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
);
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_TESS_CONTROL_PROGRAM_NV
:
135 case GL_TESS_EVALUATION_PROGRAM_NV
:
136 case GL_GEOMETRY_PROGRAM_NV
:
137 case GL_COMPUTE_PROGRAM_NV
: {
138 struct brw_program
*prog
= rzalloc(NULL
, struct brw_program
);
140 prog
->id
= get_new_program_id(brw
->screen
);
142 return _mesa_init_gl_program(&prog
->program
, target
, id
);
148 case GL_FRAGMENT_PROGRAM_ARB
: {
149 struct brw_program
*prog
;
151 struct gen4_fragment_program
*g4_prog
=
152 rzalloc(NULL
, struct gen4_fragment_program
);
153 prog
= &g4_prog
->base
;
155 prog
= rzalloc(NULL
, struct brw_program
);
159 prog
->id
= get_new_program_id(brw
->screen
);
161 return _mesa_init_gl_program(&prog
->program
, target
, id
);
168 unreachable("Unsupported target in brwNewProgram()");
172 static void brwDeleteProgram( struct gl_context
*ctx
,
173 struct gl_program
*prog
)
175 _mesa_delete_program( ctx
, prog
);
180 brwProgramStringNotify(struct gl_context
*ctx
,
182 struct gl_program
*prog
)
184 assert(target
== GL_VERTEX_PROGRAM_ARB
|| !prog
->IsPositionInvariant
);
186 struct brw_context
*brw
= brw_context(ctx
);
187 const struct brw_compiler
*compiler
= brw
->screen
->compiler
;
190 case GL_FRAGMENT_PROGRAM_ARB
: {
191 struct brw_program
*newFP
= brw_program(prog
);
192 const struct brw_program
*curFP
=
193 brw_program_const(brw
->fragment_program
);
196 brw
->ctx
.NewDriverState
|= BRW_NEW_FRAGMENT_PROGRAM
;
197 newFP
->id
= get_new_program_id(brw
->screen
);
199 brw_add_texrect_params(prog
);
201 prog
->nir
= brw_create_nir(brw
, NULL
, prog
, MESA_SHADER_FRAGMENT
, true);
203 brw_fs_precompile(ctx
, NULL
, prog
);
206 case GL_VERTEX_PROGRAM_ARB
: {
207 struct brw_program
*newVP
= brw_program(prog
);
208 const struct brw_program
*curVP
=
209 brw_program_const(brw
->vertex_program
);
212 brw
->ctx
.NewDriverState
|= BRW_NEW_VERTEX_PROGRAM
;
213 if (newVP
->program
.IsPositionInvariant
) {
214 _mesa_insert_mvp_code(ctx
, &newVP
->program
);
216 newVP
->id
= get_new_program_id(brw
->screen
);
218 /* Also tell tnl about it:
220 _tnl_program_string(ctx
, target
, prog
);
222 brw_add_texrect_params(prog
);
224 prog
->nir
= brw_create_nir(brw
, NULL
, prog
, MESA_SHADER_VERTEX
,
225 compiler
->scalar_stage
[MESA_SHADER_VERTEX
]);
227 brw_vs_precompile(ctx
, NULL
, prog
);
232 * driver->ProgramStringNotify is only called for ARB programs, fixed
233 * function vertex programs, and ir_to_mesa (which isn't used by the
234 * i965 back-end). Therefore, even after geometry shaders are added,
235 * this function should only ever be called with a target of
236 * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
238 unreachable("Unexpected target in brwProgramStringNotify");
245 brw_memory_barrier(struct gl_context
*ctx
, GLbitfield barriers
)
247 struct brw_context
*brw
= brw_context(ctx
);
248 unsigned bits
= (PIPE_CONTROL_DATA_CACHE_FLUSH
|
249 PIPE_CONTROL_NO_WRITE
|
250 PIPE_CONTROL_CS_STALL
);
251 assert(brw
->gen
>= 7 && brw
->gen
<= 9);
253 if (barriers
& (GL_VERTEX_ATTRIB_ARRAY_BARRIER_BIT
|
254 GL_ELEMENT_ARRAY_BARRIER_BIT
|
255 GL_COMMAND_BARRIER_BIT
))
256 bits
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
258 if (barriers
& GL_UNIFORM_BARRIER_BIT
)
259 bits
|= (PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
|
260 PIPE_CONTROL_CONST_CACHE_INVALIDATE
);
262 if (barriers
& GL_TEXTURE_FETCH_BARRIER_BIT
)
263 bits
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
265 if (barriers
& GL_TEXTURE_UPDATE_BARRIER_BIT
)
266 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
268 if (barriers
& GL_FRAMEBUFFER_BARRIER_BIT
)
269 bits
|= (PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
270 PIPE_CONTROL_RENDER_TARGET_FLUSH
);
272 /* Typed surface messages are handled by the render cache on IVB, so we
273 * need to flush it too.
