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 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_lower_returns(nir
);
82 nir_validate_shader(nir
);
83 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
84 nir_shader_get_entrypoint(nir
), true, false);
86 nir
= prog_to_nir(prog
, options
);
87 NIR_PASS_V(nir
, nir_lower_regs_to_ssa
); /* turn registers into SSA */
89 nir_validate_shader(nir
);
93 nir
= brw_preprocess_nir(brw
->screen
->compiler
, nir
);
95 if (stage
== MESA_SHADER_FRAGMENT
) {
96 static const struct nir_lower_wpos_ytransform_options wpos_options
= {
97 .state_tokens
= {STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
, 0, 0, 0},
98 .fs_coord_pixel_center_integer
= 1,
99 .fs_coord_origin_upper_left
= 1,
101 _mesa_add_state_reference(prog
->Parameters
,
102 (gl_state_index
*) wpos_options
.state_tokens
);
104 NIR_PASS(progress
, nir
, nir_lower_wpos_ytransform
, &wpos_options
);
107 NIR_PASS(progress
, nir
, nir_lower_system_values
);
108 NIR_PASS_V(nir
, brw_nir_lower_uniforms
, is_scalar
);
110 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
112 /* nir_shader may have been cloned so make sure shader_info is in sync */
113 if (nir
->info
!= &prog
->info
) {
114 const char *name
= prog
->info
.name
;
115 const char *label
= prog
->info
.label
;
116 prog
->info
= *nir
->info
;
117 prog
->info
.name
= name
;
118 prog
->info
.label
= label
;
122 NIR_PASS_V(nir
, nir_lower_samplers
, shader_prog
);
123 NIR_PASS_V(nir
, nir_lower_atomics
, shader_prog
);
130 get_new_program_id(struct intel_screen
*screen
)
132 static pthread_mutex_t m
= PTHREAD_MUTEX_INITIALIZER
;
133 pthread_mutex_lock(&m
);
134 unsigned id
= screen
->program_id
++;
135 pthread_mutex_unlock(&m
);
139 static struct gl_program
*brwNewProgram(struct gl_context
*ctx
, GLenum target
,
140 GLuint id
, bool is_arb_asm
)
142 struct brw_context
*brw
= brw_context(ctx
);
145 case GL_VERTEX_PROGRAM_ARB
:
146 case GL_TESS_CONTROL_PROGRAM_NV
:
147 case GL_TESS_EVALUATION_PROGRAM_NV
:
148 case GL_GEOMETRY_PROGRAM_NV
:
149 case GL_COMPUTE_PROGRAM_NV
: {
150 struct brw_program
*prog
= rzalloc(NULL
, struct brw_program
);
152 prog
->id
= get_new_program_id(brw
->screen
);
154 return _mesa_init_gl_program(&prog
->program
, target
, id
, is_arb_asm
);
160 case GL_FRAGMENT_PROGRAM_ARB
: {
161 struct brw_program
*prog
= rzalloc(NULL
, struct brw_program
);
164 prog
->id
= get_new_program_id(brw
->screen
);
166 return _mesa_init_gl_program(&prog
->program
, target
, id
, is_arb_asm
);
173 unreachable("Unsupported target in brwNewProgram()");
177 static void brwDeleteProgram( struct gl_context
*ctx
,
178 struct gl_program
*prog
)
180 _mesa_delete_program( ctx
, prog
);
185 brwProgramStringNotify(struct gl_context
*ctx
,
187 struct gl_program
*prog
)
189 assert(target
== GL_VERTEX_PROGRAM_ARB
|| !prog
->arb
.IsPositionInvariant
);
191 struct brw_context
*brw
= brw_context(ctx
);
192 const struct brw_compiler
*compiler
= brw
->screen
->compiler
;
195 case GL_FRAGMENT_PROGRAM_ARB
: {
196 struct brw_program
*newFP
= brw_program(prog
);
197 const struct brw_program
*curFP
=
198 brw_program_const(brw
->fragment_program
);
201 brw
->ctx
.NewDriverState
|= BRW_NEW_FRAGMENT_PROGRAM
;
202 newFP
->id
= get_new_program_id(brw
->screen
);
204 brw_add_texrect_params(prog
);
206 prog
->nir
= brw_create_nir(brw
, NULL
, prog
, MESA_SHADER_FRAGMENT
, true);
208 brw_fs_precompile(ctx
, prog
);
211 case GL_VERTEX_PROGRAM_ARB
: {
