2 * Copyright 2011 Adam Rak <adam.rak@streamnovation.com>
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 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
24 * Adam Rak <adam.rak@streamnovation.com>
29 #include "pipe/p_defines.h"
30 #include "pipe/p_state.h"
31 #include "pipe/p_context.h"
32 #include "util/u_blitter.h"
33 #include "util/list.h"
34 #include "util/u_transfer.h"
35 #include "util/u_surface.h"
36 #include "util/u_pack_color.h"
37 #include "util/u_memory.h"
38 #include "util/u_inlines.h"
39 #include "util/u_framebuffer.h"
40 #include "pipebuffer/pb_buffer.h"
41 #include "evergreend.h"
42 #include "r600_shader.h"
43 #include "r600_pipe.h"
44 #include "r600_formats.h"
45 #include "evergreen_compute.h"
46 #include "evergreen_compute_internal.h"
47 #include "compute_memory_pool.h"
48 #include "sb/sb_public.h"
50 #include "radeon/radeon_llvm_util.h"
52 #include "radeon/radeon_elf_util.h"
56 RAT0 is for global binding write
57 VTX1 is for global binding read
59 for wrting images RAT1...
60 for reading images TEX2...
63 TEX2... consumes the same fetch resources, that VTX2... would consume
65 CONST0 and VTX0 is for parameters
66 CONST0 is binding smaller input parameter buffer, and for constant indexing,
68 VTX0 is for indirect/non-constant indexing, or if the input is bigger than
69 the constant cache can handle
71 RAT-s are limited to 12, so we can only bind at most 11 texture for writing
72 because we reserve RAT0 for global bindings. With byteaddressing enabled,
73 we should reserve another one too.=> 10 image binding for writing max.
76 CL_DEVICE_MAX_READ_IMAGE_ARGS: 128
77 CL_DEVICE_MAX_WRITE_IMAGE_ARGS: 8
79 so 10 for writing is enough. 176 is the max for reading according to the docs
81 writable images should be listed first < 10, so their id corresponds to RAT(id+1)
82 writable images will consume TEX slots, VTX slots too because of linear indexing
86 struct r600_resource
*r600_compute_buffer_alloc_vram(struct r600_screen
*screen
,
89 struct pipe_resource
*buffer
= NULL
;
92 buffer
= pipe_buffer_create((struct pipe_screen
*) screen
,
97 return (struct r600_resource
*)buffer
;
101 static void evergreen_set_rat(struct r600_pipe_compute
*pipe
,
103 struct r600_resource
*bo
,
107 struct pipe_surface rat_templ
;
108 struct r600_surface
*surf
= NULL
;
109 struct r600_context
*rctx
= NULL
;
112 assert((size
& 3) == 0);
113 assert((start
& 0xFF) == 0);
117 COMPUTE_DBG(rctx
->screen
, "bind rat: %i \n", id
);
119 /* Create the RAT surface */
120 memset(&rat_templ
, 0, sizeof(rat_templ
));
121 rat_templ
.format
= PIPE_FORMAT_R32_UINT
;
122 rat_templ
.u
.tex
.level
= 0;
123 rat_templ
.u
.tex
.first_layer
= 0;
124 rat_templ
.u
.tex
.last_layer
= 0;
126 /* Add the RAT the list of color buffers */
127 pipe
->ctx
->framebuffer
.state
.cbufs
[id
] = pipe
->ctx
->b
.b
.create_surface(
128 (struct pipe_context
*)pipe
->ctx
,
129 (struct pipe_resource
*)bo
, &rat_templ
);
131 /* Update the number of color buffers */
132 pipe
->ctx
->framebuffer
.state
.nr_cbufs
=
133 MAX2(id
+ 1, pipe
->ctx
->framebuffer
.state
.nr_cbufs
);
135 /* Update the cb_target_mask
136 * XXX: I think this is a potential spot for bugs once we start doing
137 * GL interop. cb_target_mask may be modified in the 3D sections
139 pipe
->ctx
->compute_cb_target_mask
|= (0xf << (id
* 4));
141 surf
= (struct r600_surface
*)pipe
->ctx
->framebuffer
.state
.cbufs
[id
];
142 evergreen_init_color_surface_rat(rctx
, surf
);
145 static void evergreen_cs_set_vertex_buffer(struct r600_context
*rctx
,
148 struct pipe_resource
*buffer
)
150 struct r600_vertexbuf_state
*state
= &rctx
->cs_vertex_buffer_state
;
151 struct pipe_vertex_buffer
*vb
= &state
->vb
[vb_index
];
153 vb
->buffer_offset
= offset
;
155 vb
->user_buffer
= NULL
;
157 /* The vertex instructions in the compute shaders use the texture cache,
158 * so we need to invalidate it. */
159 rctx
->b
.flags
|= R600_CONTEXT_INV_VERTEX_CACHE
;
160 state
->enabled_mask
|= 1 << vb_index
;
161 state
->dirty_mask
|= 1 << vb_index
;
162 r600_mark_atom_dirty(rctx
, &state
->atom
);
165 static void evergreen_cs_set_constant_buffer(struct r600_context
*rctx
,
169 struct pipe_resource
*buffer
)
171 struct pipe_constant_buffer cb
;
172 cb
.buffer_size
= size
;
173 cb
.buffer_offset
= offset
;
175 cb
.user_buffer
= NULL
;
177 rctx
->b
.b
.set_constant_buffer(&rctx
->b
.b
, PIPE_SHADER_COMPUTE
, cb_index
, &cb
);
180 /* We need to define these R600 registers here, because we can't include
181 * evergreend.h and r600d.h.