275 if (brw
->gen
== 7 && !brw
->is_haswell
)
276 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
278 brw_emit_pipe_control_flush(brw
, bits
);
282 brw_blend_barrier(struct gl_context
*ctx
)
284 struct brw_context
*brw
= brw_context(ctx
);
286 if (!ctx
->Extensions
.MESA_shader_framebuffer_fetch
) {
288 brw_emit_pipe_control_flush(brw
,
289 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
290 PIPE_CONTROL_CS_STALL
);
291 brw_emit_pipe_control_flush(brw
,
292 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
);
294 brw_emit_pipe_control_flush(brw
,
295 PIPE_CONTROL_RENDER_TARGET_FLUSH
);
301 brw_add_texrect_params(struct gl_program
*prog
)
303 for (int texunit
= 0; texunit
< BRW_MAX_TEX_UNIT
; texunit
++) {
304 if (!(prog
->TexturesUsed
[texunit
] & (1 << TEXTURE_RECT_INDEX
)))
307 int tokens
[STATE_LENGTH
] = {
315 _mesa_add_state_reference(prog
->Parameters
, (gl_state_index
*)tokens
);
320 brw_get_scratch_bo(struct brw_context
*brw
,
321 drm_intel_bo
**scratch_bo
, int size
)
323 drm_intel_bo
*old_bo
= *scratch_bo
;
325 if (old_bo
&& old_bo
->size
< size
) {
326 drm_intel_bo_unreference(old_bo
);
331 *scratch_bo
= drm_intel_bo_alloc(brw
->bufmgr
, "scratch bo", size
, 4096);
336 * Reserve enough scratch space for the given stage to hold \p per_thread_size
337 * bytes times the given \p thread_count.
340 brw_alloc_stage_scratch(struct brw_context
*brw
,
341 struct brw_stage_state
*stage_state
,
342 unsigned per_thread_size
,
343 unsigned thread_count
)
345 if (stage_state
->per_thread_scratch
< per_thread_size
) {
346 stage_state
->per_thread_scratch
= per_thread_size
;
348 if (stage_state
->scratch_bo
)
349 drm_intel_bo_unreference(stage_state
->scratch_bo
);
351 stage_state
->scratch_bo
=
352 drm_intel_bo_alloc(brw
->bufmgr
, "shader scratch space",
353 per_thread_size
* thread_count
, 4096);
357 void brwInitFragProgFuncs( struct dd_function_table
*functions
)
359 assert(functions
->ProgramStringNotify
== _tnl_program_string
);
361 functions
->NewProgram
= brwNewProgram
;
362 functions
->DeleteProgram
= brwDeleteProgram
;
363 functions
->ProgramStringNotify
= brwProgramStringNotify
;
365 functions
->NewShader
= brw_new_shader
;
366 functions
->LinkShader
= brw_link_shader
;
368 functions
->MemoryBarrier
= brw_memory_barrier
;
369 functions
->BlendBarrier
= brw_blend_barrier
;
372 struct shader_times
{
379 brw_init_shader_time(struct brw_context
*brw
)
381 const int max_entries
= 2048;
382 brw
->shader_time
.bo
=
383 drm_intel_bo_alloc(brw
->bufmgr
, "shader time",
384 max_entries
* SHADER_TIME_STRIDE
* 3, 4096);
385 brw
->shader_time
.names
= rzalloc_array(brw
, const char *, max_entries
);
386 brw
->shader_time
.ids
= rzalloc_array(brw
, int, max_entries
);
387 brw
->shader_time
.types
= rzalloc_array(brw
, enum shader_time_shader_type
,
389 brw
->shader_time
.cumulative
= rzalloc_array(brw
, struct shader_times
,
391 brw
->shader_time
.max_entries
= max_entries
;
395 compare_time(const void *a
, const void *b
