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 prog
->nir
= brw_create_nir(brw
, NULL
, prog
, MESA_SHADER_FRAGMENT
, true);
147 brw_fs_precompile(ctx
, NULL
, prog
);
150 case GL_VERTEX_PROGRAM_ARB
: {
151 struct gl_vertex_program
*vprog
= (struct gl_vertex_program
*) prog
;
152 struct brw_vertex_program
*newVP
= brw_vertex_program(vprog
);
153 const struct brw_vertex_program
*curVP
=
154 brw_vertex_program_const(brw
->vertex_program
);
157 brw
->ctx
.NewDriverState
|= BRW_NEW_VERTEX_PROGRAM
;
158 if (newVP
->program
.IsPositionInvariant
) {
159 _mesa_insert_mvp_code(ctx
, &newVP
->program
);
161 newVP
->id
= get_new_program_id(brw
->intelScreen
);
163 /* Also tell tnl about it:
165 _tnl_program_string(ctx
, target
, prog
);
167 brw_add_texrect_params(prog
);
169 prog
->nir
= brw_create_nir(brw
, NULL
, prog
, MESA_SHADER_VERTEX
,
170 brw
->intelScreen
->compiler
->scalar_vs
);
172 brw_vs_precompile(ctx
, NULL
, prog
);
177 * driver->ProgramStringNotify is only called for ARB programs, fixed
178 * function vertex programs, and ir_to_mesa (which isn't used by the
179 * i965 back-end). Therefore, even after geometry shaders are added,
180 * this function should only ever be called with a target of
181 * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
183 unreachable("Unexpected target in brwProgramStringNotify");
190 brw_memory_barrier(struct gl_context
*ctx
, GLbitfield barriers
)
192 struct brw_context
*brw
= brw_context(ctx
);
193 unsigned bits
= (PIPE_CONTROL_DATA_CACHE_INVALIDATE
|
194 PIPE_CONTROL_NO_WRITE
|
195 PIPE_CONTROL_CS_STALL
);
196 assert(brw
->gen
>= 7 && brw
->gen
<= 9);
198 if (barriers
& (GL_VERTEX_ATTRIB_ARRAY_BARRIER_BIT
|
199 GL_ELEMENT_ARRAY_BARRIER_BIT
|
200 GL_COMMAND_BARRIER_BIT
))
201 bits
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
203 if (barriers
& GL_UNIFORM_BARRIER_BIT
)
204 bits
|= (PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
|
205 PIPE_CONTROL_CONST_CACHE_INVALIDATE
);
207 if (barriers
& GL_TEXTURE_FETCH_BARRIER_BIT
)
208 bits
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
210 if (barriers
& GL_TEXTURE_UPDATE_BARRIER_BIT
)
211 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
213 if (barriers
& GL_FRAMEBUFFER_BARRIER_BIT
)
214 bits
|= (PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
215 PIPE_CONTROL_RENDER_TARGET_FLUSH
);
217 /* Typed surface messages are handled by the render cache on IVB, so we
218 * need to flush it too.
220 if (brw
->gen
== 7 && !brw
->is_haswell
)
221 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
223 brw_emit_pipe_control_flush(brw
, bits
);
227 brw_add_texrect_params(struct gl_program
*prog
)
229 for (int texunit
= 0; texunit
< BRW_MAX_TEX_UNIT
; texunit
++) {
230 if (!(prog
->TexturesUsed
[texunit
] & (1 << TEXTURE_RECT_INDEX
)))
233 int tokens
[STATE_LENGTH
] = {
241 _mesa_add_state_reference(prog
->Parameters
, (gl_state_index
*)tokens
);
245 /* Per-thread scratch space is a power-of-two multiple of 1KB. */
247 brw_get_scratch_size(int size
)
251 for (i
= 1024; i
< size
; i
*= 2)
258 brw_get_scratch_bo(struct brw_context
*brw
,
259 drm_intel_bo
**scratch_bo
, int size
)
261 drm_intel_bo
*old_bo
= *scratch_bo
;
263 if (old_bo
&& old_bo
->size
< size
) {
264 drm_intel_bo_unreference(old_bo
);
269 *scratch_bo
= drm_intel_bo_alloc(brw
->bufmgr
, "scratch bo", size
, 4096);
273 void brwInitFragProgFuncs( struct dd_function_table
*functions
)
275 assert(functions
->ProgramStringNotify
== _tnl_program_string
);
277 functions
->NewProgram
= brwNewProgram
;
278 functions
->DeleteProgram
= brwDeleteProgram
;
279 functions
->ProgramStringNotify
= brwProgramStringNotify
