2 * Copyright © 2010 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Eric Anholt <eric@anholt.net>
30 #include <sys/types.h>
32 #include "main/macros.h"
33 #include "main/shaderobj.h"
34 #include "main/uniforms.h"
35 #include "program/prog_parameter.h"
36 #include "program/prog_print.h"
37 #include "program/prog_optimize.h"
38 #include "program/register_allocate.h"
39 #include "program/sampler.h"
40 #include "program/hash_table.h"
41 #include "brw_context.h"
46 #include "../glsl/glsl_types.h"
47 #include "../glsl/ir_optimization.h"
48 #include "../glsl/ir_print_visitor.h"
51 ARF
= BRW_ARCHITECTURE_REGISTER_FILE
,
52 GRF
= BRW_GENERAL_REGISTER_FILE
,
53 MRF
= BRW_MESSAGE_REGISTER_FILE
,
54 IMM
= BRW_IMMEDIATE_VALUE
,
55 FIXED_HW_REG
, /* a struct brw_reg */
56 UNIFORM
, /* prog_data->params[hw_reg] */
61 FS_OPCODE_FB_WRITE
= 256,
79 static int using_new_fs
= -1;
80 static struct brw_reg
brw_reg_from_fs_reg(class fs_reg
*reg
);
83 brw_new_shader(GLcontext
*ctx
, GLuint name
, GLuint type
)
85 struct brw_shader
*shader
;
87 shader
= talloc_zero(NULL
, struct brw_shader
);
89 shader
->base
.Type
= type
;
90 shader
->base
.Name
= name
;
91 _mesa_init_shader(ctx
, &shader
->base
);
97 struct gl_shader_program
*
98 brw_new_shader_program(GLcontext
*ctx
, GLuint name
)
100 struct brw_shader_program
*prog
;
101 prog
= talloc_zero(NULL
, struct brw_shader_program
);
103 prog
->base
.Name
= name
;
104 _mesa_init_shader_program(ctx
, &prog
->base
);
110 brw_compile_shader(GLcontext
*ctx
, struct gl_shader
*shader
)
112 if (!_mesa_ir_compile_shader(ctx
, shader
))
119 brw_link_shader(GLcontext
*ctx
, struct gl_shader_program
*prog
)
121 if (using_new_fs
== -1)
122 using_new_fs
= getenv("INTEL_NEW_FS") != NULL
;
124 for (unsigned i
= 0; i
< prog
->_NumLinkedShaders
; i
++) {
125 struct brw_shader
*shader
= (struct brw_shader
*)prog
->_LinkedShaders
[i
];
127 if (using_new_fs
&& shader
->base
.Type
== GL_FRAGMENT_SHADER
) {
128 void *mem_ctx
= talloc_new(NULL
);
132 talloc_free(shader
->ir
);
133 shader
->ir
= new(shader
) exec_list
;
134 clone_ir_list(mem_ctx
, shader
->ir
, shader
->base
.ir
);
136 do_mat_op_to_vec(shader
->ir
);
137 do_mod_to_fract(shader
->ir
);
138 do_div_to_mul_rcp(shader
->ir
);
139 do_sub_to_add_neg(shader
->ir
);
140 do_explog_to_explog2(shader
->ir
);
141 do_lower_texture_projection(shader
->ir
);
146 brw_do_channel_expressions(shader
->ir
);
147 brw_do_vector_splitting(shader
->ir
);
149 progress
= do_lower_jumps(shader
->ir
, true, true,
150 true, /* main return */
151 false, /* continue */
155 progress
= do_common_optimization(shader
->ir
, true, 32) || progress
;
157 progress
= lower_noise(shader
->ir
) || progress
;
159 lower_variable_index_to_cond_assign(shader
->ir
,
161 GL_TRUE
, /* output */
163 GL_TRUE
/* uniform */
167 validate_ir_tree(shader
->ir
);
169 reparent_ir(shader
->ir
, shader
->ir
);
170 talloc_free(mem_ctx
);
174 if (!_mesa_ir_link_shader(ctx
, prog
))
181 type_size(const struct glsl_type
*type
)
183 unsigned int size
, i
;
185 switch (type
->base_type
) {
188 case GLSL_TYPE_FLOAT
:
190 return type
->components();
191 case GLSL_TYPE_ARRAY
:
192 return type_size(type
->fields
.array
) * type
->length
;
193 case GLSL_TYPE_STRUCT
:
195 for (i
= 0; i
< type
->length
; i
++) {
196 size
+= type_size(type
->fields
.structure
[i
].type
);
199 case GLSL_TYPE_SAMPLER
:
200 /* Samplers take up no register space, since they're baked in at
205 assert(!"not reached");
212 /* Callers of this talloc-based new need not call delete. It's
213 * easier to just talloc_free 'ctx' (or any of its ancestors). */
214 static void* operator new(size_t size
, void *ctx
)
218 node
= talloc_size(ctx
, size
);
219 assert(node
!= NULL
);
227 this->reg_offset
= 0;
233 /** Generic unset register constructor. */
237 this->file
= BAD_FILE
;
240 /** Immediate value constructor. */
245 this->type
= BRW_REGISTER_TYPE_F
;
249 /** Immediate value constructor. */
254 this->type
= BRW_REGISTER_TYPE_D
;
258 /** Immediate value constructor. */
263 this->type
= BRW_REGISTER_TYPE_UD
;
267 /** Fixed brw_reg Immediate value constructor. */
268 fs_reg(struct brw_reg fixed_hw_reg
)
271 this->file
= FIXED_HW_REG
;
272 this->fixed_hw_reg
= fixed_hw_reg