183 #define R_028868_SQ_PGM_RESOURCES_VS 0x028868
184 #define R_028850_SQ_PGM_RESOURCES_PS 0x028850
188 static void r600_shader_binary_read_config(const struct radeon_shader_binary
*binary
,
189 struct r600_bytecode
*bc
,
190 uint64_t symbol_offset
,
194 const unsigned char *config
=
195 radeon_shader_binary_config_start(binary
, symbol_offset
);
197 for (i
= 0; i
< binary
->config_size_per_symbol
; i
+= 8) {
199 util_le32_to_cpu(*(uint32_t*)(config
+ i
));
201 util_le32_to_cpu(*(uint32_t*)(config
+ i
+ 4));
204 case R_028850_SQ_PGM_RESOURCES_PS
:
205 case R_028868_SQ_PGM_RESOURCES_VS
:
206 /* Evergreen / Northern Islands */
207 case R_028844_SQ_PGM_RESOURCES_PS
:
208 case R_028860_SQ_PGM_RESOURCES_VS
:
209 case R_0288D4_SQ_PGM_RESOURCES_LS
:
210 bc
->ngpr
= MAX2(bc
->ngpr
, G_028844_NUM_GPRS(value
));
211 bc
->nstack
= MAX2(bc
->nstack
, G_028844_STACK_SIZE(value
));
213 case R_02880C_DB_SHADER_CONTROL
:
214 *use_kill
= G_02880C_KILL_ENABLE(value
);
216 case R_0288E8_SQ_LDS_ALLOC
:
223 static unsigned r600_create_shader(struct r600_bytecode
*bc
,
224 const struct radeon_shader_binary
*binary
,
228 assert(binary
->code_size
% 4 == 0);
229 bc
->bytecode
= CALLOC(1, binary
->code_size
);
230 memcpy(bc
->bytecode
, binary
->code
, binary
->code_size
);
231 bc
->ndw
= binary
->code_size
/ 4;
233 r600_shader_binary_read_config(binary
, bc
, 0, use_kill
);
239 static void r600_destroy_shader(struct r600_bytecode
*bc
)
244 static void *evergreen_create_compute_state(struct pipe_context
*ctx
,
245 const const struct pipe_compute_state
*cso
)
247 struct r600_context
*rctx
= (struct r600_context
*)ctx
;
248 struct r600_pipe_compute
*shader
= CALLOC_STRUCT(r600_pipe_compute
);
250 const struct pipe_llvm_program_header
*header
;
255 COMPUTE_DBG(rctx
->screen
, "*** evergreen_create_compute_state\n");
257 code
= cso
->prog
+ sizeof(struct pipe_llvm_program_header
);
258 radeon_shader_binary_init(&shader
->binary
);
259 radeon_elf_read(code
, header
->num_bytes
, &shader
->binary
);
260 r600_create_shader(&shader
->bc
, &shader
->binary
, &use_kill
);
262 shader
->code_bo
= r600_compute_buffer_alloc_vram(rctx
->screen
,
264 p
= r600_buffer_map_sync_with_rings(&rctx
->b
, shader
->code_bo
, PIPE_TRANSFER_WRITE
);
265 memcpy(p
, shader
->bc
.bytecode
, shader
->bc
.ndw
* 4);
266 rctx
->b
.ws
->buffer_unmap(shader
->code_bo
->buf
);
270 shader
->local_size
= cso
->req_local_mem
;
271 shader
->private_size
= cso
->req_private_mem
;
272 shader
->input_size
= cso
->req_input_mem
;
277 static void evergreen_delete_compute_state(struct pipe_context
*ctx
, void *state
)
279 struct r600_context
*rctx
= (struct r600_context
*)ctx
;
280 struct r600_pipe_compute
*shader
= state
;
282 COMPUTE_DBG(rctx
->screen
, "*** evergreen_delete_compute_state\n");
287 radeon_shader_binary_clean(&shader
->binary
);
288 r600_destroy_shader(&shader
->bc
);
290 /* TODO destroy shader->code_bo, shader->const_bo
291 * we'll need something like r600_buffer_free */
295 static void evergreen_bind_compute_state(struct pipe_context
*ctx
, void *state
)
297 struct r600_context
*rctx
= (struct r600_context
*)ctx
;
299 COMPUTE_DBG(rctx
->screen
, "*** evergreen_bind_compute_state\n");
301 rctx
->cs_shader_state
.shader
= (struct r600_pipe_compute
*)state
;
304 /* The kernel parameters are stored a vtx buffer (ID=0), besides the explicit
305 * kernel parameters there are implicit parameters that need to be stored
306 * in the vertex buffer as well. Here is how these parameters are organized in
309 * DWORDS 0-2: Number of work groups in each dimension (x,y,z)
310 * DWORDS 3-5: Number of global work items in each dimension (x,y,z)
311 * DWORDS 6-8: Number of work items within each work group in each dimension
313 * DWORDS 9+ : Kernel parameters
315 static void evergreen_compute_upload_input(struct pipe_context
*ctx
,
316 const struct pipe_grid_info
*info
)
318 struct r600_context
*rctx
= (struct r600_context
*)ctx
;
319 struct r600_pipe_compute
*shader
= rctx
->cs_shader_state
.shader
;
321 /* We need to reserve 9 dwords (36 bytes) for implicit kernel
324 unsigned input_size
= shader
->input_size
+ 36;
325 uint32_t *num_work_groups_start
;
326 uint32_t *global_size_start
;
