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
,
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
,
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
);
108 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
111 NIR_PASS_V(nir
, nir_lower_samplers
, shader_prog
);
112 NIR_PASS_V(nir
, nir_lower_atomics
, shader_prog
);
119 get_new_program_id(struct intel_screen
*screen
)
121 static pthread_mutex_t m
= PTHREAD_MUTEX_INITIALIZER
;
122 pthread_mutex_lock(&m
);
123 unsigned id
= screen
->program_id
++;
124 pthread_mutex_unlock(&m
);
128 static struct gl_program
*brwNewProgram( struct gl_context
*ctx
,
132 struct brw_context
*brw
= brw_context(ctx
);
135 case GL_VERTEX_PROGRAM_ARB
:
136 case GL_TESS_CONTROL_PROGRAM_NV
:
137 case GL_TESS_EVALUATION_PROGRAM_NV
:
138 case GL_GEOMETRY_PROGRAM_NV
:
139 case GL_COMPUTE_PROGRAM_NV
: {
140 struct brw_program
*prog
= rzalloc(NULL
, struct brw_program
);
142 prog
->id
= get_new_program_id(brw
->screen
);
144 return _mesa_init_gl_program(&prog
->program
, target
, id
);
150 case GL_FRAGMENT_PROGRAM_ARB
: {
151 struct brw_program
*prog
;
153 struct gen4_fragment_program
*g4_prog
=
154 rzalloc(NULL
, struct gen4_fragment_program
);
155 prog
= &g4_prog
->base
;
157 prog
= rzalloc(NULL
, struct brw_program
);
161 prog
->id
= get_new_program_id(brw
->screen
);
163 return _mesa_init_gl_program(&prog
->program
, target
, id
);
170 unreachable("Unsupported target in brwNewProgram()");
174 static void brwDeleteProgram( struct gl_context
*ctx
,
175 struct gl_program
*prog
)
177 _mesa_delete_program( ctx
, prog
);
182 brwProgramStringNotify(struct gl_context
*ctx
,
184 struct gl_program
*prog
)
186 assert(target
== GL_VERTEX_PROGRAM_ARB
|| !prog
->IsPositionInvariant
);
188 struct brw_context
*brw
= brw_context(ctx
);
189 const struct brw_compiler
*compiler
= brw
->screen
->compiler
;
192 case GL_FRAGMENT_PROGRAM_ARB
: {
193 struct brw_program
*newFP
= brw_program(prog
);
194 const struct brw_program
*curFP
=
195 brw_program_const(brw
->fragment_program
);
198 brw
->ctx
.NewDriverState
|= BRW_NEW_FRAGMENT_PROGRAM
;
199 newFP
->id
= get_new_program_id(brw
->screen
);
201 brw_add_texrect_params(prog
);
203 prog
->nir
= brw_create_nir(brw
, NULL
, prog
, MESA_SHADER_FRAGMENT
, true);
205 brw_fs_precompile(ctx
, NULL
, prog
);
208 case GL_VERTEX_PROGRAM_ARB
: {
209 struct brw_program
*newVP
= brw_program(prog
);
210 const struct brw_program
*curVP
=
211 brw_program_const(brw
->vertex_program
);
214 brw
->ctx
.NewDriverState
|= BRW_NEW_VERTEX_PROGRAM
;
215 if (newVP
->program
.IsPositionInvariant
) {
216 _mesa_insert_mvp_code(ctx
, &newVP
->program
);
218 newVP
->id
= get_new_program_id(brw
->screen
);
220 /* Also tell tnl about it:
222 _tnl_program_string(ctx
, target
, prog
);
224 brw_add_texrect_params(prog
);
226 prog
->nir
= brw_create_nir(brw
, NULL
, prog
, MESA_SHADER_VERTEX
,
227 compiler
->scalar_stage
