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 assert(!"Unsupported target in brwNewProgram()");
122 static void brwDeleteProgram( struct gl_context
*ctx
,
123 struct gl_program
*prog
)
125 _mesa_delete_program( ctx
, prog
);
130 brwIsProgramNative(struct gl_context
*ctx
,
132 struct gl_program
*prog
)
138 brwProgramStringNotify(struct gl_context
*ctx
,
140 struct gl_program
*prog
)
142 struct brw_context
*brw
= brw_context(ctx
);
145 case GL_FRAGMENT_PROGRAM_ARB
: {
146 struct gl_fragment_program
*fprog
= (struct gl_fragment_program
*) prog
;
147 struct brw_fragment_program
*newFP
= brw_fragment_program(fprog
);
148 const struct brw_fragment_program
*curFP
=
149 brw_fragment_program_const(brw
->fragment_program
);
152 brw
->state
.dirty
.brw
|= BRW_NEW_FRAGMENT_PROGRAM
;
153 newFP
->id
= get_new_program_id(brw
->intelScreen
);
156 case GL_VERTEX_PROGRAM_ARB
: {
157 struct gl_vertex_program
*vprog
= (struct gl_vertex_program
*) prog
;
158 struct brw_vertex_program
*newVP
= brw_vertex_program(vprog
);
159 const struct brw_vertex_program
*curVP
=
160 brw_vertex_program_const(brw
->vertex_program
);
163 brw
->state
.dirty
.brw
|= BRW_NEW_VERTEX_PROGRAM
;
164 if (newVP
->program
.IsPositionInvariant
) {
165 _mesa_insert_mvp_code(ctx
, &newVP
->program
);
167 newVP
->id
= get_new_program_id(brw
->intelScreen
);
169 /* Also tell tnl about it:
171 _tnl_program_string(ctx
, target
, prog
);
176 * driver->ProgramStringNotify is only called for ARB programs, fixed
177 * function vertex programs, and ir_to_mesa (which isn't used by the
178 * i965 back-end). Therefore, even after geometry shaders are added,
179 * this function should only ever be called with a target of
180 * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
182 assert(!"Unexpected target in brwProgramStringNotify");
186 brw_add_texrect_params(prog
);
192 brw_add_texrect_params(struct gl_program
*prog
)
194 for (int texunit
= 0; texunit
< BRW_MAX_TEX_UNIT
; texunit
++) {
195 if (!(prog
->TexturesUsed
[texunit
] & (1 << TEXTURE_RECT_INDEX
)))
198 int tokens
[STATE_LENGTH
] = {
206 _mesa_add_state_reference(prog
->Parameters
, (gl_state_index
*)tokens
);
210 /* Per-thread scratch space is a power-of-two multiple of 1KB. */
212 brw_get_scratch_size(int size
)
216 for (i
= 1024; i
< size
; i
*= 2)
223 brw_get_scratch_bo(struct brw_context
*brw
,
224 drm_intel_bo
**scratch_bo
, int size
)
226 drm_intel_bo
*old_bo
= *scratch_bo
;
228 if (old_bo
&& old_bo
->size
< size
) {
229 drm_intel_bo_unreference(old_bo
);
234 *scratch_bo
= drm_intel_bo_alloc(brw
->bufmgr
, "scratch bo", size
, 4096);
238 void brwInitFragProgFuncs( struct dd_function_table
*functions
)
240 assert(functions
->ProgramStringNotify
== _tnl_program_string
);
242 functions
->BindProgram
= brwBindProgram
;
243 functions
->NewProgram
= brwNewProgram
;
244 functions
->DeleteProgram
= brwDeleteProgram
;
245 functions
->IsProgramNative
= brwIsProgramNative
;
246 functions
->ProgramStringNotify
= brwProgramStringNotify
;
248 functions
->NewShader
= brw_new_shader
;
249 functions
->NewShaderProgram
= brw_new_shader_program
;
250 functions
->LinkShader
= brw_link_shader
;
254 brw_init_shader_time(struct brw_context
*brw
)
256 const int max_entries
= 4096;
257 brw
->shader_time
.bo
= drm_intel_bo_alloc(brw
->bufmgr
, "shader time",
258 max_entries
* SHADER_TIME_STRIDE
,
260 brw
->shader_time
.shader_programs
= rzalloc_array(brw
, struct gl_shader_program
*,
262 brw
->shader_time
.programs
= rzalloc_array(brw
, struct gl_program
*,
264 brw
->shader_time
.types
= rzalloc_array(brw
, enum shader_time_shader_type
,
266 brw
->shader_time
.cumulative
= rzalloc_array(brw
, uint64_t,
268 brw
->shader_time
.max_entries
= max_entries
;
272 compare_time(const void *a
, const void *b
)
274 uint64_t * const *a_val
= a
;
275 uint64_t * const *b_val
= b
;
277 /* We don't just subtract because we're turning the value to an int. */
278 if (**a_val
< **b_val
)
280 else if (**a_val
== **b_val
)
287 get_written_and_reset(struct brw_context
*brw
, int i
,
288 uint64_t *written
, uint64_t *reset
)
290 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
291 assert(type
== ST_VS
|| type
== ST_FS8
|| type
== ST_FS16
);
293 /* Find where we recorded written and reset. */
296 for (wi
= i
; brw
->shader_time
.types
[wi
] != type
+ 1; wi
++)
299 for (ri
= i
; brw
->shader_time
.types
[ri
] != type
+ 2; ri
++)
302 *written
= brw
->shader_time
.cumulative
[wi
];
303 *reset
= brw
->shader_time
.cumulative
[ri
];
307 print_shader_time_line(const char *stage
, const char *name
,
308 int shader_num
, uint64_t time
, uint64_t total
)
310 printf("%-6s%-6s", stage
, name
);
312 if (shader_num
!= -1)
313 printf("%4d: ", shader_num
);
317 printf("%16lld (%7.2f Gcycles) %4.1f%%\n",
319 (double)time
/ 1000000000.0,
320 (double)time
/ total
* 100.0);
324 brw_report_shader_time(struct brw_context
*brw
)
326 if (!brw
->shader_time
.bo
|| !brw
->shader_time
.num_entries
)
329 uint64_t scaled
[brw
->shader_time
.num_entries
];
330 uint64_t *sorted
[brw
->shader_time
.num_entries
];
331 uint64_t total_by_type
[ST_FS16
+ 1];
332 memset(total_by_type
, 0, sizeof(total_by_type
));
334 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
335 uint64_t written
= 0, reset
= 0;
336 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
338 sorted
[i
] = &scaled
[i
];
345 case ST_FS16_WRITTEN
:
347 /* We'll handle these when along with the time. */
354 get_written_and_reset(brw
, i
, &written
, &reset
);
358 /* I sometimes want to print things that aren't the 3 shader times.
