2 Copyright (C) Intel Corp. 2006. All Rights Reserved.
3 Intel funded Tungsten Graphics 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 <keithw@vmware.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/prog_print.h"
38 #include "program/program.h"
39 #include "program/programopt.h"
41 #include "util/ralloc.h"
44 #include "brw_context.h"
48 get_new_program_id(struct intel_screen
*screen
)
50 static pthread_mutex_t m
= PTHREAD_MUTEX_INITIALIZER
;
51 pthread_mutex_lock(&m
);
52 unsigned id
= screen
->program_id
++;
53 pthread_mutex_unlock(&m
);
57 static struct gl_program
*brwNewProgram( struct gl_context
*ctx
,
61 struct brw_context
*brw
= brw_context(ctx
);
64 case GL_VERTEX_PROGRAM_ARB
: {
65 struct brw_vertex_program
*prog
= CALLOC_STRUCT(brw_vertex_program
);
67 prog
->id
= get_new_program_id(brw
->intelScreen
);
69 return _mesa_init_vertex_program( ctx
, &prog
->program
,
76 case GL_FRAGMENT_PROGRAM_ARB
: {
77 struct brw_fragment_program
*prog
= CALLOC_STRUCT(brw_fragment_program
);
79 prog
->id
= get_new_program_id(brw
->intelScreen
);
81 return _mesa_init_fragment_program( ctx
, &prog
->program
,
88 case MESA_GEOMETRY_PROGRAM
: {
89 struct brw_geometry_program
*prog
= CALLOC_STRUCT(brw_geometry_program
);
91 prog
->id
= get_new_program_id(brw
->intelScreen
);
93 return _mesa_init_geometry_program(ctx
, &prog
->program
, target
, id
);
99 case GL_COMPUTE_PROGRAM_NV
: {
100 struct brw_compute_program
*prog
= CALLOC_STRUCT(brw_compute_program
);
102 prog
->id
= get_new_program_id(brw
->intelScreen
);
104 return _mesa_init_compute_program(ctx
, &prog
->program
, target
, id
);
111 unreachable("Unsupported target in brwNewProgram()");
115 static void brwDeleteProgram( struct gl_context
*ctx
,
116 struct gl_program
*prog
)
118 _mesa_delete_program( ctx
, prog
);
123 brwProgramStringNotify(struct gl_context
*ctx
,
125 struct gl_program
*prog
)
127 struct brw_context
*brw
= brw_context(ctx
);
130 case GL_FRAGMENT_PROGRAM_ARB
: {
131 struct gl_fragment_program
*fprog
= (struct gl_fragment_program
*) prog
;
132 struct brw_fragment_program
*newFP
= brw_fragment_program(fprog
);
133 const struct brw_fragment_program
*curFP
=
134 brw_fragment_program_const(brw
->fragment_program
);
137 brw
->state
.dirty
.brw
|= BRW_NEW_FRAGMENT_PROGRAM
;
138 newFP
->id
= get_new_program_id(brw
->intelScreen
);
141 case GL_VERTEX_PROGRAM_ARB
: {
142 struct gl_vertex_program
*vprog
= (struct gl_vertex_program
*) prog
;
143 struct brw_vertex_program
*newVP
= brw_vertex_program(vprog
);
144 const struct brw_vertex_program
*curVP
=
145 brw_vertex_program_const(brw
->vertex_program
);
148 brw
->state
.dirty
.brw
|= BRW_NEW_VERTEX_PROGRAM
;
149 if (newVP
->program
.IsPositionInvariant
) {
150 _mesa_insert_mvp_code(ctx
, &newVP
->program
);
152 newVP
->id
= get_new_program_id(brw
->intelScreen
);
154 /* Also tell tnl about it:
156 _tnl_program_string(ctx
, target
, prog
);
161 * driver->ProgramStringNotify is only called for ARB programs, fixed
162 * function vertex programs, and ir_to_mesa (which isn't used by the
163 * i965 back-end). Therefore, even after geometry shaders are added,
164 * this function should only ever be called with a target of
165 * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
167 unreachable("Unexpected target in brwProgramStringNotify");
170 brw_add_texrect_params(prog
);
176 brw_add_texrect_params(struct gl_program
*prog
)
178 for (int texunit
= 0; texunit
< BRW_MAX_TEX_UNIT
; texunit
++) {
179 if (!(prog
->TexturesUsed
[texunit
] & (1 << TEXTURE_RECT_INDEX
)))
182 int tokens
[STATE_LENGTH
] = {
190 _mesa_add_state_reference(prog
->Parameters
, (gl_state_index
*)tokens
);
194 /* Per-thread scratch space is a power-of-two multiple of 1KB. */
196 brw_get_scratch_size(int size
)
200 for (i
= 1024; i
< size
; i
*= 2)
207 brw_get_scratch_bo(struct brw_context
*brw
,
208 drm_intel_bo
**scratch_bo
, int size
)
210 drm_intel_bo