;
273 this->type
= fixed_hw_reg
.type
;
276 fs_reg(enum register_file file
, int hw_reg
);
277 fs_reg(class fs_visitor
*v
, const struct glsl_type
*type
);
279 /** Register file: ARF, GRF, MRF, IMM. */
280 enum register_file file
;
281 /** virtual register number. 0 = fixed hw reg */
283 /** Offset within the virtual register. */
285 /** HW register number. Generally unset until register allocation. */
287 /** Register type. BRW_REGISTER_TYPE_* */
291 struct brw_reg fixed_hw_reg
;
293 /** Value for file == BRW_IMMMEDIATE_FILE */
301 static const fs_reg reg_undef
;
302 static const fs_reg
reg_null(ARF
, BRW_ARF_NULL
);
304 class fs_inst
: public exec_node
{
306 /* Callers of this talloc-based new need not call delete. It's
307 * easier to just talloc_free 'ctx' (or any of its ancestors). */
308 static void* operator new(size_t size
, void *ctx
)
312 node
= talloc_zero_size(ctx
, size
);
313 assert(node
!= NULL
);
320 this->opcode
= BRW_OPCODE_NOP
;
321 this->saturate
= false;
322 this->conditional_mod
= BRW_CONDITIONAL_NONE
;
323 this->predicated
= false;
327 this->shadow_compare
= false;
338 this->opcode
= opcode
;
341 fs_inst(int opcode
, fs_reg dst
, fs_reg src0
)
344 this->opcode
= opcode
;
349 fs_inst(int opcode
, fs_reg dst
, fs_reg src0
, fs_reg src1
)
352 this->opcode
= opcode
;
358 fs_inst(int opcode
, fs_reg dst
, fs_reg src0
, fs_reg src1
, fs_reg src2
)
361 this->opcode
= opcode
;
368 int opcode
; /* BRW_OPCODE_* or FS_OPCODE_* */
373 int conditional_mod
; /**< BRW_CONDITIONAL_* */
375 int mlen
; /**< SEND message length */
377 int target
; /**< MRT target. */
382 * Annotation for the generated IR. One of the two can be set.
385 const char *annotation
;
389 class fs_visitor
: public ir_visitor
393 fs_visitor(struct brw_wm_compile
*c
, struct brw_shader
*shader
)
398 this->fp
= brw
->fragment_program
;
399 this->intel
= &brw
->intel
;
400 this->ctx
= &intel
->ctx
;
401 this->mem_ctx
= talloc_new(NULL
);
402 this->shader
= shader
;
404 this->variable_ht
= hash_table_ctor(0,
405 hash_table_pointer_hash
,
406 hash_table_pointer_compare
);
408 this->frag_color
= NULL
;
409 this->frag_data
= NULL
;
410 this->frag_depth
= NULL
;
411 this->first_non_payload_grf
= 0;
413 this->current_annotation
= NULL
;
414 this->annotation_string
= NULL
;
415 this->annotation_ir
= NULL
;
416 this->base_ir
= NULL
;
418 this->virtual_grf_sizes
= NULL
;
419 this->virtual_grf_next
= 1;
420 this->virtual_grf_array_size
= 0;
421 this->virtual_grf_def
= NULL
;
422 this->virtual_grf_use
= NULL
;
426 talloc_free(this->mem_ctx
);
427 hash_table_dtor(this->variable_ht
);
430 fs_reg
*variable_storage(ir_variable
*var
);
431 int virtual_grf_alloc(int size
);
433 void visit(ir_variable
*ir
);
434 void visit(ir_assignment
*ir
);
435 void visit(ir_dereference_variable
*ir
);
436 void visit(ir_dereference_record
*ir
);
437 void visit(ir_dereference_array
*ir
);
438 void visit(ir_expression
*ir
);
439 void visit(ir_texture
*ir
);
440 void visit(ir_if
*ir
);
441 void visit(ir_constant
*ir
);
442 void visit(ir_swizzle
*ir
);
443 void visit(ir_return
*ir
);
444 void visit(ir_loop
*ir
);
445 void visit(ir_loop_jump
*ir
);
446 void visit(ir_discard
*ir
);
447 void visit(ir_call
*ir
);
448 void visit(ir_function
*ir
);
449 void visit(ir_function_signature
*ir
);
451 fs_inst
*emit(fs_inst inst
);
452 void assign_curb_setup();
453 void calculate_urb_setup();
454 void assign_urb_setup();
456 void assign_regs_trivial();
457 void calculate_live_intervals();
458 bool virtual_grf_interferes(int a
, int b
);
459 void generate_code();
460 void generate_fb_write(fs_inst
*inst
);
461 void generate_linterp(fs_inst
*inst
, struct brw_reg dst
,
462 struct brw_reg
*src
);
463 void generate_tex(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
);
464 void generate_math(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg
*src
);
465 void generate_discard(fs_inst
*inst
, struct brw_reg temp
);
466 void generate_ddx(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
);
467 void generate_ddy(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
);
469 void emit_dummy_fs();
470 void emit_fragcoord_interpolation(ir_variable
*ir
);
471 void emit_general_interpolation(ir_variable
