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
, true);
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
,
173 brw
->intelScreen
->compiler
->scalar_vs
);
176 brw_vs_precompile(ctx
, NULL
, prog
);
181 * driver->ProgramStringNotify is only called for ARB programs, fixed
182 * function vertex programs, and ir_to_mesa (which isn't used by the
183 * i965 back-end). Therefore, even after geometry shaders are added,
184 * this function should only ever be called with a target of
185 * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
187 unreachable("Unexpected target in brwProgramStringNotify");
194 brw_memory_barrier(struct gl_context
*ctx
, GLbitfield barriers
)
196 struct brw_context
*brw
= brw_context(ctx
);
197 unsigned bits
= (PIPE_CONTROL_DATA_CACHE_INVALIDATE
|
198 PIPE_CONTROL_NO_WRITE
|
199 PIPE_CONTROL_CS_STALL
);
200 assert(brw
->gen
>= 7 && brw
->gen
<= 9);
202 if (barriers
& (GL_VERTEX_ATTRIB_ARRAY_BARRIER_BIT
|
203 GL_ELEMENT_ARRAY_BARRIER_BIT
|
204 GL_COMMAND_BARRIER_BIT
))
205 bits
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
207 if (barriers
& GL_UNIFORM_BARRIER_BIT
)
208 bits
|= (PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
|
209 PIPE_CONTROL_CONST_CACHE_INVALIDATE
);
211 if (barriers
& GL_TEXTURE_FETCH_BARRIER_BIT
)
212 bits
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
214 if (barriers
& GL_TEXTURE_UPDATE_BARRIER_BIT
)
215 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
217 if (barriers
& GL_FRAMEBUFFER_BARRIER_BIT
)
218 bits
|= (PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
219 PIPE_CONTROL_RENDER_TARGET_FLUSH
);
221 /* Typed surface messages are handled by the render cache on IVB, so we
222 * need to flush it too.
224 if (brw
->gen
== 7 && !brw
->is_haswell
)
225 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
227 brw_emit_pipe_control_flush(brw
, bits
);
231 brw_add_texrect_params(struct gl_program
*prog
)
233 for (int texunit
= 0; texunit
< BRW_MAX_TEX_UNIT
; texunit
++) {
234 if (!(prog
->TexturesUsed
[texunit
] & (1 << TEXTURE_RECT_INDEX
)))
237 int tokens
[STATE_LENGTH
] = {
245 _mesa_add_state_reference(prog
->Parameters
, (gl_state_index
*)tokens
);
249 /* Per-thread scratch space is a power-of-two multiple of 1KB. */
251 brw_get_scratch_size(int size
)
255 for (i
= 1024; i
< size
; i
*= 2)
262 brw_get_scratch_bo(struct brw_context
*brw
,
263 drm_intel_bo
**scratch_bo
, int size
)
265 drm_intel_bo
*old_bo
= *scratch_bo
;
267 if (old_bo
&& old_bo
->size
< size
) {
268 drm_intel_bo_unreference(old_bo
);
273 *scratch_bo
= drm_intel_bo_alloc(brw
->bufmgr
, "scratch bo", size
, 4096);
277 void brwInitFragProgFuncs( struct dd_function_table
*functions
)
279 assert(functions
->ProgramStringNotify
== _tnl_program_string
);
281 functions
->NewProgram
= brwNewProgram
;
282 functions
->DeleteProgram
= brwDeleteProgram
;
283 functions
->ProgramStringNotify
= brwProgramStringNotify
;
285 functions
->NewShader
= brw_new_shader
;
286 functions
->LinkShader
= brw_link_shader
;
288 functions
->MemoryBarrier
= brw_memory_barrier
;
291 struct shader_times
{
298 brw_init_shader_time(struct brw_context
*brw
)
300 const int max_entries
= 2048;
301 brw
->shader_time
.bo
=
302 drm_intel_bo_alloc(brw