*old_bo
= *scratch_bo
;
212 if (old_bo
&& old_bo
->size
< size
) {
213 drm_intel_bo_unreference(old_bo
);
218 *scratch_bo
= drm_intel_bo_alloc(brw
->bufmgr
, "scratch bo", size
, 4096);
222 void brwInitFragProgFuncs( struct dd_function_table
*functions
)
224 assert(functions
->ProgramStringNotify
== _tnl_program_string
);
226 functions
->NewProgram
= brwNewProgram
;
227 functions
->DeleteProgram
= brwDeleteProgram
;
228 functions
->ProgramStringNotify
= brwProgramStringNotify
;
230 functions
->NewShader
= brw_new_shader
;
231 functions
->LinkShader
= brw_link_shader
;
235 brw_init_shader_time(struct brw_context
*brw
)
237 const int max_entries
= 4096;
238 brw
->shader_time
.bo
= drm_intel_bo_alloc(brw
->bufmgr
, "shader time",
239 max_entries
* SHADER_TIME_STRIDE
,
241 brw
->shader_time
.shader_programs
= rzalloc_array(brw
, struct gl_shader_program
*,
243 brw
->shader_time
.programs
= rzalloc_array(brw
, struct gl_program
*,
245 brw
->shader_time
.types
= rzalloc_array(brw
, enum shader_time_shader_type
,
247 brw
->shader_time
.cumulative
= rzalloc_array(brw
, uint64_t,
249 brw
->shader_time
.max_entries
= max_entries
;
253 compare_time(const void *a
, const void *b
)
255 uint64_t * const *a_val
= a
;
256 uint64_t * const *b_val
= b
;
258 /* We don't just subtract because we're turning the value to an int. */
259 if (**a_val
< **b_val
)
261 else if (**a_val
== **b_val
)
268 get_written_and_reset(struct brw_context
*brw
, int i
,
269 uint64_t *written
, uint64_t *reset
)
271 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
272 assert(type
== ST_VS
|| type
== ST_GS
|| type
== ST_FS8
|| type
== ST_FS16
);
274 /* Find where we recorded written and reset. */
277 for (wi
= i
; brw
->shader_time
.types
[wi
] != type
+ 1; wi
++)
280 for (ri
= i
; brw
->shader_time
.types
[ri
] != type
+ 2; ri
++)
283 *written
= brw
->shader_time
.cumulative
[wi
];
284 *reset
= brw
->shader_time
.cumulative
[ri
];
288 print_shader_time_line(const char *stage
, const char *name
,
289 int shader_num
, uint64_t time
, uint64_t total
)
291 fprintf(stderr
, "%-6s%-18s", stage
, name
);
293 if (shader_num
!= -1)
294 fprintf(stderr
, "%4d: ", shader_num
);
296 fprintf(stderr
, " : ");
298 fprintf(stderr
, "%16lld (%7.2f Gcycles) %4.1f%%\n",
300 (double)time
/ 1000000000.0,
301 (double)time
/ total
* 100.0);
305 brw_report_shader_time(struct brw_context
*brw
)
307 if (!brw
->shader_time
.bo
|| !brw
->shader_time
.num_entries
)
310 uint64_t scaled
[brw
->shader_time
.num_entries
];
311 uint64_t *sorted
[brw
->shader_time
.num_entries
];
312 uint64_t total_by_type
[ST_FS16
+ 1];
313 memset(total_by_type
, 0, sizeof(total_by_type
));
315 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
316 uint64_t written
= 0, reset
= 0;
317 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
319 sorted
[i
] = &scaled
[i
];
328 case ST_FS16_WRITTEN
:
330 /* We'll handle these when along with the time. */
338 get_written_and_reset(brw
, i
, &written
, &reset
);
342 /* I sometimes want to print things that aren't the 3 shader times.
343 * Just print the sum in that case.
350 uint64_t time
= brw
->shader_time
.cumulative
[i
];
352 scaled
[i
] = time
/ written
* (written
+ reset
);
362 total_by_type
[type
] += scaled
[i
];
372 fprintf(stderr
, "No shader time collected yet\n");
376 qsort(sorted
, brw
->shader_time
.num_entries
, sizeof(sorted
[0]), compare_time
);
378 fprintf(stderr
, "\n");
379 fprintf(stderr
, "type ID cycles spent %% of total\n");
380 for (int s
= 0; s
< brw
->shader_time
.num_entries
; s
++) {
381 const char *shader_name
;
383 /* Work back from the sorted pointers times to a time to print. */
384 int i
= sorted
[s
] - scaled
;
385 struct gl_shader_program
*prog
= brw
->shader_time
.shader_programs
[i
];
392 shader_num
= prog
->Name
;
394 /* The fixed function fragment shader generates GLSL IR with a Name
395 * of 0, and nothing else does.