*ir
);
472 void emit_interpolation_setup_gen4();
473 void emit_interpolation_setup_gen6();
474 fs_inst
*emit_texture_gen4(ir_texture
*ir
, fs_reg dst
, fs_reg coordinate
);
475 fs_inst
*emit_texture_gen5(ir_texture
*ir
, fs_reg dst
, fs_reg coordinate
);
476 void emit_fb_writes();
477 void emit_assignment_writes(fs_reg
&l
, fs_reg
&r
,
478 const glsl_type
*type
, bool predicated
);
480 struct brw_reg
interp_reg(int location
, int channel
);
481 int setup_uniform_values(int loc
, const glsl_type
*type
);
482 void setup_builtin_uniform_values(ir_variable
*ir
);
484 struct brw_context
*brw
;
485 const struct gl_fragment_program
*fp
;
486 struct intel_context
*intel
;
488 struct brw_wm_compile
*c
;
489 struct brw_compile
*p
;
490 struct brw_shader
*shader
;
492 exec_list instructions
;
494 int *virtual_grf_sizes
;
495 int virtual_grf_next
;
496 int virtual_grf_array_size
;
497 int *virtual_grf_def
;
498 int *virtual_grf_use
;
500 struct hash_table
*variable_ht
;
501 ir_variable
*frag_color
, *frag_data
, *frag_depth
;
502 int first_non_payload_grf
;
503 int urb_setup
[FRAG_ATTRIB_MAX
];
505 /** @{ debug annotation info */
506 const char *current_annotation
;
507 ir_instruction
*base_ir
;
508 const char **annotation_string
;
509 ir_instruction
**annotation_ir
;
514 /* Result of last visit() method. */
529 fs_visitor::virtual_grf_alloc(int size
)
531 if (virtual_grf_array_size
<= virtual_grf_next
) {
532 if (virtual_grf_array_size
== 0)
533 virtual_grf_array_size
= 16;
535 virtual_grf_array_size
*= 2;
536 virtual_grf_sizes
= talloc_realloc(mem_ctx
, virtual_grf_sizes
,
537 int, virtual_grf_array_size
);
539 /* This slot is always unused. */
540 virtual_grf_sizes
[0] = 0;
542 virtual_grf_sizes
[virtual_grf_next
] = size
;
543 return virtual_grf_next
++;
546 /** Fixed HW reg constructor. */
547 fs_reg::fs_reg(enum register_file file
, int hw_reg
)
551 this->hw_reg
= hw_reg
;
552 this->type
= BRW_REGISTER_TYPE_F
;
556 brw_type_for_base_type(const struct glsl_type
*type
)
558 switch (type
->base_type
) {
559 case GLSL_TYPE_FLOAT
:
560 return BRW_REGISTER_TYPE_F
;
563 return BRW_REGISTER_TYPE_D
;
565 return BRW_REGISTER_TYPE_UD
;
566 case GLSL_TYPE_ARRAY
:
567 case GLSL_TYPE_STRUCT
:
568 /* These should be overridden with the type of the member when
569 * dereferenced into. BRW_REGISTER_TYPE_UD seems like a likely
570 * way to trip up if we don't.
572 return BRW_REGISTER_TYPE_UD
;
574 assert(!"not reached");
575 return BRW_REGISTER_TYPE_F
;
579 /** Automatic reg constructor. */
580 fs_reg::fs_reg(class fs_visitor
*v
, const struct glsl_type
*type
)
585 this->reg
= v
->virtual_grf_alloc(type_size(type
));
586 this->reg_offset
= 0;
587 this->type
= brw_type_for_base_type(type
);
591 fs_visitor::variable_storage(ir_variable
*var
)
593 return (fs_reg
*)hash_table_find(this->variable_ht
, var
);
596 /* Our support for uniforms is piggy-backed on the struct
597 * gl_fragment_program, because that's where the values actually
598 * get stored, rather than in some global gl_shader_program uniform
602 fs_visitor::setup_uniform_values(int loc
, const glsl_type
*type
)
604 unsigned int offset
= 0;
607 if (type
->is_matrix()) {
608 const glsl_type
*column
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
609 type
->vector_elements
,
612 for (unsigned int i
= 0; i
< type
->matrix_columns
; i
++) {
613 offset
+= setup_uniform_values(loc
+ offset
, column
);
619 switch (type
->base_type
) {
620 case GLSL_TYPE_FLOAT
:
624 vec_values
= fp
->Base
.Parameters
->ParameterValues
[loc
];
625 for (unsigned int i
= 0; i
< type
->vector_elements
; i
++) {
626 c
->prog_data
.param
[c
->prog_data
.nr_params
++] = &vec_values
[i
];
630 case GLSL_TYPE_STRUCT
:
631 for (unsigned int i
= 0; i
< type
->length
; i
++) {
632 offset
+= setup_uniform_values(loc
+ offset
,
633 type
->fields
.structure
[i
].type
);
637 case GLSL_TYPE_ARRAY
:
638 for (unsigned int i
= 0; i
< type
->length
; i
++) {
639 offset
+= setup_uniform_values(loc
+ offset
, type
->fields
.array
);
643 case GLSL_TYPE_SAMPLER
:
644 /* The sampler takes up a slot, but we don't use any values from it. */
648 assert(!"not reached");
654 /* Our support for builtin uniforms is even scarier than non-builtin.
