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
);
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
);
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
->intelScreen
->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
->intelScreen
);
138 return _mesa_init_gl_program(&prog
->program
.Base
, 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
->intelScreen
);
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
->intelScreen
);
160 return _mesa_init_gl_program(&prog
->program
.Base
, 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
->intelScreen
);
171 return _mesa_init_gl_program(&prog
->program
.Base
, 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
->intelScreen
);
182 return _mesa_init_gl_program(&prog
->program
.Base
, 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
->intelScreen
);
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 struct brw_context
*brw
= brw_context(ctx
);
217 const struct brw_compiler
*compiler
= brw
->intelScreen
->compiler
;
220 case GL_FRAGMENT_PROGRAM_ARB
: {
221 struct gl_fragment_program
*fprog
= (struct gl_fragment_program
*) prog
;
222 struct brw_fragment_program
*newFP
= brw_fragment_program(fprog
);
223 const struct brw_fragment_program
*curFP
=
224 brw_fragment_program_const(brw
->fragment_program
);
227 brw
->ctx
.NewDriverState
|= BRW_NEW_FRAGMENT_PROGRAM
;
228 newFP
->id
= get_new_program_id(brw
->intelScreen
);
230 brw_add_texrect_params(prog
);
232 prog
->nir
= brw_create_nir(brw
, NULL
, prog
, MESA_SHADER_FRAGMENT
, true);
234 brw_fs_precompile(ctx
, NULL
, prog
);
237 case GL_VERTEX_PROGRAM_ARB
: {
238 struct gl_vertex_program
*vprog
= (struct gl_vertex_program
*) prog
;
239 struct brw_vertex_program
*newVP
= brw_vertex_program(vprog
);
240 const struct brw_vertex_program
*curVP
=
241 brw_vertex_program_const(brw
->vertex_program
);
244 brw
->ctx
.NewDriverState
|= BRW_NEW_VERTEX_PROGRAM
;
245 if (newVP
->program
.IsPositionInvariant
) {
246 _mesa_insert_mvp_code(ctx
, &newVP
->program
);
248 newVP
->id
= get_new_program_id(brw
->intelScreen
);
250 /* Also tell tnl about it:
252 _tnl_program_string(ctx
, target
, prog
);
254 brw_add_texrect_params(prog
);
256 prog
->nir
= brw_create_nir(brw
, NULL
, prog
, MESA_SHADER_VERTEX
,
257 compiler
->scalar_stage
[MESA_SHADER_VERTEX
]);
259 brw_vs_precompile(ctx
, NULL
, prog
);
264 * driver->ProgramStringNotify is only called for ARB programs, fixed
265 * function vertex programs, and ir_to_mesa (which isn't used by the
266 * i965 back-end). Therefore, even after geometry shaders are added,
267 * this function should only ever be called with a target of
268 * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
270 unreachable("Unexpected target in brwProgramStringNotify");
277 brw_memory_barrier(struct gl_context
*ctx
, GLbitfield barriers
)
279 struct brw_context
*brw
= brw_context(ctx
);
280 unsigned bits
= (PIPE_CONTROL_DATA_CACHE_FLUSH
|
281 PIPE_CONTROL_NO_WRITE
|
282 PIPE_CONTROL_CS_STALL
);
283 assert(brw
->gen
>= 7 && brw
->gen
<= 9);
285 if (barriers
& (GL_VERTEX_ATTRIB_ARRAY_BARRIER_BIT
|
286 GL_ELEMENT_ARRAY_BARRIER_BIT
|
287 GL_COMMAND_BARRIER_BIT
))
288 bits
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
290 if (barriers
& GL_UNIFORM_BARRIER_BIT
)
291 bits
|= (PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
|
292 PIPE_CONTROL_CONST_CACHE_INVALIDATE
);
294 if (barriers
& GL_TEXTURE_FETCH_BARRIER_BIT
)
295 bits
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
297 if (barriers
& GL_TEXTURE_UPDATE_BARRIER_BIT
)
298 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
300 if (barriers
& GL_FRAMEBUFFER_BARRIER_BIT
)
301 bits
|= (PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
302 PIPE_CONTROL_RENDER_TARGET_FLUSH
);
304 /* Typed surface messages are handled by the render cache on IVB, so we
305 * need to flush it too.
307 if (brw
->gen
== 7 && !brw
->is_haswell
)
308 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
310 brw_emit_pipe_control_flush(brw
, bits
);
314 brw_add_texrect_params(struct gl_program
*prog
)
316 for (int texunit
= 0; texunit
< BRW_MAX_TEX_UNIT
; texunit
++) {
317 if (!(prog
->TexturesUsed
[texunit
] & (1 << TEXTURE_RECT_INDEX
)))
320 int tokens
[STATE_LENGTH
] = {
328 _mesa_add_state_reference(prog
->Parameters
, (gl_state_index
*)tokens
);
333 brw_get_scratch_bo(struct brw_context
*brw
,
334 drm_intel_bo
**scratch_bo
, int size
)
336 drm_intel_bo
*old_bo
= *scratch_bo
;
338 if (old_bo
&& old_bo
->size
< size
) {
339 drm_intel_bo_unreference(old_bo
);
344 *scratch_bo
= drm_intel_bo_alloc(brw
->bufmgr
, "scratch bo", size
, 4096);
349 * Reserve enough scratch space for the given stage to hold \p per_thread_size
350 * bytes times the given \p thread_count.
