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 "program/prog_parameter.h"
35 #include "program/prog_print.h"
36 #include "program/prog_to_nir.h"
37 #include "program/program.h"
38 #include "program/programopt.h"
40 #include "util/ralloc.h"
41 #include "compiler/glsl/ir.h"
42 #include "compiler/glsl/glsl_to_nir.h"
44 #include "brw_program.h"
45 #include "brw_context.h"
46 #include "brw_shader.h"
48 #include "intel_batchbuffer.h"
51 brw_nir_lower_uniforms(nir_shader
*nir
, bool is_scalar
)
54 nir_assign_var_locations(&nir
->uniforms
, &nir
->num_uniforms
, 0,
55 type_size_scalar_bytes
);
56 nir_lower_io(nir
, nir_var_uniform
, type_size_scalar_bytes
, 0);
58 nir_assign_var_locations(&nir
->uniforms
, &nir
->num_uniforms
, 0,
59 type_size_vec4_bytes
);
60 nir_lower_io(nir
, nir_var_uniform
, type_size_vec4_bytes
, 0);
65 brw_create_nir(struct brw_context
*brw
,
66 const struct gl_shader_program
*shader_prog
,
67 const struct gl_program
*prog
,
68 gl_shader_stage stage
,
71 struct gl_context
*ctx
= &brw
->ctx
;
72 const nir_shader_compiler_options
*options
=
73 ctx
->Const
.ShaderCompilerOptions
[stage
].NirOptions
;
77 /* First, lower the GLSL IR or Mesa IR to NIR */
79 nir
= glsl_to_nir(shader_prog
, stage
, options
);
80 nir_remove_dead_variables(nir
, nir_var_shader_in
| nir_var_shader_out
);
81 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
82 nir_shader_get_entrypoint(nir
), true, false);
84 nir
= prog_to_nir(prog
, options
);
85 NIR_PASS_V(nir
, nir_convert_to_ssa
); /* turn registers into SSA */
87 nir_validate_shader(nir
);
91 nir
= brw_preprocess_nir(brw
->screen
->compiler
, nir
);
93 if (stage
== MESA_SHADER_FRAGMENT
) {
94 static const struct nir_lower_wpos_ytransform_options wpos_options
= {
95 .state_tokens
= {STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
, 0, 0, 0},
96 .fs_coord_pixel_center_integer
= 1,
97 .fs_coord_origin_upper_left
= 1,
99 _mesa_add_state_reference(prog
->Parameters
,
100 (gl_state_index
*) wpos_options
.state_tokens
);
102 NIR_PASS(progress
, nir
, nir_lower_wpos_ytransform
, &wpos_options
);
105 NIR_PASS(progress
, nir
, nir_lower_system_values
);
106 NIR_PASS_V(nir
, brw_nir_lower_uniforms
, is_scalar
);
109 NIR_PASS_V(nir
, nir_lower_samplers
, shader_prog
);
110 NIR_PASS_V(nir
, nir_lower_atomics
, shader_prog
);
117 get_new_program_id(struct intel_screen
*screen
)
119 static pthread_mutex_t m
= PTHREAD_MUTEX_INITIALIZER
;
120 pthread_mutex_lock(&m
);
121 unsigned id
= screen
->program_id
++;
122 pthread_mutex_unlock(&m
);
126 static struct gl_program
*brwNewProgram( struct gl_context
*ctx
,
130 struct brw_context
*brw
= brw_context(ctx
);
133 case GL_VERTEX_PROGRAM_ARB
:
134 case GL_GEOMETRY_PROGRAM_NV
: {
135 struct brw_program
*prog
= CALLOC_STRUCT(brw_program
);
137 prog
->id
= get_new_program_id(brw
->screen
);
139 return _mesa_init_gl_program(&prog
->program
, target
, id
);
145 case GL_FRAGMENT_PROGRAM_ARB
: {
146 struct brw_fragment_program
*prog
;
148 struct gen4_fragment_program
*g4_prog
=
149 CALLOC_STRUCT(gen4_fragment_program
);
150 prog
= &g4_prog
->base
;
152 prog
= CALLOC_STRUCT(brw_fragment_program
);
156 prog
->id
= get_new_program_id(brw
->screen
);
158 return _mesa_init_gl_program(&prog
->program
, target
, id
);
164 case GL_TESS_CONTROL_PROGRAM_NV
: {
165 struct brw_tess_ctrl_program
*prog
= CALLOC_STRUCT(brw_tess_ctrl_program
);
167 prog
->id
= get_new_program_id(brw
->screen
);
169 return _mesa_init_gl_program(&prog
->program
, target
, id
);
175 case GL_TESS_EVALUATION_PROGRAM_NV
: {