655 * It sits on top of the PROG_STATE_VAR parameters that are
656 * automatically updated from GL context state.
659 fs_visitor::setup_builtin_uniform_values(ir_variable
*ir
)
661 const struct gl_builtin_uniform_desc
*statevar
= NULL
;
663 for (unsigned int i
= 0; _mesa_builtin_uniform_desc
[i
].name
; i
++) {
664 statevar
= &_mesa_builtin_uniform_desc
[i
];
665 if (strcmp(ir
->name
, _mesa_builtin_uniform_desc
[i
].name
) == 0)
669 if (!statevar
->name
) {
671 printf("Failed to find builtin uniform `%s'\n", ir
->name
);
676 if (ir
->type
->is_array()) {
677 array_count
= ir
->type
->length
;
682 for (int a
= 0; a
< array_count
; a
++) {
683 for (unsigned int i
= 0; i
< statevar
->num_elements
; i
++) {
684 struct gl_builtin_uniform_element
*element
= &statevar
->elements
[i
];
685 int tokens
[STATE_LENGTH
];
687 memcpy(tokens
, element
->tokens
, sizeof(element
->tokens
));
688 if (ir
->type
->is_array()) {
692 /* This state reference has already been setup by ir_to_mesa,
693 * but we'll get the same index back here.
695 int index
= _mesa_add_state_reference(this->fp
->Base
.Parameters
,
696 (gl_state_index
*)tokens
);
697 float *vec_values
= this->fp
->Base
.Parameters
->ParameterValues
[index
];
699 /* Add each of the unique swizzles of the element as a
700 * parameter. This'll end up matching the expected layout of
701 * the array/matrix/structure we're trying to fill in.
704 for (unsigned int i
= 0; i
< 4; i
++) {
705 int swiz
= GET_SWZ(element
->swizzle
, i
);
706 if (swiz
== last_swiz
)
710 c
->prog_data
.param
[c
->prog_data
.nr_params
++] = &vec_values
[swiz
];
717 fs_visitor::emit_fragcoord_interpolation(ir_variable
*ir
)
719 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(this, ir
->type
);
721 fs_reg neg_y
= this->pixel_y
;
725 if (ir
->pixel_center_integer
) {
726 emit(fs_inst(BRW_OPCODE_MOV
, wpos
, this->pixel_x
));
728 emit(fs_inst(BRW_OPCODE_ADD
, wpos
, this->pixel_x
, fs_reg(0.5f
)));
733 if (ir
->origin_upper_left
&& ir
->pixel_center_integer
) {
734 emit(fs_inst(BRW_OPCODE_MOV
, wpos
, this->pixel_y
));
736 fs_reg pixel_y
= this->pixel_y
;
737 float offset
= (ir
->pixel_center_integer
? 0.0 : 0.5);
739 if (!ir
->origin_upper_left
) {
740 pixel_y
.negate
= true;
741 offset
+= c
->key
.drawable_height
- 1.0;
744 emit(fs_inst(BRW_OPCODE_ADD
, wpos
, pixel_y
, fs_reg(offset
)));
749 emit(fs_inst(FS_OPCODE_LINTERP
, wpos
, this->delta_x
, this->delta_y
,
750 interp_reg(FRAG_ATTRIB_WPOS
, 2)));
753 /* gl_FragCoord.w: Already set up in emit_interpolation */
754 emit(fs_inst(BRW_OPCODE_MOV
, wpos
, this->wpos_w
));
756 hash_table_insert(this->variable_ht
, reg
, ir
);
761 fs_visitor::emit_general_interpolation(ir_variable
*ir
)
763 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(this, ir
->type
);
764 /* Interpolation is always in floating point regs. */
765 reg
->type
= BRW_REGISTER_TYPE_F
;
768 unsigned int array_elements
;
769 const glsl_type
*type
;
771 if (ir
->type
->is_array()) {
772 array_elements
= ir
->type
->length
;
773 if (array_elements
== 0) {
776 type
= ir
->type
->fields
.array
;
782 int location
= ir
->location
;
783 for (unsigned int i
= 0; i
< array_elements
; i
++) {
784 for (unsigned int j
= 0; j
< type
->matrix_columns
; j
++) {
785 if (urb_setup
[location
] == -1) {
786 /* If there's no incoming setup data for this slot, don't
787 * emit interpolation for it.
