1 /**************************************************************************
3 * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
30 * SPU command processing code
37 #include "pipe/p_defines.h"
39 #include "spu_command.h"
41 #include "spu_render.h"
42 #include "spu_per_fragment_op.h"
43 #include "spu_texture.h"
45 #include "spu_vertex_shader.h"
46 #include "spu_dcache.h"
47 #include "cell/common.h"
50 struct spu_vs_context draw
;
54 * Buffers containing dynamically generated SPU code:
56 static unsigned char attribute_fetch_code_buffer
[136 * PIPE_MAX_ATTRIBS
]
62 align(int value
, int alignment
)
64 return (value
+ alignment
- 1) & ~(alignment
- 1);
70 * Tell the PPU that this SPU has finished copying a buffer to
71 * local store and that it may be reused by the PPU.
72 * This is done by writting a 16-byte batch-buffer-status block back into
73 * main memory (in cell_context->buffer_status[]).
76 release_buffer(uint buffer
)
78 /* Evidently, using less than a 16-byte status doesn't work reliably */
79 static const uint status
[4] ALIGN16_ATTRIB
80 = {CELL_BUFFER_STATUS_FREE
, 0, 0, 0};
82 const uint index
= 4 * (spu
.init
.id
* CELL_NUM_BUFFERS
+ buffer
);
83 uint
*dst
= spu
.init
.buffer_status
+ index
;
85 ASSERT(buffer
< CELL_NUM_BUFFERS
);
87 mfc_put((void *) &status
, /* src in local memory */
88 (unsigned int) dst
, /* dst in main memory */
89 sizeof(status
), /* size */
90 TAG_MISC
, /* tag is unimportant */
97 cmd_clear_surface(const struct cell_command_clear_surface
*clear
)
99 D_PRINTF(CELL_DEBUG_CMD
, "CLEAR SURF %u to 0x%08x\n", clear
->surface
, clear
->value
);
101 if (clear
->surface
== 0) {
102 spu
.fb
.color_clear_value
= clear
->value
;
103 if (spu
.init
.debug_flags
& CELL_DEBUG_CHECKER
) {
104 uint x
= (spu
.init
.id
<< 4) | (spu
.init
.id
<< 12) |
105 (spu
.init
.id
<< 20) | (spu
.init
.id
<< 28);
106 spu
.fb
.color_clear_value
^= x
;
110 spu
.fb
.depth_clear_value
= clear
->value
;
116 /* Simply set all tiles' status to CLEAR.
117 * When we actually begin rendering into a tile, we'll initialize it to
118 * the clear value. If any tiles go untouched during the frame,
119 * really_clear_tiles() will set them to the clear value.
121 if (clear
->surface
== 0) {
122 memset(spu
.ctile_status
, TILE_STATUS_CLEAR
, sizeof(spu
.ctile_status
));
125 memset(spu
.ztile_status
, TILE_STATUS_CLEAR
, sizeof(spu
.ztile_status
));
131 * This path clears the whole framebuffer to the clear color right now.
135 printf("SPU: %s num=%d w=%d h=%d\n",
136 __FUNCTION__, num_tiles, spu.fb.width_tiles, spu.fb.height_tiles);
139 /* init a single tile to the clear value */
140 if (clear
->surface
== 0) {
141 clear_c_tile(&spu
.ctile
);
144 clear_z_tile(&spu
.ztile
);
147 /* walk over my tiles, writing the 'clear' tile's data */
149 const uint num_tiles
= spu
.fb
.width_tiles
* spu
.fb
.height_tiles
;
151 for (i
= spu
.init
.id
; i
< num_tiles
; i
+= spu
.init
.num_spus
) {
152 uint tx
= i
% spu
.fb
.width_tiles
;
153 uint ty
= i
/ spu
.fb
.width_tiles
;
154 if (clear
->surface
== 0)
155 put_tile(tx
, ty
, &spu
.ctile
, TAG_SURFACE_CLEAR
, 0);
157 put_tile(tx
, ty
, &spu
.ztile
, TAG_SURFACE_CLEAR
, 1);
161 if (spu
.init
.debug_flags
& CELL_DEBUG_SYNC
) {
162 wait_on_mask(1 << TAG_SURFACE_CLEAR
);
165 #endif /* CLEAR_OPT */
167 D_PRINTF(CELL_DEBUG_CMD
, "CLEAR SURF done\n");
172 cmd_release_verts(const struct cell_command_release_verts
*release
)
174 D_PRINTF(CELL_DEBUG_CMD
, "RELEASE VERTS %u\n", release
->vertex_buf
);
175 ASSERT(release
->vertex_buf
!= ~0U);
176 release_buffer(release
->vertex_buf
);
181 * Process a CELL_CMD_STATE_FRAGMENT_OPS command.