176 struct brw_tess_eval_program
*prog
= CALLOC_STRUCT(brw_tess_eval_program
);
178 prog
->id
= get_new_program_id(brw
->screen
);
180 return _mesa_init_gl_program(&prog
->program
, target
, id
);
186 case GL_COMPUTE_PROGRAM_NV
: {
187 struct brw_compute_program
*prog
= CALLOC_STRUCT(brw_compute_program
);
189 prog
->id
= get_new_program_id(brw
->screen
);
191 return _mesa_init_gl_program(&prog
->program
, target
, id
);
198 unreachable("Unsupported target in brwNewProgram()");
202 static void brwDeleteProgram( struct gl_context
*ctx
,
203 struct gl_program
*prog
)
205 _mesa_delete_program( ctx
, prog
);
210 brwProgramStringNotify(struct gl_context
*ctx
,
212 struct gl_program
*prog
)
214 assert(target
== GL_VERTEX_PROGRAM_ARB
|| !prog
->IsPositionInvariant
);
216 struct brw_context
*brw
= brw_context(ctx
);
217 const struct brw_compiler
*compiler
= brw
->screen
->compiler
;
220 case GL_FRAGMENT_PROGRAM_ARB
: {
221 struct brw_fragment_program
*newFP
= brw_fragment_program(prog
);
222 const struct brw_fragment_program
*curFP
=
223 brw_fragment_program_const(brw
->fragment_program
);
226 brw
->ctx
.NewDriverState
|= BRW_NEW_FRAGMENT_PROGRAM
;
227 newFP
->id
= get_new_program_id(brw
->screen
);
229 brw_add_texrect_params(prog
);
231 prog
->nir
= brw_create_nir(brw
, NULL
, prog
, MESA_SHADER_FRAGMENT
, true);
233 brw_fs_precompile(ctx
, NULL
, prog
);
236 case GL_VERTEX_PROGRAM_ARB
: {
237 struct brw_program
*newVP
= brw_program(prog
);
238 const struct brw_program
*curVP
=
239 brw_program_const(brw
->vertex_program
);
242 brw
->ctx
.NewDriverState
|= BRW_NEW_VERTEX_PROGRAM
;
243 if (newVP
->program
.IsPositionInvariant
) {
244 _mesa_insert_mvp_code(ctx
, &newVP
->program
);
246 newVP
->id
= get_new_program_id(brw
->screen
);
248 /* Also tell tnl about it:
250 _tnl_program_string(ctx
, target
, prog
);
252 brw_add_texrect_params(prog
);
254 prog
->nir
= brw_create_nir(brw
, NULL
, prog
, MESA_SHADER_VERTEX
,
255 compiler
->scalar_stage
[MESA_SHADER_VERTEX
]);
257 brw_vs_precompile(ctx
, NULL
, prog
);
262 * driver->ProgramStringNotify is only called for ARB programs, fixed
263 * function vertex programs, and ir_to_mesa (which isn't used by the
264 * i965 back-end). Therefore, even after geometry shaders are added,
265 * this function should only ever be called with a target of
266 * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
268 unreachable("Unexpected target in brwProgramStringNotify");
275 brw_memory_barrier(struct gl_context
*ctx
, GLbitfield barriers
)
277 struct brw_context
*brw
= brw_context(ctx
);
278 unsigned bits
= (PIPE_CONTROL_DATA_CACHE_FLUSH
|
279 PIPE_CONTROL_NO_WRITE
|
280 PIPE_CONTROL_CS_STALL
);
281 assert(brw
->gen
>= 7 && brw
->gen
<= 9);
283 if (barriers
& (GL_VERTEX_ATTRIB_ARRAY_BARRIER_BIT
|
284 GL_ELEMENT_ARRAY_BARRIER_BIT
|
285 GL_COMMAND_BARRIER_BIT
))
286 bits
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
288 if (barriers
& GL_UNIFORM_BARRIER_BIT
)
289 bits
|= (PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
|
290 PIPE_CONTROL_CONST_CACHE_INVALIDATE
);
292 if (barriers
& GL_TEXTURE_FETCH_BARRIER_BIT
)
293 bits
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
295 if (barriers
& GL_TEXTURE_UPDATE_BARRIER_BIT
)
296 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
298 if (barriers
& GL_FRAMEBUFFER_BARRIER_BIT
)
299 bits
|= (PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
300 PIPE_CONTROL_RENDER_TARGET_FLUSH
);
302 /* Typed surface messages are handled by the render cache on IVB, so we
303 * need to flush it too.