789 attr
.reg_offset
+= type
->vector_elements
;
794 for (unsigned int c
= 0; c
< type
->vector_elements
; c
++) {
795 struct brw_reg interp
= interp_reg(location
, c
);
796 emit(fs_inst(FS_OPCODE_LINTERP
,
803 attr
.reg_offset
-= type
->vector_elements
;
805 for (unsigned int c
= 0; c
< type
->vector_elements
; c
++) {
806 emit(fs_inst(BRW_OPCODE_MUL
,
816 hash_table_insert(this->variable_ht
, reg
, ir
);
820 fs_visitor::visit(ir_variable
*ir
)
824 if (variable_storage(ir
))
827 if (strcmp(ir
->name
, "gl_FragColor") == 0) {
828 this->frag_color
= ir
;
829 } else if (strcmp(ir
->name
, "gl_FragData") == 0) {
830 this->frag_data
= ir
;
831 } else if (strcmp(ir
->name
, "gl_FragDepth") == 0) {
832 this->frag_depth
= ir
;
835 if (ir
->mode
== ir_var_in
) {
836 if (!strcmp(ir
->name
, "gl_FragCoord")) {
837 emit_fragcoord_interpolation(ir
);
839 } else if (!strcmp(ir
->name
, "gl_FrontFacing")) {
840 reg
= new(this->mem_ctx
) fs_reg(this, ir
->type
);
841 struct brw_reg r1_6ud
= retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_UD
);
842 /* bit 31 is "primitive is back face", so checking < (1 << 31) gives
845 fs_inst
*inst
= emit(fs_inst(BRW_OPCODE_CMP
,
849 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
850 emit(fs_inst(BRW_OPCODE_AND
, *reg
, *reg
, fs_reg(1u)));
852 emit_general_interpolation(ir
);
857 if (ir
->mode
== ir_var_uniform
) {
858 int param_index
= c
->prog_data
.nr_params
;
860 if (!strncmp(ir
->name
, "gl_", 3)) {
861 setup_builtin_uniform_values(ir
);
863 setup_uniform_values(ir
->location
, ir
->type
);
866 reg
= new(this->mem_ctx
) fs_reg(UNIFORM
, param_index
);
870 reg
= new(this->mem_ctx
) fs_reg(this, ir
->type
);
872 hash_table_insert(this->variable_ht
, reg
, ir
);
876 fs_visitor::visit(ir_dereference_variable
*ir
)
878 fs_reg
*reg
= variable_storage(ir
->var
);
883 fs_visitor::visit(ir_dereference_record
*ir
)
885 const glsl_type
*struct_type
= ir
->record
->type
;
887 ir
->record
->accept(this);
889 unsigned int offset
= 0;
890 for (unsigned int i
= 0; i
< struct_type
->length
; i
++) {
891 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
893 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
895 this->result
.reg_offset
+= offset
;
896 this->result
.type
= brw_type_for_base_type(ir
->type
);
900 fs_visitor::visit(ir_dereference_array
*ir
)
905 ir
->array
->accept(this);
906 index
= ir
->array_index
->as_constant();
908 element_size
= type_size(ir
->type
);
909 this->result
.type
= brw_type_for_base_type(ir
->type
);
912 assert(this->result
.file
== UNIFORM
||
913 (this->result
.file
== GRF
&&
914 this->result
.reg
!= 0));
915 this->result
.reg_offset
+= index
->value
.i
[0] * element_size
;
917 assert(!"FINISHME: non-constant array element");
922 fs_visitor::visit(ir_expression
*ir
)
924 unsigned int operand
;
929 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
930 ir
->operands
[operand
]->accept(this);
931 if (this->result
.file
== BAD_FILE
) {
933 printf("Failed to get tree for expression operand:\n");
934 ir
->operands
[operand
]->accept(&v
);
937 op
[operand
] = this->result
;
939 /* Matrix expression operands should have been broken down to vector
940 * operations already.
942 assert(!ir
->operands
[operand
]->type
->is_matrix());
943 /* And then those vector operands should have been broken down to scalar.
945 assert(!ir
->operands
[operand
]->type
->is_vector());
948 /* Storage for our result. If our result goes into an assignment, it will
949 * just get copy-propagated out, so no worries.