->bufmgr
, "shader time",
303 max_entries
* SHADER_TIME_STRIDE
* 3, 4096);
304 brw
->shader_time
.names
= rzalloc_array(brw
, const char *, max_entries
);
305 brw
->shader_time
.ids
= rzalloc_array(brw
, int, max_entries
);
306 brw
->shader_time
.types
= rzalloc_array(brw
, enum shader_time_shader_type
,
308 brw
->shader_time
.cumulative
= rzalloc_array(brw
, struct shader_times
,
310 brw
->shader_time
.max_entries
= max_entries
;
314 compare_time(const void *a
, const void *b
)
316 uint64_t * const *a_val
= a
;
317 uint64_t * const *b_val
= b
;
319 /* We don't just subtract because we're turning the value to an int. */
320 if (**a_val
< **b_val
)
322 else if (**a_val
== **b_val
)
329 print_shader_time_line(const char *stage
, const char *name
,
330 int shader_num
, uint64_t time
, uint64_t total
)
332 fprintf(stderr
, "%-6s%-18s", stage
, name
);
335 fprintf(stderr
, "%4d: ", shader_num
);
337 fprintf(stderr
, " : ");
339 fprintf(stderr
, "%16lld (%7.2f Gcycles) %4.1f%%\n",
341 (double)time
/ 1000000000.0,
342 (double)time
/ total
* 100.0);
346 brw_report_shader_time(struct brw_context
*brw
)
348 if (!brw
->shader_time
.bo
|| !brw
->shader_time
.num_entries
)
351 uint64_t scaled
[brw
->shader_time
.num_entries
];
352 uint64_t *sorted
[brw
->shader_time
.num_entries
];
353 uint64_t total_by_type
[ST_CS
+ 1];
354 memset(total_by_type
, 0, sizeof(total_by_type
));
356 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
357 uint64_t written
= 0, reset
= 0;
358 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
360 sorted
[i
] = &scaled
[i
];
368 written
= brw
->shader_time
.cumulative
[i
].written
;
369 reset
= brw
->shader_time
.cumulative
[i
].reset
;
373 /* I sometimes want to print things that aren't the 3 shader times.
374 * Just print the sum in that case.
381 uint64_t time
= brw
->shader_time
.cumulative
[i
].time
;
383 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
++) {
414 /* Work back from the sorted pointers times to a time to print. */
415 int i
= sorted
[s
] - scaled
;
420 int shader_num
= brw
->shader_time
.ids
[i
];
421 const char *shader_name
= brw
->shader_time
.names
[i
];
423 switch (brw
->shader_time
.types
[i
]) {
444 print_shader_time_line(stage
, shader_name
, shader_num
,
448 fprintf(stderr
, "\n");
449 print_shader_time_line("total", "vs", 0, total_by_type
[ST_VS
], total
);
450 print_shader_time_line("total", "gs", 0, total_by_type
[ST_GS
], total
);
451 print_shader_time_line("total", "fs8", 0, total_by_type
[ST_FS8
], total
);
452 print_shader_time_line("total", "fs16", 0, total_by_type
[ST_FS16
], total
);
453 print_shader_time_line("total", "cs", 0, total_by_type
[ST_CS
], 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);
467 void *bo_map
= brw
->shader_time
.bo
->virtual;
469 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
470 uint32_t *times
= bo_map
+ i
* 3 * SHADER_TIME_STRIDE
;
472 brw
->shader_time
.cumulative
[i
].time
+= times
[SHADER_TIME_STRIDE
* 0 / 4];
473 brw
->shader_time
.cumulative
[i
].written
+= times
[SHADER_TIME_STRIDE
* 1 / 4];
474 brw
->shader_time
.cumulative
[i
].reset
+= times
[SHADER_TIME_STRIDE
* 2 / 4];
477 /* Zero the BO out to clear it out for our next collection.