)
397 uint64_t * const *a_val
= a
;
398 uint64_t * const *b_val
= b
;
400 /* We don't just subtract because we're turning the value to an int. */
401 if (**a_val
< **b_val
)
403 else if (**a_val
== **b_val
)
410 print_shader_time_line(const char *stage
, const char *name
,
411 int shader_num
, uint64_t time
, uint64_t total
)
413 fprintf(stderr
, "%-6s%-18s", stage
, name
);
416 fprintf(stderr
, "%4d: ", shader_num
);
418 fprintf(stderr
, " : ");
420 fprintf(stderr
, "%16lld (%7.2f Gcycles) %4.1f%%\n",
422 (double)time
/ 1000000000.0,
423 (double)time
/ total
* 100.0);
427 brw_report_shader_time(struct brw_context
*brw
)
429 if (!brw
->shader_time
.bo
|| !brw
->shader_time
.num_entries
)
432 uint64_t scaled
[brw
->shader_time
.num_entries
];
433 uint64_t *sorted
[brw
->shader_time
.num_entries
];
434 uint64_t total_by_type
[ST_CS
+ 1];
435 memset(total_by_type
, 0, sizeof(total_by_type
));
437 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
438 uint64_t written
= 0, reset
= 0;
439 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
441 sorted
[i
] = &scaled
[i
];
451 written
= brw
->shader_time
.cumulative
[i
].written
;
452 reset
= brw
->shader_time
.cumulative
[i
].reset
;
456 /* I sometimes want to print things that aren't the 3 shader times.
457 * Just print the sum in that case.
464 uint64_t time
= brw
->shader_time
.cumulative
[i
].time
;
466 scaled
[i
] = time
/ written
* (written
+ reset
);
479 total_by_type
[type
] += scaled
[i
];
489 fprintf(stderr
, "No shader time collected yet\n");
493 qsort(sorted
, brw
->shader_time
.num_entries
, sizeof(sorted
[0]), compare_time
);
495 fprintf(stderr
, "\n");
496 fprintf(stderr
, "type ID cycles spent %% of total\n");
497 for (int s
= 0; s
< brw
->shader_time
.num_entries
; s
++) {
499 /* Work back from the sorted pointers times to a time to print. */
500 int i
= sorted
[s
] - scaled
;
505 int shader_num
= brw
->shader_time
.ids
[i
];
506 const char *shader_name
= brw
->shader_time
.names
[i
];
508 switch (brw
->shader_time
.types
[i
]) {
535 print_shader_time_line(stage
, shader_name
, shader_num
,
539 fprintf(stderr
, "\n");
540 print_shader_time_line("total", "vs", 0, total_by_type
[ST_VS
], total
);
541 print_shader_time_line("total", "tcs", 0, total_by_type
[ST_TCS
], total
);
542 print_shader_time_line("total", "tes", 0, total_by_type
[ST_TES
], total
);
543 print_shader_time_line("total", "gs", 0, total_by_type
[ST_GS
], total
);
544 print_shader_time_line("total", "fs8", 0, total_by_type
[ST_FS8
], total
);
545 print_shader_time_line("total", "fs16", 0, total_by_type
[ST_FS16
], total
);
546 print_shader_time_line("total", "cs", 0, total_by_type
[ST_CS
], total
);
550 brw_collect_shader_time(struct brw_context
*brw
)
552 if (!brw
->shader_time
.bo
)
555 /* This probably stalls on the last rendering. We could fix that by
556 * delaying reading the reports, but it doesn't look like it's a big
557 * overhead compared to the cost of tracking the time in the first place.