951 this->result
= fs_reg(this, ir
->type
);
953 switch (ir
->operation
) {
954 case ir_unop_logic_not
:
955 emit(fs_inst(BRW_OPCODE_ADD
, this->result
, op
[0], fs_reg(-1)));
958 op
[0].negate
= !op
[0].negate
;
959 this->result
= op
[0];
963 this->result
= op
[0];
966 temp
= fs_reg(this, ir
->type
);
968 emit(fs_inst(BRW_OPCODE_MOV
, this->result
, fs_reg(0.0f
)));
970 inst
= emit(fs_inst(BRW_OPCODE_CMP
, reg_null
, op
[0], fs_reg(0.0f
)));
971 inst
->conditional_mod
= BRW_CONDITIONAL_G
;
972 inst
= emit(fs_inst(BRW_OPCODE_MOV
, this->result
, fs_reg(1.0f
)));
973 inst
->predicated
= true;
975 inst
= emit(fs_inst(BRW_OPCODE_CMP
, reg_null
, op
[0], fs_reg(0.0f
)));
976 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
977 inst
= emit(fs_inst(BRW_OPCODE_MOV
, this->result
, fs_reg(-1.0f
)));
978 inst
->predicated
= true;
982 emit(fs_inst(FS_OPCODE_RCP
, this->result
, op
[0]));
986 emit(fs_inst(FS_OPCODE_EXP2
, this->result
, op
[0]));
989 emit(fs_inst(FS_OPCODE_LOG2
, this->result
, op
[0]));
993 assert(!"not reached: should be handled by ir_explog_to_explog2");
996 emit(fs_inst(FS_OPCODE_SIN
, this->result
, op
[0]));
999 emit(fs_inst(FS_OPCODE_COS
, this->result
, op
[0]));
1003 emit(fs_inst(FS_OPCODE_DDX
, this->result
, op
[0]));
1006 emit(fs_inst(FS_OPCODE_DDY
, this->result
, op
[0]));
1010 emit(fs_inst(BRW_OPCODE_ADD
, this->result
, op
[0], op
[1]));
1013 assert(!"not reached: should be handled by ir_sub_to_add_neg");
1017 emit(fs_inst(BRW_OPCODE_MUL
, this->result
, op
[0], op
[1]));
1020 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1023 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1027 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
1028 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
1029 emit(fs_inst(BRW_OPCODE_AND
, this->result
, this->result
, fs_reg(0x1)));
1031 case ir_binop_greater
:
1032 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
1033 inst
->conditional_mod
= BRW_CONDITIONAL_G
;
1034 emit(fs_inst(BRW_OPCODE_AND
, this->result
, this->result
, fs_reg(0x1)));
1036 case ir_binop_lequal
:
1037 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
1038 inst
->conditional_mod
= BRW_CONDITIONAL_LE
;
1039 emit(fs_inst(BRW_OPCODE_AND
, this->result
, this->result
, fs_reg(0x1)));
1041 case ir_binop_gequal
:
1042 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
1043 inst
->conditional_mod
= BRW_CONDITIONAL_GE
;
1044 emit(fs_inst(BRW_OPCODE_AND
, this->result
, this->result
, fs_reg(0x1)));
1046 case ir_binop_equal
:
1047 case ir_binop_all_equal
: /* same as nequal for scalars */
1048 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
1049 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
1050 emit(fs_inst(BRW_OPCODE_AND
, this->result
, this->result
, fs_reg(0x1)));
1052 case ir_binop_nequal
:
1053 case ir_binop_any_nequal
: /* same as nequal for scalars */
1054 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
1055 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
1056 emit(fs_inst(BRW_OPCODE_AND
, this->result
, this->result
, fs_reg(0x1)));
1059 case ir_binop_logic_xor
:
1060 emit(fs_inst(BRW_OPCODE_XOR
, this->result
, op
[0], op
[1]));
1063 case ir_binop_logic_or
:
1064 emit(fs_inst(BRW_OPCODE_OR
, this->result
, op
[0], op
[1]));
1067 case ir_binop_logic_and
:
1068 emit(fs_inst(BRW_OPCODE_AND
, this->result
, op
[0], op
[1]));
1072 case ir_binop_cross
:
1074 assert(!"not reached: should be handled by brw_fs_channel_expressions");
1078 assert(!"not reached: should be handled by lower_noise");
1082 emit(fs_inst(FS_OPCODE_SQRT
, this->result
, op
[0]));
1086 emit(fs_inst(FS_OPCODE_RSQ
, this->result
, op
[0]));
1092 emit(fs_inst(BRW_OPCODE_MOV
, this->result
, op
[0]));
1095 emit(fs_inst(BRW_OPCODE_MOV
, this->result
, op
[0]));
1099 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], fs_reg(0.0f
)));
1100 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
1103 emit(fs_inst(BRW_OPCODE_RNDD
, this->result
, op
[0]));
1106 op
[0].negate
= ~op
[0].negate
;
1107 inst
= emit(fs_inst(BRW_OPCODE_RNDD
, this->result
, op
[0]));
1108 this->result
.negate
= true;
1111 inst
= emit(fs_inst(BRW_OPCODE_RNDD
, this->result
, op
[0]));
1114 inst
= emit(fs_inst(BRW_OPCODE_FRC
, this->result
, op
[0]));
1118 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
1119 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
1121 inst
= emit(fs_inst(BRW_OPCODE_SEL
, this->result
, op
[0], op
[1]));
1122 inst
->predicated
= true;
1125 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
1126 inst
->conditional_mod
= BRW_CONDITIONAL_G
;
1128 inst
= emit(fs_inst(BRW_OPCODE_SEL
, this->result
, op
[0], op
[1]));
1129 inst
->predicated
= true;
1133 inst
= emit(fs_inst(FS_OPCODE_POW
, this->result
, op
[0], op
[1]));
1136 case ir_unop_bit_not
:
1138 case ir_binop_lshift
:
1139 case ir_binop_rshift
:
1140 case ir_binop_bit_and
