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
= CALLOC_STRUCT(brw_fragment_program
);
147 prog
->id
= get_new_program_id(brw
->screen
);
149 return _mesa_init_gl_program(&prog
->program
.Base
, target
, id
);
155 case GL_GEOMETRY_PROGRAM_NV
: {
156 struct brw_geometry_program
*prog
= CALLOC_STRUCT(brw_geometry_program
);
158 prog
->id
= get_new_program_id(brw
->screen
);
160 return _mesa_init_gl_program(&prog
->program
, target
, id
);
166 case GL_TESS_CONTROL_PROGRAM_NV
: {
167 struct brw_tess_ctrl_program
*prog
= CALLOC_STRUCT(brw_tess_ctrl_program
);
169 prog
->id
= get_new_program_id(brw
->screen
);
171 return _mesa_init_gl_program(&prog
->program
, target
, id
);
177 case GL_TESS_EVALUATION_PROGRAM_NV
: {
178 struct brw_tess_eval_program
*prog
= CALLOC_STRUCT(brw_tess_eval_program
);
180 prog
->id
= get_new_program_id(brw
->screen
);
182 return _mesa_init_gl_program(&prog
->program
, target
, id
);
188 case GL_COMPUTE_PROGRAM_NV
: {
189 struct brw_compute_program
*prog
= CALLOC_STRUCT(brw_compute_program
);
191 prog
->id
= get_new_program_id(brw
->screen
);
193 return _mesa_init_gl_program(&prog
->program
.Base
, target
, id
);
200 unreachable("Unsupported target in brwNewProgram()");
204 static void brwDeleteProgram( struct gl_context
*ctx
,
205 struct gl_program
*prog
)
207 _mesa_delete_program( ctx
, prog
);
212 brwProgramStringNotify(struct gl_context
*ctx
,
214 struct gl_program
*prog
)
216 assert(target
== GL_VERTEX_PROGRAM_ARB
|| !prog
->IsPositionInvariant
);
218 struct brw_context
*brw
= brw_context(ctx
);
219 const struct brw_compiler
*compiler
= brw
->screen
->compiler
;
222 case GL_FRAGMENT_PROGRAM_ARB
: {
223 struct gl_fragment_program
*fprog
= (struct gl_fragment_program
*) prog
;
224 struct brw_fragment_program
*newFP
= brw_fragment_program(fprog
);
225 const struct brw_fragment_program
*curFP
=
226 brw_fragment_program_const(brw
->fragment_program
);
229 brw
->ctx
.NewDriverState
|= BRW_NEW_FRAGMENT_PROGRAM
;
230 newFP
->id
= get_new_program_id(brw
->screen
);
232 brw_add_texrect_params(prog
);
234 prog
->nir
= brw_create_nir(brw
, NULL
, prog
, MESA_SHADER_FRAGMENT
, true);
236 brw_fs_precompile(ctx
, NULL
, prog
);
239 case GL_VERTEX_PROGRAM_ARB
: {
240 struct brw_vertex_program
*newVP
= brw_vertex_program(prog
);
241 const struct brw_vertex_program
*curVP
=
242 brw_vertex_program_const(brw
->vertex_program
);
245 brw
->ctx
.NewDriverState
|= BRW_NEW_VERTEX_PROGRAM
;
246 if (newVP
->program
.IsPositionInvariant
) {
247 _mesa_insert_mvp_code(ctx
, &newVP
->program
);
249 newVP
->id
= get_new_program_id(brw
->screen
);
251 /* Also tell tnl about it:
253 _tnl_program_string(ctx
, target
, prog
);
255 brw_add_texrect_params(prog
);
257 prog
->nir
= brw_create_nir(brw
, NULL
, prog
, MESA_SHADER_VERTEX
,
258 compiler
->scalar_stage
[MESA_SHADER_VERTEX
]);
260 brw_vs_precompile(ctx