212 struct brw_program
*newVP
= brw_program(prog
);
213 const struct brw_program
*curVP
=
214 brw_program_const(brw
->vertex_program
);
217 brw
->ctx
.NewDriverState
|= BRW_NEW_VERTEX_PROGRAM
;
218 if (newVP
->program
.arb
.IsPositionInvariant
) {
219 _mesa_insert_mvp_code(ctx
, &newVP
->program
);
221 newVP
->id
= get_new_program_id(brw
->screen
);
223 /* Also tell tnl about it:
225 _tnl_program_string(ctx
, target
, prog
);
227 brw_add_texrect_params(prog
);
229 prog
->nir
= brw_create_nir(brw
, NULL
, prog
, MESA_SHADER_VERTEX
,
230 compiler
->scalar_stage
[MESA_SHADER_VERTEX
]);
232 brw_vs_precompile(ctx
, prog
);
237 * driver->ProgramStringNotify is only called for ARB programs, fixed
238 * function vertex programs, and ir_to_mesa (which isn't used by the
239 * i965 back-end). Therefore, even after geometry shaders are added,
240 * this function should only ever be called with a target of
241 * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
243 unreachable("Unexpected target in brwProgramStringNotify");
250 brw_memory_barrier(struct gl_context
*ctx
, GLbitfield barriers
)
252 struct brw_context
*brw
= brw_context(ctx
);
253 unsigned bits
= (PIPE_CONTROL_DATA_CACHE_FLUSH
|
254 PIPE_CONTROL_NO_WRITE
|
255 PIPE_CONTROL_CS_STALL
);
256 assert(brw
->gen
>= 7 && brw
->gen
<= 9);
258 if (barriers
& (GL_VERTEX_ATTRIB_ARRAY_BARRIER_BIT
|
259 GL_ELEMENT_ARRAY_BARRIER_BIT
|
260 GL_COMMAND_BARRIER_BIT
))
261 bits
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
263 if (barriers
& GL_UNIFORM_BARRIER_BIT
)
264 bits
|= (PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
|
265 PIPE_CONTROL_CONST_CACHE_INVALIDATE
);
267 if (barriers
& GL_TEXTURE_FETCH_BARRIER_BIT
)
268 bits
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
270 if (barriers
& GL_TEXTURE_UPDATE_BARRIER_BIT
)
271 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
273 if (barriers
& GL_FRAMEBUFFER_BARRIER_BIT
)
274 bits
|= (PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
275 PIPE_CONTROL_RENDER_TARGET_FLUSH
);
277 /* Typed surface messages are handled by the render cache on IVB, so we
278 * need to flush it too.
280 if (brw
->gen
== 7 && !brw
->is_haswell
)
281 bits
|= PIPE_CONTROL_RENDER_TARGET_FLUSH
;
283 brw_emit_pipe_control_flush(brw
, bits
);
287 brw_blend_barrier(struct gl_context
*ctx
)
289 struct brw_context
*brw
= brw_context(ctx
);
291 if (!ctx
->Extensions
.MESA_shader_framebuffer_fetch
) {
293 brw_emit_pipe_control_flush(brw
,
294 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
295 PIPE_CONTROL_CS_STALL
);
296 brw_emit_pipe_control_flush(brw
,
297 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
);
299 brw_emit_pipe_control_flush(brw
,
300 PIPE_CONTROL_RENDER_TARGET_FLUSH
);
306 brw_add_texrect_params(struct gl_program
*prog
)
308 for (int texunit
= 0; texunit
< BRW_MAX_TEX_UNIT
; texunit
++) {
309 if (!(prog
->TexturesUsed
[texunit
] & (1 << TEXTURE_RECT_INDEX
)))
312 int tokens
[STATE_LENGTH
] = {
320 _mesa_add_state_reference(prog
->Parameters
, (gl_state_index
*)tokens
);
325 brw_get_scratch_bo(struct brw_context
*brw
,
326 drm_intel_bo
**scratch_bo
, int size
)
328 drm_intel_bo
*old_bo
= *scratch_bo
;
330 if (old_bo
&& old_bo
->size
< size
) {
331 drm_intel_bo_unreference(old_bo
);
336 *scratch_bo
= drm_intel_bo_alloc(brw
->bufmgr
, "scratch bo", size
, 4096);
341 * Reserve enough scratch space for the given stage to hold \p per_thread_size
342 * bytes times the given \p thread_count.