;
281 functions
->NewShader
= brw_new_shader
;
282 functions
->LinkShader
= brw_link_shader
;
284 functions
->MemoryBarrier
= brw_memory_barrier
;
287 struct shader_times
{
294 brw_init_shader_time(struct brw_context
*brw
)
296 const int max_entries
= 2048;
297 brw
->shader_time
.bo
=
298 drm_intel_bo_alloc(brw
->bufmgr
, "shader time",
299 max_entries
* SHADER_TIME_STRIDE
* 3, 4096);
300 brw
->shader_time
.names
= rzalloc_array(brw
, const char *, max_entries
);
301 brw
->shader_time
.ids
= rzalloc_array(brw
, int, max_entries
);
302 brw
->shader_time
.types
= rzalloc_array(brw
, enum shader_time_shader_type
,
304 brw
->shader_time
.cumulative
= rzalloc_array(brw
, struct shader_times
,
306 brw
->shader_time
.max_entries
= max_entries
;
310 compare_time(const void *a
, const void *b
)
312 uint64_t * const *a_val
= a
;
313 uint64_t * const *b_val
= b
;
315 /* We don't just subtract because we're turning the value to an int. */
316 if (**a_val
< **b_val
)
318 else if (**a_val
== **b_val
)
325 print_shader_time_line(const char *stage
, const char *name
,
326 int shader_num
, uint64_t time
, uint64_t total
)
328 fprintf(stderr
, "%-6s%-18s", stage
, name
);
331 fprintf(stderr
, "%4d: ", shader_num
);
333 fprintf(stderr
, " : ");
335 fprintf(stderr
, "%16lld (%7.2f Gcycles) %4.1f%%\n",
337 (double)time
/ 1000000000.0,
338 (double)time
/ total
* 100.0);
342 brw_report_shader_time(struct brw_context
*brw
)
344 if (!brw
->shader_time
.bo
|| !brw
->shader_time
.num_entries
)
347 uint64_t scaled
[brw
->shader_time
.num_entries
];
348 uint64_t *sorted
[brw
->shader_time
.num_entries
];
349 uint64_t total_by_type
[ST_CS
+ 1];
350 memset(total_by_type
, 0, sizeof(total_by_type
));
352 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
353 uint64_t written
= 0, reset
= 0;
354 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
356 sorted
[i
] = &scaled
[i
];
364 written
= brw
->shader_time
.cumulative
[i
].written
;
365 reset
= brw
->shader_time
.cumulative
[i
].reset
;
369 /* I sometimes want to print things that aren't the 3 shader times.
370 * Just print the sum in that case.
377 uint64_t time
= brw
->shader_time
.cumulative
[i
].time
;
379 scaled
[i
] = time
/ written
* (written
+ reset
);
390 total_by_type
[type
] += scaled
[i
];
400 fprintf(stderr
, "No shader time collected yet\n");
404 qsort(sorted
, brw
->shader_time
.num_entries
, sizeof(sorted
[0]), compare_time
);
406 fprintf(stderr
, "\n");
407 fprintf(stderr
, "type ID cycles spent %% of total\n");
408 for (int s
= 0; s
< brw
->shader_time
.num_entries
; s
++) {
410 /* Work back from the sorted pointers times to a time to print. */
411 int i
= sorted
[s
] - scaled
;
416 int shader_num
= brw
->shader_time
.ids
[i
];
417 const char *shader_name
= brw
->shader_time
.names
[i
];
419 switch (brw
->shader_time
.types
[i
]) {
440 print_shader_time_line(stage
, shader_name
, shader_num
,
444 fprintf(stderr
, "\n");
445 print_shader_time_line("total", "vs", 0, total_by_type
[ST_VS
], total
);
446 print_shader_time_line("total", "gs", 0, total_by_type
[ST_GS
], total
);
447 print_shader_time_line("total", "fs8", 0, total_by_type
[ST_FS8
], total
);
448 print_shader_time_line("total", "fs16", 0, total_by_type
[ST_FS16
], total
);
449 print_shader_time_line("total", "cs", 0, total_by_type
[ST_CS
], total
);
453 brw_collect_shader_time(struct brw_context
*brw
)
455 if (!brw
->shader_time
.bo
)
458 /* This probably stalls on the last rendering. We could fix that by
459 * delaying reading the reports, but it doesn't look like it's a big
460 * overhead compared to the cost of tracking the time in the first place.