, NULL
, prog
);
265 * driver->ProgramStringNotify is only called for ARB programs, fixed
266 * function vertex programs, and ir_to_mesa (which isn't used by the
267 * i965 back-end). Therefore, even after geometry shaders are added,
268 * this function should only ever be called with a target of
269 * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
271 unreachable("Unexpected target in brwProgramStringNotify");
278 brw_memory_barrier(struct gl_context
*ctx
, GLbitfield barriers
)
280 struct brw_context
*brw
= brw_context(ctx
);
281 unsigned bits
= (PIPE_CONTROL_DATA_CACHE_FLUSH
|
282 PIPE_CONTROL_NO_WRITE
|
283 PIPE_CONTROL_CS_STALL
);
284 assert(brw
->gen
>= 7 && brw
->gen
<= 9);
286 if (barriers
& (GL_VERTEX_ATTRIB_ARRAY_BARRIER_BIT
|
287 GL_ELEMENT_ARRAY_BARRIER_BIT
|
288 GL_COMMAND_BARRIER_BIT
))
289 bits
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
291 if (barriers
& GL_UNIFORM_BARRIER_BIT
)
292 bits
|= (PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
|
293 PIPE_CONTROL_CONST_CACHE_INVALIDATE
);
295 if (barriers
& GL_TEXTURE_FETCH_BARRIER_BIT
)
296 bits
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
298 if (barriers
& GL_TEXTURE_UPDATE_BARRIER_BIT
)
299 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
301 if (barriers
& GL_FRAMEBUFFER_BARRIER_BIT
)
302 bits
|= (PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
303 PIPE_CONTROL_RENDER_TARGET_FLUSH
);
305 /* Typed surface messages are handled by the render cache on IVB, so we
306 * need to flush it too.
308 if (brw
->gen
== 7 && !brw
->is_haswell
)
309 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
311 brw_emit_pipe_control_flush(brw
, bits
);
315 brw_blend_barrier(struct gl_context
*ctx
)
317 struct brw_context
*brw
= brw_context(ctx
);
319 if (!ctx
->Extensions
.MESA_shader_framebuffer_fetch
) {
321 brw_emit_pipe_control_flush(brw
,
322 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
323 PIPE_CONTROL_CS_STALL
);
324 brw_emit_pipe_control_flush(brw
,
325 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
);
327 brw_emit_pipe_control_flush(brw
,
328 PIPE_CONTROL_RENDER_TARGET_FLUSH
);
334 brw_add_texrect_params(struct gl_program
*prog
)
336 for (int texunit
= 0; texunit
< BRW_MAX_TEX_UNIT
; texunit
++) {
337 if (!(prog
->TexturesUsed
[texunit
] & (1 << TEXTURE_RECT_INDEX
)))
340 int tokens
[STATE_LENGTH
] = {
348 _mesa_add_state_reference(prog
->Parameters
, (gl_state_index
*)tokens
);
353 brw_get_scratch_bo(struct brw_context
*brw
,
354 drm_intel_bo
**scratch_bo
, int size
)
356 drm_intel_bo
*old_bo
= *scratch_bo
;
358 if (old_bo
&& old_bo
->size
< size
) {
359 drm_intel_bo_unreference(old_bo
);
364 *scratch_bo
= drm_intel_bo_alloc(brw
->bufmgr
, "scratch bo", size
, 4096);
369 * Reserve enough scratch space for the given stage to hold \p per_thread_size
370 * bytes times the given \p thread_count.