182 * This involves installing new fragment ops SPU code.
183 * If this function is never called, we'll use a regular C fallback function
184 * for fragment processing.
187 cmd_state_fragment_ops(const struct cell_command_fragment_ops
*fops
)
189 static int warned
= 0;
191 D_PRINTF(CELL_DEBUG_CMD
, "CMD_STATE_FRAGMENT_OPS\n");
192 /* Copy SPU code from batch buffer to spu buffer */
193 memcpy(spu
.fragment_ops_code
, fops
->code
, SPU_MAX_FRAGMENT_OPS_INSTS
* 4);
194 /* Copy state info (for fallback case only) */
195 memcpy(&spu
.depth_stencil_alpha
, &fops
->dsa
, sizeof(fops
->dsa
));
196 memcpy(&spu
.blend
, &fops
->blend
, sizeof(fops
->blend
));
197 memcpy(&spu
.blend_color
, &fops
->blend_color
, sizeof(fops
->blend_color
));
199 /* Parity twist! For now, always use the fallback code by default,
200 * only switching to codegen when specifically requested. This
201 * allows us to develop freely without risking taking down the
204 * Later, the parity of this check will be reversed, so that
205 * codegen is *always* used, unless we specifically indicate that
208 * Eventually, the option will be removed completely, because in
209 * final code we'll always use codegen and won't even provide the
210 * raw state records that the fallback code requires.
212 if ((spu
.init
.debug_flags
& CELL_DEBUG_FRAGMENT_OP_FALLBACK
) == 0) {
213 spu
.fragment_ops
= (spu_fragment_ops_func
) spu
.fragment_ops_code
;
216 /* otherwise, the default fallback code remains in place */
218 fprintf(stderr
, "Cell Warning: using fallback per-fragment code\n");
223 spu
.read_depth
= spu
.depth_stencil_alpha
.depth
.enabled
;
224 spu
.read_stencil
= spu
.depth_stencil_alpha
.stencil
[0].enabled
;
229 cmd_state_fragment_program(const struct cell_command_fragment_program
*fp
)
231 D_PRINTF(CELL_DEBUG_CMD
, "CMD_STATE_FRAGMENT_PROGRAM\n");
232 /* Copy SPU code from batch buffer to spu buffer */
233 memcpy(spu
.fragment_program_code
, fp
->code
,
234 SPU_MAX_FRAGMENT_PROGRAM_INSTS
* 4);
236 /* Point function pointer at new code */
237 spu
.fragment_program
= (spu_fragment_program_func
)spu
.fragment_program_code
;
243 cmd_state_fs_constants(const uint64_t *buffer
, uint pos
)
245 const uint num_const
= buffer
[pos
+ 1];
246 const float *constants
= (const float *) &buffer
[pos
+ 2];
249 D_PRINTF(CELL_DEBUG_CMD
, "CMD_STATE_FS_CONSTANTS (%u)\n", num_const
);
251 /* Expand each float to float[4] for SOA execution */
252 for (i
= 0; i
< num_const
; i
++) {
253 D_PRINTF(CELL_DEBUG_CMD
, " const[%u] = %f\n", i
, constants
[i
]);
254 spu
.constants
[i
] = spu_splats(constants
[i
]);
257 /* return new buffer pos (in 8-byte words) */
258 return pos
+ 2 + num_const
/ 2;
263 cmd_state_framebuffer(const struct cell_command_framebuffer
*cmd
)
265 D_PRINTF(CELL_DEBUG_CMD
, "FRAMEBUFFER: %d x %d at %p, cformat 0x%x zformat 0x%x\n",
272 ASSERT_ALIGN16(cmd
->color_start