559 drm_intel_bo_map(brw
->shader_time
.bo
, true);
560 void *bo_map
= brw
->shader_time
.bo
->virtual;
562 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
563 uint32_t *times
= bo_map
+ i
* 3 * SHADER_TIME_STRIDE
;
565 brw
->shader_time
.cumulative
[i
].time
+= times
[SHADER_TIME_STRIDE
* 0 / 4];
566 brw
->shader_time
.cumulative
[i
].written
+= times
[SHADER_TIME_STRIDE
* 1 / 4];
567 brw
->shader_time
.cumulative
[i
].reset
+= times
[SHADER_TIME_STRIDE
* 2 / 4];
570 /* Zero the BO out to clear it out for our next collection.
572 memset(bo_map
, 0, brw
->shader_time
.bo
->size
);
573 drm_intel_bo_unmap(brw
->shader_time
.bo
);
577 brw_collect_and_report_shader_time(struct brw_context
*brw
)
579 brw_collect_shader_time(brw
);
581 if (brw
->shader_time
.report_time
== 0 ||
582 get_time() - brw
->shader_time
.report_time
>= 1.0) {
583 brw_report_shader_time(brw
);
584 brw
->shader_time
.report_time
= get_time();
589 * Chooses an index in the shader_time buffer and sets up tracking information
592 * Note that this holds on to references to the underlying programs, which may
593 * change their lifetimes compared to normal operation.
596 brw_get_shader_time_index(struct brw_context
*brw
,
597 struct gl_shader_program
*shader_prog
,
598 struct gl_program
*prog
,
599 enum shader_time_shader_type type
)
601 int shader_time_index
= brw
->shader_time
.num_entries
++;
602 assert(shader_time_index
< brw
->shader_time
.max_entries
);
603 brw
->shader_time
.types
[shader_time_index
] = type
;
605 int id
= shader_prog
? shader_prog
->Name
: prog
->Id
;
609 } else if (!shader_prog
) {
611 } else if (shader_prog
->Label
) {
612 name
= ralloc_strdup(brw
->shader_time
.names
, shader_prog
->Label
);
617 brw
->shader_time
.names
[shader_time_index
] = name
;
618 brw
->shader_time
.ids
[shader_time_index
] = id
;
620 return shader_time_index
;
624 brw_destroy_shader_time(struct brw_context
*brw
)
626 drm_intel_bo_unreference(brw
->shader_time
.bo
);
627 brw
->shader_time
.bo
= NULL
;
631 brw_stage_prog_data_free(const void *p
)
633 struct brw_stage_prog_data
*prog_data
= (struct brw_stage_prog_data
*)p
;
635 ralloc_free(prog_data
->param
);
636 ralloc_free(prog_data
->pull_param
);
637 ralloc_free(prog_data
->image_param
);
641 brw_dump_ir(const char *stage
, struct gl_shader_program
*shader_prog
,
642 struct gl_linked_shader
*shader
, struct gl_program
*prog
)
647 "GLSL IR for native %s shader %d:\n",
648 stage
, shader_prog
->Name
);
649 _mesa_print_ir(stderr
, shader
->ir
, NULL
);
650 fprintf(stderr
, "\n\n");
653 fprintf(stderr
, "ARB_%s_program %d ir for native %s shader\n",
654 stage
, prog
->Id
, stage
);
655 _mesa_print_program(prog
);
660 brw_setup_tex_for_precompile(struct brw_context
*brw
,
661 struct brw_sampler_prog_key_data
*tex
,
662 struct gl_program
*prog
)
664 const bool has_shader_channel_select
= brw
->is_haswell
|| brw
->gen
>= 8;
665 unsigned sampler_count
= util_last_bit(prog
->SamplersUsed
);
666 for (unsigned i
= 0; i
< sampler_count
; i
++) {
667 if (!has_shader_channel_select
&& (prog
->ShadowSamplers
& (1 << i
))) {
668 /* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
670 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_ONE
);
672 /* Color sampler: assume no swizzling. */
673 tex
->swizzles
[i
] = SWIZZLE_XYZW
;