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 "shader/prog_parameter.h"
39 #include "shader/prog_statevars.h"
40 #include "intel_batchbuffer.h"
41 #include "brw_context.h"
42 #include "brw_defines.h"
43 #include "brw_state.h"
47 /* Partition the CURBE between the various users of constant values:
49 static void calculate_curbe_offsets( struct brw_context
*brw
)
51 GLcontext
*ctx
= &brw
->intel
.ctx
;
52 /* CACHE_NEW_WM_PROG */
53 GLuint nr_fp_regs
= (brw
->wm
.prog_data
->nr_params
+ 15) / 16;
55 /* BRW_NEW_VERTEX_PROGRAM */
56 struct brw_vertex_program
*vp
= brw_vertex_program(brw
->vertex_program
);
57 GLuint nr_vp_regs
= (vp
->program
.Base
.Parameters
->NumParameters
* 4 + 15) / 16;
58 GLuint nr_clip_regs
= 0;
62 if (ctx
->Transform
.ClipPlanesEnabled
) {
63 GLuint nr_planes
= 6 + brw_count_bits(ctx
->Transform
.ClipPlanesEnabled
);
64 nr_clip_regs
= (nr_planes
* 4 + 15) / 16;
68 total_regs
= nr_fp_regs
+ nr_vp_regs
+ nr_clip_regs
;
70 /* This can happen - what to do? Probably rather than falling
71 * back, the best thing to do is emit programs which code the
72 * constants as immediate values. Could do this either as a static
73 * cap on WM and VS, or adaptively.
75 * Unfortunately, this is currently dependent on the results of the
76 * program generation process (in the case of wm), so this would
77 * introduce the need to re-generate programs in the event of a
78 * curbe allocation failure.
80 /* Max size is 32 - just large enough to
81 * hold the 128 parameters allowed by
82 * the fragment and vertex program
83 * api's. It's not clear what happens
84 * when both VP and FP want to use 128
87 assert(total_regs
<= 32);
91 if (nr_fp_regs
> brw
->curbe
.wm_size
||
92 nr_vp_regs
> brw
->curbe
.vs_size
||
93 nr_clip_regs
!= brw
->curbe
.clip_size
||
94 (total_regs
< brw
->curbe
.total_size
/ 4 &&
95 brw
->curbe
.total_size
> 16)) {
99 /* Calculate a new layout:
102 brw
->curbe
.wm_start
= reg
;
103 brw
->curbe
.wm_size
= nr_fp_regs
; reg
+= nr_fp_regs
;
104 brw
->curbe
.clip_start
= reg
;
105 brw
->curbe
.clip_size
= nr_clip_regs
; reg
+= nr_clip_regs
;
106 brw
->curbe
.vs_start
= reg
;
107 brw
->curbe
.vs_size
= nr_vp_regs
; reg
+= nr_vp_regs
;
108 brw
->curbe
.total_size
= reg
;
111 _mesa_printf("curbe wm %d+%d clip %d+%d vs %d+%d\n",
114 brw
->curbe
.clip_start
,
115 brw
->curbe
.clip_size
,
117 brw
->curbe
.vs_size
);
119 brw
->state
.dirty
.brw
|= BRW_NEW_CURBE_OFFSETS
;
124 const struct brw_tracked_state brw_curbe_offsets
= {
126 .mesa
= _NEW_TRANSFORM
,
127 .brw
= BRW_NEW_VERTEX_PROGRAM
,
128 .cache
= CACHE_NEW_WM_PROG
130 .prepare
= calculate_curbe_offsets
136 /* Define the number of curbes within CS's urb allocation. Multiple
137 * urb entries -> multiple curbes. These will be used by
138 * fixed-function hardware in a double-buffering scheme to avoid a
139 * pipeline stall each time the contents of the curbe is changed.
141 void brw_upload_cs_urb_state(struct brw_context
*brw
)
143 struct brw_cs_urb_state cs_urb
;
144 memset(&cs_urb
, 0, sizeof(cs_urb
));
146 /* It appears that this is the state packet for the CS unit, ie. the
147 * urb entries detailed here are housed in the CS range from the
150 cs_urb
.header
.opcode
= CMD_CS_URB_STATE
;
151 cs_urb
.header
.length
= sizeof(cs_urb
)/4 - 2;
153 /* BRW_NEW_URB_FENCE */
154 cs_urb
.bits0
.nr_urb_entries
= brw
->urb
.nr_cs_entries
;
155 cs_urb
.bits0
.urb_entry_size
= brw
->urb
.csize
- 1;
157 assert(brw
->urb
.nr_cs_entries
);
158 BRW_CACHED_BATCH_STRUCT(brw
, &cs_urb
);
161 static GLfloat fixed_plane
[6][4] = {
170 /* Upload a new set of constants. Too much variability to go into the
171 * cache mechanism, but maybe would benefit from a comparison against
172 * the current uploaded set of constants.