359 * Just print the sum in that case.
366 uint64_t time
= brw
->shader_time
.cumulative
[i
];
368 scaled
[i
] = time
/ written
* (written
+ reset
);
377 total_by_type
[type
] += scaled
[i
];
387 printf("No shader time collected yet\n");
391 qsort(sorted
, brw
->shader_time
.num_entries
, sizeof(sorted
[0]), compare_time
);
394 printf("type ID cycles spent %% of total\n");
395 for (int s
= 0; s
< brw
->shader_time
.num_entries
; s
++) {
396 const char *shader_name
;
398 /* Work back from the sorted pointers times to a time to print. */
399 int i
= sorted
[s
] - scaled
;
405 if (brw
->shader_time
.shader_programs
[i
]) {
406 shader_num
= brw
->shader_time
.shader_programs
[i
]->Name
;
408 /* The fixed function fragment shader generates GLSL IR with a Name
409 * of 0, and nothing else does.
411 if (shader_num
== 0 &&
412 (brw
->shader_time
.types
[i
] == ST_FS8
||
413 brw
->shader_time
.types
[i
] == ST_FS16
)) {
417 shader_name
= "glsl";
419 } else if (brw
->shader_time
.programs
[i
]) {
420 shader_num
= brw
->shader_time
.programs
[i
]->Id
;
421 if (shader_num
== 0) {
425 shader_name
= "prog";
428 shader_name
= "other";
431 switch (brw
->shader_time
.types
[i
]) {
446 print_shader_time_line(stage
, shader_name
, shader_num
,
451 print_shader_time_line("total", "vs", -1, total_by_type
[ST_VS
], total
);
452 print_shader_time_line("total", "fs8", -1, total_by_type
[ST_FS8
], total
);
453 print_shader_time_line("total", "fs16", -1, total_by_type
[ST_FS16
], total
);
457 brw_collect_shader_time(struct brw_context
*brw
)
459 if (!brw
->shader_time
.bo
)
462 /* This probably stalls on the last rendering. We could fix that by
463 * delaying reading the reports, but it doesn't look like it's a big
464 * overhead compared to the cost of tracking the time in the first place.
466 drm_intel_bo_map(brw
->shader_time
.bo
, true);
468 uint32_t *times
= brw
->shader_time
.bo
->virtual;
470 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
471 brw
->shader_time
.cumulative
[i
] += times
[i
* SHADER_TIME_STRIDE
/ 4];
474 /* Zero the BO out to clear it out for our next collection.
476 memset(times
, 0, brw
->shader_time
.bo
->size
);
477 drm_intel_bo_unmap(brw
->shader_time
.bo
);
481 brw_collect_and_report_shader_time(struct brw_context
*brw
)
483 brw_collect_shader_time(brw
);
485 if (brw
->shader_time
.report_time
== 0 ||
486 get_time() - brw
->shader_time
.report_time
>= 1.0) {
487 brw_report_shader_time(brw
);
488 brw
->shader_time
.report_time
= get_time();
493 * Chooses an index in the shader_time buffer and sets up tracking information
496 * Note that this holds on to references to the underlying programs, which may
497 * change their lifetimes compared to normal operation.
500 brw_get_shader_time_index(struct brw_context
*brw
,
501 struct gl_shader_program
*shader_prog
,
502 struct gl_program
*prog
,
503 enum shader_time_shader_type type
)
505 struct gl_context
*ctx
= &brw
->ctx
;
507 int shader_time_index
= brw
->shader_time
.num_entries
++;
508 assert(shader_time_index
< brw
->shader_time
.max_entries
);
509 brw
->shader_time
.types
[shader_time_index
] = type
;
511 _mesa_reference_shader_program(ctx
,
512 &brw
->shader_time
.shader_programs
[shader_time_index
],
515 _mesa_reference_program(ctx
,
516 &brw
->shader_time
.programs
[shader_time_index
],
519 return shader_time_index
;
523 brw_destroy_shader_time(struct brw_context
*brw
)
525 drm_intel_bo_unreference(brw
->shader_time
.bo
);
526 brw
->shader_time
.bo
= NULL
;