:
1141 case ir_binop_bit_xor
:
1142 case ir_binop_bit_or
:
1143 assert(!"GLSL 1.30 features unsupported");
1149 fs_visitor::emit_assignment_writes(fs_reg
&l
, fs_reg
&r
,
1150 const glsl_type
*type
, bool predicated
)
1152 switch (type
->base_type
) {
1153 case GLSL_TYPE_FLOAT
:
1154 case GLSL_TYPE_UINT
:
1156 case GLSL_TYPE_BOOL
:
1157 for (unsigned int i
= 0; i
< type
->components(); i
++) {
1158 l
.type
= brw_type_for_base_type(type
);
1159 r
.type
= brw_type_for_base_type(type
);
1161 fs_inst
*inst
= emit(fs_inst(BRW_OPCODE_MOV
, l
, r
));
1162 inst
->predicated
= predicated
;
1168 case GLSL_TYPE_ARRAY
:
1169 for (unsigned int i
= 0; i
< type
->length
; i
++) {
1170 emit_assignment_writes(l
, r
, type
->fields
.array
, predicated
);
1173 case GLSL_TYPE_STRUCT
:
1174 for (unsigned int i
= 0; i
< type
->length
; i
++) {
1175 emit_assignment_writes(l
, r
, type
->fields
.structure
[i
].type
,
1180 case GLSL_TYPE_SAMPLER
:
1184 assert(!"not reached");
1190 fs_visitor::visit(ir_assignment
*ir
)
1195 /* FINISHME: arrays on the lhs */
1196 ir
->lhs
->accept(this);
1199 ir
->rhs
->accept(this);
1202 assert(l
.file
!= BAD_FILE
);
1203 assert(r
.file
!= BAD_FILE
);
1205 if (ir
->condition
) {
1206 /* Get the condition bool into the predicate. */
1207 ir
->condition
->accept(this);
1208 inst
= emit(fs_inst(BRW_OPCODE_CMP
, reg_null
, this->result
, fs_reg(0)));
1209 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
1212 if (ir
->lhs
->type
->is_scalar() ||
1213 ir
->lhs
->type
->is_vector()) {
1214 for (int i
= 0; i
< ir
->lhs
->type
->vector_elements
; i
++) {
1215 if (ir
->write_mask
& (1 << i
)) {
1216 inst
= emit(fs_inst(BRW_OPCODE_MOV
, l
, r
));
1218 inst
->predicated
= true;
1224 emit_assignment_writes(l
, r
, ir
->lhs
->type
, ir
->condition
!= NULL
);
1229 fs_visitor::emit_texture_gen4(ir_texture
*ir
, fs_reg dst
, fs_reg coordinate
)
1233 bool simd16
= false;
1236 if (ir
->shadow_comparitor
) {
1237 for (mlen
= 0; mlen
< ir
->coordinate
->type
->vector_elements
; mlen
++) {
1238 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
),
1240 coordinate
.reg_offset
++;
1242 /* gen4's SIMD8 sampler always has the slots for u,v,r present. */
1245 if (ir
->op
== ir_tex
) {
1246 /* There's no plain shadow compare message, so we use shadow
1247 * compare with a bias of 0.0.
1249 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
),
1252 } else if (ir
->op
== ir_txb
) {
1253 ir
->lod_info
.bias
->accept(this);
1254 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
),
1258 assert(ir
->op
== ir_txl
);
1259 ir
->lod_info
.lod
->accept(this);
1260 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
),
1265 ir
->shadow_comparitor
->accept(this);
1266 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
), this->result
));
1268 } else if (ir
->op
== ir_tex
) {
1269 for (mlen
= 0; mlen
< ir
->coordinate
->type
->vector_elements
; mlen
++) {
1270 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
),
1272 coordinate
.reg_offset
++;
1274 /* gen4's SIMD8 sampler always has the slots for u,v,r present. */
1277 /* Oh joy. gen4 doesn't have SIMD8 non-shadow-compare bias/lod
1278 * instructions. We'll need to do SIMD16 here.
1280 assert(ir
->op
== ir_txb
|| ir
->op
== ir_txl
);
1282 for (mlen
= 0; mlen
< ir
->coordinate
->type
->vector_elements
* 2;) {
1283 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
),
1285 coordinate
.reg_offset
++;
1288 /* The unused upper half. */
1292 /* lod/bias appears after u/v/r. */
1295 if (ir
->op
== ir_txb
) {
1296 ir
->lod_info
.bias
->accept(this);
1297 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
),
1301 ir
->lod_info
.lod
->accept(this);
1302 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
),
1307 /* The unused upper half. */
1310 /* Now, since we're doing simd16, the return is 2 interleaved
1311 * vec4s where the odd-indexed ones are junk. We'll need to move
1312 * this weirdness around to the expected layout.
1316 dst
= fs_reg(this, glsl_type::get_array_instance(glsl_type::vec4_type
,
1318 dst
.type
= BRW_REGISTER_TYPE_F
;
1321 fs_inst
*inst
= NULL
;
1324 inst
= emit(fs_inst(FS_OPCODE_TEX
, dst
, fs_reg(MRF
, base_mrf
)));
1327 inst
= emit(fs_inst(FS_OPCODE_TXB
, dst
, fs_reg(MRF
, base_mrf
)));
1330 inst
= emit(fs_inst(FS_OPCODE_TXL
, dst
, fs_reg(MRF
, base_mrf
)));
1334 assert(!"GLSL 1.30 features unsupported");
1340 for (int i
= 0; i
< 4; i
++) {
1341 emit(fs_inst(BRW_OPCODE_MOV
, orig_dst
, dst
));
1342 orig_dst
.reg_offset
++;
1343 dst
.reg_offset
+= 2;
1351 fs_visitor::emit_texture_gen5(ir_texture
*ir
, fs_reg dst
, fs_reg coordinate
)
1353 /* gen5's SIMD8 sampler has slots for u, v, r, array index, then
1354 * optional parameters like shadow comparitor or LOD bias. If
1355 * optional parameters aren't present, those base slots are
1356 * optional and don't need to be included in the message.