462 drm_intel_bo_map(brw
->shader_time
.bo
, true);
463 void *bo_map
= brw
->shader_time
.bo
->virtual;
465 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
466 uint32_t *times
= bo_map
+ i
* 3 * SHADER_TIME_STRIDE
;
468 brw
->shader_time
.cumulative
[i
].time
+= times
[SHADER_TIME_STRIDE
* 0 / 4];
469 brw
->shader_time
.cumulative
[i
].written
+= times
[SHADER_TIME_STRIDE
* 1 / 4];
470 brw
->shader_time
.cumulative
[i
].reset
+= times
[SHADER_TIME_STRIDE
* 2 / 4];
473 /* Zero the BO out to clear it out for our next collection.
475 memset(bo_map
, 0, brw
->shader_time
.bo
->size
);
476 drm_intel_bo_unmap(brw
->shader_time
.bo
);
480 brw_collect_and_report_shader_time(struct brw_context
*brw
)
482 brw_collect_shader_time(brw
);
484 if (brw
->shader_time
.report_time
== 0 ||
485 get_time() - brw
->shader_time
.report_time
>= 1.0) {
486 brw_report_shader_time(brw
);
487 brw
->shader_time
.report_time
= get_time();
492 * Chooses an index in the shader_time buffer and sets up tracking information
495 * Note that this holds on to references to the underlying programs, which may
496 * change their lifetimes compared to normal operation.
499 brw_get_shader_time_index(struct brw_context
*brw
,
500 struct gl_shader_program
*shader_prog
,
501 struct gl_program
*prog
,
502 enum shader_time_shader_type type
)
504 int shader_time_index
= brw
->shader_time
.num_entries
++;
505 assert(shader_time_index
< brw
->shader_time
.max_entries
);
506 brw
->shader_time
.types
[shader_time_index
] = type
;
508 int id
= shader_prog
? shader_prog
->Name
: prog
->Id
;
512 } else if (!shader_prog
) {
514 } else if (shader_prog
->Label
) {
515 name
= ralloc_strdup(brw
->shader_time
.names
, shader_prog
->Label
);
520 brw
->shader_time
.names
[shader_time_index
] = name
;
521 brw
->shader_time
.ids
[shader_time_index
] = id
;
523 return shader_time_index
;
527 brw_destroy_shader_time(struct brw_context
*brw
)
529 drm_intel_bo_unreference(brw
->shader_time
.bo
);
530 brw
->shader_time
.bo
= NULL
;
534 brw_mark_surface_used(struct brw_stage_prog_data
*prog_data
,
537 assert(surf_index
< BRW_MAX_SURFACES
);
539 prog_data
->binding_table
.size_bytes
=
540 MAX2(prog_data
->binding_table
.size_bytes
, (surf_index
+ 1) * 4);
544 brw_stage_prog_data_free(const void *p
)
546 struct brw_stage_prog_data
*prog_data
= (struct brw_stage_prog_data
*)p
;
548 ralloc_free(prog_data
->param
);
549 ralloc_free(prog_data
->pull_param
);
550 ralloc_free(prog_data
->image_param
);
554 brw_dump_ir(const char *stage
, struct gl_shader_program
*shader_prog
,
555 struct gl_shader
*shader
, struct gl_program
*prog
)
560 "GLSL IR for native %s shader %d:\n",
561 stage
, shader_prog
->Name
);
562 _mesa_print_ir(stderr
, shader
->ir
, NULL
);
563 fprintf(stderr
, "\n\n");
566 fprintf(stderr
, "ARB_%s_program %d ir for native %s shader\n",
567 stage
, prog
->Id
, stage
);
568 _mesa_print_program(prog
);
573 brw_setup_tex_for_precompile(struct brw_context
*brw
,
574 struct brw_sampler_prog_key_data
*tex
,
575 struct gl_program
*prog
)
577 const bool has_shader_channel_select
= brw
->is_haswell
|| brw
->gen
>= 8;
578 unsigned sampler_count
= _mesa_fls(prog
->SamplersUsed
);
579 for (unsigned i
= 0; i
< sampler_count
; i
++) {
580 if (!has_shader_channel_select
&& (prog
->ShadowSamplers
& (1 << i
))) {
581 /* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
583 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_ONE
);
585 /* Color sampler: assume no swizzling. */
586 tex
->swizzles
[i
] = SWIZZLE_XYZW
;