2 Copyright (C) Intel Corp. 2006. All Rights Reserved.
3 Intel funded Tungsten Graphics (http://www.tungstengraphics.com) to
4 develop this 3D driver.
6 Permission is hereby granted, free of charge, to any person obtaining
7 a 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, sublicense, 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
16 portions of the Software.
18 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
19 EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
21 IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
22 LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
23 OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
24 WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **********************************************************************/
29 * Keith Whitwell <keith@tungstengraphics.com>
34 #include "main/glheader.h"
35 #include "main/context.h"
36 #include "main/macros.h"
37 #include "main/enums.h"
38 #include "program/prog_parameter.h"
39 #include "program/prog_print.h"
40 #include "program/prog_statevars.h"
41 #include "intel_batchbuffer.h"
42 #include "intel_regions.h"
43 #include "brw_context.h"
44 #include "brw_defines.h"
45 #include "brw_state.h"
50 * Partition the CURBE between the various users of constant values:
51 * Note that vertex and fragment shaders can now fetch constants out
52 * of constant buffers. We no longer allocatea block of the GRF for
53 * constants. That greatly reduces the demand for space in the CURBE.
54 * Some of the comments within are dated...
56 static void calculate_curbe_offsets( struct brw_context
*brw
)
58 GLcontext
*ctx
= &brw
->intel
.ctx
;
59 /* CACHE_NEW_WM_PROG */
60 const GLuint nr_fp_regs
= (brw
->wm
.prog_data
->nr_params
+ 15) / 16;
62 /* BRW_NEW_VERTEX_PROGRAM */
63 const GLuint nr_vp_regs
= (brw
->vs
.prog_data
->nr_params
+ 15) / 16;
64 GLuint nr_clip_regs
= 0;
68 if (ctx
->Transform
.ClipPlanesEnabled
) {
69 GLuint nr_planes
= 6 + brw_count_bits(ctx
->Transform
.ClipPlanesEnabled
);
70 nr_clip_regs
= (nr_planes
* 4 + 15) / 16;
74 total_regs
= nr_fp_regs
+ nr_vp_regs
+ nr_clip_regs
;
76 /* This can happen - what to do? Probably rather than falling
77 * back, the best thing to do is emit programs which code the
78 * constants as immediate values. Could do this either as a static
79 * cap on WM and VS, or adaptively.
81 * Unfortunately, this is currently dependent on the results of the
82 * program generation process (in the case of wm), so this would
83 * introduce the need to re-generate programs in the event of a
84 * curbe allocation failure.
86 /* Max size is 32 - just large enough to
87 * hold the 128 parameters allowed by
88 * the fragment and vertex program
89 * api's. It's not clear what happens
90 * when both VP and FP want to use 128
93 assert(total_regs
<= 32);
97 if (nr_fp_regs
> brw
->curbe
.wm_size
||
98 nr_vp_regs
> brw
->curbe
.vs_size
||
99 nr_clip_regs
!= brw
->curbe
.clip_size
||
100 (total_regs
< brw
->curbe
.total_size
/ 4 &&
101 brw
->curbe
.total_size
> 16)) {
105 /* Calculate a new layout:
108 brw
->curbe
.wm_start
= reg
;
109 brw
->curbe
.wm_size
= nr_fp_regs
; reg
+= nr_fp_regs
;
110 brw
->curbe
.clip_start
= reg
;
111 brw
->curbe
.clip_size
= nr_clip_regs
; reg
+= nr_clip_regs
;
112 brw
->curbe
.vs_start
= reg
;
113 brw
->curbe
.vs_size
= nr_vp_regs
; reg
+= nr_vp_regs
;
114 brw
->curbe
.total_size
= reg
;
117 printf("curbe wm %d+%d clip %d+%d vs %d+%d\n",
120 brw
->curbe
.clip_start
,
121 brw
->curbe
.clip_size
,
123 brw
->curbe
.vs_size
);
125 brw
->state
.dirty
.brw
|= BRW_NEW_CURBE_OFFSETS
;
130 const struct brw_tracked_state brw_curbe_offsets
= {
132 .mesa
= _NEW_TRANSFORM
,
133 .brw
= BRW_NEW_VERTEX_PROGRAM
| BRW_NEW_CONTEXT
,
134 .cache
= CACHE_NEW_WM_PROG
136 .prepare
= calculate_curbe_offsets
142 /* Define the number of curbes within CS's urb allocation. Multiple
143 * urb entries -> multiple curbes. These will be used by
144 * fixed-function hardware in a double-buffering scheme to avoid a
145 * pipeline stall each time the contents of the curbe is changed.