353 brw_alloc_stage_scratch(struct brw_context
*brw
,
354 struct brw_stage_state
*stage_state
,
355 unsigned per_thread_size
,
356 unsigned thread_count
)
358 if (stage_state
->per_thread_scratch
< per_thread_size
) {
359 stage_state
->per_thread_scratch
= per_thread_size
;
361 if (stage_state
->scratch_bo
)
362 drm_intel_bo_unreference(stage_state
->scratch_bo
);
364 stage_state
->scratch_bo
=
365 drm_intel_bo_alloc(brw
->bufmgr
, "shader scratch space",
366 per_thread_size
* thread_count
, 4096);
370 void brwInitFragProgFuncs( struct dd_function_table
*functions
)
372 assert(functions
->ProgramStringNotify
== _tnl_program_string
);
374 functions
->NewProgram
= brwNewProgram
;
375 functions
->DeleteProgram
= brwDeleteProgram
;
376 functions
->ProgramStringNotify
= brwProgramStringNotify
;
378 functions
->NewShader
= brw_new_shader
;
379 functions
->LinkShader
= brw_link_shader
;
381 functions
->MemoryBarrier
= brw_memory_barrier
;
384 struct shader_times
{
391 brw_init_shader_time(struct brw_context
*brw
)
393 const int max_entries
= 2048;
394 brw
->shader_time
.bo
=
395 drm_intel_bo_alloc(brw
->bufmgr
, "shader time",
396 max_entries
* SHADER_TIME_STRIDE
* 3, 4096);
397 brw
->shader_time
.names
= rzalloc_array(brw
, const char *, max_entries
);
398 brw
->shader_time
.ids
= rzalloc_array(brw
, int, max_entries
);
399 brw
->shader_time
.types
= rzalloc_array(brw
, enum shader_time_shader_type
,
401 brw
->shader_time
.cumulative
= rzalloc_array(brw
, struct shader_times
,
403 brw
->shader_time
.max_entries
= max_entries
;
407 compare_time(const void *a
, const void *b
)
409 uint64_t * const *a_val
= a
;
410 uint64_t * const *b_val
= b
;
412 /* We don't just subtract because we're turning the value to an int. */
413 if (**a_val
< **b_val
)
415 else if (**a_val
== **b_val
)
422 print_shader_time_line(const char *stage
, const char *name
,
423 int shader_num
, uint64_t time
, uint64_t total
)
425 fprintf(stderr
, "%-6s%-18s", stage
, name
);
428 fprintf(stderr
, "%4d: ", shader_num
);
430 fprintf(stderr
, " : ");
432 fprintf(stderr
, "%16lld (%7.2f Gcycles) %4.1f%%\n",
434 (double)time
/ 1000000000.0,
435 (double)time
/ total
* 100.0);
439 brw_report_shader_time(struct brw_context
*brw
)
441 if (!brw
->shader_time
.bo
|| !brw
->shader_time
.num_entries
)
444 uint64_t scaled
[brw
->shader_time
.num_entries
];
445 uint64_t *sorted
[brw
->shader_time
.num_entries
];
446 uint64_t total_by_type
[ST_CS
+ 1];
447 memset(total_by_type
, 0, sizeof(total_by_type
));
449 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
450 uint64_t written
= 0, reset
= 0;
451 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
453 sorted
[i
] = &scaled
[i
];
463 written
= brw
->shader_time
.cumulative
[i
].written
;
464 reset
= brw
->shader_time
.cumulative
[i
].reset
;
468 /* I sometimes want to print things that aren't the 3 shader times.
469 * Just print the sum in that case.
476 uint64_t time
= brw
->shader_time
.cumulative
[i
].time
;
478 scaled
[i
] = time
/ written
* (written
+ reset
);
491 total_by_type
[type
] += scaled
[i
];
501 fprintf(stderr
, "No shader time collected yet\n");
505 qsort(sorted
, brw
->shader_time
.num_entries
, sizeof(sorted
[0]), compare_time
);
507 fprintf(stderr
, "\n");
508 fprintf(stderr
, "type ID cycles spent %% of total\n");
509 for (int s
= 0; s
< brw
->shader_time
.num_entries
; s
++) {
511 /* Work back from the sorted pointers times to a time to print. */
512 int i
= sorted
[s
] - scaled
;
517 int shader_num
= brw
->shader_time
.ids
[i
];
518 const char *shader_name
= brw
->shader_time
.names
[i
];
520 switch (brw
->shader_time
.types
[i
]) {
547 print_shader_time_line(stage
, shader_name
, shader_num
,
551 fprintf(stderr
, "\n");
552 print_shader_time_line("total", "vs", 0, total_by_type
[ST_VS
], total
);
553 print_shader_time_line("total", "tcs", 0, total_by_type
[ST_TCS
], total
);
554 print_shader_time_line("total", "tes", 0, total_by_type
[ST_TES
], total
);
555 print_shader_time_line("total", "gs", 0, total_by_type
[ST_GS
], total
);
556 print_shader_time_line("total", "fs8", 0, total_by_type
[ST_FS8
], total
);
557 print_shader_time_line("total", "fs16", 0, total_by_type
[ST_FS16
], total
);
558 print_shader_time_line("total", "cs", 0, total_by_type
[ST_CS
], total
);
562 brw_collect_shader_time(struct brw_context
*brw
)
564 if (!brw
->shader_time
.bo
)
567 /* This probably stalls on the last rendering. We could fix that by
568 * delaying reading the reports, but it doesn't look like it's a big
569 * overhead compared to the cost of tracking the time in the first place.
