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 "glsl/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 void brwBindProgram( struct gl_context
*ctx
,
59 struct gl_program
*prog
)
61 struct brw_context
*brw
= brw_context(ctx
);
64 case GL_VERTEX_PROGRAM_ARB
:
65 brw
->state
.dirty
.brw
|= BRW_NEW_VERTEX_PROGRAM
;
67 case MESA_GEOMETRY_PROGRAM
:
68 brw
->state
.dirty
.brw
|= BRW_NEW_GEOMETRY_PROGRAM
;
70 case GL_FRAGMENT_PROGRAM_ARB
:
71 brw
->state
.dirty
.brw
|= BRW_NEW_FRAGMENT_PROGRAM
;
76 static struct gl_program
*brwNewProgram( struct gl_context
*ctx
,
80 struct brw_context
*brw
= brw_context(ctx
);
83 case GL_VERTEX_PROGRAM_ARB
: {
84 struct brw_vertex_program
*prog
= CALLOC_STRUCT(brw_vertex_program
);
86 prog
->id
= get_new_program_id(brw
->intelScreen
);
88 return _mesa_init_vertex_program( ctx
, &prog
->program
,
95 case GL_FRAGMENT_PROGRAM_ARB
: {
96 struct brw_fragment_program
*prog
= CALLOC_STRUCT(brw_fragment_program
);
98 prog
->id
= get_new_program_id(brw
->intelScreen
);
100 return _mesa_init_fragment_program( ctx
, &prog
->program
,
107 case MESA_GEOMETRY_PROGRAM
: {
108 struct brw_geometry_program
*prog
= CALLOC_STRUCT(brw_geometry_program
);
110 prog
->id
= get_new_program_id(brw
->intelScreen
);
112 return _mesa_init_geometry_program(ctx
, &prog
->program
, target
, id
);
118 case GL_COMPUTE_PROGRAM_NV
: {
119 struct brw_compute_program
*prog
= CALLOC_STRUCT(brw_compute_program
);
121 prog
->id
= get_new_program_id(brw
->intelScreen
);
123 return _mesa_init_compute_program(ctx
, &prog
->program
, target
, id
);
130 assert(!"Unsupported target in brwNewProgram()");
135 static void brwDeleteProgram( struct gl_context
*ctx
,
136 struct gl_program
*prog
)
138 _mesa_delete_program( ctx
, prog
);
143 brwIsProgramNative(struct gl_context
*ctx
,
145 struct gl_program
*prog
)
151 brwProgramStringNotify(struct gl_context
*ctx
,
153 struct gl_program
*prog
)
155 struct brw_context
*brw
= brw_context(ctx
);
158 case GL_FRAGMENT_PROGRAM_ARB
: {
159 struct gl_fragment_program
*fprog
= (struct gl_fragment_program
*) prog
;
160 struct brw_fragment_program
*newFP
= brw_fragment_program(fprog
);
161 const struct brw_fragment_program
*curFP
=
162 brw_fragment_program_const(brw
->fragment_program
);
165 brw
->state
.dirty
.brw
|= BRW_NEW_FRAGMENT_PROGRAM
;
166 newFP
->id
= get_new_program_id(brw
->intelScreen
);
169 case GL_VERTEX_PROGRAM_ARB
: {
170 struct gl_vertex_program
*vprog
= (struct gl_vertex_program
*) prog
;
171 struct brw_vertex_program
*newVP
= brw_vertex_program(vprog
);
172 const struct brw_vertex_program
*curVP
=
173 brw_vertex_program_const(brw
->vertex_program
);
176 brw
->state
.dirty
.brw
|= BRW_NEW_VERTEX_PROGRAM
;
177 if (newVP
->program
.IsPositionInvariant
) {
178 _mesa_insert_mvp_code(ctx
, &newVP
->program
);
180 newVP
->id
= get_new_program_id(brw
->intelScreen
);
182 /* Also tell tnl about it:
184 _tnl_program_string(ctx
, target
, prog
);
189 * driver->ProgramStringNotify is only called for ARB programs, fixed
190 * function vertex programs, and ir_to_mesa (which isn't used by the
191 * i965 back-end). Therefore, even after geometry shaders are added,
192 * this function should only ever be called with a target of
193 * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
195 assert(!"Unexpected target in brwProgramStringNotify");
199 brw_add_texrect_params(prog
);
205 brw_add_texrect_params(struct gl_program
*prog
)
207 for (int texunit
= 0; texunit
< BRW_MAX_TEX_UNIT
; texunit
++) {
208 if (!(prog
->TexturesUsed
[texunit
] & (1 << TEXTURE_RECT_INDEX
)))
211 int tokens
[STATE_LENGTH
] = {
219 _mesa_add_state_reference(prog
->Parameters
, (gl_state_index
*)tokens
);
223 /* Per-thread scratch space is a power-of-two multiple of 1KB. */
225 brw_get_scratch_size(int size
)
229 for (i
= 1024; i
< size
; i
*= 2)
236 brw_get_scratch_bo(struct brw_context
*brw
,
237 drm_intel_bo
**scratch_bo
, int size
)
239 drm_intel_bo
*old_bo
= *scratch_bo
;
241 if (old_bo
&& old_bo
->size
< size