147 void brw_upload_cs_urb_state(struct brw_context
*brw
)
149 struct brw_cs_urb_state cs_urb
;
150 memset(&cs_urb
, 0, sizeof(cs_urb
));
152 /* It appears that this is the state packet for the CS unit, ie. the
153 * urb entries detailed here are housed in the CS range from the
156 cs_urb
.header
.opcode
= CMD_CS_URB_STATE
;
157 cs_urb
.header
.length
= sizeof(cs_urb
)/4 - 2;
159 /* BRW_NEW_URB_FENCE */
160 cs_urb
.bits0
.nr_urb_entries
= brw
->urb
.nr_cs_entries
;
161 cs_urb
.bits0
.urb_entry_size
= brw
->urb
.csize
- 1;
163 assert(brw
->urb
.nr_cs_entries
);
164 BRW_CACHED_BATCH_STRUCT(brw
, &cs_urb
);
167 static GLfloat fixed_plane
[6][4] = {
176 /* Upload a new set of constants. Too much variability to go into the
177 * cache mechanism, but maybe would benefit from a comparison against
178 * the current uploaded set of constants.
180 static void prepare_constant_buffer(struct brw_context
*brw
)
182 GLcontext
*ctx
= &brw
->intel
.ctx
;
183 const struct brw_vertex_program
*vp
=
184 brw_vertex_program_const(brw
->vertex_program
);
185 const struct brw_fragment_program
*fp
=
186 brw_fragment_program_const(brw
->fragment_program
);
187 const GLuint sz
= brw
->curbe
.total_size
;
188 const GLuint bufsz
= sz
* 16 * sizeof(GLfloat
);
193 brw
->curbe
.last_bufsz
= 0;
197 buf
= brw
->curbe
.next_buf
;
199 /* fragment shader constants */
200 if (brw
->curbe
.wm_size
) {
201 GLuint offset
= brw
->curbe
.wm_start
* 16;
203 _mesa_load_state_parameters(ctx
, fp
->program
.Base
.Parameters
);
205 /* copy float constants */
206 for (i
= 0; i
< brw
->wm
.prog_data
->nr_params
; i
++)
207 buf
[offset
+ i
] = *brw
->wm
.prog_data
->param
[i
];
211 /* The clipplanes are actually delivered to both CLIP and VS units.
212 * VS uses them to calculate the outcode bitmasks.
214 if (brw
->curbe
.clip_size
) {
215 GLuint offset
= brw
->curbe
.clip_start
* 16;
218 /* If any planes are going this way, send them all this way:
220 for (i
= 0; i
< 6; i
++) {
221 buf
[offset
+ i
* 4 + 0] = fixed_plane
[i
][0];
222 buf
[offset
+ i
* 4 + 1] = fixed_plane
[i
][1];
223 buf
[offset
+ i
* 4 + 2] = fixed_plane
[i
][2];
224 buf
[offset
+ i
* 4 + 3] = fixed_plane
[i
][3];
227 /* Clip planes: _NEW_TRANSFORM plus _NEW_PROJECTION to get to
230 assert(MAX_CLIP_PLANES
== 6);
231 for (j
= 0; j
< MAX_CLIP_PLANES
; j
++) {
232 if (ctx
->Transform
.ClipPlanesEnabled
& (1<<j
)) {
233 buf
[offset
+ i
* 4 + 0] = ctx
->Transform
._ClipUserPlane
[j
][0];
234 buf
[offset
+ i
* 4 + 1] = ctx
->Transform
._ClipUserPlane
[j
][1];
235 buf
[offset
+ i
* 4 + 2] = ctx
->Transform
._ClipUserPlane
[j
][2];
236 buf
[offset
+ i
* 4 + 3] = ctx
->Transform
._ClipUserPlane
[j
][3];
242 /* vertex shader constants */
243 if (brw
->curbe
.vs_size
) {
244 GLuint offset
= brw
->curbe
.vs_start
* 16;
245 GLuint nr
= brw
->vs
.prog_data
->nr_params
/ 4;
247 if (brw
->vertex_program
->IsNVProgram
)
248 _mesa_load_tracked_matrices(ctx
);
250 /* Updates the ParamaterValues[i] pointers for all parameters of the
251 * basic type of PROGRAM_STATE_VAR.
253 _mesa_load_state_parameters(ctx
, vp
->program
.Base
.Parameters
);
255 if (vp
->use_const_buffer
) {
256 /* Load the subset of push constants that will get used when
257 * we also have a pull constant buffer.