327 uint32_t *local_size_start
;
328 uint32_t *kernel_parameters_start
;
330 struct pipe_transfer
*transfer
= NULL
;
332 if (shader
->input_size
== 0) {
336 if (!shader
->kernel_param
) {
337 /* Add space for the grid dimensions */
338 shader
->kernel_param
= (struct r600_resource
*)
339 pipe_buffer_create(ctx
->screen
, PIPE_BIND_CUSTOM
,
340 PIPE_USAGE_IMMUTABLE
, input_size
);
343 u_box_1d(0, input_size
, &box
);
344 num_work_groups_start
= ctx
->transfer_map(ctx
,
345 (struct pipe_resource
*)shader
->kernel_param
,
346 0, PIPE_TRANSFER_WRITE
| PIPE_TRANSFER_DISCARD_RANGE
,
348 global_size_start
= num_work_groups_start
+ (3 * (sizeof(uint
) /4));
349 local_size_start
= global_size_start
+ (3 * (sizeof(uint
)) / 4);
350 kernel_parameters_start
= local_size_start
+ (3 * (sizeof(uint
)) / 4);
352 /* Copy the work group size */
353 memcpy(num_work_groups_start
, info
->grid
, 3 * sizeof(uint
));
355 /* Copy the global size */
356 for (i
= 0; i
< 3; i
++) {
357 global_size_start
[i
] = info
->grid
[i
] * info
->block
[i
];
360 /* Copy the local dimensions */
361 memcpy(local_size_start
, info
->block
, 3 * sizeof(uint
));
363 /* Copy the kernel inputs */
364 memcpy(kernel_parameters_start
, info
->input
, shader
->input_size
);
366 for (i
= 0; i
< (input_size
/ 4); i
++) {
367 COMPUTE_DBG(rctx
->screen
, "input %i : %u\n", i
,
368 ((unsigned*)num_work_groups_start
)[i
]);
371 ctx
->transfer_unmap(ctx
, transfer
);
373 /* ID=0 is reserved for the parameters */
374 evergreen_cs_set_constant_buffer(rctx
, 0, 0, input_size
,
375 (struct pipe_resource
*)shader
->kernel_param
);
378 static void evergreen_emit_dispatch(struct r600_context
*rctx
,
379 const struct pipe_grid_info
*info
)
382 struct radeon_winsys_cs
*cs
= rctx
->b
.gfx
.cs
;
383 struct r600_pipe_compute
*shader
= rctx
->cs_shader_state
.shader
;
385 unsigned num_pipes
= rctx
->screen
->b
.info
.r600_max_quad_pipes
;
386 unsigned wave_divisor
= (16 * num_pipes
);
389 unsigned lds_size
= shader
->local_size
/ 4 +
393 /* Calculate group_size/grid_size */
394 for (i
= 0; i
< 3; i
++) {
395 group_size
*= info
->block
[i
];
398 for (i
= 0; i
< 3; i
++) {
399 grid_size
*= info
->grid
[i
];
402 /* num_waves = ceil((tg_size.x * tg_size.y, tg_size.z) / (16 * num_pipes)) */
403 num_waves
= (info
->block
[0] * info
->block
[1] * info
->block
[2] +
404 wave_divisor
- 1) / wave_divisor
;
406 COMPUTE_DBG(rctx
->screen
, "Using %u pipes, "
407 "%u wavefronts per thread block, "
408 "allocating %u dwords lds.\n",
409 num_pipes
, num_waves
, lds_size
);
411 radeon_set_config_reg(cs
, R_008970_VGT_NUM_INDICES
, group_size
);
413 radeon_set_config_reg_seq(cs
, R_00899C_VGT_COMPUTE_START_X
, 3);
414 radeon_emit(cs
, 0); /* R_00899C_VGT_COMPUTE_START_X */
415 radeon_emit(cs
, 0); /* R_0089A0_VGT_COMPUTE_START_Y */
416 radeon_emit(cs
, 0); /* R_0089A4_VGT_COMPUTE_START_Z */
418 radeon_set_config_reg(cs
, R_0089AC_VGT_COMPUTE_THREAD_GROUP_SIZE
,
421 radeon_compute_set_context_reg_seq(cs
, R_0286EC_SPI_COMPUTE_NUM_THREAD_X
, 3);
422 radeon_emit(cs
, info
->block
[0]); /* R_0286EC_SPI_COMPUTE_NUM_THREAD_X */
423 radeon_emit(cs
, info
->block
[1]); /* R_0286F0_SPI_COMPUTE_NUM_THREAD_Y */
424 radeon_emit(cs
, info
->block
[2]); /* R_0286F4_SPI_COMPUTE_NUM_THREAD_Z */
426 if (rctx
->b
.chip_class
< CAYMAN
) {
427 assert(lds_size
<= 8192);
429 /* Cayman appears to have a slightly smaller limit, see the
430 * value of CM_R_0286FC_SPI_LDS_MGMT.NUM_LS_LDS */
431 assert(lds_size
<= 8160);
434 radeon_compute_set_context_reg(cs
, R_0288E8_SQ_LDS_ALLOC
,
435 lds_size
| (num_waves
<< 14));
437 /* Dispatch packet */
438 radeon_emit(cs
, PKT3C(PKT3_DISPATCH_DIRECT
, 3, 0));
439 radeon_emit(cs
, info
->grid
[0]);
440 radeon_emit(cs
, info
->grid
[1]);
441 radeon_emit(cs
, info
->grid
[2]);
442 /* VGT_DISPATCH_INITIATOR = COMPUTE_SHADER_EN */
446 static void compute_emit_cs(struct r600_context
*rctx
,
447 const struct pipe_grid_info
*info
)
449 struct radeon_winsys_cs
*cs
= rctx
->b
.gfx
.cs
;
452 /* make sure that the gfx ring is only one active */
453 if (rctx
->b
.dma
.cs
&& rctx
->b
.dma
.cs
->cdw
) {
454 rctx
->b
.dma
.flush(rctx
, RADEON_FLUSH_ASYNC
, NULL
);
457 /* Initialize all the compute-related registers.