) {
242 drm_intel_bo_unreference(old_bo
);
247 *scratch_bo
= drm_intel_bo_alloc(brw
->bufmgr
, "scratch bo", size
, 4096);
251 void brwInitFragProgFuncs( struct dd_function_table
*functions
)
253 assert(functions
->ProgramStringNotify
== _tnl_program_string
);
255 functions
->BindProgram
= brwBindProgram
;
256 functions
->NewProgram
= brwNewProgram
;
257 functions
->DeleteProgram
= brwDeleteProgram
;
258 functions
->IsProgramNative
= brwIsProgramNative
;
259 functions
->ProgramStringNotify
= brwProgramStringNotify
;
261 functions
->NewShader
= brw_new_shader
;
262 functions
->NewShaderProgram
= brw_new_shader_program
;
263 functions
->LinkShader
= brw_link_shader
;
267 brw_init_shader_time(struct brw_context
*brw
)
269 const int max_entries
= 4096;
270 brw
->shader_time
.bo
= drm_intel_bo_alloc(brw
->bufmgr
, "shader time",
271 max_entries
* SHADER_TIME_STRIDE
,
273 brw
->shader_time
.shader_programs
= rzalloc_array(brw
, struct gl_shader_program
*,
275 brw
->shader_time
.programs
= rzalloc_array(brw
, struct gl_program
*,
277 brw
->shader_time
.types
= rzalloc_array(brw
, enum shader_time_shader_type
,
279 brw
->shader_time
.cumulative
= rzalloc_array(brw
, uint64_t,
281 brw
->shader_time
.max_entries
= max_entries
;
285 compare_time(const void *a
, const void *b
)
287 uint64_t * const *a_val
= a
;
288 uint64_t * const *b_val
= b
;
290 /* We don't just subtract because we're turning the value to an int. */
291 if (**a_val
< **b_val
)
293 else if (**a_val
== **b_val
)
300 get_written_and_reset(struct brw_context
*brw
, int i
,
301 uint64_t *written
, uint64_t *reset
)
303 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
304 assert(type
== ST_VS
|| type
== ST_GS
|| type
== ST_FS8
|| type
== ST_FS16
);
306 /* Find where we recorded written and reset. */
309 for (wi
= i
; brw
->shader_time
.types
[wi
] != type
+ 1; wi
++)
312 for (ri
= i
; brw
->shader_time
.types
[ri
] != type
+ 2; ri
++)
315 *written
= brw
->shader_time
.cumulative
[wi
];
316 *reset
= brw
->shader_time
.cumulative
[ri
];
320 print_shader_time_line(const char *stage
, const char *name
,
321 int shader_num
, uint64_t time
, uint64_t total
)
323 fprintf(stderr
, "%-6s%-18s", stage
, name
);
325 if (shader_num
!= -1)
326 fprintf(stderr
, "%4d: ", shader_num
);
328 fprintf(stderr
, " : ");
330 fprintf(stderr
, "%16lld (%7.2f Gcycles) %4.1f%%\n",
332 (double)time
/ 1000000000.0,
333 (double)time
/ total
* 100.0);
337 brw_report_shader_time(struct brw_context
*brw
)
339 if (!brw
->shader_time
.bo
|| !brw
->shader_time
.num_entries
)
342 uint64_t scaled
[brw
->shader_time
.num_entries
];
343 uint64_t *sorted
[brw
->shader_time
.num_entries
];
344 uint64_t total_by_type
[ST_FS16
+ 1];
345 memset(total_by_type
, 0, sizeof(total_by_type
));
347 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
348 uint64_t written
= 0, reset
= 0;
349 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
351 sorted
[i
] = &scaled
[i
];
360 case ST_FS16_WRITTEN
:
362 /* We'll handle these when along with the time. */
370 get_written_and_reset(brw
, i
, &written
, &reset
);
374 /* I sometimes want to print things that aren't the 3 shader times.
375 * Just print the sum in that case.
382 uint64_t time
= brw
->shader_time
.cumulative
[i
];
384 scaled
[i
] = time
/ written
* (written
+ reset
);
394 total_by_type
[type
] += scaled
[i
];
404 fprintf(stderr
, "No shader time collected yet\n");
408 qsort(sorted
, brw
->shader_time
.num_entries
, sizeof(sorted
[0]), compare_time
);
410 fprintf(stderr
, "\n");
411 fprintf(stderr
, "type ID cycles spent %% of total\n");
412 for (int s
= 0; s
< brw
->shader_time
.num_entries
; s
++) {
413 const char *shader_name
;
415 /* Work back from the sorted pointers times to a time to print. */
416 int i
= sorted
[s
] - scaled
;
417 struct gl_shader_program
*prog
= brw
->shader_time
.shader_programs
[i
];
424 shader_num
= prog
->Name
;
426 /* The fixed function fragment shader generates GLSL IR with a Name
427 * of 0, and nothing else does.