[MESA_SHADER_VERTEX
]);
229 brw_vs_precompile(ctx
, NULL
, prog
);
234 * driver->ProgramStringNotify is only called for ARB programs, fixed
235 * function vertex programs, and ir_to_mesa (which isn't used by the
236 * i965 back-end). Therefore, even after geometry shaders are added,
237 * this function should only ever be called with a target of
238 * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
240 unreachable("Unexpected target in brwProgramStringNotify");
247 brw_memory_barrier(struct gl_context
*ctx
, GLbitfield barriers
)
249 struct brw_context
*brw
= brw_context(ctx
);
250 unsigned bits
= (PIPE_CONTROL_DATA_CACHE_FLUSH
|
251 PIPE_CONTROL_NO_WRITE
|
252 PIPE_CONTROL_CS_STALL
);
253 assert(brw
->gen
>= 7 && brw
->gen
<= 9);
255 if (barriers
& (GL_VERTEX_ATTRIB_ARRAY_BARRIER_BIT
|
256 GL_ELEMENT_ARRAY_BARRIER_BIT
|
257 GL_COMMAND_BARRIER_BIT
))
258 bits
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
260 if (barriers
& GL_UNIFORM_BARRIER_BIT
)
261 bits
|= (PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
|
262 PIPE_CONTROL_CONST_CACHE_INVALIDATE
);
264 if (barriers
& GL_TEXTURE_FETCH_BARRIER_BIT
)
265 bits
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
267 if (barriers
& GL_TEXTURE_UPDATE_BARRIER_BIT
)
268 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
270 if (barriers
& GL_FRAMEBUFFER_BARRIER_BIT
)
271 bits
|= (PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
272 PIPE_CONTROL_RENDER_TARGET_FLUSH
);
274 /* Typed surface messages are handled by the render cache on IVB, so we
275 * need to flush it too.
277 if (brw
->gen
== 7 && !brw
->is_haswell
)
278 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
280 brw_emit_pipe_control_flush(brw
, bits
);
284 brw_blend_barrier(struct gl_context
*ctx
)
286 struct brw_context
*brw
= brw_context(ctx
);
288 if (!ctx
->Extensions
.MESA_shader_framebuffer_fetch
) {
290 brw_emit_pipe_control_flush(brw
,
291 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
292 PIPE_CONTROL_CS_STALL
);
293 brw_emit_pipe_control_flush(brw
,
294 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
);
296 brw_emit_pipe_control_flush(brw
,
297 PIPE_CONTROL_RENDER_TARGET_FLUSH
);
303 brw_add_texrect_params(struct gl_program
*prog
)
305 for (int texunit
= 0; texunit
< BRW_MAX_TEX_UNIT
; texunit
++) {
306 if (!(prog
->TexturesUsed
[texunit
] & (1 << TEXTURE_RECT_INDEX
)))
309 int tokens
[STATE_LENGTH
] = {
317 _mesa_add_state_reference(prog
->Parameters
, (gl_state_index
*)tokens
);
322 brw_get_scratch_bo(struct brw_context
*brw
,
323 drm_intel_bo
**scratch_bo
, int size
)
325 drm_intel_bo
*old_bo
= *scratch_bo
;
327 if (old_bo
&& old_bo
->size
< size
) {
328 drm_intel_bo_unreference(old_bo
);
333 *scratch_bo
= drm_intel_bo_alloc(brw
->bufmgr
, "scratch bo", size
, 4096);
338 * Reserve enough scratch space for the given stage to hold \p per_thread_size
339 * bytes times the given \p thread_count.
