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 struct brw_vertex_program
*prog
= CALLOC_STRUCT(brw_vertex_program
);
136 prog
->id
= get_new_program_id(brw
->screen
);
138 return _mesa_init_gl_program(&prog
->program
, target
, id
);
144 case GL_FRAGMENT_PROGRAM_ARB
: {
145 struct brw_fragment_program
*prog
;
147 struct gen4_fragment_program
*g4_prog
=
148 CALLOC_STRUCT(gen4_fragment_program
);
149 prog
= &g4_prog
->base
;
151 prog
= CALLOC_STRUCT(brw_fragment_program
);
155 prog
->id
= get_new_program_id(brw
->screen
);
157 return _mesa_init_gl_program(&prog
->program
, target
, id
);
163 case GL_GEOMETRY_PROGRAM_NV
: {
164 struct brw_geometry_program
*prog
= CALLOC_STRUCT(brw_geometry_program
);
166 prog
->id
= get_new_program_id(brw
->screen
);
168 return _mesa_init_gl_program(&prog
->program
, target
, id
);
174 case GL_TESS_CONTROL_PROGRAM_NV
: {
175 struct brw_tess_ctrl_program
*prog
= CALLOC_STRUCT(brw_tess_ctrl_program
);
177 prog
->id
= get_new_program_id(brw
->screen
);
179 return _mesa_init_gl_program(&prog
->program
, target
, id
);
185 case GL_TESS_EVALUATION_PROGRAM_NV
: {
186 struct brw_tess_eval_program
*prog
= CALLOC_STRUCT(brw_tess_eval_program
);
188 prog
->id
= get_new_program_id(brw
->screen
);
190 return _mesa_init_gl_program(&prog
->program
, target
, id
);
196 case GL_COMPUTE_PROGRAM_NV
: {
197 struct brw_compute_program
*prog
= CALLOC_STRUCT(brw_compute_program
);
199 prog
->id
= get_new_program_id(brw
->screen
);
201 return _mesa_init_gl_program(&prog
->program
, target
, id
);
208 unreachable("Unsupported target in brwNewProgram()");
212 static void brwDeleteProgram( struct gl_context
*ctx
,
213 struct gl_program
*prog
)
215 _mesa_delete_program( ctx
, prog
);
220 brwProgramStringNotify(struct gl_context
*ctx
,
222 struct gl_program
*prog
)
224 assert(target
== GL_VERTEX_PROGRAM_ARB
|| !prog
->IsPositionInvariant
);
226 struct brw_context
*brw
= brw_context(ctx
);
227 const struct brw_compiler
*compiler
= brw
->screen
->compiler
;
230 case GL_FRAGMENT_PROGRAM_ARB
: {
231 struct brw_fragment_program
*newFP
= brw_fragment_program(prog
);
232 const struct brw_fragment_program
*curFP
=
233 brw_fragment_program_const(brw
->fragment_program
);
236 brw
->ctx
.NewDriverState
|= BRW_NEW_FRAGMENT_PROGRAM
;
237 newFP
->id
= get_new_program_id(brw
->screen
);
239 brw_add_texrect_params(prog
);
241 prog
->nir
= brw_create_nir(brw
, NULL
, prog
, MESA_SHADER_FRAGMENT
, true);
243 brw_fs_precompile(ctx
, NULL
, prog
);
246 case GL_VERTEX_PROGRAM_ARB
: {
247 struct brw_vertex_program
*newVP
= brw_vertex_program(prog
);
248 const struct brw_vertex_program
*curVP
=
249 brw_vertex_program_const(brw
->vertex_program
);
252 brw
->ctx
.NewDriverState
|= BRW_NEW_VERTEX_PROGRAM
;
253 if (newVP
->program
.IsPositionInvariant
) {
254 _mesa_insert_mvp_code(ctx
, &newVP
->program
);
256 newVP
->id
= get_new_program_id(brw
->screen
);
258 /* Also tell tnl about it:
260 _tnl_program_string(ctx
, target
, prog
);
262 brw_add_texrect_params(prog
);