459 * See evergreen_init_atom_start_compute_cs() in this file for the list
460 * of registers initialized by the start_compute_cs_cmd atom.
462 r600_emit_command_buffer(cs
, &rctx
->start_compute_cs_cmd
);
464 /* emit config state */
465 if (rctx
->b
.chip_class
== EVERGREEN
)
466 r600_emit_atom(rctx
, &rctx
->config_state
.atom
);
468 rctx
->b
.flags
|= R600_CONTEXT_WAIT_3D_IDLE
| R600_CONTEXT_FLUSH_AND_INV
;
469 r600_flush_emit(rctx
);
471 /* Emit colorbuffers. */
472 /* XXX support more than 8 colorbuffers (the offsets are not a multiple of 0x3C for CB8-11) */
473 for (i
= 0; i
< 8 && i
< rctx
->framebuffer
.state
.nr_cbufs
; i
++) {
474 struct r600_surface
*cb
= (struct r600_surface
*)rctx
->framebuffer
.state
.cbufs
[i
];
475 unsigned reloc
= radeon_add_to_buffer_list(&rctx
->b
, &rctx
->b
.gfx
,
476 (struct r600_resource
*)cb
->base
.texture
,
477 RADEON_USAGE_READWRITE
,
478 RADEON_PRIO_SHADER_RW_BUFFER
);
480 radeon_compute_set_context_reg_seq(cs
, R_028C60_CB_COLOR0_BASE
+ i
* 0x3C, 7);
481 radeon_emit(cs
, cb
->cb_color_base
); /* R_028C60_CB_COLOR0_BASE */
482 radeon_emit(cs
, cb
->cb_color_pitch
); /* R_028C64_CB_COLOR0_PITCH */
483 radeon_emit(cs
, cb
->cb_color_slice
); /* R_028C68_CB_COLOR0_SLICE */
484 radeon_emit(cs
, cb
->cb_color_view
); /* R_028C6C_CB_COLOR0_VIEW */
485 radeon_emit(cs
, cb
->cb_color_info
); /* R_028C70_CB_COLOR0_INFO */
486 radeon_emit(cs
, cb
->cb_color_attrib
); /* R_028C74_CB_COLOR0_ATTRIB */
487 radeon_emit(cs
, cb
->cb_color_dim
); /* R_028C78_CB_COLOR0_DIM */
489 radeon_emit(cs
, PKT3(PKT3_NOP
, 0, 0)); /* R_028C60_CB_COLOR0_BASE */
490 radeon_emit(cs
, reloc
);
492 radeon_emit(cs
, PKT3(PKT3_NOP
, 0, 0)); /* R_028C74_CB_COLOR0_ATTRIB */
493 radeon_emit(cs
, reloc
);
496 radeon_compute_set_context_reg(cs
, R_028C70_CB_COLOR0_INFO
+ i
* 0x3C,
497 S_028C70_FORMAT(V_028C70_COLOR_INVALID
));
499 radeon_compute_set_context_reg(cs
, R_028E50_CB_COLOR8_INFO
+ (i
- 8) * 0x1C,
500 S_028C70_FORMAT(V_028C70_COLOR_INVALID
));
502 /* Set CB_TARGET_MASK XXX: Use cb_misc_state */
503 radeon_compute_set_context_reg(cs
, R_028238_CB_TARGET_MASK
,
504 rctx
->compute_cb_target_mask
);
507 /* Emit vertex buffer state */
508 rctx
->cs_vertex_buffer_state
.atom
.num_dw
= 12 * util_bitcount(rctx
->cs_vertex_buffer_state
.dirty_mask
);
509 r600_emit_atom(rctx
, &rctx
->cs_vertex_buffer_state
.atom
);
511 /* Emit constant buffer state */
512 r600_emit_atom(rctx
, &rctx
->constbuf_state
[PIPE_SHADER_COMPUTE
].atom
);
514 /* Emit sampler state */
515 r600_emit_atom(rctx
, &rctx
->samplers
[PIPE_SHADER_COMPUTE
].states
.atom
);
517 /* Emit sampler view (texture resource) state */
518 r600_emit_atom(rctx
, &rctx
->samplers
[PIPE_SHADER_COMPUTE
].views
.atom
);
520 /* Emit compute shader state */
521 r600_emit_atom(rctx
, &rctx
->cs_shader_state
.atom
);
523 /* Emit dispatch state and dispatch packet */
524 evergreen_emit_dispatch(rctx
, info
);
526 /* XXX evergreen_flush_emit() hardcodes the CP_COHER_SIZE to 0xffffffff
528 rctx
->b
.flags
|= R600_CONTEXT_INV_CONST_CACHE
|
529 R600_CONTEXT_INV_VERTEX_CACHE
|
530 R600_CONTEXT_INV_TEX_CACHE
;
531 r600_flush_emit(rctx
);
534 if (rctx
->b
.chip_class
>= CAYMAN
) {
535 cs
->buf
[cs
->cdw
++] = PKT3(PKT3_EVENT_WRITE
, 0, 0);
536 cs
->buf
[cs
->cdw
++] = EVENT_TYPE(EVENT_TYPE_CS_PARTIAL_FLUSH
) | EVENT_INDEX(4);
537 /* DEALLOC_STATE prevents the GPU from hanging when a
538 * SURFACE_SYNC packet is emitted some time after a DISPATCH_DIRECT
539 * with any of the CB*_DEST_BASE_ENA or DB_DEST_BASE_ENA bits set.
