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"
45 #include "brw_shader.h"
48 #include "intel_batchbuffer.h"
51 get_new_program_id(struct intel_screen
*screen
)
53 static pthread_mutex_t m
= PTHREAD_MUTEX_INITIALIZER
;
54 pthread_mutex_lock(&m
);
55 unsigned id
= screen
->program_id
++;
56 pthread_mutex_unlock(&m
);
60 static struct gl_program
*brwNewProgram( struct gl_context
*ctx
,
64 struct brw_context
*brw
= brw_context(ctx
);
67 case GL_VERTEX_PROGRAM_ARB
: {
68 struct brw_vertex_program
*prog
= CALLOC_STRUCT(brw_vertex_program
);
70 prog
->id
= get_new_program_id(brw
->intelScreen
);
72 return _mesa_init_vertex_program( ctx
, &prog
->program
,
79 case GL_FRAGMENT_PROGRAM_ARB
: {
80 struct brw_fragment_program
*prog
= CALLOC_STRUCT(brw_fragment_program
);
82 prog
->id
= get_new_program_id(brw
->intelScreen
);
84 return _mesa_init_fragment_program( ctx
, &prog
->program
,
91 case GL_GEOMETRY_PROGRAM_NV
: {
92 struct brw_geometry_program
*prog
= CALLOC_STRUCT(brw_geometry_program
);
94 prog
->id
= get_new_program_id(brw
->intelScreen
);
96 return _mesa_init_geometry_program(ctx
, &prog
->program
, target
, id
);
102 case GL_COMPUTE_PROGRAM_NV
: {
103 struct brw_compute_program
*prog
= CALLOC_STRUCT(brw_compute_program
);
105 prog
->id
= get_new_program_id(brw
->intelScreen
);
107 return _mesa_init_compute_program(ctx
, &prog
->program
, target
, id
);
114 unreachable("Unsupported target in brwNewProgram()");
118 static void brwDeleteProgram( struct gl_context
*ctx
,
119 struct gl_program
*prog
)
121 _mesa_delete_program( ctx
, prog
);
126 brwProgramStringNotify(struct gl_context
*ctx
,
128 struct gl_program
*prog
)
130 struct brw_context
*brw
= brw_context(ctx
);
133 case GL_FRAGMENT_PROGRAM_ARB
: {
134 struct gl_fragment_program
*fprog
= (struct gl_fragment_program
*) prog
;
135 struct brw_fragment_program
*newFP
= brw_fragment_program(fprog
);
136 const struct brw_fragment_program
*curFP
=
137 brw_fragment_program_const(brw
->fragment_program
);
140 brw
->ctx
.NewDriverState
|= BRW_NEW_FRAGMENT_PROGRAM
;
141 newFP
->id
= get_new_program_id(brw
->intelScreen
);
143 brw_add_texrect_params(prog
);
145 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_FRAGMENT
].NirOptions
) {
146 prog
->nir
= brw_create_nir(brw
, NULL
, prog
, MESA_SHADER_FRAGMENT
);
149 brw_fs_precompile(ctx
, NULL
, prog
);
152 case GL_VERTEX_PROGRAM_ARB
: {
153 struct gl_vertex_program
*vprog
= (struct gl_vertex_program
*) prog
;
154 struct brw_vertex_program
*newVP
= brw_vertex_program(vprog
);
155 const struct brw_vertex_program
*curVP
=
156 brw_vertex_program_const(brw
->vertex_program
);
159 brw
->ctx
.NewDriverState
|= BRW_NEW_VERTEX_PROGRAM
;
160 if (newVP
->program
.IsPositionInvariant
) {
161 _mesa_insert_mvp_code(ctx
, &newVP
->program
);
163 newVP
->id
= get_new_program_id(brw
->intelScreen
);
165 /* Also tell tnl about it:
167 _tnl_program_string(ctx
, target
, prog
);
169 brw_add_texrect_params(prog
);
171 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].NirOptions
) {
172 prog
->nir
= brw_create_nir(brw
, NULL
, prog
, MESA_SHADER_VERTEX
);
175 brw_vs_precompile(ctx
, NULL
, prog
);
180 * driver->ProgramStringNotify is only called for ARB programs, fixed
181 * function vertex programs, and ir_to_mesa (which isn't used by the
182 * i965 back-end). Therefore, even after geometry shaders are added,
183 * this function should only ever be called with a target of
184 * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
186 unreachable("Unexpected target in brwProgramStringNotify");
193 brw_memory_barrier(struct gl_context
*ctx
, GLbitfield barriers
)
195 struct brw_context
*brw
= brw_context(ctx
);
196 unsigned bits
= (PIPE_CONTROL_DATA_CACHE_INVALIDATE
|
197 PIPE_CONTROL_NO_WRITE
|
198 PIPE_CONTROL_CS_STALL
);
199 assert(brw
->gen
>= 7 && brw
->gen
<= 8);
201 if (barriers
& (GL_VERTEX_ATTRIB_ARRAY_BARRIER_BIT
|
202 GL_ELEMENT_ARRAY_BARRIER_BIT
|
203 GL_COMMAND_BARRIER_BIT
))
204 bits
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
206 if (barriers
& GL_UNIFORM_BARRIER_BIT
)
207 bits
|= (PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
|
208 PIPE_CONTROL_CONST_CACHE_INVALIDATE
);
210 if (barriers
& GL_TEXTURE_FETCH_BARRIER_BIT
)
211 bits
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
213 if (barriers
& GL_TEXTURE_UPDATE_BARRIER_BIT
)
214 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
216 if (barriers
& GL_FRAMEBUFFER_BARRIER_BIT
)
217 bits
|= (PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
218 PIPE_CONTROL_RENDER_TARGET_FLUSH
);
220 /* Typed surface messages are handled by the render cache on IVB, so we
221 * need to flush it too.
