
| 3 * 4);
528 BEGIN_1IC0(push
, NVE4_CP(UPLOAD_EXEC
), 1 + 8);
529 PUSH_DATA (push
, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR
| (0x20 << 1));
530 PUSH_DATAp(push
, info
->block
, 3);
531 PUSH_DATAp(push
, info
->grid
, 3);
534 PUSH_DATA (push
, info
->work_dim
);
536 BEGIN_NVC0(push
, NVE4_CP(FLUSH
), 1);
537 PUSH_DATA (push
, NVE4_COMPUTE_FLUSH_CB
);
540 static inline uint8_t
541 nve4_compute_derive_cache_split(struct nvc0_context
*nvc0
, uint32_t shared_size
)
543 if (shared_size
> (32 << 10))
544 return NVC0_3D_CACHE_SPLIT_48K_SHARED_16K_L1
;
545 if (shared_size
> (16 << 10))
546 return NVE4_3D_CACHE_SPLIT_32K_SHARED_32K_L1
;
547 return NVC1_3D_CACHE_SPLIT_16K_SHARED_48K_L1
;
551 nve4_compute_setup_launch_desc(struct nvc0_context
*nvc0
,
552 struct nve4_cp_launch_desc
*desc
,
553 const struct pipe_grid_info
*info
)
555 const struct nvc0_screen
*screen
= nvc0
->screen
;
556 const struct nvc0_program
*cp
= nvc0
->compprog
;
558 nve4_cp_launch_desc_init_default(desc
);
560 desc
->entry
= nvc0_program_symbol_offset(cp
, info
->pc
);
562 desc
->griddim_x
= info
->grid
[0];
563 desc
->griddim_y
= info
->grid
[1];
564 desc
->griddim_z
= info
->grid
[2];
565 desc
->blockdim_x
= info
->block
[0];
566 desc
->blockdim_y
= info
->block
[1];
567 desc
->blockdim_z
= info
->block
[2];
569 desc
->shared_size
= align(cp
->cp
.smem_size
, 0x100);
570 desc
->local_size_p
= (cp
->hdr
[1] & 0xfffff0) + align(cp
->cp
.lmem_size
, 0x10);
571 desc
->local_size_n
= 0;
572 desc
->cstack_size
= 0x800;
573 desc
->cache_split
= nve4_compute_derive_cache_split(nvc0
, cp
->cp
.smem_size
);
575 desc
->gpr_alloc
= cp
->num_gprs
;
576 desc
->bar_alloc
= cp
->num_barriers
;
578 // Only bind user uniforms and the driver constant buffer through the
579 // launch descriptor because UBOs are sticked to the driver cb to avoid the
580 // limitation of 8 CBs.
581 if (nvc0
->constbuf
[5][0].user
|| cp
->parm_size
) {
582 nve4_cp_launch_desc_set_cb(desc
, 0, screen
->uniform_bo
,
583 NVC0_CB_USR_INFO(5), 1 << 16);
585 nve4_cp_launch_desc_set_cb(desc
, 7, screen
->uniform_bo
,
586 NVC0_CB_AUX_INFO(5), 1 << 11);
590 gp100_compute_setup_launch_desc(struct nvc0_context
*nvc0
,
591 struct gp100_cp_launch_desc
*desc
,
592 const struct pipe_grid_info
*info
)
594 const struct nvc0_screen
*screen
= nvc0
->screen
;
595 const struct nvc0_program
*cp
= nvc0
->compprog
;
597 gp100_cp_launch_desc_init_default(desc
);
599 desc
->entry
= nvc0_program_symbol_offset(cp
, info
->pc
);
601 desc
->griddim_x
= info
->grid
[0];
602 desc
->griddim_y
= info
->grid
[1];
603 desc
->griddim_z
= info
->grid
[2];
604 desc
->blockdim_x
= info
->block
[0];
605 desc
->blockdim_y
= info
->block
[1];
606 desc
->blockdim_z
= info
->block
[2];
608 desc
->shared_size
= align(cp
->cp
.smem_size
, 0x100);
609 desc
->local_size_p
= (cp
->hdr
[1] & 0xfffff0) + align(cp
->cp
.lmem_size
, 0x10);
610 desc
->local_size_n
= 0;
611 desc
->cstack_size
= 0x800;
613 desc
->gpr_alloc
= cp
->num_gprs
;
614 desc
->bar_alloc
= cp
->num_barriers
;
616 // Only bind user uniforms and the driver constant buffer through the
617 // launch descriptor because UBOs are sticked to the driver cb to avoid the
618 // limitation of 8 CBs.