345 brw_alloc_stage_scratch(struct brw_context
*brw
,
346 struct brw_stage_state
*stage_state
,
347 unsigned per_thread_size
,
348 unsigned thread_count
)
350 if (stage_state
->per_thread_scratch
< per_thread_size
) {
351 stage_state
->per_thread_scratch
= per_thread_size
;
353 if (stage_state
->scratch_bo
)
354 drm_intel_bo_unreference(stage_state
->scratch_bo
);
356 stage_state
->scratch_bo
=
357 drm_intel_bo_alloc(brw
->bufmgr
, "shader scratch space",
358 per_thread_size
* thread_count
, 4096);
362 void brwInitFragProgFuncs( struct dd_function_table
*functions
)
364 assert(functions
->ProgramStringNotify
== _tnl_program_string
);
366 functions
->NewProgram
= brwNewProgram
;
367 functions
->DeleteProgram
= brwDeleteProgram
;
368 functions
->ProgramStringNotify
= brwProgramStringNotify
;
370 functions
->LinkShader
= brw_link_shader
;
372 functions
->MemoryBarrier
= brw_memory_barrier
;
373 functions
->BlendBarrier
= brw_blend_barrier
;
376 struct shader_times
{
383 brw_init_shader_time(struct brw_context
*brw
)
385 const int max_entries
= 2048;
386 brw
->shader_time
.bo
=
387 drm_intel_bo_alloc(brw
->bufmgr
, "shader time",
388 max_entries
* SHADER_TIME_STRIDE
* 3, 4096);
389 brw
->shader_time
.names
= rzalloc_array(brw
, const char *, max_entries
);
390 brw
->shader_time
.ids
= rzalloc_array(brw
, int, max_entries
);
391 brw
->shader_time
.types
= rzalloc_array(brw
, enum shader_time_shader_type
,
393 brw
->shader_time
.cumulative
= rzalloc_array(brw
, struct shader_times
,
395 brw
->shader_time
.max_entries
= max_entries
;
399 compare_time(const void *a
, const void *b
)
401 uint64_t * const *a_val
= a
;
402 uint64_t * const *b_val
= b
;
404 /* We don't just subtract because we're turning the value to an int. */
405 if (**a_val
< **b_val
)
407 else if (**a_val
== **b_val
)
414 print_shader_time_line(const char *stage
, const char *name
,
415 int shader_num
, uint64_t time
, uint64_t total
)
417 fprintf(stderr
, "%-6s%-18s", stage
, name
);
420 fprintf(stderr
, "%4d: ", shader_num
);
422 fprintf(stderr
, " : ");
424 fprintf(stderr
, "%16lld (%7.2f Gcycles) %4.1f%%\n",
426 (double)time
/ 1000000000.0,
427 (double)time
/ total
* 100.0);
431 brw_report_shader_time(struct brw_context
*brw
)
433 if (!brw
->shader_time
.bo
|| !brw
->shader_time
.num_entries
)
436 uint64_t scaled
[brw
->shader_time
.num_entries
];
437 uint64_t *sorted
[brw
->shader_time
.num_entries
];
438 uint64_t total_by_type
[ST_CS
+ 1];
439 memset(total_by_type
, 0, sizeof(total_by_type
));
441 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
442 uint64_t written
= 0, reset
= 0;
443 enum shader_time_shader_type type
= brw
->shader_time
.types
[i
];
445 sorted
[i
] = &scaled
[i
];
455 written
= brw
->shader_time
.cumulative
[i
].written
;
456 reset
= brw
->shader_time
.cumulative
[i
].reset
;
460 /* I sometimes want to print things that aren't the 3 shader times.
461 * Just print the sum in that case.