373 brw_alloc_stage_scratch(struct brw_context
*brw
,
374 struct brw_stage_state
*stage_state
,
375 unsigned per_thread_size
,
376 unsigned thread_count
)
378 if (stage_state
->per_thread_scratch
< per_thread_size
) {
379 stage_state
->per_thread_scratch
= per_thread_size
;
381 if (stage_state
->scratch_bo
)
382 drm_intel_bo_unreference(stage_state
->scratch_bo
);
384 stage_state
->scratch_bo
=
385 drm_intel_bo_alloc(brw
->bufmgr
, "shader scratch space",
386 per_thread_size
* thread_count
, 4096);
390 void brwInitFragProgFuncs( struct dd_function_table
*functions
)
392 assert(functions
->ProgramStringNotify
== _tnl_program_string
);
394 functions
->NewProgram
= brwNewProgram
;
395 functions
->DeleteProgram
= brwDeleteProgram
;
396 functions
->ProgramStringNotify
= brwProgramStringNotify
;
398 functions
->NewShader
= brw_new_shader
;
399 functions
->LinkShader
= brw_link_shader
;
401 functions
->MemoryBarrier
= brw_memory_barrier
;
402 functions
->BlendBarrier
= brw_blend_barrier
;
405 struct shader_times
{
412 brw_init_shader_time(struct brw_context
*brw
)
414 const int max_entries
= 2048;
415 brw
->shader_time
.bo
=
416 drm_intel_bo_alloc(brw
->bufmgr
, "shader time",
417 max_entries
* SHADER_TIME_STRIDE
* 3, 4096);
418 brw
->shader_time
.names
= rzalloc_array(brw
, const char *, max_entries
);
419 brw
->shader_time
.ids
= rzalloc_array(brw
, int, max_entries
);
420 brw
->shader_time
.types
= rzalloc_array(brw
, enum shader_time_shader_type
,
422 brw
->shader_time
.cumulative
= rzalloc_array(brw
, struct shader_times
,
424 brw
->shader_time
.max_entries
= max_entries
;
428 compare_time(const void *a
, const void *b
)
430 uint64_t * const *a_val
= a
;
431 uint64_t * const *b_val
= b
;
433 /* We don't just subtract because we're turning the value to an int. */
434 if (**a_val
< **b_val
)
436 else if (**a_val
== **b_val
)
443 print_shader_time_line(const char *stage
, const char *name
,
444 int shader_num
, uint64_t time
, uint64_t total
)
446 fprintf(stderr
, "%-6s%-18s", stage
, name
);
449 fprintf(stderr
, "%4d: ", shader_num
);
451 fprintf(stderr
, " : ");
453 fprintf(stderr
, "%16lld (%7.2f Gcycles) %4.1f%%\n",
455 (double)time
/ 1000000000.0,
456 (double)time
/ total
* 100.0);
460 brw_report_shader_time(struct brw_context
*brw
)
462 if (!brw
->shader_time
.bo
|| !brw
->shader_time
.num_entries
)
465 uint64_t scaled
[brw
->shader_time
.num_entries
];
466 uint64_t *sorted
[brw
->shader_time
.num_entries
];
467 uint64_t total_by_type
[ST_CS
+ 1];
468 memset(total_by_type
, 0, sizeof(total_by_type
));
470 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
471 uint64_t written
= 0, reset
= 0;
472 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
474 sorted
[i
] = &scaled
[i
];
484 written
= brw
->shader_time
.cumulative
[i
].written
;
485 reset
= brw
->shader_time
.cumulative
[i
].reset
;
489 /* I sometimes want to print things that aren't the 3 shader times.
490 * Just print the sum in that case.
497 uint64_t time
= brw
->shader_time
.cumulative
[i
].time
;
499 scaled
[i
] = time
/ written
* (written
+ reset
);
512 total_by_type
[type
] += scaled
[i
];
522 fprintf(stderr
, "No shader time collected yet\n");
526 qsort(sorted
, brw
->shader_time
.num_entries
, sizeof(sorted
[0]), compare_time
);
528 fprintf(stderr
, "\n");
529 fprintf(stderr
, "type ID cycles spent %% of total\n");
530 for (int s
= 0; s
< brw
->shader_time
.num_entries
; s
++) {
532 /* Work back from the sorted pointers times to a time to print. */
533 int i
= sorted
[s
] - scaled
;
538 int shader_num
= brw
->shader_time
.ids
[i
];
539 const char *shader_name
= brw
->shader_time
.names
[i
];
541 switch (brw
->shader_time
.types
[i
]) {
568 print_shader_time_line(stage
, shader_name
, shader_num
,
572 fprintf(stderr
, "\n");
573 print_shader_time_line("total", "vs", 0, total_by_type
[ST_VS
], total
);
574 print_shader_time_line("total", "tcs", 0, total_by_type
[ST_TCS
], total
);
575 print_shader_time_line("total", "tes", 0, total_by_type
[ST_TES
], total
);
576 print_shader_time_line("total", "gs", 0, total_by_type
[ST_GS
], total
);
577 print_shader_time_line("total", "fs8", 0, total_by_type
[ST_FS8
], total
);
578 print_shader_time_line("total", "fs16", 0, total_by_type
[ST_FS16
], total
);
579 print_shader_time_line("total", "cs", 0, total_by_type
[ST_CS
], total
);
583 brw_collect_shader_time(struct brw_context
*brw
)
585 if (!brw
->shader_time
.bo
)
588 /* This probably stalls on the last rendering. We could fix that by
589 * delaying reading the reports, but it doesn't look like it's a big
590 * overhead compared to the cost of tracking the time in the first place.