619 if (nvc0
->constbuf
[5][0].user
|| cp
->parm_size
) {
620 gp100_cp_launch_desc_set_cb(desc
, 0, screen
->uniform_bo
,
621 NVC0_CB_USR_INFO(5), 1 << 16);
623 gp100_cp_launch_desc_set_cb(desc
, 7, screen
->uniform_bo
,
624 NVC0_CB_AUX_INFO(5), 1 << 11);
628 nve4_compute_alloc_launch_desc(struct nouveau_context
*nv
,
629 struct nouveau_bo
**pbo
, uint64_t *pgpuaddr
)
631 uint8_t *ptr
= nouveau_scratch_get(nv
, 512, pgpuaddr
, pbo
);
634 if (*pgpuaddr
& 255) {
635 unsigned adj
= 256 - (*pgpuaddr
& 255);
643 nve4_upload_indirect_desc(struct nouveau_pushbuf
*push
,
644 struct nv04_resource
*res
, uint64_t gpuaddr
,
645 uint32_t length
, uint32_t bo_offset
)
647 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH
), 2);
648 PUSH_DATAh(push
, gpuaddr
);
649 PUSH_DATA (push
, gpuaddr
);
650 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_LINE_LENGTH_IN
), 2);
651 PUSH_DATA (push
, length
);
654 nouveau_pushbuf_space(push
, 32, 0, 1);
655 PUSH_REFN(push
, res
->bo
, NOUVEAU_BO_RD
| res
->domain
);
657 BEGIN_1IC0(push
, NVE4_CP(UPLOAD_EXEC
), 1 + (length
/ 4));
658 PUSH_DATA (push
, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR
| (0x08 << 1));
659 nouveau_pushbuf_data(push
, res
->bo
, bo_offset
,
660 NVC0_IB_ENTRY_1_NO_PREFETCH
| length
);
664 nve4_launch_grid(struct pipe_context
*pipe
, const struct pipe_grid_info
*info
)
666 struct nvc0_context
*nvc0
= nvc0_context(pipe
);
667 struct nouveau_pushbuf
*push
= nvc0
->base
.pushbuf
;
669 uint64_t desc_gpuaddr
;
670 struct nouveau_bo
*desc_bo
;
673 desc
= nve4_compute_alloc_launch_desc(&nvc0
->base
, &desc_bo
, &desc_gpuaddr
);
678 BCTX_REFN_bo(nvc0
->bufctx_cp
, CP_DESC
, NOUVEAU_BO_GART
| NOUVEAU_BO_RD
,
681 ret
= !nve4_state_validate_cp(nvc0
, ~0);
685 if (nvc0
->screen
->compute
->oclass
>= GP100_COMPUTE_CLASS
)
686 gp100_compute_setup_launch_desc(nvc0
, desc
, info
);
688 nve4_compute_setup_launch_desc(nvc0
, desc
, info
);
690 nve4_compute_upload_input(nvc0
, info
);
693 if (debug_get_num_option("NV50_PROG_DEBUG", 0)) {
694 if (nvc0
->screen
->compute
->oclass
>= GP100_COMPUTE_CLASS
)
695 gp100_compute_dump_launch_desc(desc
);
697 nve4_compute_dump_launch_desc(desc
);
701 if (unlikely(info
->indirect
)) {
702 struct nv04_resource
*res
= nv04_resource(info
->indirect
);
703 uint32_t offset
= res
->offset
+ info
->indirect_offset
;
705 /* upload the descriptor */
706 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH
), 2);
707 PUSH_DATAh(push
, desc_gpuaddr
);
708 PUSH_DATA (push
, desc_gpuaddr
);
709 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_LINE_LENGTH_IN