264 prog
->nir
= brw_create_nir(brw
, NULL
, prog
, MESA_SHADER_VERTEX
,
265 compiler
->scalar_stage
[MESA_SHADER_VERTEX
]);
267 brw_vs_precompile(ctx
, NULL
, prog
);
272 * driver->ProgramStringNotify is only called for ARB programs, fixed
273 * function vertex programs, and ir_to_mesa (which isn't used by the
274 * i965 back-end). Therefore, even after geometry shaders are added,
275 * this function should only ever be called with a target of
276 * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
278 unreachable("Unexpected target in brwProgramStringNotify");
285 brw_memory_barrier(struct gl_context
*ctx
, GLbitfield barriers
)
287 struct brw_context
*brw
= brw_context(ctx
);
288 unsigned bits
= (PIPE_CONTROL_DATA_CACHE_FLUSH
|
289 PIPE_CONTROL_NO_WRITE
|
290 PIPE_CONTROL_CS_STALL
);
291 assert(brw
->gen
>= 7 && brw
->gen
<= 9);
293 if (barriers
& (GL_VERTEX_ATTRIB_ARRAY_BARRIER_BIT
|
294 GL_ELEMENT_ARRAY_BARRIER_BIT
|
295 GL_COMMAND_BARRIER_BIT
))
296 bits
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
298 if (barriers
& GL_UNIFORM_BARRIER_BIT
)
299 bits
|= (PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
|
300 PIPE_CONTROL_CONST_CACHE_INVALIDATE
);
302 if (barriers
& GL_TEXTURE_FETCH_BARRIER_BIT
)
303 bits
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
305 if (barriers
& GL_TEXTURE_UPDATE_BARRIER_BIT
)
306 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
308 if (barriers
& GL_FRAMEBUFFER_BARRIER_BIT
)
309 bits
|= (PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
310 PIPE_CONTROL_RENDER_TARGET_FLUSH
);
312 /* Typed surface messages are handled by the render cache on IVB, so we
313 * need to flush it too.
315 if (brw
->gen
== 7 && !brw
->is_haswell
)
316 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
318 brw_emit_pipe_control_flush(brw
, bits
);
322 brw_blend_barrier(struct gl_context
*ctx
)
324 struct brw_context
*brw
= brw_context(ctx
);
326 if (!ctx
->Extensions
.MESA_shader_framebuffer_fetch
) {
328 brw_emit_pipe_control_flush(brw
,
329 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
330 PIPE_CONTROL_CS_STALL
);
331 brw_emit_pipe_control_flush(brw
,
332 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
);
334 brw_emit_pipe_control_flush(brw
,
335 PIPE_CONTROL_RENDER_TARGET_FLUSH
);
341 brw_add_texrect_params(struct gl_program
*prog
)
343 for (int texunit
= 0; texunit
< BRW_MAX_TEX_UNIT
; texunit
++) {
344 if (!(prog
->TexturesUsed
[texunit
] & (1 << TEXTURE_RECT_INDEX
)))
347 int tokens
[STATE_LENGTH
] = {
355 _mesa_add_state_reference(prog
->Parameters
, (gl_state_index
*)tokens
);
360 brw_get_scratch_bo(struct brw_context
*brw
,
361 drm_intel_bo
**scratch_bo
, int size
)
363 drm_intel_bo
*old_bo
= *scratch_bo
;
365 if (old_bo
&& old_bo
->size
< size
) {
366 drm_intel_bo_unreference(old_bo
);
371 *scratch_bo
= drm_intel_bo_alloc(brw
->bufmgr
, "scratch bo", size
, 4096);
376 * Reserve enough scratch space for the given stage to hold \p per_thread_size
377 * bytes times the given \p thread_count.