223 if (brw
->gen
== 7 && !brw
->is_haswell
)
224 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
226 brw_emit_pipe_control_flush(brw
, bits
);
230 brw_add_texrect_params(struct gl_program
*prog
)
232 for (int texunit
= 0; texunit
< BRW_MAX_TEX_UNIT
; texunit
++) {
233 if (!(prog
->TexturesUsed
[texunit
] & (1 << TEXTURE_RECT_INDEX
)))
236 int tokens
[STATE_LENGTH
] = {
244 _mesa_add_state_reference(prog
->Parameters
, (gl_state_index
*)tokens
);
248 /* Per-thread scratch space is a power-of-two multiple of 1KB. */
250 brw_get_scratch_size(int size
)
254 for (i
= 1024; i
< size
; i
*= 2)
261 brw_get_scratch_bo(struct brw_context
*brw
,
262 drm_intel_bo
**scratch_bo
, int size
)
264 drm_intel_bo
*old_bo
= *scratch_bo
;
266 if (old_bo
&& old_bo
->size
< size
) {
267 drm_intel_bo_unreference(old_bo
);
272 *scratch_bo
= drm_intel_bo_alloc(brw
->bufmgr
, "scratch bo", size
, 4096);
276 void brwInitFragProgFuncs( struct dd_function_table
*functions
)
278 assert(functions
->ProgramStringNotify
== _tnl_program_string
);
280 functions
->NewProgram
= brwNewProgram
;
281 functions
->DeleteProgram
= brwDeleteProgram
;
282 functions
->ProgramStringNotify
= brwProgramStringNotify
;
284 functions
->NewShader
= brw_new_shader
;
285 functions
->LinkShader
= brw_link_shader
;
287 functions
->MemoryBarrier
= brw_memory_barrier
;
290 struct shader_times
{
297 brw_init_shader_time(struct brw_context
*brw
)
299 const int max_entries
= 2048;
300 brw
->shader_time
.bo
=
301 drm_intel_bo_alloc(brw
->bufmgr
, "shader time",
302 max_entries
* SHADER_TIME_STRIDE
* 3, 4096);
303 brw
->shader_time
.names
= rzalloc_array(brw
, const char *, max_entries
);
304 brw
->shader_time
.ids
= rzalloc_array(brw
, int, max_entries
);
305 brw
->shader_time
.types
= rzalloc_array(brw
, enum shader_time_shader_type
,
307 brw
->shader_time
.cumulative
= rzalloc_array(brw
, struct shader_times
,
309 brw
->shader_time
.max_entries
= max_entries
;
313 compare_time(const void *a
, const void *b
)
315 uint64_t * const *a_val
= a
;
316 uint64_t * const *b_val
= b
;
318 /* We don't just subtract because we're turning the value to an int. */
319 if (**a_val
< **b_val
)
321 else if (**a_val
== **b_val
)
328 print_shader_time_line(const char *stage
, const char *name
,
329 int shader_num
, uint64_t time
, uint64_t total
)
331 fprintf(stderr
, "%-6s%-18s", stage
, name
);
334 fprintf(stderr
, "%4d: ", shader_num
);
336 fprintf(stderr
, " : ");
338 fprintf(stderr
, "%16lld (%7.2f Gcycles) %4.1f%%\n",
340 (double)time
/ 1000000000.0,
341 (double)time
/ total
* 100.0);
345 brw_report_shader_time(struct brw_context
*brw
)
347 if (!brw
->shader_time
.bo
|| !brw
->shader_time
.num_entries
)
350 uint64_t scaled
[brw
->shader_time
.num_entries
];
351 uint64_t *sorted
[brw
->shader_time
.num_entries
];
352 uint64_t total_by_type
[ST_CS
+ 1];
353 memset(total_by_type
, 0, sizeof(total_by_type
));
355 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
356 uint64_t written
= 0, reset
= 0;
357 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
359 sorted
[i
] = &scaled
[i
];
367 written
= brw
->shader_time
.cumulative
[i
].written
;
368 reset
= brw
->shader_time
.cumulative
[i
].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
].time
;
382 scaled
[i
] = time
/ written
* (written
+ reset
);
393 total_by_type
[type
] += scaled
[i
];
403 fprintf(stderr
, "No shader time collected yet\n");
407 qsort(sorted
, brw
->shader_time
.num_entries
, sizeof(sorted
[0]), compare_time
);
409 fprintf(stderr
, "\n");
410 fprintf(stderr
, "type ID cycles spent %% of total\n");
411 for (int s
= 0; s
< brw
->shader_time
.num_entries
; s
++) {
413 /* Work back from the sorted pointers times to a time to print. */
414 int i
= sorted
[s
] - scaled
;