541 cs
->buf
[cs
->cdw
++] = PKT3C(PKT3_DEALLOC_STATE
, 0, 0);
542 cs
->buf
[cs
->cdw
++] = 0;
546 COMPUTE_DBG(rctx
->screen
, "cdw: %i\n", cs
->cdw
);
547 for (i
= 0; i
< cs
->cdw
; i
++) {
548 COMPUTE_DBG(rctx
->screen
, "%4i : 0x%08X\n", i
, cs
->buf
[i
]);
556 * Emit function for r600_cs_shader_state atom
558 void evergreen_emit_cs_shader(struct r600_context
*rctx
,
559 struct r600_atom
*atom
)
561 struct r600_cs_shader_state
*state
=
562 (struct r600_cs_shader_state
*)atom
;
563 struct r600_pipe_compute
*shader
= state
->shader
;
564 struct radeon_winsys_cs
*cs
= rctx
->b
.gfx
.cs
;
566 struct r600_resource
*code_bo
;
567 unsigned ngpr
, nstack
;
569 code_bo
= shader
->code_bo
;
570 va
= shader
->code_bo
->gpu_address
+ state
->pc
;
571 ngpr
= shader
->bc
.ngpr
;
572 nstack
= shader
->bc
.nstack
;
574 radeon_compute_set_context_reg_seq(cs
, R_0288D0_SQ_PGM_START_LS
, 3);
575 radeon_emit(cs
, va
>> 8); /* R_0288D0_SQ_PGM_START_LS */
576 radeon_emit(cs
, /* R_0288D4_SQ_PGM_RESOURCES_LS */
577 S_0288D4_NUM_GPRS(ngpr
)
578 | S_0288D4_STACK_SIZE(nstack
));
579 radeon_emit(cs
, 0); /* R_0288D8_SQ_PGM_RESOURCES_LS_2 */
581 radeon_emit(cs
, PKT3C(PKT3_NOP
, 0, 0));
582 radeon_emit(cs
, radeon_add_to_buffer_list(&rctx
->b
, &rctx
->b
.gfx
,
583 code_bo
, RADEON_USAGE_READ
,
584 RADEON_PRIO_USER_SHADER
));
587 static void evergreen_launch_grid(struct pipe_context
*ctx
,
588 const struct pipe_grid_info
*info
)
590 struct r600_context
*rctx
= (struct r600_context
*)ctx
;
592 struct r600_pipe_compute
*shader
= rctx
->cs_shader_state
.shader
;
595 rctx
->cs_shader_state
.pc
= info
->pc
;
596 /* Get the config information for this kernel. */
597 r600_shader_binary_read_config(&shader
->binary
, &shader
->bc
,
598 info
->pc
, &use_kill
);
601 COMPUTE_DBG(rctx
->screen
, "*** evergreen_launch_grid: pc = %u\n", info
->pc
);
604 evergreen_compute_upload_input(ctx
, info
);
605 compute_emit_cs(rctx
, info
);
608 static void evergreen_set_compute_resources(struct pipe_context
*ctx
,
609 unsigned start
, unsigned count
,
610 struct pipe_surface
**surfaces
)
612 struct r600_context
*rctx
= (struct r600_context
*)ctx
;
613 struct r600_surface
**resources
= (struct r600_surface
**)surfaces
;
615 COMPUTE_DBG(rctx
->screen
, "*** evergreen_set_compute_resources: start = %u count = %u\n",
618 for (unsigned i
= 0; i
< count
; i
++) {
619 /* The First two vertex buffers are reserved for parameters and
621 unsigned vtx_id
= 2 + i
;
623 struct r600_resource_global
*buffer
=
624 (struct r600_resource_global
*)
625 resources
[i
]->base
.texture
;
626 if (resources
[i
]->base
.writable
) {
629 evergreen_set_rat(rctx
->cs_shader_state
.shader
, i
+1,
630 (struct r600_resource
*)resources
[i
]->base
.texture
,
631 buffer
->chunk
->start_in_dw
*4,
632 resources
[i
]->base
.texture
->width0
);
635 evergreen_cs_set_vertex_buffer(rctx
, vtx_id
,
636 buffer
->chunk
->start_in_dw
* 4,
637 resources
[i
]->base
.texture
);
642 static void evergreen_set_global_binding(struct pipe_context
*ctx
,
643 unsigned first
, unsigned n
,
644 struct pipe_resource
**resources
,
647 struct r600_context
*rctx
= (struct r600_context
*)ctx
;
648 struct compute_memory_pool
*pool
= rctx
->screen
->global_pool
;
649 struct r600_resource_global
**buffers
=
650 (struct r600_resource_global
**)resources
;
653 COMPUTE_DBG(rctx
->screen
, "*** evergreen_set_global_binding first = %u n = %u\n",
661 /* We mark these items for promotion to the pool if they
