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>
33 #include "main/imports.h"
34 #include "main/enums.h"
35 #include "main/shaderobj.h"
36 #include "program/prog_parameter.h"
37 #include "program/program.h"
38 #include "program/programopt.h"
40 #include "glsl/ralloc.h"
42 #include "brw_context.h"
46 get_new_program_id(struct intel_screen
*screen
)
48 static pthread_mutex_t m
= PTHREAD_MUTEX_INITIALIZER
;
49 pthread_mutex_lock(&m
);
50 unsigned id
= screen
->program_id
++;
51 pthread_mutex_unlock(&m
);
55 static void brwBindProgram( struct gl_context
*ctx
,
57 struct gl_program
*prog
)
59 struct brw_context
*brw
= brw_context(ctx
);
62 case GL_VERTEX_PROGRAM_ARB
:
63 brw
->state
.dirty
.brw
|= BRW_NEW_VERTEX_PROGRAM
;
65 case MESA_GEOMETRY_PROGRAM
:
66 brw
->state
.dirty
.brw
|= BRW_NEW_GEOMETRY_PROGRAM
;
68 case GL_FRAGMENT_PROGRAM_ARB
:
69 brw
->state
.dirty
.brw
|= BRW_NEW_FRAGMENT_PROGRAM
;
74 static struct gl_program
*brwNewProgram( struct gl_context
*ctx
,
78 struct brw_context
*brw
= brw_context(ctx
);
81 case GL_VERTEX_PROGRAM_ARB
: {
82 struct brw_vertex_program
*prog
= CALLOC_STRUCT(brw_vertex_program
);
84 prog
->id
= get_new_program_id(brw
->intelScreen
);
86 return _mesa_init_vertex_program( ctx
, &prog
->program
,
93 case GL_FRAGMENT_PROGRAM_ARB
: {
94 struct brw_fragment_program
*prog
= CALLOC_STRUCT(brw_fragment_program
);
96 prog
->id
= get_new_program_id(brw
->intelScreen
);
98 return _mesa_init_fragment_program( ctx
, &prog
->program
,
105 case MESA_GEOMETRY_PROGRAM
: {
106 struct brw_geometry_program
*prog
= CALLOC_STRUCT(brw_geometry_program
);
108 prog
->id
= get_new_program_id(brw
->intelScreen
);
110 return _mesa_init_geometry_program(ctx
, &prog
->program
, target
, id
);
117 return _mesa_new_program(ctx
, target
, id
);
121 static void brwDeleteProgram( struct gl_context
*ctx
,
122 struct gl_program
*prog
)
124 _mesa_delete_program( ctx
, prog
);
129 brwIsProgramNative(struct gl_context
*ctx
,
131 struct gl_program
*prog
)
137 brwProgramStringNotify(struct gl_context
*ctx
,
139 struct gl_program
*prog
)
141 struct brw_context
*brw
= brw_context(ctx
);
144 case GL_FRAGMENT_PROGRAM_ARB
: {
145 struct gl_fragment_program
*fprog
= (struct gl_fragment_program
*) prog
;
146 struct brw_fragment_program
*newFP
= brw_fragment_program(fprog
);
147 const struct brw_fragment_program
*curFP
=
148 brw_fragment_program_const(brw
->fragment_program
);
151 brw
->state
.dirty
.brw
|= BRW_NEW_FRAGMENT_PROGRAM
;
152 newFP
->id
= get_new_program_id(brw
->intelScreen
);
155 case GL_VERTEX_PROGRAM_ARB
: {
156 struct gl_vertex_program
*vprog
= (struct gl_vertex_program
*) prog
;
157 struct brw_vertex_program
*newVP
= brw_vertex_program(vprog
);
158 const struct brw_vertex_program
*curVP
=
159 brw_vertex_program_const(brw
->vertex_program
);
162 brw
->state
.dirty
.brw
|= BRW_NEW_VERTEX_PROGRAM
;
163 if (newVP
->program
.IsPositionInvariant
) {
164 _mesa_insert_mvp_code(ctx
, &newVP
->program
);
166 newVP
->id
= get_new_program_id(brw
->intelScreen
);
168 /* Also tell tnl about it:
170 _tnl_program_string(ctx
, target
, prog
);
175 * driver->ProgramStringNotify is only called for ARB programs, fixed
176 * function vertex programs, and ir_to_mesa (which isn't used by the
177 * i965 back-end). Therefore, even after geometry shaders are added,
178 * this function should only ever be called with a target of
179 * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
181 assert(!"Unexpected target in brwProgramStringNotify");
185 brw_add_texrect_params(prog
);
191 brw_add_texrect_params(struct gl_program
*prog
)
193 for (int texunit
= 0; texunit