174 static void prepare_constant_buffer(struct brw_context
*brw
)
176 GLcontext
*ctx
= &brw
->intel
.ctx
;
177 const struct brw_vertex_program
*vp
=
178 brw_vertex_program_const(brw
->vertex_program
);
179 const struct brw_fragment_program
*fp
=
180 brw_fragment_program_const(brw
->fragment_program
);
181 const GLuint sz
= brw
->curbe
.total_size
;
182 const GLuint bufsz
= sz
* 16 * sizeof(GLfloat
);
186 /* Update our own dependency flags. This works because this
187 * function will also be called whenever fp or vp changes.
189 brw
->curbe
.tracked_state
.dirty
.mesa
= (_NEW_TRANSFORM
|_NEW_PROJECTION
);
190 brw
->curbe
.tracked_state
.dirty
.mesa
|= vp
->program
.Base
.Parameters
->StateFlags
;
191 brw
->curbe
.tracked_state
.dirty
.mesa
|= fp
->program
.Base
.Parameters
->StateFlags
;
194 if (brw
->curbe
.last_buf
) {
195 free(brw
->curbe
.last_buf
);
196 brw
->curbe
.last_buf
= NULL
;
197 brw
->curbe
.last_bufsz
= 0;
202 buf
= (GLfloat
*) _mesa_calloc(bufsz
);
204 /* fragment shader constants */
205 if (brw
->curbe
.wm_size
) {
206 GLuint offset
= brw
->curbe
.wm_start
* 16;
208 _mesa_load_state_parameters(ctx
, fp
->program
.Base
.Parameters
);
210 /* copy float constants */
211 for (i
= 0; i
< brw
->wm
.prog_data
->nr_params
; i
++)
212 buf
[offset
+ i
] = *brw
->wm
.prog_data
->param
[i
];
216 /* The clipplanes are actually delivered to both CLIP and VS units.
217 * VS uses them to calculate the outcode bitmasks.
219 if (brw
->curbe
.clip_size
) {
220 GLuint offset
= brw
->curbe
.clip_start
* 16;
223 /* If any planes are going this way, send them all this way:
225 for (i
= 0; i
< 6; i
++) {
226 buf
[offset
+ i
* 4 + 0] = fixed_plane
[i
][0];
227 buf
[offset
+ i
* 4 + 1] = fixed_plane
[i
][1];
228 buf
[offset
+ i
* 4 + 2] = fixed_plane
[i
][2];
229 buf
[offset
+ i
* 4 + 3] = fixed_plane
[i
][3];
232 /* Clip planes: _NEW_TRANSFORM plus _NEW_PROJECTION to get to
235 assert(MAX_CLIP_PLANES
== 6);
236 for (j
= 0; j
< MAX_CLIP_PLANES
; j
++) {
237 if (ctx
->Transform
.ClipPlanesEnabled
& (1<<j
)) {
238 buf
[offset
+ i
* 4 + 0] = ctx
->Transform
._ClipUserPlane
[j
][0];
239 buf
[offset
+ i
* 4 + 1] = ctx
->Transform
._ClipUserPlane
[j
][1];
240 buf
[offset
+ i
* 4 + 2] = ctx
->Transform
._ClipUserPlane
[j
][2];
241 buf
[offset
+ i
* 4 + 3] = ctx
->Transform
._ClipUserPlane
[j
][3];
247 /* vertex shader constants */
248 if (brw
->curbe
.vs_size
) {
249 GLuint offset
= brw
->curbe
.vs_start
* 16;
250 GLuint nr
= vp
->program
.Base
.Parameters
->NumParameters
;
252 _mesa_load_state_parameters(ctx
, vp
->program
.Base
.Parameters
);
254 for (i
= 0; i
< nr
; i
++) {
255 const GLfloat
*value
= vp
->program
.Base
.Parameters
->ParameterValues
[i
];
256 buf
[offset
+ i
* 4 + 0] = value
[0];
257 buf
[offset
+ i
* 4 + 1] = value
[1];
258 buf
[offset
+ i
* 4 + 2] = value
[2];
259 buf
[offset
+ i
* 4 + 3] = value
[3];
264 for (i
= 0; i
< sz
*16; i
+=4)
265 _mesa_printf("curbe %d.%d: %f %f %f %f\n", i
/8, i
&4,
266 buf
[i
+0], buf
[i
+1], buf
[i
+2], buf
[i
+3]);
268 _mesa_printf("last_buf %p buf %p sz %d/%d cmp %d\n",
269 brw
->curbe
.last_buf
, buf
,
270 bufsz
, brw
->curbe