592 drm_intel_bo_map(brw
->shader_time
.bo
, true);
593 void *bo_map
= brw
->shader_time
.bo
->virtual;
595 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
596 uint32_t *times
= bo_map
+ i
* 3 * SHADER_TIME_STRIDE
;
598 brw
->shader_time
.cumulative
[i
].time
+= times
[SHADER_TIME_STRIDE
* 0 / 4];
599 brw
->shader_time
.cumulative
[i
].written
+= times
[SHADER_TIME_STRIDE
* 1 / 4];
600 brw
->shader_time
.cumulative
[i
].reset
+= times
[SHADER_TIME_STRIDE
* 2 / 4];
603 /* Zero the BO out to clear it out for our next collection.
605 memset(bo_map
, 0, brw
->shader_time
.bo
->size
);
606 drm_intel_bo_unmap(brw
->shader_time
.bo
);
610 brw_collect_and_report_shader_time(struct brw_context
*brw
)
612 brw_collect_shader_time(brw
);
614 if (brw
->shader_time
.report_time
== 0 ||
615 get_time() - brw
->shader_time
.report_time
>= 1.0) {
616 brw_report_shader_time(brw
);
617 brw
->shader_time
.report_time
= get_time();
622 * Chooses an index in the shader_time buffer and sets up tracking information
625 * Note that this holds on to references to the underlying programs, which may
626 * change their lifetimes compared to normal operation.
629 brw_get_shader_time_index(struct brw_context
*brw
,
630 struct gl_shader_program
*shader_prog
,
631 struct gl_program
*prog
,
632 enum shader_time_shader_type type
)
634 int shader_time_index
= brw
->shader_time
.num_entries
++;
635 assert(shader_time_index
< brw
->shader_time
.max_entries
);
636 brw
->shader_time
.types
[shader_time_index
] = type
;
638 int id
= shader_prog
? shader_prog
->Name
: prog
->Id
;
642 } else if (!shader_prog
) {
644 } else if (shader_prog
->Label
) {
645 name
= ralloc_strdup(brw
->shader_time
.names
, shader_prog
->Label
);
650 brw
->shader_time
.names
[shader_time_index
] = name
;
651 brw
->shader_time
.ids
[shader_time_index
] = id
;
653 return shader_time_index
;
657 brw_destroy_shader_time(struct brw_context
*brw
)
659 drm_intel_bo_unreference(brw
->shader_time
.bo
);
660 brw
->shader_time
.bo
= NULL
;
664 brw_stage_prog_data_free(const void *p
)
666 struct brw_stage_prog_data
*prog_data
= (struct brw_stage_prog_data
*)p
;
668 ralloc_free(prog_data
->param
);
669 ralloc_free(prog_data
->pull_param
);
670 ralloc_free(prog_data
->image_param
);
674 brw_dump_ir(const char *stage
, struct gl_shader_program
*shader_prog
,
675 struct gl_linked_shader
*shader
, struct gl_program
*prog
)
680 "GLSL IR for native %s shader %d:\n",
681 stage
, shader_prog
->Name
);
682 _mesa_print_ir(stderr
, shader
->ir
, NULL
);
683 fprintf(stderr
, "\n\n");
686 fprintf(stderr
, "ARB_%s_program %d ir for native %s shader\n",
687 stage
, prog
->Id
, stage
);
688 _mesa_print_program(prog
);
693 brw_setup_tex_for_precompile(struct brw_context
*brw
,
694 struct brw_sampler_prog_key_data
*tex
,
695 struct gl_program
*prog
)
697 const bool has_shader_channel_select
= brw
->is_haswell
|| brw
->gen
>= 8;
698 unsigned sampler_count
= util_last_bit(prog
->SamplersUsed
);
699 for (unsigned i
= 0; i
< sampler_count
; i
++) {
700 if (!has_shader_channel_select
&& (prog
->ShadowSamplers
& (1 << i
))) {
701 /* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
703 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_ONE
);
705 /* Color sampler: assume no swizzling. */
706 tex
->swizzles
[i
] = SWIZZLE_XYZW
;