479 memset(bo_map
, 0, brw
->shader_time
.bo
->size
);
480 drm_intel_bo_unmap(brw
->shader_time
.bo
);
484 brw_collect_and_report_shader_time(struct brw_context
*brw
)
486 brw_collect_shader_time(brw
);
488 if (brw
->shader_time
.report_time
== 0 ||
489 get_time() - brw
->shader_time
.report_time
>= 1.0) {
490 brw_report_shader_time(brw
);
491 brw
->shader_time
.report_time
= get_time();
496 * Chooses an index in the shader_time buffer and sets up tracking information
499 * Note that this holds on to references to the underlying programs, which may
500 * change their lifetimes compared to normal operation.
503 brw_get_shader_time_index(struct brw_context
*brw
,
504 struct gl_shader_program
*shader_prog
,
505 struct gl_program
*prog
,
506 enum shader_time_shader_type type
)
508 int shader_time_index
= brw
->shader_time
.num_entries
++;
509 assert(shader_time_index
< brw
->shader_time
.max_entries
);
510 brw
->shader_time
.types
[shader_time_index
] = type
;
512 int id
= shader_prog
? shader_prog
->Name
: prog
->Id
;
516 } else if (!shader_prog
) {
518 } else if (shader_prog
->Label
) {
519 name
= ralloc_strdup(brw
->shader_time
.names
, shader_prog
->Label
);
524 brw
->shader_time
.names
[shader_time_index
] = name
;
525 brw
->shader_time
.ids
[shader_time_index
] = id
;
527 return shader_time_index
;
531 brw_destroy_shader_time(struct brw_context
*brw
)
533 drm_intel_bo_unreference(brw
->shader_time
.bo
);
534 brw
->shader_time
.bo
= NULL
;
538 brw_mark_surface_used(struct brw_stage_prog_data
*prog_data
,
541 assert(surf_index
< BRW_MAX_SURFACES
);
543 prog_data
->binding_table
.size_bytes
=
544 MAX2(prog_data
->binding_table
.size_bytes
, (surf_index
+ 1) * 4);
548 brw_stage_prog_data_compare(const struct brw_stage_prog_data
*a
,
549 const struct brw_stage_prog_data
*b
)
551 /* Compare all the struct up to the pointers. */
552 if (memcmp(a
, b
, offsetof(struct brw_stage_prog_data
, param
)))
555 if (memcmp(a
->param
, b
->param
, a
->nr_params
* sizeof(void *)))
558 if (memcmp(a
->pull_param
, b
->pull_param
, a
->nr_pull_params
* sizeof(void *)))
565 brw_stage_prog_data_free(const void *p
)
567 struct brw_stage_prog_data
*prog_data
= (struct brw_stage_prog_data
*)p
;
569 ralloc_free(prog_data
->param
);
570 ralloc_free(prog_data
->pull_param
);
574 brw_dump_ir(const char *stage
, struct gl_shader_program
*shader_prog
,
575 struct gl_shader
*shader
, struct gl_program
*prog
)
580 "GLSL IR for native %s shader %d:\n",
581 stage
, shader_prog
->Name
);
582 _mesa_print_ir(stderr
, shader
->ir
, NULL
);
583 fprintf(stderr
, "\n\n");
586 fprintf(stderr
, "ARB_%s_program %d ir for native %s shader\n",
587 stage
, prog
->Id
, stage
);
588 _mesa_print_program(prog
);
593 brw_setup_tex_for_precompile(struct brw_context
*brw
,
594 struct brw_sampler_prog_key_data
*tex
,
595 struct gl_program
*prog
)
597 const bool has_shader_channel_select
= brw
->is_haswell
|| brw
->gen
>= 8;
598 unsigned sampler_count
= _mesa_fls(prog
->SamplersUsed
);
599 for (unsigned i
= 0; i
< sampler_count
; i
++) {
600 if (!has_shader_channel_select
&& (prog
->ShadowSamplers
& (1 << i
))) {
601 /* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
603 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_ONE
);
605 /* Color sampler: assume no swizzling. */
606 tex
->swizzles
[i
] = SWIZZLE_XYZW
;