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/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
);
116 case GL_COMPUTE_PROGRAM_NV
: {
117 struct brw_compute_program
*prog
= CALLOC_STRUCT(brw_compute_program
);
119 prog
->id
= get_new_program_id(brw
->intelScreen
);
121 return _mesa_init_compute_program(ctx
, &prog
->program
, target
, id
);
128 assert(!"Unsupported target in brwNewProgram()");
133 static void brwDeleteProgram( struct gl_context
*ctx
,
134 struct gl_program
*prog
)
136 _mesa_delete_program( ctx
, prog
);
141 brwIsProgramNative(struct gl_context
*ctx
,
143 struct gl_program
*prog
)
149 brwProgramStringNotify(struct gl_context
*ctx
,
151 struct gl_program
*prog
)
153 struct brw_context
*brw
= brw_context(ctx
);
156 case GL_FRAGMENT_PROGRAM_ARB
: {
157 struct gl_fragment_program
*fprog
= (struct gl_fragment_program
*) prog
;
158 struct brw_fragment_program
*newFP
= brw_fragment_program(fprog
);
159 const struct brw_fragment_program
*curFP
=
160 brw_fragment_program_const(brw
->fragment_program
);
163 brw
->state
.dirty
.brw
|= BRW_NEW_FRAGMENT_PROGRAM
;
164 newFP
->id
= get_new_program_id(brw
->intelScreen
);
167 case GL_VERTEX_PROGRAM_ARB
: {
168 struct gl_vertex_program
*vprog
= (struct gl_vertex_program
*) prog
;
169 struct brw_vertex_program
*newVP
= brw_vertex_program(vprog
);
170 const struct brw_vertex_program
*curVP
=
171 brw_vertex_program_const(brw
->vertex_program
);
174 brw
->state
.dirty
.brw
|= BRW_NEW_VERTEX_PROGRAM
;
175 if (newVP
->program
.IsPositionInvariant
) {
176 _mesa_insert_mvp_code(ctx
, &newVP
->program
);
178 newVP
->id
= get_new_program_id(brw
->intelScreen
);
180 /* Also tell tnl about it:
182 _tnl_program_string(ctx
, target
, prog
);
187 * driver->ProgramStringNotify is only called for ARB programs, fixed
188 * function vertex programs, and ir_to_mesa (which isn't used by the
189 * i965 back-end). Therefore, even after geometry shaders are added,
190 * this function should only ever be called with a target of
191 * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
193 assert(!"Unexpected target in brwProgramStringNotify");
197 brw_add_texrect_params(prog
);
203 brw_add_texrect_params(struct gl_program
*prog
)
205 for (int texunit
= 0; texunit
< BRW_MAX_TEX_UNIT
; texunit
++) {
206 if (!(prog
->TexturesUsed
[texunit
] & (1 << TEXTURE_RECT_INDEX
)))
209 int tokens
[STATE_LENGTH
] = {
217 _mesa_add_state_reference(prog
->Parameters
, (gl_state_index
*)tokens
);
221 /* Per-thread scratch space is a power-of-two multiple of 1KB. */
223 brw_get_scratch_size(int size
)
227 for (i
= 1024; i
< size
; i
*= 2)
234 brw_get_scratch_bo(struct brw_context
*brw
,
235 drm_intel_bo
**scratch_bo
, int size
)
237 drm_intel_bo
*old_bo
= *scratch_bo
;
239 if (old_bo
&& old_bo
->size
< size
) {
240 drm_intel_bo_unreference(old_bo
);
245 *scratch_bo
= drm_intel_bo_alloc(brw
->bufmgr
, "scratch bo", size
, 4096);
249 void brwInitFragProgFuncs( struct dd_function_table
*functions
)
251 assert(functions
->ProgramStringNotify
== _tnl_program_string
);
253 functions
->BindProgram
= brwBindProgram
;
254 functions
->NewProgram
= brwNewProgram
;
255 functions
->DeleteProgram
= brwDeleteProgram
;
256 functions
->IsProgramNative
= brwIsProgramNative
;
257 functions
->ProgramStringNotify
= brwProgramStringNotify
;
259 functions
->NewShader