662 * aren't already there */
663 for (i
= first
; i
< first
+ n
; i
++) {
664 struct compute_memory_item
*item
= buffers
[i
]->chunk
;
666 if (!is_item_in_pool(item
))
667 buffers
[i
]->chunk
->status
|= ITEM_FOR_PROMOTING
;
670 if (compute_memory_finalize_pending(pool
, ctx
) == -1) {
675 for (i
= first
; i
< first
+ n
; i
++)
677 uint32_t buffer_offset
;
679 assert(resources
[i
]->target
== PIPE_BUFFER
);
680 assert(resources
[i
]->bind
& PIPE_BIND_GLOBAL
);
682 buffer_offset
= util_le32_to_cpu(*(handles
[i
]));
683 handle
= buffer_offset
+ buffers
[i
]->chunk
->start_in_dw
* 4;
685 *(handles
[i
]) = util_cpu_to_le32(handle
);
688 evergreen_set_rat(rctx
->cs_shader_state
.shader
, 0, pool
->bo
, 0, pool
->size_in_dw
* 4);
689 evergreen_cs_set_vertex_buffer(rctx
, 1, 0,
690 (struct pipe_resource
*)pool
->bo
);
694 * This function initializes all the compute specific registers that need to
695 * be initialized for each compute command stream. Registers that are common
696 * to both compute and 3D will be initialized at the beginning of each compute
697 * command stream by the start_cs_cmd atom. However, since the SET_CONTEXT_REG
698 * packet requires that the shader type bit be set, we must initialize all
699 * context registers needed for compute in this function. The registers
700 * initialized by the start_cs_cmd atom can be found in evergreen_state.c in the
701 * functions evergreen_init_atom_start_cs or cayman_init_atom_start_cs depending
704 void evergreen_init_atom_start_compute_cs(struct r600_context
*rctx
)
706 struct r600_command_buffer
*cb
= &rctx
->start_compute_cs_cmd
;
708 int num_stack_entries
;
710 /* since all required registers are initialized in the
711 * start_compute_cs_cmd atom, we can EMIT_EARLY here.
713 r600_init_command_buffer(cb
, 256);
714 cb
->pkt_flags
= RADEON_CP_PACKET3_COMPUTE_MODE
;
716 /* This must be first. */
717 r600_store_value(cb
, PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
718 r600_store_value(cb
, 0x80000000);
719 r600_store_value(cb
, 0x80000000);
721 /* We're setting config registers here. */
722 r600_store_value(cb
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
723 r600_store_value(cb
, EVENT_TYPE(EVENT_TYPE_CS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
725 switch (rctx
->b
.family
) {
729 num_stack_entries
= 256;
733 num_stack_entries
= 256;
737 num_stack_entries
= 512;
742 num_stack_entries
= 512;
746 num_stack_entries
= 256;
750 num_stack_entries
= 256;
754 num_stack_entries
= 512;
758 num_stack_entries
= 512;
762 num_stack_entries
= 256;
766 num_stack_entries
= 256;
770 /* Config Registers */
771 if (rctx
->b
.chip_class
< CAYMAN
)
772 evergreen_init_common_regs(rctx
, cb
, rctx
->b
.chip_class
, rctx
->b
.family
,
773 rctx
->screen
->b
.info
.drm_minor
);
775 cayman_init_common_regs(cb
, rctx
->b
.chip_class
, rctx
->b
.family
,
776 rctx
->screen
->b
.info
.drm_minor
);
778 /* The primitive type always needs to be POINTLIST for compute. */
779 r600_store_config_reg(cb
, R_008958_VGT_PRIMITIVE_TYPE
,
780 V_008958_DI_PT_POINTLIST
);
782 if (rctx
->b
.chip_class
< CAYMAN
) {
784 /* These registers control which simds can be used by each stage.
785 * The default for these registers is 0xffffffff, which means
786 * all simds are available for each stage. It's possible we may
787 * want to play around with these in the future, but for now
788 * the default value is fine.