< BRW_MAX_TEX_UNIT
; texunit
++) {
194 if (!(prog
->TexturesUsed
[texunit
] & (1 << TEXTURE_RECT_INDEX
)))
197 int tokens
[STATE_LENGTH
] = {
205 _mesa_add_state_reference(prog
->Parameters
, (gl_state_index
*)tokens
);
209 /* Per-thread scratch space is a power-of-two multiple of 1KB. */
211 brw_get_scratch_size(int size
)
215 for (i
= 1024; i
< size
; i
*= 2)
222 brw_get_scratch_bo(struct brw_context
*brw
,
223 drm_intel_bo
**scratch_bo
, int size
)
225 drm_intel_bo
*old_bo
= *scratch_bo
;
227 if (old_bo
&& old_bo
->size
< size
) {
228 drm_intel_bo_unreference(old_bo
);
233 *scratch_bo
= drm_intel_bo_alloc(brw
->bufmgr
, "scratch bo", size
, 4096);
237 void brwInitFragProgFuncs( struct dd_function_table
*functions
)
239 assert(functions
->ProgramStringNotify
== _tnl_program_string
);
241 functions
->BindProgram
= brwBindProgram
;
242 functions
->NewProgram
= brwNewProgram
;
243 functions
->DeleteProgram
= brwDeleteProgram
;
244 functions
->IsProgramNative
= brwIsProgramNative
;
245 functions
->ProgramStringNotify
= brwProgramStringNotify
;
247 functions
->NewShader
= brw_new_shader
;
248 functions
->NewShaderProgram
= brw_new_shader_program
;
249 functions
->LinkShader
= brw_link_shader
;
253 brw_init_shader_time(struct brw_context
*brw
)
255 const int max_entries
= 4096;
256 brw
->shader_time
.bo
= drm_intel_bo_alloc(brw
->bufmgr
, "shader time",
257 max_entries
* SHADER_TIME_STRIDE
,
259 brw
->shader_time
.shader_programs
= rzalloc_array(brw
, struct gl_shader_program
*,
261 brw
->shader_time
.programs
= rzalloc_array(brw
, struct gl_program
*,
263 brw
->shader_time
.types
= rzalloc_array(brw
, enum shader_time_shader_type
,
265 brw
->shader_time
.cumulative
= rzalloc_array(brw
, uint64_t,
267 brw
->shader_time
.max_entries
= max_entries
;
271 compare_time(const void *a
, const void *b
)
273 uint64_t * const *a_val
= a
;
274 uint64_t * const *b_val
= b
;
276 /* We don't just subtract because we're turning the value to an int. */
277 if (**a_val
< **b_val
)
279 else if (**a_val
== **b_val
)
286 get_written_and_reset(struct brw_context
*brw
, int i
,
287 uint64_t *written
, uint64_t *reset
)
289 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
290 assert(type
== ST_VS
|| type
== ST_FS8
|| type
== ST_FS16
);
292 /* Find where we recorded written and reset. */
295 for (wi
= i
; brw
->shader_time
.types
[wi
] != type
+ 1; wi
++)
298 for (ri
= i
; brw
->shader_time
.types
[ri
] != type
+ 2; ri
++)
301 *written
= brw
->shader_time
.cumulative
[wi
];
302 *reset
= brw
->shader_time
.cumulative
[ri
];
306 print_shader_time_line(const char *stage
, const char *name
,
307 int shader_num
, uint64_t time
, uint64_t total
)
309 printf("%-6s%-6s", stage
, name
);
311 if (shader_num
!= -1)
312 printf("%4d: ", shader_num
);
316 printf("%16lld (%7.2f Gcycles) %4.1f%%\n",
318 (double)time
/ 1000000000.0,
319 (double)time
/ total
* 100.0);
323 brw_report_shader_time(struct brw_context
*brw
)
325 if (!brw
->shader_time
.bo
|| !brw
->shader_time
.num_entries
)
328 uint64_t scaled
[brw
->shader_time
.num_entries
];
329 uint64_t *sorted
[brw
->shader_time
.num_entries
];
330 uint64_t total_by_type
[ST_FS16
+ 1];
331 memset(total_by_type
, 0, sizeof(total_by_type
));
333 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
334 uint64_t written
= 0, reset
= 0;
335 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
337 sorted
[i
] = &scaled
[i
];
344 case ST_FS16_WRITTEN
:
346 /* We'll handle these when along with the time. */
353 get_written_and_reset(brw
, i
, &written
, &reset
);
357 /* I sometimes want to print things that aren't the 3 shader times.