419 int shader_num
= brw
->shader_time
.ids
[i
];
420 const char *shader_name
= brw
->shader_time
.names
[i
];
422 switch (brw
->shader_time
.types
[i
]) {
443 print_shader_time_line(stage
, shader_name
, shader_num
,
447 fprintf(stderr
, "\n");
448 print_shader_time_line("total", "vs", 0, total_by_type
[ST_VS
], total
);
449 print_shader_time_line("total", "gs", 0, total_by_type
[ST_GS
], total
);
450 print_shader_time_line("total", "fs8", 0, total_by_type
[ST_FS8
], total
);
451 print_shader_time_line("total", "fs16", 0, total_by_type
[ST_FS16
], total
);
452 print_shader_time_line("total", "cs", 0, total_by_type
[ST_CS
], 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);
466 void *bo_map
= brw
->shader_time
.bo
->virtual;
468 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
469 uint32_t *times
= bo_map
+ i
* 3 * SHADER_TIME_STRIDE
;
471 brw
->shader_time
.cumulative
[i
].time
+= times
[SHADER_TIME_STRIDE
* 0 / 4];
472 brw
->shader_time
.cumulative
[i
].written
+= times
[SHADER_TIME_STRIDE
* 1 / 4];
473 brw
->shader_time
.cumulative
[i
].reset
+= times
[SHADER_TIME_STRIDE
* 2 / 4];
476 /* Zero the BO out to clear it out for our next collection.
478 memset(bo_map
, 0, brw
->shader_time
.bo
->size
);
479 drm_intel_bo_unmap(brw
->shader_time
.bo
);
483 brw_collect_and_report_shader_time(struct brw_context
*brw
)
485 brw_collect_shader_time(brw
);
487 if (brw
->shader_time
.report_time
== 0 ||
488 get_time() - brw
->shader_time
.report_time
>= 1.0) {
489 brw_report_shader_time(brw
);
490 brw
->shader_time
.report_time
= get_time();
495 * Chooses an index in the shader_time buffer and sets up tracking information
498 * Note that this holds on to references to the underlying programs, which may
499 * change their lifetimes compared to normal operation.
502 brw_get_shader_time_index(struct brw_context
*brw
,
503 struct gl_shader_program
*shader_prog
,
504 struct gl_program
*prog
,
505 enum shader_time_shader_type type
)
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 int id
= shader_prog
? shader_prog
->Name
: prog
->Id
;
515 } else if (!shader_prog
) {
517 } else if (shader_prog
->Label
) {
518 name
= ralloc_strdup(brw
->shader_time
.names
, shader_prog
->Label
);
523 brw
->shader_time
.names
[shader_time_index
] = name
;
524 brw
->shader_time
.ids
[shader_time_index
] = id
;
526 return shader_time_index
;
530 brw_destroy_shader_time(struct brw_context
*brw
)
532 drm_intel_bo_unreference(brw
->shader_time
.bo
);
533 brw
->shader_time
.bo
= NULL
;
537 brw_mark_surface_used(struct brw_stage_prog_data
*prog_data
,
540 assert(surf_index
< BRW_MAX_SURFACES
);
542 prog_data
->binding_table
.size_bytes
=
543 MAX2(prog_data
->binding_table
.size_bytes
, (surf_index
+ 1) * 4);
547 brw_stage_prog_data_compare(const struct brw_stage_prog_data
*a
,
548 const struct brw_stage_prog_data
*b
)
550 /* Compare all the struct up to the pointers. */
551 if (memcmp(a
, b
, offsetof(struct brw_stage_prog_data
, param
)))
554 if (memcmp(a
->param
, b
->param
, a
->nr_params
* sizeof(void *)))
557 if (memcmp(a
->pull_param
, b
->pull_param
, a
->nr_pull_params
* sizeof(void *)))
564 brw_stage_prog_data_free(const void *p
)
566 struct brw_stage_prog_data
*prog_data
= (struct brw_stage_prog_data
*)p
;
568 ralloc_free(prog_data
->param
);
569 ralloc_free(prog_data
->pull_param
);
573 brw_dump_ir(const char *stage
, struct gl_shader_program
*shader_prog
,
574 struct gl_shader
*shader
, struct gl_program
*prog
)
579 "GLSL IR for native %s shader %d:\n",
580 stage
, shader_prog
->Name
);
581 _mesa_print_ir(stderr
, shader
->ir
, NULL
);
582 fprintf(stderr
, "\n\n");
585 fprintf(stderr
, "ARB_%s_program %d ir for native %s shader\n",
586 stage
, prog
->Id
, stage
);
587 _mesa_print_program(prog
);