790 * R_008E20_SQ_STATIC_THREAD_MGMT1
791 * R_008E24_SQ_STATIC_THREAD_MGMT2
792 * R_008E28_SQ_STATIC_THREAD_MGMT3
795 /* XXX: We may need to adjust the thread and stack resource
796 * values for 3D/compute interop */
798 r600_store_config_reg_seq(cb
, R_008C18_SQ_THREAD_RESOURCE_MGMT_1
, 5);
800 /* R_008C18_SQ_THREAD_RESOURCE_MGMT_1
801 * Set the number of threads used by the PS/VS/GS/ES stage to
804 r600_store_value(cb
, 0);
806 /* R_008C1C_SQ_THREAD_RESOURCE_MGMT_2
807 * Set the number of threads used by the CS (aka LS) stage to
808 * the maximum number of threads and set the number of threads
809 * for the HS stage to 0. */
810 r600_store_value(cb
, S_008C1C_NUM_LS_THREADS(num_threads
));
812 /* R_008C20_SQ_STACK_RESOURCE_MGMT_1
813 * Set the Control Flow stack entries to 0 for PS/VS stages */
814 r600_store_value(cb
, 0);
816 /* R_008C24_SQ_STACK_RESOURCE_MGMT_2
817 * Set the Control Flow stack entries to 0 for GS/ES stages */
818 r600_store_value(cb
, 0);
820 /* R_008C28_SQ_STACK_RESOURCE_MGMT_3
821 * Set the Contol Flow stack entries to 0 for the HS stage, and
822 * set it to the maximum value for the CS (aka LS) stage. */
824 S_008C28_NUM_LS_STACK_ENTRIES(num_stack_entries
));
826 /* Give the compute shader all the available LDS space.
827 * NOTE: This only sets the maximum number of dwords that a compute
828 * shader can allocate. When a shader is executed, we still need to
829 * allocate the appropriate amount of LDS dwords using the
830 * CM_R_0288E8_SQ_LDS_ALLOC register.
832 if (rctx
->b
.chip_class
< CAYMAN
) {
833 r600_store_config_reg(cb
, R_008E2C_SQ_LDS_RESOURCE_MGMT
,
834 S_008E2C_NUM_PS_LDS(0x0000) | S_008E2C_NUM_LS_LDS(8192));
836 r600_store_context_reg(cb
, CM_R_0286FC_SPI_LDS_MGMT
,
837 S_0286FC_NUM_PS_LDS(0) |
838 S_0286FC_NUM_LS_LDS(255)); /* 255 * 32 = 8160 dwords */
841 /* Context Registers */
843 if (rctx
->b
.chip_class
< CAYMAN
) {
844 /* workaround for hw issues with dyn gpr - must set all limits
845 * to 240 instead of 0, 0x1e == 240 / 8
847 r600_store_context_reg(cb
, R_028838_SQ_DYN_GPR_RESOURCE_LIMIT_1
,
848 S_028838_PS_GPRS(0x1e) |
849 S_028838_VS_GPRS(0x1e) |
850 S_028838_GS_GPRS(0x1e) |
851 S_028838_ES_GPRS(0x1e) |
852 S_028838_HS_GPRS(0x1e) |
853 S_028838_LS_GPRS(0x1e));
856 /* XXX: Investigate setting bit 15, which is FAST_COMPUTE_MODE */
857 r600_store_context_reg(cb
, R_028A40_VGT_GS_MODE
,
858 S_028A40_COMPUTE_MODE(1) | S_028A40_PARTIAL_THD_AT_EOI(1));
860 r600_store_context_reg(cb
, R_028B54_VGT_SHADER_STAGES_EN
, 2/*CS_ON*/);
862 r600_store_context_reg(cb
, R_0286E8_SPI_COMPUTE_INPUT_CNTL
,
863 S_0286E8_TID_IN_GROUP_ENA
865 | S_0286E8_DISABLE_INDEX_PACK
)
868 /* The LOOP_CONST registers are an optimizations for loops that allows
869 * you to store the initial counter, increment value, and maximum
870 * counter value in a register so that hardware can calculate the
871 * correct number of iterations for the loop, so that you don't need
872 * to have the loop counter in your shader code. We don't currently use
873 * this optimization, so we must keep track of the counter in the
874 * shader and use a break instruction to exit loops. However, the
875 * hardware will still uses this register to determine when to exit a
876 * loop, so we need to initialize the counter to 0, set the increment
877 * value to 1 and the maximum counter value to the 4095 (0xfff) which
878 * is the maximum value allowed. This gives us a maximum of 4096
879 * iterations for our loops, but hopefully our break instruction will
880 * execute before some time before the 4096th iteration.