305 if (brw
->gen
== 7 && !brw
->is_haswell
)
306 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
308 brw_emit_pipe_control_flush(brw
, bits
);
312 brw_blend_barrier(struct gl_context
*ctx
)
314 struct brw_context
*brw
= brw_context(ctx
);
316 if (!ctx
->Extensions
.MESA_shader_framebuffer_fetch
) {
318 brw_emit_pipe_control_flush(brw
,
319 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
320 PIPE_CONTROL_CS_STALL
);
321 brw_emit_pipe_control_flush(brw
,
322 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
);
324 brw_emit_pipe_control_flush(brw
,
325 PIPE_CONTROL_RENDER_TARGET_FLUSH
);
331 brw_add_texrect_params(struct gl_program
*prog
)
333 for (int texunit
= 0; texunit
< BRW_MAX_TEX_UNIT
; texunit
++) {
334 if (!(prog
->TexturesUsed
[texunit
] & (1 << TEXTURE_RECT_INDEX
)))
337 int tokens
[STATE_LENGTH
] = {
345 _mesa_add_state_reference(prog
->Parameters
, (gl_state_index
*)tokens
);
350 brw_get_scratch_bo(struct brw_context
*brw
,
351 drm_intel_bo
**scratch_bo
, int size
)
353 drm_intel_bo
*old_bo
= *scratch_bo
;
355 if (old_bo
&& old_bo
->size
< size
) {
356 drm_intel_bo_unreference(old_bo
);
361 *scratch_bo
= drm_intel_bo_alloc(brw
->bufmgr
, "scratch bo", size
, 4096);
366 * Reserve enough scratch space for the given stage to hold \p per_thread_size
367 * bytes times the given \p thread_count.
370 brw_alloc_stage_scratch(struct brw_context
*brw
,
371 struct brw_stage_state
*stage_state
,
372 unsigned per_thread_size
,
373 unsigned thread_count
)
375 if (stage_state
->per_thread_scratch
< per_thread_size
) {
376 stage_state
->per_thread_scratch
= per_thread_size
;
378 if (stage_state
->scratch_bo
)
379 drm_intel_bo_unreference(stage_state
->scratch_bo
);
381 stage_state
->scratch_bo
=
382 drm_intel_bo_alloc(brw
->bufmgr
, "shader scratch space",
383 per_thread_size
* thread_count
, 4096);
387 void brwInitFragProgFuncs( struct dd_function_table
*functions
)
389 assert(functions
->ProgramStringNotify
== _tnl_program_string
);
391 functions
->NewProgram
= brwNewProgram
;
392 functions
->DeleteProgram
= brwDeleteProgram
;
393 functions
->ProgramStringNotify
= brwProgramStringNotify
;
395 functions
->NewShader
= brw_new_shader
;
396 functions
->LinkShader
= brw_link_shader
;
398 functions
->MemoryBarrier
= brw_memory_barrier
;
399 functions
->BlendBarrier
= brw_blend_barrier
;
402 struct shader_times
{
409 brw_init_shader_time(struct brw_context
*brw
)
411 const int max_entries
= 2048;
412 brw
->shader_time
.bo
=
413 drm_intel_bo_alloc(brw
->bufmgr
, "shader time",
414 max_entries
* SHADER_TIME_STRIDE
* 3, 4096);
415 brw
->shader_time
.names
= rzalloc_array(brw
, const char *, max_entries
);
416 brw
->shader_time
.ids
= rzalloc_array(brw
, int, max_entries
);
417 brw
->shader_time
.types
= rzalloc_array(brw
, enum shader_time_shader_type
,
419 brw
->shader_time
.cumulative
= rzalloc_array(brw
, struct shader_times
,
421 brw
->shader_time
.max_entries
= max_entries
;
425 compare_time(const void *a
, const void *b