);
273 ASSERT_ALIGN16(cmd
->depth_start
);
275 spu
.fb
.color_start
= cmd
->color_start
;
276 spu
.fb
.depth_start
= cmd
->depth_start
;
277 spu
.fb
.color_format
= cmd
->color_format
;
278 spu
.fb
.depth_format
= cmd
->depth_format
;
279 spu
.fb
.width
= cmd
->width
;
280 spu
.fb
.height
= cmd
->height
;
281 spu
.fb
.width_tiles
= (spu
.fb
.width
+ TILE_SIZE
- 1) / TILE_SIZE
;
282 spu
.fb
.height_tiles
= (spu
.fb
.height
+ TILE_SIZE
- 1) / TILE_SIZE
;
284 switch (spu
.fb
.depth_format
) {
285 case PIPE_FORMAT_Z32_UNORM
:
287 spu
.fb
.zscale
= (float) 0xffffffffu
;
289 case PIPE_FORMAT_Z24S8_UNORM
:
290 case PIPE_FORMAT_S8Z24_UNORM
:
291 case PIPE_FORMAT_Z24X8_UNORM
:
292 case PIPE_FORMAT_X8Z24_UNORM
:
294 spu
.fb
.zscale
= (float) 0x00ffffffu
;
296 case PIPE_FORMAT_Z16_UNORM
:
298 spu
.fb
.zscale
= (float) 0xffffu
;
308 * Tex texture mask_s/t and scale_s/t fields depend on the texture size and
309 * sampler wrap modes.
312 update_tex_masks(struct spu_texture
*texture
,
313 const struct pipe_sampler_state
*sampler
,
318 for (i
= 0; i
< CELL_MAX_TEXTURE_LEVELS
; i
++) {
319 int width
= texture
->level
[i
].width
;
320 int height
= texture
->level
[i
].height
;
322 if (sampler
->wrap_s
== PIPE_TEX_WRAP_REPEAT
)
323 texture
->level
[i
].mask_s
= spu_splats(width
- 1);
325 texture
->level
[i
].mask_s
= spu_splats(~0);
327 if (sampler
->wrap_t
== PIPE_TEX_WRAP_REPEAT
)
328 texture
->level
[i
].mask_t
= spu_splats(height
- 1);
330 texture
->level
[i
].mask_t
= spu_splats(~0);
332 if (sampler
->normalized_coords
) {
333 texture
->level
[i
].scale_s
= spu_splats((float) width
);
334 texture
->level
[i
].scale_t
= spu_splats((float) height
);
337 texture
->level
[i
].scale_s
= spu_splats(1.0f
);
338 texture
->level
[i
].scale_t
= spu_splats(1.0f
);
342 /* XXX temporary hack */
343 if (texture
->target
== PIPE_TEXTURE_CUBE
) {
344 spu
.sample_texture4
[unit
] = sample_texture4_cube
;
350 cmd_state_sampler(const struct cell_command_sampler
*sampler
)
352 uint unit
= sampler
->unit
;
354 D_PRINTF(CELL_DEBUG_CMD
, "SAMPLER [%u]\n", unit
);
356 spu
.sampler
[unit
] = sampler
->state
;
358 switch (spu
.sampler
[unit
].min_img_filter
) {
359 case PIPE_TEX_FILTER_LINEAR
:
360 spu
.min_sample_texture4
[unit
] = sample_texture4_bilinear
;
362 case PIPE_TEX_FILTER_ANISO
:
363 /* fall-through, for now */
364 case PIPE_TEX_FILTER_NEAREST
:
365 spu
.min_sample_texture4
[unit
] = sample_texture4_nearest
;
371 switch (spu
.sampler
[sampler
->unit
].mag_img_filter
) {
372 case PIPE_TEX_FILTER_LINEAR
:
373 spu
.mag_sample_texture4
[unit
] = sample_texture4_bilinear
;
375 case PIPE_TEX_FILTER_ANISO
:
376 /* fall-through, for now */
377 case PIPE_TEX_FILTER_NEAREST
:
378 spu
.mag_sample_texture4
[unit
] = sample_texture4_nearest
;
384 switch (spu
.sampler
[sampler
->unit
].min_mip_filter
) {
385 case PIPE_TEX_MIPFILTER_NEAREST
:
386 case PIPE_TEX_MIPFILTER_LINEAR
:
387 spu
.sample_texture4
[unit
] = sample_texture4_lod
;
389 case PIPE_TEX_MIPFILTER_NONE
:
390 spu
.sample_texture4
[unit
] = spu
.mag_sample_texture4
[unit
];
396 update_tex_masks(&spu
.texture
[unit
], &spu
.sampler
[unit
], unit
);
401 cmd_state_texture(const struct cell_command_texture
*texture
)
403 const uint unit
= texture
->unit
;
406 D_PRINTF(CELL_DEBUG_CMD
, "TEXTURE [%u]\n", texture
->unit
);
408 spu
.texture
[unit
].max_level
= 0;
409 spu
.texture
[unit
].target
= texture
->target
;
411 for (i
= 0; i
< CELL_MAX_TEXTURE_LEVELS
; i
++) {
412 uint width
= texture
->width
[i
];
413 uint height
= texture
->height
[i
];
414 uint depth
= texture
->depth
[i
];
416 D_PRINTF(CELL_DEBUG_CMD
, " LEVEL %u: at %p size[0] %u x %u\n", i
,
417 texture
->start
[i
], texture
->width
[i
], texture
->height
[i
]);
419 spu
.texture
[unit
].level
[i
].start
= texture
->start
[i
];
420 spu
.texture
[unit
].level
[i
].width
= width
;
421 spu
.texture
[unit
].level
[i
].height
= height
;
422 spu
.texture
[unit
].level
[i
].depth
= depth
;
424 spu
.texture
[unit
].level
[i
].tiles_per_row
=
425 (width
+ TILE_SIZE
- 1) / TILE_SIZE
;
427 spu
.texture
[unit
].level
[i
].bytes_per_image
=
428 4 * align(width
, TILE_SIZE
) * align(height
, TILE_SIZE
) * depth
;
430 spu
.texture
[unit
].level
[i
].max_s
= spu_splats((int) width
- 1);
431 spu
.texture
[unit
].level
[i
].max_t
= spu_splats((int) height
- 1);
433 if (texture
->start
[i
])
434 spu
.texture
[unit
].max_level
= i
;
437 update_tex_masks(&spu
.texture
[unit
], &spu
.sampler
[unit
], unit
);
442 cmd_state_vertex_info(const struct vertex_info
*vinfo
)
444 D_PRINTF(CELL_DEBUG_CMD
, "VERTEX_INFO num_attribs=%u\n", vinfo
->num_attribs
);
445 ASSERT(vinfo
->num_attribs
>= 1);
446 ASSERT(vinfo
->num_attribs
<= 8);
447 memcpy(&spu
.vertex_info
, vinfo
, sizeof(*vinfo
));
452 cmd_state_vs_array_info(const struct cell_array_info
*vs_info
)
454 const unsigned attr
= vs_info
->attr
;
456 ASSERT(attr
< PIPE_MAX_ATTRIBS
);
457 draw
.vertex_fetch
.src_ptr
[attr
] = vs_info
->base
;
458 draw
.vertex_fetch
.pitch
[attr
] = vs_info
->pitch
;
459 draw
.vertex_fetch
.size
[attr
] = vs_info
->size
;
460 draw
.vertex_fetch
.code_offset
[attr
] = vs_info
->function_offset
;
461 draw
.vertex_fetch
.dirty
= 1;
466 cmd_state_attrib_fetch(const struct cell_attribute_fetch_code
*code
)
468 mfc_get(attribute_fetch_code_buffer
,
469 (unsigned int) code
->base
, /* src */
474 wait_on_mask(1 << TAG_BATCH_BUFFER
);
476 draw
.vertex_fetch
.code
= attribute_fetch_code_buffer
;
483 D_PRINTF(CELL_DEBUG_CMD
, "FINISH\n");
484 really_clear_tiles(0);
485 /* wait for all outstanding DMAs to finish */
486 mfc_write_tag_mask(~0);
487 mfc_read_tag_status_all();
488 /* send mbox message to PPU */
489 spu_write_out_mbox(CELL_CMD_FINISH
);
494 * Execute a batch of commands which was sent to us by the PPU.