), 2);
710 PUSH_DATA (push
, 256);
712 BEGIN_1IC0(push
, NVE4_CP(UPLOAD_EXEC
), 1 + (256 / 4));
713 PUSH_DATA (push
, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR
| (0x08 << 1));
714 PUSH_DATAp(push
, (const uint32_t *)desc
, 256 / 4);
716 if (nvc0
->screen
->compute
->oclass
>= GP100_COMPUTE_CLASS
) {
717 nve4_upload_indirect_desc(push
, res
, desc_gpuaddr
+ 48, 12, offset
);
719 /* overwrite griddim_x and griddim_y as two 32-bits integers even
720 * if griddim_y must be a 16-bits integer */
721 nve4_upload_indirect_desc(push
, res
, desc_gpuaddr
+ 48, 8, offset
);
723 /* overwrite the 16 high bits of griddim_y with griddim_z because
724 * we need (z << 16) | x */
725 nve4_upload_indirect_desc(push
, res
, desc_gpuaddr
+ 54, 4, offset
+ 8);
729 /* upload descriptor and flush */
730 BEGIN_NVC0(push
, NVE4_CP(LAUNCH_DESC_ADDRESS
), 1);
731 PUSH_DATA (push
, desc_gpuaddr
>> 8);
732 BEGIN_NVC0(push
, NVE4_CP(LAUNCH
), 1);
733 PUSH_DATA (push
, 0x3);
734 BEGIN_NVC0(push
, SUBC_CP(NV50_GRAPH_SERIALIZE
), 1);
739 NOUVEAU_ERR("Failed to launch grid !\n");
740 nouveau_scratch_done(&nvc0
->base
);
741 nouveau_bufctx_reset(nvc0
->bufctx_cp
, NVC0_BIND_CP_DESC
);
745 #define NVE4_TIC_ENTRY_INVALID 0x000fffff
748 nve4_compute_validate_textures(struct nvc0_context
*nvc0
)
750 struct nouveau_bo
*txc
= nvc0
->screen
->txc
;
751 struct nouveau_pushbuf
*push
= nvc0
->base
.pushbuf
;
752 const unsigned s
= 5;
754 uint32_t commands
[2][32];
755 unsigned n
[2] = { 0, 0 };
757 for (i
= 0; i
< nvc0
->num_textures
[s
]; ++i
) {
758 struct nv50_tic_entry
*tic
= nv50_tic_entry(nvc0
->textures
[s
][i
]);
759 struct nv04_resource
*res
;
760 const bool dirty
= !!(nvc0
->textures_dirty
[s
] & (1 << i
));
763 nvc0
->tex_handles
[s
][i
] |= NVE4_TIC_ENTRY_INVALID
;
766 res
= nv04_resource(tic
->pipe
.texture
);
767 nvc0_update_tic(nvc0
, tic
, res
);
770 tic
->id
= nvc0_screen_tic_alloc(nvc0
->screen
, tic
);
772 PUSH_SPACE(push
, 16);
773 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH
), 2);
774 PUSH_DATAh(push
, txc
->offset
+ (tic
->id
* 32));
775 PUSH_DATA (push
, txc
->offset
+ (tic
->id
* 32));
776 BEGIN_NVC0(push
, NVE4_CP(UPLOAD_LINE_LENGTH_IN
), 2);
777 PUSH_DATA (push
, 32);
779 BEGIN_1IC0(push
, NVE4_CP(UPLOAD_EXEC
), 9);
780 PUSH_DATA (push
, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR
| (0x20 << 1));
781 PUSH_DATAp(push
, &tic
->tic
[0], 8);
783 commands
[0][n
[0]++] = (tic
->id
<< 4) | 1;
785 if (res
->status
& NOUVEAU_BUFFER_STATUS_GPU_WRITING
) {
786 commands
[1][n
[1]++] = (tic
->id