259 for (i
= 0; i
< vp
->program
.Base
.Parameters
->NumParameters
; i
++) {
260 if (brw
->vs
.constant_map
[i
] != -1) {
261 assert(brw
->vs
.constant_map
[i
] <= nr
);
262 memcpy(buf
+ offset
+ brw
->vs
.constant_map
[i
] * 4,
263 vp
->program
.Base
.Parameters
->ParameterValues
[i
],
268 for (i
= 0; i
< nr
; i
++) {
269 memcpy(buf
+ offset
+ i
* 4,
270 vp
->program
.Base
.Parameters
->ParameterValues
[i
],
277 for (i
= 0; i
< sz
*16; i
+=4)
278 printf("curbe %d.%d: %f %f %f %f\n", i
/8, i
&4,
279 buf
[i
+0], buf
[i
+1], buf
[i
+2], buf
[i
+3]);
281 printf("last_buf %p buf %p sz %d/%d cmp %d\n",
282 brw
->curbe
.last_buf
, buf
,
283 bufsz
, brw
->curbe
.last_bufsz
,
284 brw
->curbe
.last_buf
? memcmp(buf
, brw
->curbe
.last_buf
, bufsz
) : -1);
287 if (brw
->curbe
.curbe_bo
!= NULL
&&
288 bufsz
== brw
->curbe
.last_bufsz
&&
289 memcmp(buf
, brw
->curbe
.last_buf
, bufsz
) == 0) {
290 /* constants have not changed */
292 /* Update the record of what our last set of constants was. We
293 * don't just flip the pointers because we don't fill in the
294 * data in the padding between the entries.
296 memcpy(brw
->curbe
.last_buf
, buf
, bufsz
);
297 brw
->curbe
.last_bufsz
= bufsz
;
299 if (brw
->curbe
.curbe_bo
!= NULL
&&
300 brw
->curbe
.curbe_next_offset
+ bufsz
> brw
->curbe
.curbe_bo
->size
)
302 drm_intel_gem_bo_unmap_gtt(brw
->curbe
.curbe_bo
);
303 drm_intel_bo_unreference(brw
->curbe
.curbe_bo
);
304 brw
->curbe
.curbe_bo
= NULL
;
307 if (brw
->curbe
.curbe_bo
== NULL
) {
308 /* Allocate a single page for CURBE entries for this batchbuffer.
309 * They're generally around 64b.
311 brw
->curbe
.curbe_bo
= drm_intel_bo_alloc(brw
->intel
.bufmgr
, "CURBE",
313 brw
->curbe
.curbe_next_offset
= 0;
314 drm_intel_gem_bo_map_gtt(brw
->curbe
.curbe_bo
);
315 assert(bufsz
< 4096);
318 brw
->curbe
.curbe_offset
= brw
->curbe
.curbe_next_offset
;
319 brw
->curbe
.curbe_next_offset
+= bufsz
;
320 brw
->curbe
.curbe_next_offset
= ALIGN(brw
->curbe
.curbe_next_offset
, 64);
322 /* Copy data to the buffer:
324 memcpy(brw
->curbe
.curbe_bo
->virtual + brw
->curbe
.curbe_offset
,
329 brw_add_validated_bo(brw
, brw
->curbe
.curbe_bo
);
331 /* Because this provokes an action (ie copy the constants into the
332 * URB), it shouldn't be shortcircuited if identical to the
333 * previous time - because eg. the urb destination may have
334 * changed, or the urb contents different to last time.
336 * Note that the data referred to is actually copied internally,
337 * not just used in place according to passed pointer.
339 * It appears that the CS unit takes care of using each available
340 * URB entry (Const URB Entry == CURBE) in turn, and issuing
341 * flushes as necessary when doublebuffering of CURBEs isn't
346 static void emit_constant_buffer(struct brw_context
*brw
)
348 struct intel_context
*intel
= &brw
->intel
;
349 GLuint sz
= brw
->curbe
.total_size
;
353 OUT_BATCH((CMD_CONST_BUFFER
<< 16) | (2 - 2));
356 OUT_BATCH((CMD_CONST_BUFFER
<< 16) | (1 << 8) | (2 - 2));
357 OUT_RELOC(brw
->curbe
.curbe_bo
,
358 I915_GEM_DOMAIN_INSTRUCTION
, 0,
359 (sz
- 1) + brw
->curbe
.curbe_offset
);
364 /* This tracked state is unique in that the state it monitors varies
365 * dynamically depending on the parameters tracked by the fragment and
366 * vertex programs. This is the template used as a starting point,
367 * each context will maintain a copy of this internally and update as
370 const struct brw_tracked_state brw_constant_buffer
= {
372 .mesa
= _NEW_PROGRAM_CONSTANTS
,
373 .brw
= (BRW_NEW_FRAGMENT_PROGRAM
|
374 BRW_NEW_VERTEX_PROGRAM
|
375 BRW_NEW_URB_FENCE
| /* Implicit - hardware requires this, not used above */
376 BRW_NEW_PSP
| /* Implicit - hardware requires this, not used above */
377 BRW_NEW_CURBE_OFFSETS
|
379 .cache
= (CACHE_NEW_WM_PROG
)
381 .prepare
= prepare_constant_buffer
,
382 .emit
= emit_constant_buffer
,