495 * See the cell_emit_state.c code to see where the commands come from.
497 * The opcode param encodes the location of the buffer and its size.
500 cmd_batch(uint opcode
)
502 const uint buf
= (opcode
>> 8) & 0xff;
503 uint size
= (opcode
>> 16);
504 uint64_t buffer
[CELL_BUFFER_SIZE
/ 8] ALIGN16_ATTRIB
;
505 const unsigned usize
= size
/ sizeof(buffer
[0]);
508 D_PRINTF(CELL_DEBUG_CMD
, "BATCH buffer %u, len %u, from %p\n",
509 buf
, size
, spu
.init
.buffers
[buf
]);
511 ASSERT((opcode
& CELL_CMD_OPCODE_MASK
) == CELL_CMD_BATCH
);
513 ASSERT_ALIGN16(spu
.init
.buffers
[buf
]);
515 size
= ROUNDUP16(size
);
517 ASSERT_ALIGN16(spu
.init
.buffers
[buf
]);
519 mfc_get(buffer
, /* dest */
520 (unsigned int) spu
.init
.buffers
[buf
], /* src */
525 wait_on_mask(1 << TAG_BATCH_BUFFER
);
527 /* Tell PPU we're done copying the buffer to local store */
528 D_PRINTF(CELL_DEBUG_CMD
, "release batch buf %u\n", buf
);
532 * Loop over commands in the batch buffer
534 for (pos
= 0; pos
< usize
; /* no incr */) {
535 switch (buffer
[pos
]) {
539 case CELL_CMD_CLEAR_SURFACE
:
541 struct cell_command_clear_surface
*clr
542 = (struct cell_command_clear_surface
*) &buffer
[pos
];
543 cmd_clear_surface(clr
);
544 pos
+= sizeof(*clr
) / 8;
547 case CELL_CMD_RENDER
:
549 struct cell_command_render
*render
550 = (struct cell_command_render
*) &buffer
[pos
];
552 cmd_render(render
, &pos_incr
);
557 * state-update commands
559 case CELL_CMD_STATE_FRAMEBUFFER
:
561 struct cell_command_framebuffer
*fb
562 = (struct cell_command_framebuffer
*) &buffer
[pos
];
563 cmd_state_framebuffer(fb
);
564 pos
+= sizeof(*fb
) / 8;
567 case CELL_CMD_STATE_FRAGMENT_OPS
:
569 struct cell_command_fragment_ops
*fops
570 = (struct cell_command_fragment_ops
*) &buffer
[pos
];
571 cmd_state_fragment_ops(fops
);
572 pos
+= sizeof(*fops
) / 8;
575 case CELL_CMD_STATE_FRAGMENT_PROGRAM
:
577 struct cell_command_fragment_program
*fp
578 = (struct cell_command_fragment_program
*) &buffer
[pos
];
579 cmd_state_fragment_program(fp
);
580 pos
+= sizeof(*fp
) / 8;
583 case CELL_CMD_STATE_FS_CONSTANTS
:
584 pos
= cmd_state_fs_constants(buffer
, pos
);
586 case CELL_CMD_STATE_RASTERIZER
:
588 struct cell_command_rasterizer
*rast
=
589 (struct cell_command_rasterizer
*) &buffer
[pos
];
590 spu
.rasterizer
= rast
->rasterizer
;
591 pos
+= sizeof(*rast
) / 8;
594 case CELL_CMD_STATE_SAMPLER
:
596 struct cell_command_sampler
*sampler
597 = (struct cell_command_sampler
*) &buffer
[pos