.last_bufsz
,
271 brw
->curbe
.last_buf
? memcmp(buf
, brw
->curbe
.last_buf
, bufsz
) : -1);
274 if (brw
->curbe
.curbe_bo
!= NULL
&&
275 brw
->curbe
.last_buf
&&
276 bufsz
== brw
->curbe
.last_bufsz
&&
277 memcmp(buf
, brw
->curbe
.last_buf
, bufsz
) == 0) {
278 /* constants have not changed */
282 /* constants have changed */
283 if (brw
->curbe
.last_buf
)
284 _mesa_free(brw
->curbe
.last_buf
);
286 brw
->curbe
.last_buf
= buf
;
287 brw
->curbe
.last_bufsz
= bufsz
;
289 if (brw
->curbe
.curbe_bo
!= NULL
&&
290 (brw
->curbe
.need_new_bo
||
291 brw
->curbe
.curbe_next_offset
+ bufsz
> brw
->curbe
.curbe_bo
->size
))
293 dri_bo_unreference(brw
->curbe
.curbe_bo
);
294 brw
->curbe
.curbe_bo
= NULL
;
297 if (brw
->curbe
.curbe_bo
== NULL
) {
298 /* Allocate a single page for CURBE entries for this batchbuffer.
299 * They're generally around 64b.
301 brw
->curbe
.curbe_bo
= dri_bo_alloc(brw
->intel
.bufmgr
, "CURBE",
303 brw
->curbe
.curbe_next_offset
= 0;
306 brw
->curbe
.curbe_offset
= brw
->curbe
.curbe_next_offset
;
307 brw
->curbe
.curbe_next_offset
+= bufsz
;
308 brw
->curbe
.curbe_next_offset
= ALIGN(brw
->curbe
.curbe_next_offset
, 64);
310 /* Copy data to the buffer:
312 dri_bo_subdata(brw
->curbe
.curbe_bo
, brw
->curbe
.curbe_offset
, bufsz
, buf
);
315 brw_add_validated_bo(brw
, brw
->curbe
.curbe_bo
);
317 /* Because this provokes an action (ie copy the constants into the
318 * URB), it shouldn't be shortcircuited if identical to the
319 * previous time - because eg. the urb destination may have
320 * changed, or the urb contents different to last time.
322 * Note that the data referred to is actually copied internally,
323 * not just used in place according to passed pointer.
325 * It appears that the CS unit takes care of using each available
326 * URB entry (Const URB Entry == CURBE) in turn, and issuing
327 * flushes as necessary when doublebuffering of CURBEs isn't
333 static void emit_constant_buffer(struct brw_context
*brw
)
335 struct intel_context
*intel
= &brw
->intel
;
336 GLuint sz
= brw
->curbe
.total_size
;
338 BEGIN_BATCH(2, IGNORE_CLIPRECTS
);
340 OUT_BATCH((CMD_CONST_BUFFER
<< 16) | (2 - 2));
343 OUT_BATCH((CMD_CONST_BUFFER
<< 16) | (1 << 8) | (2 - 2));
344 OUT_RELOC(brw
->curbe
.curbe_bo
,
345 I915_GEM_DOMAIN_INSTRUCTION
, 0,
346 (sz
- 1) + brw
->curbe
.curbe_offset
);
351 /* This tracked state is unique in that the state it monitors varies
352 * dynamically depending on the parameters tracked by the fragment and
353 * vertex programs. This is the template used as a starting point,
354 * each context will maintain a copy of this internally and update as
357 const struct brw_tracked_state brw_constant_buffer
= {
359 .mesa
= (_NEW_TRANSFORM
|_NEW_PROJECTION
), /* plus fp and vp flags */
360 .brw
= (BRW_NEW_FRAGMENT_PROGRAM
|
361 BRW_NEW_VERTEX_PROGRAM
|
362 BRW_NEW_URB_FENCE
| /* Implicit - hardware requires this, not used above */
363 BRW_NEW_PSP
| /* Implicit - hardware requires this, not used above */
364 BRW_NEW_CURBE_OFFSETS
|
366 .cache
= (CACHE_NEW_WM_PROG
)
368 .prepare
= prepare_constant_buffer
,
369 .emit
= emit_constant_buffer
,