)
427 uint64_t * const *a_val
= a
;
428 uint64_t * const *b_val
= b
;
430 /* We don't just subtract because we're turning the value to an int. */
431 if (**a_val
< **b_val
)
433 else if (**a_val
== **b_val
)
440 print_shader_time_line(const char *stage
, const char *name
,
441 int shader_num
, uint64_t time
, uint64_t total
)
443 fprintf(stderr
, "%-6s%-18s", stage
, name
);
446 fprintf(stderr
, "%4d: ", shader_num
);
448 fprintf(stderr
, " : ");
450 fprintf(stderr
, "%16lld (%7.2f Gcycles) %4.1f%%\n",
452 (double)time
/ 1000000000.0,
453 (double)time
/ total
* 100.0);
457 brw_report_shader_time(struct brw_context
*brw
)
459 if (!brw
->shader_time
.bo
|| !brw
->shader_time
.num_entries
)
462 uint64_t scaled
[brw
->shader_time
.num_entries
];
463 uint64_t *sorted
[brw
->shader_time
.num_entries
];
464 uint64_t total_by_type
[ST_CS
+ 1];
465 memset(total_by_type
, 0, sizeof(total_by_type
));
467 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
468 uint64_t written
= 0, reset
= 0;
469 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
471 sorted
[i
] = &scaled
[i
];
481 written
= brw
->shader_time
.cumulative
[i
].written
;
482 reset
= brw
->shader_time
.cumulative
[i
].reset
;
486 /* I sometimes want to print things that aren't the 3 shader times.
487 * Just print the sum in that case.
494 uint64_t time
= brw
->shader_time
.cumulative
[i
].time
;
496 scaled
[i
] = time
/ written
* (written
+ reset
);
509 total_by_type
[type
] += scaled
[i
];
519 fprintf(stderr
, "No shader time collected yet\n");
523 qsort(sorted
, brw
->shader_time
.num_entries
, sizeof(sorted
[0]), compare_time
);
525 fprintf(stderr
, "\n");
526 fprintf(stderr
, "type ID cycles spent %% of total\n");
527 for (int s
= 0; s
< brw
->shader_time
.num_entries
; s
++) {
529 /* Work back from the sorted pointers times to a time to print. */
530 int i
= sorted
[s
] - scaled
;
535 int shader_num
= brw
->shader_time
.ids
[i
];
536 const char *shader_name
= brw
->shader_time
.names
[i
];
538 switch (brw
->shader_time
.types
[i
]) {
565 print_shader_time_line(stage
, shader_name
, shader_num
,
569 fprintf(stderr
, "\n");
570 print_shader_time_line("total", "vs", 0, total_by_type
[ST_VS
], total
);
571 print_shader_time_line("total", "tcs", 0, total_by_type
[ST_TCS
], total
);
572 print_shader_time_line("total", "tes", 0, total_by_type
[ST_TES
], total
);
573 print_shader_time_line("total", "gs", 0, total_by_type
[ST_GS
], total
);
574 print_shader_time_line("total", "fs8", 0, total_by_type
[ST_FS8
], total
);
575 print_shader_time_line("total", "fs16", 0, total_by_type
[ST_FS16
], total
);
576 print_shader_time_line("total", "cs", 0, total_by_type
[ST_CS
], total
);
580 brw_collect_shader_time(struct brw_context
*brw
)
582 if (!brw
->shader_time
.bo
)
585 /* This probably stalls on the last rendering. We could fix that by
586 * delaying reading the reports, but it doesn't look like it's a big
587 * overhead compared to the cost of tracking the time in the first place.