468 uint64_t time
= brw
->shader_time
.cumulative
[i
].time
;
470 scaled
[i
] = time
/ written
* (written
+ reset
);
483 total_by_type
[type
] += scaled
[i
];
493 fprintf(stderr
, "No shader time collected yet\n");
497 qsort(sorted
, brw
->shader_time
.num_entries
, sizeof(sorted
[0]), compare_time
);
499 fprintf(stderr
, "\n");
500 fprintf(stderr
, "type ID cycles spent %% of total\n");
501 for (int s
= 0; s
< brw
->shader_time
.num_entries
; s
++) {
503 /* Work back from the sorted pointers times to a time to print. */
504 int i
= sorted
[s
] - scaled
;
509 int shader_num
= brw
->shader_time
.ids
[i
];
510 const char *shader_name
= brw
->shader_time
.names
[i
];
512 switch (brw
->shader_time
.types
[i
]) {
539 print_shader_time_line(stage
, shader_name
, shader_num
,
543 fprintf(stderr
, "\n");
544 print_shader_time_line("total", "vs", 0, total_by_type
[ST_VS
], total
);
545 print_shader_time_line("total", "tcs", 0, total_by_type
[ST_TCS
], total
);
546 print_shader_time_line("total", "tes", 0, total_by_type
[ST_TES
], total
);
547 print_shader_time_line("total", "gs", 0, total_by_type
[ST_GS
], total
);
548 print_shader_time_line("total", "fs8", 0, total_by_type
[ST_FS8
], total
);
549 print_shader_time_line("total", "fs16", 0, total_by_type
[ST_FS16
], total
);
550 print_shader_time_line("total", "cs", 0, total_by_type
[ST_CS
], total
);
554 brw_collect_shader_time(struct brw_context
*brw
)
556 if (!brw
->shader_time
.bo
)
559 /* This probably stalls on the last rendering. We could fix that by
560 * delaying reading the reports, but it doesn't look like it's a big
561 * overhead compared to the cost of tracking the time in the first place.
563 drm_intel_bo_map(brw
->shader_time
.bo
, true);
564 void *bo_map
= brw
->shader_time
.bo
->virtual;
566 for (int i
= 0; i
< brw
->shader_time
.num_entries
; i
++) {
567 uint32_t *times
= bo_map
+ i
* 3 * SHADER_TIME_STRIDE
;
569 brw
->shader_time
.cumulative
[i
].time
+= times
[SHADER_TIME_STRIDE
* 0 / 4];
570 brw
->shader_time
.cumulative
[i
].written
+= times
[SHADER_TIME_STRIDE
* 1 / 4];
571 brw
->shader_time
.cumulative
[i
].reset
+= times
[SHADER_TIME_STRIDE
* 2 / 4];
574 /* Zero the BO out to clear it out for our next collection.
576 memset(bo_map
, 0, brw
->shader_time
.bo
->size
);
577 drm_intel_bo_unmap(brw
->shader_time
.bo
);
581 brw_collect_and_report_shader_time(struct brw_context
*brw
)
583 brw_collect_shader_time(brw
);
585 if (brw
->shader_time
.report_time
== 0 ||
586 get_time() - brw
->shader_time
.report_time
>= 1.0) {
587 brw_report_shader_time(brw
);
588 brw
->shader_time
.report_time
= get_time();
593 * Chooses an index in the shader_time buffer and sets up tracking information
596 * Note that this holds on to references to the underlying programs, which may
597 * change their lifetimes compared to normal operation.
600 brw_get_shader_time_index(struct brw_context
*brw
, struct gl_program
*prog
,
601 enum shader_time_shader_type type
, bool is_glsl_sh
)
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
;
610 } else if (is_glsl_sh
) {
611 name
= prog
->info
.label
?
612 ralloc_strdup(brw
->shader_time
.names
, prog
->info
.label
) : "glsl";
617 brw
->shader_time
.names
[shader_time_index
] = name
;
618 brw
->shader_time
.ids
[shader_time_index
] = prog
->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_arb_asm(const char *stage
, struct gl_program
*prog
)
643 fprintf(stderr
, "ARB_%s_program %d ir for native %s shader\n",
644 stage
, prog
->Id
, stage
);
645 _mesa_print_program(prog
);
649 brw_setup_tex_for_precompile(struct brw_context
*brw
,
650 struct brw_sampler_prog_key_data
*tex
,
651 struct gl_program
*prog
)
653 const bool has_shader_channel_select
= brw
->is_haswell
|| brw
->gen
>= 8;
654 unsigned sampler_count
= util_last_bit(prog
->SamplersUsed
);
655 for (unsigned i
= 0; i
< sampler_count
; i
++) {
656 if (!has_shader_channel_select
&& (prog
->ShadowSamplers
& (1 << i
))) {
657 /* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
659 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_ONE
);
661 /* Color sampler: assume no swizzling. */
662 tex
->swizzles
[i
] = SWIZZLE_XYZW
;