= brw_new_shader
;
260 functions
->NewShaderProgram
= brw_new_shader_program
;
261 functions
->LinkShader
= brw_link_shader
;
265 brw_init_shader_time(struct brw_context
*brw
)
267 const int max_entries
= 4096;
268 brw
->shader_time
.bo
= drm_intel_bo_alloc(brw
->bufmgr
, "shader time",
269 max_entries
* SHADER_TIME_STRIDE
,
271 brw
->shader_time
.shader_programs
= rzalloc_array(brw
, struct gl_shader_program
*,
273 brw
->shader_time
.programs
= rzalloc_array(brw
, struct gl_program
*,
275 brw
->shader_time
.types
= rzalloc_array(brw
, enum shader_time_shader_type
,
277 brw
->shader_time
.cumulative
= rzalloc_array(brw
, uint64_t,
279 brw
->shader_time
.max_entries
= max_entries
;
283 compare_time(const void *a
, const void *b
)
285 uint64_t * const *a_val
= a
;
286 uint64_t * const *b_val
= b
;
288 /* We don't just subtract because we're turning the value to an int. */
289 if (**a_val
< **b_val
)
291 else if (**a_val
== **b_val
)
298 get_written_and_reset(struct brw_context
*brw
, int i
,
299 uint64_t *written
, uint64_t *reset
)
301 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
302 assert(type
== ST_VS
|| type
== ST_GS
|| type
== ST_FS8
|| type
== ST_FS16
);
304 /* Find where we recorded written and reset. */
307 for (wi
= i
; brw
->shader_time
.types
[wi
] != type
+ 1; wi
++)
310 for (ri
= i
; brw
->shader_time
.types
[ri
] != type
+ 2; ri
++)
313 *written
= brw
->shader_time
.cumulative
[wi
];
314 *reset
= brw
->shader_time
.cumulative
[ri
];
318 print_shader_time_line(const char *stage
, const char *name
,
319 int shader_num
, uint64_t time
, uint64_t total
)
321 printf("%-6s%-6s", stage
, name
);
323 if (shader_num
!= -1)
324 printf("%4d: ", shader_num
);
328 printf("%16lld (%7.2f Gcycles) %4.1f%%\n",
330 (double)time
/ 1000000000.0,
331 (double)time
/ total
* 100.0);
335 brw_report_shader_time(struct brw_context
*brw
)
337 if (!brw
->shader_time
.bo
|| !brw
->shader_time
.num_entries
)
340 uint64_t scaled
[brw
->shader_time
.num_entries
];
341 uint64_t *sorted
[brw
->shader_time
.num_entries
];
342 uint64_t total_by_type
[ST_FS16
+ 1];
343 memset(total_by_type
, 0, sizeof(total_by_type
));
345 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
346 uint64_t written
= 0, reset
= 0;
347 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
349 sorted
[i
] = &scaled
[i
];
358 case ST_FS16_WRITTEN
:
360 /* We'll handle these when along with the time. */
368 get_written_and_reset(brw
, i
, &written
, &reset
);
372 /* I sometimes want to print things that aren't the 3 shader times.
373 * Just print the sum in that case.
380 uint64_t time
= brw
->shader_time
.cumulative
[i
];
382 scaled
[i
] = time
/ written
* (written
+ reset
);
392 total_by_type
[type
] += scaled
[i
];
402 printf("No shader time collected yet\n");
406 qsort(sorted
, brw
->shader_time
.num_entries
, sizeof(sorted
[0]), compare_time
);
409 printf("type ID cycles spent %% of total\n");
410 for (int s
= 0; s
< brw
->shader_time
.num_entries
; s
++) {
411 const char *shader_name
;
413 /* Work back from the sorted pointers times to a time to print. */
414 int i
= sorted
[s
] - scaled
;
420 if (brw
->shader_time
.shader_programs
[i
]) {
421 shader_num
= brw
->shader_time
.shader_programs
[i
]->Name
;
423 /* The fixed function fragment shader generates GLSL IR with a Name
424 * of 0, and nothing else does.