];
598 cmd_state_sampler(sampler
);
599 pos
+= sizeof(*sampler
) / 8;
602 case CELL_CMD_STATE_TEXTURE
:
604 struct cell_command_texture
*texture
605 = (struct cell_command_texture
*) &buffer
[pos
];
606 cmd_state_texture(texture
);
607 pos
+= sizeof(*texture
) / 8;
610 case CELL_CMD_STATE_VERTEX_INFO
:
611 cmd_state_vertex_info((struct vertex_info
*) &buffer
[pos
+1]);
612 pos
+= (1 + ROUNDUP8(sizeof(struct vertex_info
)) / 8);
614 case CELL_CMD_STATE_VIEWPORT
:
615 (void) memcpy(& draw
.viewport
, &buffer
[pos
+1],
616 sizeof(struct pipe_viewport_state
));
617 pos
+= (1 + ROUNDUP8(sizeof(struct pipe_viewport_state
)) / 8);
619 case CELL_CMD_STATE_UNIFORMS
:
620 draw
.constants
= (const float (*)[4]) (uintptr_t) buffer
[pos
+ 1];
623 case CELL_CMD_STATE_VS_ARRAY_INFO
:
624 cmd_state_vs_array_info((struct cell_array_info
*) &buffer
[pos
+1]);
625 pos
+= (1 + ROUNDUP8(sizeof(struct cell_array_info
)) / 8);
627 case CELL_CMD_STATE_BIND_VS
:
629 spu_bind_vertex_shader(&draw
,
630 (struct cell_shader_info
*) &buffer
[pos
+1]);
632 pos
+= (1 + ROUNDUP8(sizeof(struct cell_shader_info
)) / 8);
634 case CELL_CMD_STATE_ATTRIB_FETCH
:
635 cmd_state_attrib_fetch((struct cell_attribute_fetch_code
*)
637 pos
+= (1 + ROUNDUP8(sizeof(struct cell_attribute_fetch_code
)) / 8);
642 case CELL_CMD_FINISH
:
646 case CELL_CMD_RELEASE_VERTS
:
648 struct cell_command_release_verts
*release
649 = (struct cell_command_release_verts
*) &buffer
[pos
];
650 cmd_release_verts(release
);
651 pos
+= sizeof(*release
) / 8;
654 case CELL_CMD_FLUSH_BUFFER_RANGE
: {
655 struct cell_buffer_range
*br
= (struct cell_buffer_range
*)
658 spu_dcache_mark_dirty((unsigned) br
->base
, br
->size
);
659 pos
+= (1 + ROUNDUP8(sizeof(struct cell_buffer_range
)) / 8);
663 printf("SPU %u: bad opcode: 0x%llx\n", spu
.init
.id
, buffer
[pos
]);
669 D_PRINTF(CELL_DEBUG_CMD
, "BATCH complete\n");
675 * Main loop for SPEs: Get a command, execute it, repeat.
682 D_PRINTF(CELL_DEBUG_CMD
, "Enter command loop\n");
687 D_PRINTF(CELL_DEBUG_CMD
, "Wait for cmd...\n");
689 /* read/wait from mailbox */
690 opcode
= (unsigned int) spu_read_in_mbox();
691 D_PRINTF(CELL_DEBUG_CMD
, "got cmd 0x%x\n", opcode
);
693 switch (opcode
& CELL_CMD_OPCODE_MASK
) {
695 D_PRINTF(CELL_DEBUG_CMD
, "EXIT\n");
698 case CELL_CMD_VS_EXECUTE
:
700 spu_execute_vertex_shader(&draw
, &cmd
.vs
);
707 printf("Bad opcode 0x%x!\n", opcode
& CELL_CMD_OPCODE_MASK
);
712 D_PRINTF(CELL_DEBUG_CMD
, "Exit command loop\n");