358 * Just print the sum in that case.
365 uint64_t time
= brw
->shader_time
.cumulative
[i
];
367 scaled
[i
] = time
/ written
* (written
+ reset
);
376 total_by_type
[type
] += scaled
[i
];
386 printf("No shader time collected yet\n");
390 qsort(sorted
, brw
->shader_time
.num_entries
, sizeof(sorted
[0]), compare_time
);
393 printf("type ID cycles spent %% of total\n");
394 for (int s
= 0; s
< brw
->shader_time
.num_entries
; s
++) {
395 const char *shader_name
;
397 /* Work back from the sorted pointers times to a time to print. */
398 int i
= sorted
[s
] - scaled
;
404 if (brw
->shader_time
.shader_programs
[i
]) {
405 shader_num
= brw
->shader_time
.shader_programs
[i
]->Name
;
407 /* The fixed function fragment shader generates GLSL IR with a Name
408 * of 0, and nothing else does.
410 if (shader_num
== 0 &&
411 (brw
->shader_time
.types
[i
] == ST_FS8
||
412 brw
->shader_time
.types
[i
] == ST_FS16
)) {
416 shader_name
= "glsl";
418 } else if (brw
->shader_time
.programs
[i
]) {
419 shader_num
= brw
->shader_time
.programs
[i
]->Id
;
420 if (shader_num
== 0) {
424 shader_name
= "prog";
427 shader_name
= "other";
430 switch (brw
->shader_time
.types
[i
]) {
445 print_shader_time_line(stage
, shader_name
, shader_num
,
450 print_shader_time_line("total", "vs", -1, total_by_type
[ST_VS
], total
);
451 print_shader_time_line("total", "fs8", -1, total_by_type
[ST_FS8
], total
);
452 print_shader_time_line("total", "fs16", -1, total_by_type
[ST_FS16
], total
);
456 brw_collect_shader_time(struct brw_context
*brw
)
458 if (!brw
->shader_time
.bo
)
461 /* This probably stalls on the last rendering. We could fix that by
462 * delaying reading the reports, but it doesn't look like it's a big
463 * overhead compared to the cost of tracking the time in the first place.
465 drm_intel_bo_map(brw
->shader_time
.bo
, true);
467 uint32_t *times
= brw
->shader_time
.bo
->virtual;
469 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
470 brw
->shader_time
.cumulative
[i
] += times
[i
* SHADER_TIME_STRIDE
/ 4];
473 /* Zero the BO out to clear it out for our next collection.
475 memset(times
, 0, brw
->shader_time
.bo
->size
);
476 drm_intel_bo_unmap(brw
->shader_time
.bo
);
480 brw_collect_and_report_shader_time(struct brw_context
*brw
)
482 brw_collect_shader_time(brw
);
484 if (brw
->shader_time
.report_time
== 0 ||
485 get_time() - brw
->shader_time
.report_time
>= 1.0) {
486 brw_report_shader_time(brw
);
487 brw
->shader_time
.report_time
= get_time();
492 * Chooses an index in the shader_time buffer and sets up tracking information
495 * Note that this holds on to references to the underlying programs, which may
496 * change their lifetimes compared to normal operation.
499 brw_get_shader_time_index(struct brw_context
*brw
,
500 struct gl_shader_program
*shader_prog
,
501 struct gl_program
*prog
,
502 enum shader_time_shader_type type
)
504 struct gl_context
*ctx
= &brw
->ctx
;
506 int shader_time_index
= brw
->shader_time
.num_entries
++;
507 assert(shader_time_index
< brw
->shader_time
.max_entries
);
508 brw
->shader_time
.types
[shader_time_index
] = type
;
510 _mesa_reference_shader_program(ctx
,
511 &brw
->shader_time
.shader_programs
[shader_time_index
],
514 _mesa_reference_program(ctx
,
515 &brw
->shader_time
.programs
[shader_time_index
],
518 return shader_time_index
;
522 brw_destroy_shader_time(struct brw_context
*brw
)
524 drm_intel_bo_unreference(brw
->shader_time
.bo
);
525 brw
->shader_time
.bo
= NULL
;