426 if (shader_num
== 0 &&
427 (brw
->shader_time
.types
[i
] == ST_FS8
||
428 brw
->shader_time
.types
[i
] == ST_FS16
)) {
432 shader_name
= "glsl";
434 } else if (brw
->shader_time
.programs
[i
]) {
435 shader_num
= brw
->shader_time
.programs
[i
]->Id
;
436 if (shader_num
== 0) {
440 shader_name
= "prog";
443 shader_name
= "other";
446 switch (brw
->shader_time
.types
[i
]) {
464 print_shader_time_line(stage
, shader_name
, shader_num
,
469 print_shader_time_line("total", "vs", -1, total_by_type
[ST_VS
], total
);
470 print_shader_time_line("total", "gs", -1, total_by_type
[ST_GS
], total
);
471 print_shader_time_line("total", "fs8", -1, total_by_type
[ST_FS8
], total
);
472 print_shader_time_line("total", "fs16", -1, total_by_type
[ST_FS16
], total
);
476 brw_collect_shader_time(struct brw_context
*brw
)
478 if (!brw
->shader_time
.bo
)
481 /* This probably stalls on the last rendering. We could fix that by
482 * delaying reading the reports, but it doesn't look like it's a big
483 * overhead compared to the cost of tracking the time in the first place.
485 drm_intel_bo_map(brw
->shader_time
.bo
, true);
487 uint32_t *times
= brw
->shader_time
.bo
->virtual;
489 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
490 brw
->shader_time
.cumulative
[i
] += times
[i
* SHADER_TIME_STRIDE
/ 4];
493 /* Zero the BO out to clear it out for our next collection.
495 memset(times
, 0, brw
->shader_time
.bo
->size
);
496 drm_intel_bo_unmap(brw
->shader_time
.bo
);
500 brw_collect_and_report_shader_time(struct brw_context
*brw
)
502 brw_collect_shader_time(brw
);
504 if (brw
->shader_time
.report_time
== 0 ||
505 get_time() - brw
->shader_time
.report_time
>= 1.0) {
506 brw_report_shader_time(brw
);
507 brw
->shader_time
.report_time
= get_time();
512 * Chooses an index in the shader_time buffer and sets up tracking information
515 * Note that this holds on to references to the underlying programs, which may
516 * change their lifetimes compared to normal operation.
519 brw_get_shader_time_index(struct brw_context
*brw
,
520 struct gl_shader_program
*shader_prog
,
521 struct gl_program
*prog
,
522 enum shader_time_shader_type type
)
524 struct gl_context
*ctx
= &brw
->ctx
;
526 int shader_time_index
= brw
->shader_time
.num_entries
++;
527 assert(shader_time_index
< brw
->shader_time
.max_entries
);
528 brw
->shader_time
.types
[shader_time_index
] = type
;
530 _mesa_reference_shader_program(ctx
,
531 &brw
->shader_time
.shader_programs
[shader_time_index
],
534 _mesa_reference_program(ctx
,
535 &brw
->shader_time
.programs
[shader_time_index
],
538 return shader_time_index
;
542 brw_destroy_shader_time(struct brw_context
*brw
)
544 drm_intel_bo_unreference(brw
->shader_time
.bo
);
545 brw
->shader_time
.bo
= NULL
;
549 brw_mark_surface_used(struct brw_stage_prog_data
*prog_data
,
552 assert(surf_index
< BRW_MAX_SURFACES
);
554 prog_data
->binding_table
.size_bytes
=
555 MAX2(prog_data
->binding_table
.size_bytes
, (surf_index
+ 1) * 4);
559 brw_stage_prog_data_compare(const struct brw_stage_prog_data
*a
,
560 const struct brw_stage_prog_data
*b
)
562 /* Compare all the struct up to the pointers. */
563 if (memcmp(a
, b
, offsetof(struct brw_stage_prog_data
, param
)))
566 if (memcmp(a
->param
, b
->param
, a
->nr_params
* sizeof(void *)))
569 if (memcmp(a
->pull_param
, b
->pull_param
, a
->nr_pull_params
* sizeof(void *)))
576 brw_stage_prog_data_free(const void *p
)
578 struct brw_stage_prog_data
*prog_data
= (struct brw_stage_prog_data
*)p
;
580 ralloc_free(prog_data
->param
);
581 ralloc_free(prog_data
->pull_param
);