<< 4) | 1;
788 nvc0
->screen
->tic
.lock
[tic
->id
/ 32] |= 1 << (tic
->id
% 32);
790 res
->status
&= ~NOUVEAU_BUFFER_STATUS_GPU_WRITING
;
791 res
->status
|= NOUVEAU_BUFFER_STATUS_GPU_READING
;
793 nvc0
->tex_handles
[s
][i
] &= ~NVE4_TIC_ENTRY_INVALID
;
794 nvc0
->tex_handles
[s
][i
] |= tic
->id
;
796 BCTX_REFN(nvc0
->bufctx_cp
, CP_TEX(i
), res
, RD
);
798 for (; i
< nvc0
->state
.num_textures
[s
]; ++i
) {
799 nvc0
->tex_handles
[s
][i
] |= NVE4_TIC_ENTRY_INVALID
;
800 nvc0
->textures_dirty
[s
] |= 1 << i
;
804 BEGIN_NIC0(push
, NVE4_CP(TIC_FLUSH
), n
[0]);
805 PUSH_DATAp(push
, commands
[0], n
[0]);
808 BEGIN_NIC0(push
, NVE4_CP(TEX_CACHE_CTL
), n
[1]);
809 PUSH_DATAp(push
, commands
[1], n
[1]);
812 nvc0
->state
.num_textures
[s
] = nvc0
->num_textures
[s
];
814 /* Invalidate all 3D textures because they are aliased. */
815 for (int s
= 0; s
< 5; s
++) {
816 for (int i
= 0; i
< nvc0
->num_textures
[s
]; i
++)
817 nouveau_bufctx_reset(nvc0
->bufctx_3d
, NVC0_BIND_3D_TEX(s
, i
));
818 nvc0
->textures_dirty
[s
] = ~0;
820 nvc0
->dirty_3d
|= NVC0_NEW_3D_TEXTURES
;
825 static const char *nve4_cache_split_name(unsigned value
)
828 case NVC1_3D_CACHE_SPLIT_16K_SHARED_48K_L1
: return "16K_SHARED_48K_L1";
829 case NVE4_3D_CACHE_SPLIT_32K_SHARED_32K_L1
: return "32K_SHARED_32K_L1";
830 case NVC0_3D_CACHE_SPLIT_48K_SHARED_16K_L1
: return "48K_SHARED_16K_L1";
837 nve4_compute_dump_launch_desc(const struct nve4_cp_launch_desc
*desc
)
839 const uint32_t *data
= (const uint32_t *)desc
;
843 debug_printf("COMPUTE LAUNCH DESCRIPTOR:\n");
845 for (i
= 0; i
< sizeof(*desc
); i
+= 4) {
847 debug_printf("[%x]: 0x%08x\n", i
, data
[i
/ 4]);
851 debug_printf("...\n");
856 debug_printf("entry = 0x%x\n", desc
->entry
);
857 debug_printf("grid dimensions = %ux%ux%u\n",
858 desc
->griddim_x
, desc
->griddim_y
, desc
->griddim_z
);
859 debug_printf("block dimensions = %ux%ux%u\n",
860 desc
->blockdim_x
, desc
->blockdim_y
, desc
->blockdim_z
);
861 debug_printf("s[] size: 0x%x\n", desc
->shared_size
);
862 debug_printf("l[] size: -0x%x / +0x%x\n",
863 desc
->local_size_n
, desc
->local_size_p
);
864 debug_printf("stack size: 0x%x\n", desc
->cstack_size
);
865 debug_printf("barrier count: %u\n", desc
->bar_alloc
);
866 debug_printf("$r count: %u\n", desc
->gpr_alloc
);
867 debug_printf("cache split: %s\n", nve4_cache_split_name(desc
->cache_split
));
868 debug_printf("linked tsc: %d\n", desc
->linked_tsc
);
870 for (i
= 0; i
< 8; ++i
) {
872 uint32_t size
= desc
->cb
[i
].size
;
873 bool valid
= !!(desc
->cb_mask
& (1 << i
));
875 address