398 shader_name
= prog
->Label
;
399 } else if (shader_num
== 0 &&
400 (brw
->shader_time
.types
[i
] == ST_FS8
||
401 brw
->shader_time
.types
[i
] == ST_FS16
)) {
405 shader_name
= "glsl";
407 } else if (brw
->shader_time
.programs
[i
]) {
408 shader_num
= brw
->shader_time
.programs
[i
]->Id
;
409 if (shader_num
== 0) {
413 shader_name
= "prog";
416 shader_name
= "other";
419 switch (brw
->shader_time
.types
[i
]) {
437 print_shader_time_line(stage
, shader_name
, shader_num
,
441 fprintf(stderr
, "\n");
442 print_shader_time_line("total", "vs", -1, total_by_type
[ST_VS
], total
);
443 print_shader_time_line("total", "gs", -1, total_by_type
[ST_GS
], total
);
444 print_shader_time_line("total", "fs8", -1, total_by_type
[ST_FS8
], total
);
445 print_shader_time_line("total", "fs16", -1, total_by_type
[ST_FS16
], total
);
449 brw_collect_shader_time(struct brw_context
*brw
)
451 if (!brw
->shader_time
.bo
)
454 /* This probably stalls on the last rendering. We could fix that by
455 * delaying reading the reports, but it doesn't look like it's a big
456 * overhead compared to the cost of tracking the time in the first place.
458 drm_intel_bo_map(brw
->shader_time
.bo
, true);
460 uint32_t *times
= brw
->shader_time
.bo
->virtual;
462 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
463 brw
->shader_time
.cumulative
[i
] += times
[i
* SHADER_TIME_STRIDE
/ 4];
466 /* Zero the BO out to clear it out for our next collection.
468 memset(times
, 0, brw
->shader_time
.bo
->size
);
469 drm_intel_bo_unmap(brw
->shader_time
.bo
);
473 brw_collect_and_report_shader_time(struct brw_context
*brw
)
475 brw_collect_shader_time(brw
);
477 if (brw
->shader_time
.report_time
== 0 ||
478 get_time() - brw
->shader_time
.report_time
>= 1.0) {
479 brw_report_shader_time(brw
);
480 brw
->shader_time
.report_time
= get_time();
485 * Chooses an index in the shader_time buffer and sets up tracking information
488 * Note that this holds on to references to the underlying programs, which may
489 * change their lifetimes compared to normal operation.
492 brw_get_shader_time_index(struct brw_context
*brw
,
493 struct gl_shader_program
*shader_prog
,
494 struct gl_program
*prog
,
495 enum shader_time_shader_type type
)
497 struct gl_context
*ctx
= &brw
->ctx
;
499 int shader_time_index
= brw
->shader_time
.num_entries
++;
500 assert(shader_time_index
< brw
->shader_time
.max_entries
);
501 brw
->shader_time
.types
[shader_time_index
] = type
;
503 _mesa_reference_shader_program(ctx
,
504 &brw
->shader_time
.shader_programs
[shader_time_index
],
507 _mesa_reference_program(ctx
,
508 &brw
->shader_time
.programs
[shader_time_index
],
511 return shader_time_index
;
515 brw_destroy_shader_time(struct brw_context
*brw
)
517 drm_intel_bo_unreference(brw
->shader_time
.bo
);
518 brw
->shader_time
.bo
= NULL
;
522 brw_mark_surface_used(struct brw_stage_prog_data
*prog_data
,
525 assert(surf_index
< BRW_MAX_SURFACES
);
527 prog_data
->binding_table
.size_bytes
=
528 MAX2(prog_data
->binding_table
.size_bytes
, (surf_index
+ 1) * 4);
532 brw_stage_prog_data_compare(const struct brw_stage_prog_data
*a
,
533 const struct brw_stage_prog_data
*b
)
535 /* Compare all the struct up to the pointers. */
536 if (memcmp(a
, b
, offsetof(struct brw_stage_prog_data
, param
)))
539 if (memcmp(a
->param
, b
->param
, a
->nr_params
* sizeof(void *)))
542 if (memcmp(a
->pull_param
, b
->pull_param
, a
->nr_pull_params
* sizeof(void *)))
549 brw_stage_prog_data_free(const void *p
)
551 struct brw_stage_prog_data
*prog_data
= (struct brw_stage_prog_data
*)p
;
553 ralloc_free(prog_data
->param
);
554 ralloc_free(prog_data
->pull_param
);
558 brw_dump_ir(struct brw_context
*brw
, const char *stage
,
559 struct gl_shader_program
*shader_prog
,
560 struct gl_shader
*shader
, struct gl_program
*prog
)
564 "GLSL IR for native %s shader %d:\n", stage
, shader_prog
->Name
);
565 _mesa_print_ir(stderr
, shader
->ir
, NULL
);
566 fprintf(stderr
, "\n\n");
568 fprintf(stderr
, "ARB_%s_program %d ir for native %s shader\n",
569 stage
, prog
->Id
, stage
);
570 _mesa_print_program(prog
);