342 brw_alloc_stage_scratch(struct brw_context
*brw
,
343 struct brw_stage_state
*stage_state
,
344 unsigned per_thread_size
,
345 unsigned thread_count
)
347 if (stage_state
->per_thread_scratch
< per_thread_size
) {
348 stage_state
->per_thread_scratch
= per_thread_size
;
350 if (stage_state
->scratch_bo
)
351 drm_intel_bo_unreference(stage_state
->scratch_bo
);
353 stage_state
->scratch_bo
=
354 drm_intel_bo_alloc(brw
->bufmgr
, "shader scratch space",
355 per_thread_size
* thread_count
, 4096);
359 void brwInitFragProgFuncs( struct dd_function_table
*functions
)
361 assert(functions
->ProgramStringNotify
== _tnl_program_string
);
363 functions
->NewProgram
= brwNewProgram
;
364 functions
->DeleteProgram
= brwDeleteProgram
;
365 functions
->ProgramStringNotify
= brwProgramStringNotify
;
367 functions
->NewShader
= brw_new_shader
;
368 functions
->LinkShader
= brw_link_shader
;
370 functions
->MemoryBarrier
= brw_memory_barrier
;
371 functions
->BlendBarrier
= brw_blend_barrier
;
374 struct shader_times
{
381 brw_init_shader_time(struct brw_context
*brw
)
383 const int max_entries
= 2048;
384 brw
->shader_time
.bo
=
385 drm_intel_bo_alloc(brw
->bufmgr
, "shader time",
386 max_entries
* SHADER_TIME_STRIDE
* 3, 4096);
387 brw
->shader_time
.names
= rzalloc_array(brw
, const char *, max_entries
);
388 brw
->shader_time
.ids
= rzalloc_array(brw
, int, max_entries
);
389 brw
->shader_time
.types
= rzalloc_array(brw
, enum shader_time_shader_type
,
391 brw
->shader_time
.cumulative
= rzalloc_array(brw
, struct shader_times
,
393 brw
->shader_time
.max_entries
= max_entries
;
397 compare_time(const void *a
, const void *b
)
399 uint64_t * const *a_val
= a
;
400 uint64_t * const *b_val
= b
;
402 /* We don't just subtract because we're turning the value to an int. */
403 if (**a_val
< **b_val
)
405 else if (**a_val
== **b_val
)
412 print_shader_time_line(const char *stage
, const char *name
,
413 int shader_num
, uint64_t time
, uint64_t total
)
415 fprintf(stderr
, "%-6s%-18s", stage
, name
);
418 fprintf(stderr
, "%4d: ", shader_num
);
420 fprintf(stderr
, " : ");
422 fprintf(stderr
, "%16lld (%7.2f Gcycles) %4.1f%%\n",
424 (double)time
/ 1000000000.0,
425 (double)time
/ total
* 100.0);
429 brw_report_shader_time(struct brw_context
*brw
)
431 if (!brw
->shader_time
.bo
|| !brw
->shader_time
.num_entries
)
434 uint64_t scaled
[brw
->shader_time
.num_entries
];
435 uint64_t *sorted
[brw
->shader_time
.num_entries
];
436 uint64_t total_by_type
[ST_CS
+ 1];
437 memset(total_by_type
, 0, sizeof(total_by_type
));
439 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
440 uint64_t written
= 0, reset
= 0;
441 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
443 sorted
[i
] = &scaled
[i
];
453 written
= brw
->shader_time
.cumulative
[i
].written
;
454 reset
= brw
->shader_time
.cumulative
[i
].reset
;
458 /* I sometimes want to print things that aren't the 3 shader times.
459 * Just print the sum in that case.
466 uint64_t time
= brw
->shader_time
.cumulative
[i
].time
;
468 scaled
[i
] = time
/ written
* (written
+ reset
);
481 total_by_type
[type
] += scaled
[i
];
491 fprintf(stderr
, "No shader time collected yet\n");
495 qsort(sorted
, brw
->shader_time
.num_entries
, sizeof(sorted
[0]), compare_time
);
497 fprintf(stderr
, "\n");
498 fprintf(stderr
, "type ID cycles spent %% of total\n");
499 for (int s
= 0; s
< brw
->shader_time
.num_entries
; s
++) {
501 /* Work back from the sorted pointers times to a time to print. */
502 int i
= sorted
[s
] - scaled
;
507 int shader_num
= brw
->shader_time
.ids
[i
];
508 const char *shader_name
= brw
->shader_time
.names
[i
];
510 switch (brw
->shader_time
.types
[i
]) {
537 print_shader_time_line(stage
, shader_name
, shader_num
,
541 fprintf(stderr
, "\n");
542 print_shader_time_line("total", "vs", 0, total_by_type
[ST_VS
], total
);
543 print_shader_time_line("total", "tcs", 0, total_by_type
[ST_TCS
], total
);
544 print_shader_time_line("total", "tes", 0, total_by_type
[ST_TES
], total
);
545 print_shader_time_line("total", "gs", 0, total_by_type
[ST_GS
], total
);
546 print_shader_time_line("total", "fs8", 0, total_by_type
[ST_FS8
], total
);
547 print_shader_time_line("total", "fs16", 0, total_by_type
[ST_FS16
], total
);
548 print_shader_time_line("total", "cs", 0, total_by_type
[ST_CS
], total
);
552 brw_collect_shader_time(struct brw_context
*brw
)
554 if (!brw
->shader_time
.bo
)
557 /* This probably stalls on the last rendering. We could fix that by
558 * delaying reading the reports, but it doesn't look like it's a big
559 * overhead compared to the cost of tracking the time in the first place.