= ((uint64_t)desc
->cb
[i
].address_h
<< 32) | desc
->cb
[i
].address_l
;
877 if (!valid
&& !address
&& !size
)
879 debug_printf("CB[%u]: address = 0x%"PRIx64
", size 0x%x%s\n",
880 i
, address
, size
, valid
? "" : " (invalid)");
885 gp100_compute_dump_launch_desc(const struct gp100_cp_launch_desc
*desc
)
887 const uint32_t *data
= (const uint32_t *)desc
;
891 debug_printf("COMPUTE LAUNCH DESCRIPTOR:\n");
893 for (i
= 0; i
< sizeof(*desc
); i
+= 4) {
895 debug_printf("[%x]: 0x%08x\n", i
, data
[i
/ 4]);
899 debug_printf("...\n");
904 debug_printf("entry = 0x%x\n", desc
->entry
);
905 debug_printf("grid dimensions = %ux%ux%u\n",
906 desc
->griddim_x
, desc
->griddim_y
, desc
->griddim_z
);
907 debug_printf("block dimensions = %ux%ux%u\n",
908 desc
->blockdim_x
, desc
->blockdim_y
, desc
->blockdim_z
);
909 debug_printf("s[] size: 0x%x\n", desc
->shared_size
);
910 debug_printf("l[] size: -0x%x / +0x%x\n",
911 desc
->local_size_n
, desc
->local_size_p
);
912 debug_printf("stack size: 0x%x\n", desc
->cstack_size
);
913 debug_printf("barrier count: %u\n", desc
->bar_alloc
);
914 debug_printf("$r count: %u\n", desc
->gpr_alloc
);
915 debug_printf("linked tsc: %d\n", desc
->linked_tsc
);
917 for (i
= 0; i
< 8; ++i
) {
919 uint32_t size
= desc
->cb
[i
].size_sh4
<< 4;
920 bool valid
= !!(desc
->cb_mask
& (1 << i
));
922 address
= ((uint64_t)desc
->cb
[i
].address_h
<< 32) | desc
->cb
[i
].address_l
;
924 if (!valid
&& !address
&& !size
)
926 debug_printf("CB[%u]: address = 0x%"PRIx64
", size 0x%x%s\n",
927 i
, address
, size
, valid
? "" : " (invalid)");
932 #ifdef NOUVEAU_NVE4_MP_TRAP_HANDLER
934 nve4_compute_trap_info(struct nvc0_context
*nvc0
)
936 struct nvc0_screen
*screen
= nvc0
->screen
;
937 struct nouveau_bo
*bo
= screen
->parm
;
939 volatile struct nve4_mp_trap_info
*info
;
942 ret
= nouveau_bo_map(bo
, NOUVEAU_BO_RDWR
, nvc0
->base
.client
);
945 map
= (uint8_t *)bo
->map
;
946 info
= (volatile struct nve4_mp_trap_info
*)(map
+ NVE4_CP_PARAM_TRAP_INFO
);
949 debug_printf("trapstat = %08x\n", info
->trapstat
);
950 debug_printf("warperr = %08x\n", info
->warperr
);
951 debug_printf("PC = %x\n", info
->pc
);
952 debug_printf("tid = %u %u %u\n",
953 info
->tid
[0], info
->tid
[1], info
->tid
[2]);
954 debug_printf("ctaid = %u %u %u\n",
955 info
->ctaid
[0], info
->ctaid
[1], info
->ctaid
[2]);
956 for (i
= 0; i
<= 63; ++i
)
957 debug_printf("$r%i = %08x\n", i
, info
->r
[i
]);
958 for (i
= 0; i
<= 6; ++i
)
959 debug_printf("$p%i = %i\n", i
, (info
->flags
>> i
) & 1);
960 debug_printf("$c = %x\n", info
->flags
>> 12);