430 shader_name
= prog
->Label
;
431 } else if (shader_num
== 0 &&
432 (brw
->shader_time
.types
[i
] == ST_FS8
||
433 brw
->shader_time
.types
[i
] == ST_FS16
)) {
437 shader_name
= "glsl";
439 } else if (brw
->shader_time
.programs
[i
]) {
440 shader_num
= brw
->shader_time
.programs
[i
]->Id
;
441 if (shader_num
== 0) {
445 shader_name
= "prog";
448 shader_name
= "other";
451 switch (brw
->shader_time
.types
[i
]) {
469 print_shader_time_line(stage
, shader_name
, shader_num
,
473 fprintf(stderr
, "\n");
474 print_shader_time_line("total", "vs", -1, total_by_type
[ST_VS
], total
);
475 print_shader_time_line("total", "gs", -1, total_by_type
[ST_GS
], total
);
476 print_shader_time_line("total", "fs8", -1, total_by_type
[ST_FS8
], total
);
477 print_shader_time_line("total", "fs16", -1, total_by_type
[ST_FS16
], total
);
481 brw_collect_shader_time(struct brw_context
*brw
)
483 if (!brw
->shader_time
.bo
)
486 /* This probably stalls on the last rendering. We could fix that by
487 * delaying reading the reports, but it doesn't look like it's a big
488 * overhead compared to the cost of tracking the time in the first place.
490 drm_intel_bo_map(brw
->shader_time
.bo
, true);
492 uint32_t *times
= brw
->shader_time
.bo
->virtual;
494 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
495 brw
->shader_time
.cumulative
[i
] += times
[i
* SHADER_TIME_STRIDE
/ 4];
498 /* Zero the BO out to clear it out for our next collection.
500 memset(times
, 0, brw
->shader_time
.bo
->size
);
501 drm_intel_bo_unmap(brw
->shader_time
.bo
);
505 brw_collect_and_report_shader_time(struct brw_context
*brw
)
507 brw_collect_shader_time(brw
);
509 if (brw
->shader_time
.report_time
== 0 ||
510 get_time() - brw
->shader_time
.report_time
>= 1.0) {
511 brw_report_shader_time(brw
);
512 brw
->shader_time
.report_time
= get_time();
517 * Chooses an index in the shader_time buffer and sets up tracking information
520 * Note that this holds on to references to the underlying programs, which may
521 * change their lifetimes compared to normal operation.
524 brw_get_shader_time_index(struct brw_context
*brw
,
525 struct gl_shader_program
*shader_prog
,
526 struct gl_program
*prog
,
527 enum shader_time_shader_type type
)
529 struct gl_context
*ctx
= &brw
->ctx
;
531 int shader_time_index
= brw
->shader_time
.num_entries
++;
532 assert(shader_time_index
< brw
->shader_time
.max_entries
);
533 brw
->shader_time
.types
[shader_time_index
] = type
;
535 _mesa_reference_shader_program(ctx
,
536 &brw
->shader_time
.shader_programs
[shader_time_index
],
539 _mesa_reference_program(ctx
,
540 &brw
->shader_time
.programs
[shader_time_index
],
543 return shader_time_index
;
547 brw_destroy_shader_time(struct brw_context
*brw
)
549 drm_intel_bo_unreference(brw
->shader_time
.bo
);
550 brw
->shader_time
.bo
= NULL
;
554 brw_mark_surface_used(struct brw_stage_prog_data
*prog_data
,
557 assert(surf_index
< BRW_MAX_SURFACES
);
559 prog_data
->binding_table
.size_bytes
=
560 MAX2(prog_data
->binding_table
.size_bytes
, (surf_index
+ 1) * 4);
564 brw_stage_prog_data_compare(const struct brw_stage_prog_data
*a
,
565 const struct brw_stage_prog_data
*b
)
567 /* Compare all the struct up to the pointers. */
568 if (memcmp(a
, b
, offsetof(struct brw_stage_prog_data
, param
)))
571 if (memcmp(a
->param
, b
->param
, a
->nr_params
* sizeof(void *)))
574 if (memcmp(a
->pull_param
, b
->pull_param
, a
->nr_pull_params
* sizeof(void *)))
581 brw_stage_prog_data_free(const void *p
)
583 struct brw_stage_prog_data
*prog_data
= (struct brw_stage_prog_data
*)p
;
585 ralloc_free(prog_data
->param
);
586 ralloc_free(prog_data
->pull_param
);
590 brw_dump_ir(struct brw_context
*brw
, const char *stage
,
591 struct gl_shader_program
*shader_prog
,
592 struct gl_shader
*shader
, struct gl_program
*prog
)
596 "GLSL IR for native %s shader %d:\n", stage
, shader_prog
->Name
);
597 _mesa_print_ir(stderr
, shader
->ir
, NULL
);
598 fprintf(stderr
, "\n\n");
600 fprintf(stderr
, "ARB_%s_program %d ir for native %s shader\n",
601 stage
, prog
->Id
, stage
);
602 _mesa_print_program(prog
);