380 brw_alloc_stage_scratch(struct brw_context
*brw
,
381 struct brw_stage_state
*stage_state
,
382 unsigned per_thread_size
,
383 unsigned thread_count
)
385 if (stage_state
->per_thread_scratch
< per_thread_size
) {
386 stage_state
->per_thread_scratch
= per_thread_size
;
388 if (stage_state
->scratch_bo
)
389 drm_intel_bo_unreference(stage_state
->scratch_bo
);
391 stage_state
->scratch_bo
=
392 drm_intel_bo_alloc(brw
->bufmgr
, "shader scratch space",
393 per_thread_size
* thread_count
, 4096);
397 void brwInitFragProgFuncs( struct dd_function_table
*functions
)
399 assert(functions
->ProgramStringNotify
== _tnl_program_string
);
401 functions
->NewProgram
= brwNewProgram
;
402 functions
->DeleteProgram
= brwDeleteProgram
;
403 functions
->ProgramStringNotify
= brwProgramStringNotify
;
405 functions
->NewShader
= brw_new_shader
;
406 functions
->LinkShader
= brw_link_shader
;
408 functions
->MemoryBarrier
= brw_memory_barrier
;
409 functions
->BlendBarrier
= brw_blend_barrier
;
412 struct shader_times
{
419 brw_init_shader_time(struct brw_context
*brw
)
421 const int max_entries
= 2048;
422 brw
->shader_time
.bo
=
423 drm_intel_bo_alloc(brw
->bufmgr
, "shader time",
424 max_entries
* SHADER_TIME_STRIDE
* 3, 4096);
425 brw
->shader_time
.names
= rzalloc_array(brw
, const char *, max_entries
);
426 brw
->shader_time
.ids
= rzalloc_array(brw
, int, max_entries
);
427 brw
->shader_time
.types
= rzalloc_array(brw
, enum shader_time_shader_type
,
429 brw
->shader_time
.cumulative
= rzalloc_array(brw
, struct shader_times
,
431 brw
->shader_time
.max_entries
= max_entries
;
435 compare_time(const void *a
, const void *b
)
437 uint64_t * const *a_val
= a
;
438 uint64_t * const *b_val
= b
;
440 /* We don't just subtract because we're turning the value to an int. */
441 if (**a_val
< **b_val
)
443 else if (**a_val
== **b_val
)
450 print_shader_time_line(const char *stage
, const char *name
,
451 int shader_num
, uint64_t time
, uint64_t total
)
453 fprintf(stderr
, "%-6s%-18s", stage
, name
);
456 fprintf(stderr
, "%4d: ", shader_num
);
458 fprintf(stderr
, " : ");
460 fprintf(stderr
, "%16lld (%7.2f Gcycles) %4.1f%%\n",
462 (double)time
/ 1000000000.0,
463 (double)time
/ total
* 100.0);
467 brw_report_shader_time(struct brw_context
*brw
)
469 if (!brw
->shader_time
.bo
|| !brw
->shader_time
.num_entries
)
472 uint64_t scaled
[brw
->shader_time
.num_entries
];
473 uint64_t *sorted
[brw
->shader_time
.num_entries
];
474 uint64_t total_by_type
[ST_CS
+ 1];
475 memset(total_by_type
, 0, sizeof(total_by_type
));
477 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
478 uint64_t written
= 0, reset
= 0;
479 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
481 sorted
[i
] = &scaled
[i
];
491 written
= brw
->shader_time
.cumulative
[i
].written
;
492 reset
= brw
->shader_time
.cumulative
[i
].reset
;
496 /* I sometimes want to print things that aren't the 3 shader times.
497 * Just print the sum in that case.
504 uint64_t time
= brw
->shader_time
.cumulative
[i
].time
;
506 scaled
[i
] = time
/ written
* (written
+ reset
);
519 total_by_type
[type
] += scaled
[i
];
529 fprintf(stderr
, "No shader time collected yet\n");
533 qsort(sorted
, brw
->shader_time
.num_entries
, sizeof(sorted
[0]), compare_time
);
535 fprintf(stderr
, "\n");
536 fprintf(stderr
, "type ID cycles spent %% of total\n");
537 for (int s
= 0; s
< brw
->shader_time
.num_entries
; s
++) {
539 /* Work back from the sorted pointers times to a time to print. */
540 int i
= sorted
[s
] - scaled
;
545 int shader_num
= brw
->shader_time
.ids
[i
];
546 const char *shader_name
= brw
->shader_time
.names
[i
];
548 switch (brw
->shader_time
.types
[i
]) {
575 print_shader_time_line(stage
, shader_name
, shader_num
,
579 fprintf(stderr
, "\n");
580 print_shader_time_line("total", "vs", 0, total_by_type
[ST_VS
], total
);
581 print_shader_time_line("total", "tcs", 0, total_by_type
[ST_TCS
], total
);
582 print_shader_time_line("total", "tes", 0, total_by_type
[ST_TES
], total
);
583 print_shader_time_line("total", "gs", 0, total_by_type
[ST_GS
], total
);
584 print_shader_time_line("total", "fs8", 0, total_by_type
[ST_FS8
], total
);
585 print_shader_time_line("total", "fs16", 0, total_by_type
[ST_FS16
], total
);
586 print_shader_time_line("total", "cs", 0, total_by_type
[ST_CS
], total
);
590 brw_collect_shader_time(struct brw_context
*brw
)
592 if (!brw
->shader_time
.bo
)
595 /* This probably stalls on the last rendering. We could fix that by
596 * delaying reading the reports, but it doesn't look like it's a big
597 * overhead compared to the cost of tracking the time in the first place.