589 drm_intel_bo_map(brw
->shader_time
.bo
, true);
590 void *bo_map
= brw
->shader_time
.bo
->virtual;
592 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
593 uint32_t *times
= bo_map
+ i
* 3 * SHADER_TIME_STRIDE
;
595 brw
->shader_time
.cumulative
[i
].time
+= times
[SHADER_TIME_STRIDE
* 0 / 4];
596 brw
->shader_time
.cumulative
[i
].written
+= times
[SHADER_TIME_STRIDE
* 1 / 4];
597 brw
->shader_time
.cumulative
[i
].reset
+= times
[SHADER_TIME_STRIDE
* 2 / 4];
600 /* Zero the BO out to clear it out for our next collection.
602 memset(bo_map
, 0, brw
->shader_time
.bo
->size
);
603 drm_intel_bo_unmap(brw
->shader_time
.bo
);
607 brw_collect_and_report_shader_time(struct brw_context
*brw
)
609 brw_collect_shader_time(brw
);
611 if (brw
->shader_time
.report_time
== 0 ||
612 get_time() - brw
->shader_time
.report_time
>= 1.0) {
613 brw_report_shader_time(brw
);
614 brw
->shader_time
.report_time
= get_time();
619 * Chooses an index in the shader_time buffer and sets up tracking information
622 * Note that this holds on to references to the underlying programs, which may
623 * change their lifetimes compared to normal operation.
626 brw_get_shader_time_index(struct brw_context
*brw
,
627 struct gl_shader_program
*shader_prog
,
628 struct gl_program
*prog
,
629 enum shader_time_shader_type type
)
631 int shader_time_index
= brw
->shader_time
.num_entries
++;
632 assert(shader_time_index
< brw
->shader_time
.max_entries
);
633 brw
->shader_time
.types
[shader_time_index
] = type
;
635 int id
= shader_prog
? shader_prog
->Name
: prog
->Id
;
639 } else if (!shader_prog
) {
641 } else if (shader_prog
->Label
) {
642 name
= ralloc_strdup(brw
->shader_time
.names
, shader_prog
->Label
);
647 brw
->shader_time
.names
[shader_time_index
] = name
;
648 brw
->shader_time
.ids
[shader_time_index
] = id
;
650 return shader_time_index
;
654 brw_destroy_shader_time(struct brw_context
*brw
)
656 drm_intel_bo_unreference(brw
->shader_time
.bo
);
657 brw
->shader_time
.bo
= NULL
;
661 brw_stage_prog_data_free(const void *p
)
663 struct brw_stage_prog_data
*prog_data
= (struct brw_stage_prog_data
*)p
;
665 ralloc_free(prog_data
->param
);
666 ralloc_free(prog_data
->pull_param
);
667 ralloc_free(prog_data
->image_param
);
671 brw_dump_ir(const char *stage
, struct gl_shader_program
*shader_prog
,
672 struct gl_linked_shader
*shader
, struct gl_program
*prog
)
677 "GLSL IR for native %s shader %d:\n",
678 stage
, shader_prog
->Name
);
679 _mesa_print_ir(stderr
, shader
->ir
, NULL
);
680 fprintf(stderr
, "\n\n");
683 fprintf(stderr
, "ARB_%s_program %d ir for native %s shader\n",
684 stage
, prog
->Id
, stage
);
685 _mesa_print_program(prog
);
690 brw_setup_tex_for_precompile(struct brw_context
*brw
,
691 struct brw_sampler_prog_key_data
*tex
,
692 struct gl_program
*prog
)
694 const bool has_shader_channel_select
= brw
->is_haswell
|| brw
->gen
>= 8;
695 unsigned sampler_count
= util_last_bit(prog
->SamplersUsed
);
696 for (unsigned i
= 0; i
< sampler_count
; i
++) {
697 if (!has_shader_channel_select
&& (prog
->ShadowSamplers
& (1 << i
))) {
698 /* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
700 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_ONE
);
702 /* Color sampler: assume no swizzling. */
703 tex
->swizzles
[i
] = SWIZZLE_XYZW
;