571 drm_intel_bo_map(brw
->shader_time
.bo
, true);
572 void *bo_map
= brw
->shader_time
.bo
->virtual;
574 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
575 uint32_t *times
= bo_map
+ i
* 3 * SHADER_TIME_STRIDE
;
577 brw
->shader_time
.cumulative
[i
].time
+= times
[SHADER_TIME_STRIDE
* 0 / 4];
578 brw
->shader_time
.cumulative
[i
].written
+= times
[SHADER_TIME_STRIDE
* 1 / 4];
579 brw
->shader_time
.cumulative
[i
].reset
+= times
[SHADER_TIME_STRIDE
* 2 / 4];
582 /* Zero the BO out to clear it out for our next collection.
584 memset(bo_map
, 0, brw
->shader_time
.bo
->size
);
585 drm_intel_bo_unmap(brw
->shader_time
.bo
);
589 brw_collect_and_report_shader_time(struct brw_context
*brw
)
591 brw_collect_shader_time(brw
);
593 if (brw
->shader_time
.report_time
== 0 ||
594 get_time() - brw
->shader_time
.report_time
>= 1.0) {
595 brw_report_shader_time(brw
);
596 brw
->shader_time
.report_time
= get_time();
601 * Chooses an index in the shader_time buffer and sets up tracking information
604 * Note that this holds on to references to the underlying programs, which may
605 * change their lifetimes compared to normal operation.
608 brw_get_shader_time_index(struct brw_context
*brw
,
609 struct gl_shader_program
*shader_prog
,
610 struct gl_program
*prog
,
611 enum shader_time_shader_type type
)
613 int shader_time_index
= brw
->shader_time
.num_entries
++;
614 assert(shader_time_index
< brw
->shader_time
.max_entries
);
615 brw
->shader_time
.types
[shader_time_index
] = type
;
617 int id
= shader_prog
? shader_prog
->Name
: prog
->Id
;
621 } else if (!shader_prog
) {
623 } else if (shader_prog
->Label
) {
624 name
= ralloc_strdup(brw
->shader_time
.names
, shader_prog
->Label
);
629 brw
->shader_time
.names
[shader_time_index
] = name
;
630 brw
->shader_time
.ids
[shader_time_index
] = id
;
632 return shader_time_index
;
636 brw_destroy_shader_time(struct brw_context
*brw
)
638 drm_intel_bo_unreference(brw
->shader_time
.bo
);
639 brw
->shader_time
.bo
= NULL
;
643 brw_stage_prog_data_free(const void *p
)
645 struct brw_stage_prog_data
*prog_data
= (struct brw_stage_prog_data
*)p
;
647 ralloc_free(prog_data
->param
);
648 ralloc_free(prog_data
->pull_param
);
649 ralloc_free(prog_data
->image_param
);
653 brw_dump_ir(const char *stage
, struct gl_shader_program
*shader_prog
,
654 struct gl_shader
*shader
, struct gl_program
*prog
)
659 "GLSL IR for native %s shader %d:\n",
660 stage
, shader_prog
->Name
);
661 _mesa_print_ir(stderr
, shader
->ir
, NULL
);
662 fprintf(stderr
, "\n\n");
665 fprintf(stderr
, "ARB_%s_program %d ir for native %s shader\n",
666 stage
, prog
->Id
, stage
);
667 _mesa_print_program(prog
);
672 brw_setup_tex_for_precompile(struct brw_context
*brw
,
673 struct brw_sampler_prog_key_data
*tex
,
674 struct gl_program
*prog
)
676 const bool has_shader_channel_select
= brw
->is_haswell
|| brw
->gen
>= 8;
677 unsigned sampler_count
= _mesa_fls(prog
->SamplersUsed
);
678 for (unsigned i
= 0; i
< sampler_count
; i
++) {
679 if (!has_shader_channel_select
&& (prog
->ShadowSamplers
& (1 << i
))) {
680 /* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
682 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_ONE
);
684 /* Color sampler: assume no swizzling. */
685 tex
->swizzles
[i
] = SWIZZLE_XYZW
;