561 drm_intel_bo_map(brw
->shader_time
.bo
, true);
562 void *bo_map
= brw
->shader_time
.bo
->virtual;
564 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
565 uint32_t *times
= bo_map
+ i
* 3 * SHADER_TIME_STRIDE
;
567 brw
->shader_time
.cumulative
[i
].time
+= times
[SHADER_TIME_STRIDE
* 0 / 4];
568 brw
->shader_time
.cumulative
[i
].written
+= times
[SHADER_TIME_STRIDE
* 1 / 4];
569 brw
->shader_time
.cumulative
[i
].reset
+= times
[SHADER_TIME_STRIDE
* 2 / 4];
572 /* Zero the BO out to clear it out for our next collection.
574 memset(bo_map
, 0, brw
->shader_time
.bo
->size
);
575 drm_intel_bo_unmap(brw
->shader_time
.bo
);
579 brw_collect_and_report_shader_time(struct brw_context
*brw
)
581 brw_collect_shader_time(brw
);
583 if (brw
->shader_time
.report_time
== 0 ||
584 get_time() - brw
->shader_time
.report_time
>= 1.0) {
585 brw_report_shader_time(brw
);
586 brw
->shader_time
.report_time
= get_time();
591 * Chooses an index in the shader_time buffer and sets up tracking information
594 * Note that this holds on to references to the underlying programs, which may
595 * change their lifetimes compared to normal operation.
598 brw_get_shader_time_index(struct brw_context
*brw
,
599 struct gl_shader_program
*shader_prog
,
600 struct gl_program
*prog
,
601 enum shader_time_shader_type type
)
603 int shader_time_index
= brw
->shader_time
.num_entries
++;
604 assert(shader_time_index
< brw
->shader_time
.max_entries
);
605 brw
->shader_time
.types
[shader_time_index
] = type
;
607 int id
= shader_prog
? shader_prog
->Name
: prog
->Id
;
611 } else if (!shader_prog
) {
613 } else if (shader_prog
->Label
) {
614 name
= ralloc_strdup(brw
->shader_time
.names
, shader_prog
->Label
);
619 brw
->shader_time
.names
[shader_time_index
] = name
;
620 brw
->shader_time
.ids
[shader_time_index
] = id
;
622 return shader_time_index
;
626 brw_destroy_shader_time(struct brw_context
*brw
)
628 drm_intel_bo_unreference(brw
->shader_time
.bo
);
629 brw
->shader_time
.bo
= NULL
;
633 brw_stage_prog_data_free(const void *p
)
635 struct brw_stage_prog_data
*prog_data
= (struct brw_stage_prog_data
*)p
;
637 ralloc_free(prog_data
->param
);
638 ralloc_free(prog_data
->pull_param
);
639 ralloc_free(prog_data
->image_param
);
643 brw_dump_ir(const char *stage
, struct gl_shader_program
*shader_prog
,
644 struct gl_linked_shader
*shader
, struct gl_program
*prog
)
649 "GLSL IR for native %s shader %d:\n",
650 stage
, shader_prog
->Name
);
651 _mesa_print_ir(stderr
, shader
->ir
, NULL
);
652 fprintf(stderr
, "\n\n");
655 fprintf(stderr
, "ARB_%s_program %d ir for native %s shader\n",
656 stage
, prog
->Id
, stage
);
657 _mesa_print_program(prog
);
662 brw_setup_tex_for_precompile(struct brw_context
*brw
,
663 struct brw_sampler_prog_key_data
*tex
,
664 struct gl_program
*prog
)
666 const bool has_shader_channel_select
= brw
->is_haswell
|| brw
->gen
>= 8;
667 unsigned sampler_count
= util_last_bit(prog
->SamplersUsed
);
668 for (unsigned i
= 0; i
< sampler_count
; i
++) {
669 if (!has_shader_channel_select
&& (prog
->ShadowSamplers
& (1 << i
))) {
670 /* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
672 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_ONE
);
674 /* Color sampler: assume no swizzling. */
675 tex
->swizzles
[i
] = SWIZZLE_XYZW
;