599 drm_intel_bo_map(brw
->shader_time
.bo
, true);
600 void *bo_map
= brw
->shader_time
.bo
->virtual;
602 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
603 uint32_t *times
= bo_map
+ i
* 3 * SHADER_TIME_STRIDE
;
605 brw
->shader_time
.cumulative
[i
].time
+= times
[SHADER_TIME_STRIDE
* 0 / 4];
606 brw
->shader_time
.cumulative
[i
].written
+= times
[SHADER_TIME_STRIDE
* 1 / 4];
607 brw
->shader_time
.cumulative
[i
].reset
+= times
[SHADER_TIME_STRIDE
* 2 / 4];
610 /* Zero the BO out to clear it out for our next collection.
612 memset(bo_map
, 0, brw
->shader_time
.bo
->size
);
613 drm_intel_bo_unmap(brw
->shader_time
.bo
);
617 brw_collect_and_report_shader_time(struct brw_context
*brw
)
619 brw_collect_shader_time(brw
);
621 if (brw
->shader_time
.report_time
== 0 ||
622 get_time() - brw
->shader_time
.report_time
>= 1.0) {
623 brw_report_shader_time(brw
);
624 brw
->shader_time
.report_time
= get_time();
629 * Chooses an index in the shader_time buffer and sets up tracking information
632 * Note that this holds on to references to the underlying programs, which may
633 * change their lifetimes compared to normal operation.
636 brw_get_shader_time_index(struct brw_context
*brw
,
637 struct gl_shader_program
*shader_prog
,
638 struct gl_program
*prog
,
639 enum shader_time_shader_type type
)
641 int shader_time_index
= brw
->shader_time
.num_entries
++;
642 assert(shader_time_index
< brw
->shader_time
.max_entries
);
643 brw
->shader_time
.types
[shader_time_index
] = type
;
645 int id
= shader_prog
? shader_prog
->Name
: prog
->Id
;
649 } else if (!shader_prog
) {
651 } else if (shader_prog
->Label
) {
652 name
= ralloc_strdup(brw
->shader_time
.names
, shader_prog
->Label
);
657 brw
->shader_time
.names
[shader_time_index
] = name
;
658 brw
->shader_time
.ids
[shader_time_index
] = id
;
660 return shader_time_index
;
664 brw_destroy_shader_time(struct brw_context
*brw
)
666 drm_intel_bo_unreference(brw
->shader_time
.bo
);
667 brw
->shader_time
.bo
= NULL
;
671 brw_stage_prog_data_free(const void *p
)
673 struct brw_stage_prog_data
*prog_data
= (struct brw_stage_prog_data
*)p
;
675 ralloc_free(prog_data
->param
);
676 ralloc_free(prog_data
->pull_param
);
677 ralloc_free(prog_data
->image_param
);
681 brw_dump_ir(const char *stage
, struct gl_shader_program
*shader_prog
,
682 struct gl_linked_shader
*shader
, struct gl_program
*prog
)
687 "GLSL IR for native %s shader %d:\n",
688 stage
, shader_prog
->Name
);
689 _mesa_print_ir(stderr
, shader
->ir
, NULL
);
690 fprintf(stderr
, "\n\n");
693 fprintf(stderr
, "ARB_%s_program %d ir for native %s shader\n",
694 stage
, prog
->Id
, stage
);
695 _mesa_print_program(prog
);
700 brw_setup_tex_for_precompile(struct brw_context
*brw
,
701 struct brw_sampler_prog_key_data
*tex
,
702 struct gl_program
*prog
)
704 const bool has_shader_channel_select
= brw
->is_haswell
|| brw
->gen
>= 8;
705 unsigned sampler_count
= util_last_bit(prog
->SamplersUsed
);
706 for (unsigned i
= 0; i
< sampler_count
; i
++) {
707 if (!has_shader_channel_select
&& (prog
->ShadowSamplers
& (1 << i
))) {
708 /* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
710 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_ONE
);
712 /* Color sampler: assume no swizzling. */
713 tex
->swizzles
[i
] = SWIZZLE_XYZW
;