882 eg_store_loop_const(cb
, R_03A200_SQ_LOOP_CONST_0
+ (160 * 4), 0x1000FFF);
885 void evergreen_init_compute_state_functions(struct r600_context
*rctx
)
887 rctx
->b
.b
.create_compute_state
= evergreen_create_compute_state
;
888 rctx
->b
.b
.delete_compute_state
= evergreen_delete_compute_state
;
889 rctx
->b
.b
.bind_compute_state
= evergreen_bind_compute_state
;
890 // rctx->context.create_sampler_view = evergreen_compute_create_sampler_view;
891 rctx
->b
.b
.set_compute_resources
= evergreen_set_compute_resources
;
892 rctx
->b
.b
.set_global_binding
= evergreen_set_global_binding
;
893 rctx
->b
.b
.launch_grid
= evergreen_launch_grid
;
897 static void *r600_compute_global_transfer_map(struct pipe_context
*ctx
,
898 struct pipe_resource
*resource
,
901 const struct pipe_box
*box
,
902 struct pipe_transfer
**ptransfer
)
904 struct r600_context
*rctx
= (struct r600_context
*)ctx
;
905 struct compute_memory_pool
*pool
= rctx
->screen
->global_pool
;
906 struct r600_resource_global
* buffer
=
907 (struct r600_resource_global
*)resource
;
909 struct compute_memory_item
*item
= buffer
->chunk
;
910 struct pipe_resource
*dst
= NULL
;
911 unsigned offset
= box
->x
;
913 if (is_item_in_pool(item
)) {
914 compute_memory_demote_item(pool
, item
, ctx
);
917 if (item
->real_buffer
== NULL
) {
919 r600_compute_buffer_alloc_vram(pool
->screen
, item
->size_in_dw
* 4);
923 dst
= (struct pipe_resource
*)item
->real_buffer
;
925 if (usage
& PIPE_TRANSFER_READ
)
926 buffer
->chunk
->status
|= ITEM_MAPPED_FOR_READING
;
928 COMPUTE_DBG(rctx
->screen
, "* r600_compute_global_transfer_map()\n"
929 "level = %u, usage = %u, box(x = %u, y = %u, z = %u "
930 "width = %u, height = %u, depth = %u)\n", level
, usage
,
931 box
->x
, box
->y
, box
->z
, box
->width
, box
->height
,
933 COMPUTE_DBG(rctx
->screen
, "Buffer id = %"PRIi64
" offset = "
934 "%u (box.x)\n", item
->id
, box
->x
);
937 assert(resource
->target
== PIPE_BUFFER
);
938 assert(resource
->bind
& PIPE_BIND_GLOBAL
);
943 ///TODO: do it better, mapping is not possible if the pool is too big
944 return pipe_buffer_map_range(ctx
, dst
,
945 offset
, box
->width
, usage
, ptransfer
);
948 static void r600_compute_global_transfer_unmap(struct pipe_context
*ctx
,
949 struct pipe_transfer
*transfer
)
951 /* struct r600_resource_global are not real resources, they just map
952 * to an offset within the compute memory pool. The function
953 * r600_compute_global_transfer_map() maps the memory pool
954 * resource rather than the struct r600_resource_global passed to
955 * it as an argument and then initalizes ptransfer->resource with
956 * the memory pool resource (via pipe_buffer_map_range).
957 * When transfer_unmap is called it uses the memory pool's
958 * vtable which calls r600_buffer_transfer_map() rather than
961 assert (!"This function should not be called");
964 static void r600_compute_global_transfer_flush_region(struct pipe_context
*ctx
,
965 struct pipe_transfer
*transfer
,
966 const struct pipe_box
*box
)
971 static void r600_compute_global_transfer_inline_write(struct pipe_context
*pipe
,
972 struct pipe_resource
*resource
,
975 const struct pipe_box
*box
,
978 unsigned layer_stride
)
983 static void r600_compute_global_buffer_destroy(struct pipe_screen
*screen
,
984 struct pipe_resource
*res
)
986 struct r600_resource_global
* buffer
= NULL
;
987 struct r600_screen
* rscreen
= NULL
;
989 assert(res
->target
== PIPE_BUFFER
);
990 assert(res
->bind
& PIPE_BIND_GLOBAL
);
992 buffer
= (struct r600_resource_global
*)res
;
993 rscreen
= (struct r600_screen
*)screen
;
995 compute_memory_free(rscreen
->global_pool
, buffer
->chunk
->id
);
997 buffer
->chunk
= NULL
;
1001 static const struct u_resource_vtbl r600_global_buffer_vtbl
=
1003 u_default_resource_get_handle
, /* get_handle */
1004 r600_compute_global_buffer_destroy
, /* resource_destroy */
1005 r600_compute_global_transfer_map
, /* transfer_map */
1006 r600_compute_global_transfer_flush_region
,/* transfer_flush_region */
1007 r600_compute_global_transfer_unmap
, /* transfer_unmap */
1008 r600_compute_global_transfer_inline_write
/* transfer_inline_write */
1011 struct pipe_resource
*r600_compute_global_buffer_create(struct pipe_screen
*screen
,
1012 const struct pipe_resource
*templ
)
1014 struct r600_resource_global
* result
= NULL
;
1015 struct r600_screen
* rscreen
= NULL
;
1018 assert(templ
->target
== PIPE_BUFFER
);
1019 assert(templ
->bind
& PIPE_BIND_GLOBAL
);
1020 assert(templ
->array_size
== 1 || templ
->array_size
== 0);
1021 assert(templ
->depth0
== 1 || templ
->depth0
== 0);
1022 assert(templ
->height0
== 1 || templ
->height0
== 0);
1024 result
= (struct r600_resource_global
*)
1025 CALLOC(sizeof(struct r600_resource_global
), 1);
1026 rscreen
= (struct r600_screen
*)screen
;
1028 COMPUTE_DBG(rscreen
, "*** r600_compute_global_buffer_create\n");
1029 COMPUTE_DBG(rscreen
, "width = %u array_size = %u\n", templ
->width0
,
1032 result
->base
.b
.vtbl
= &r600_global_buffer_vtbl
;
1033 result
->base
.b
.b
= *templ
;
1034 result
->base
.b
.b
.screen
= screen
;
1035 pipe_reference_init(&result
->base
.b
.b
.reference
, 1);
1037 size_in_dw
= (templ
->width0
+3) / 4;
1039 result
->chunk
= compute_memory_alloc(rscreen
->global_pool
, size_in_dw
);
1041 if (result
->chunk
== NULL
)
1047 return &result
->base
.b
.b
;