1358 * We don't fill in the unnecessary slots regardless, which may
1359 * look surprising in the disassembly.
1364 for (mlen
= 0; mlen
< ir
->coordinate
->type
->vector_elements
; mlen
++) {
1365 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
), coordinate
));
1366 coordinate
.reg_offset
++;
1369 if (ir
->shadow_comparitor
) {
1370 mlen
= MAX2(mlen
, 4);
1372 ir
->shadow_comparitor
->accept(this);
1373 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
), this->result
));
1377 fs_inst
*inst
= NULL
;
1380 inst
= emit(fs_inst(FS_OPCODE_TEX
, dst
, fs_reg(MRF
, base_mrf
)));
1383 ir
->lod_info
.bias
->accept(this);
1384 mlen
= MAX2(mlen
, 4);
1385 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
), this->result
));
1388 inst
= emit(fs_inst(FS_OPCODE_TXB
, dst
, fs_reg(MRF
, base_mrf
)));
1391 ir
->lod_info
.lod
->accept(this);
1392 mlen
= MAX2(mlen
, 4);
1393 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
), this->result
));
1396 inst
= emit(fs_inst(FS_OPCODE_TXL
, dst
, fs_reg(MRF
, base_mrf
)));
1400 assert(!"GLSL 1.30 features unsupported");
1409 fs_visitor::visit(ir_texture
*ir
)
1411 fs_inst
*inst
= NULL
;
1413 ir
->coordinate
->accept(this);
1414 fs_reg coordinate
= this->result
;
1416 /* Should be lowered by do_lower_texture_projection */
1417 assert(!ir
->projector
);
1419 /* Writemasking doesn't eliminate channels on SIMD8 texture
1420 * samples, so don't worry about them.
1422 fs_reg dst
= fs_reg(this, glsl_type::vec4_type
);
1424 if (intel
->gen
< 5) {
1425 inst
= emit_texture_gen4(ir
, dst
, coordinate
);
1427 inst
= emit_texture_gen5(ir
, dst
, coordinate
);
1431 _mesa_get_sampler_uniform_value(ir
->sampler
,
1432 ctx
->Shader
.CurrentProgram
,
1433 &brw
->fragment_program
->Base
);
1434 inst
->sampler
= c
->fp
->program
.Base
.SamplerUnits
[inst
->sampler
];
1438 if (ir
->shadow_comparitor
)
1439 inst
->shadow_compare
= true;
1441 if (c
->key
.tex_swizzles
[inst
->sampler
] != SWIZZLE_NOOP
) {
1442 fs_reg swizzle_dst
= fs_reg(this, glsl_type::vec4_type
);
1444 for (int i
= 0; i
< 4; i
++) {
1445 int swiz
= GET_SWZ(c
->key
.tex_swizzles
[inst
->sampler
], i
);
1446 fs_reg l
= swizzle_dst
;
1449 if (swiz
== SWIZZLE_ZERO
) {
1450 emit(fs_inst(BRW_OPCODE_MOV
, l
, fs_reg(0.0f
)));
1451 } else if (swiz
== SWIZZLE_ONE
) {
1452 emit(fs_inst(BRW_OPCODE_MOV
, l
, fs_reg(1.0f
)));
1455 r
.reg_offset
+= GET_SWZ(c
->key
.tex_swizzles
[inst
->sampler
], i
);
1456 emit(fs_inst(BRW_OPCODE_MOV
, l
, r
));
1459 this->result
= swizzle_dst
;
1464 fs_visitor::visit(ir_swizzle
*ir
)
1466 ir
->val
->accept(this);
1467 fs_reg val
= this->result
;
1469 if (ir
->type
->vector_elements
== 1) {
1470 this->result
.reg_offset
+= ir
->mask
.x
;
1474 fs_reg result
= fs_reg(this, ir
->type
);
1475 this->result
= result
;
1477 for (unsigned int i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1478 fs_reg channel
= val
;
1496 channel
.reg_offset
+= swiz
;
1497 emit(fs_inst(BRW_OPCODE_MOV
, result
, channel
));
1498 result
.reg_offset
++;
1503 fs_visitor::visit(ir_discard
*ir
)
1505 fs_reg temp
= fs_reg(this, glsl_type::uint_type
);
1507 assert(ir
->condition
== NULL
); /* FINISHME */
1509 emit(fs_inst(FS_OPCODE_DISCARD
, temp
, temp
));
1513 fs_visitor::visit(ir_constant
*ir
)
1515 fs_reg
reg(this, ir
->type
);
1518 for (unsigned int i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1519 switch (ir
->type
->base_type
) {
1520 case GLSL_TYPE_FLOAT
:
1521 emit(fs_inst(BRW_OPCODE_MOV
, reg
, fs_reg(ir
->value
.f
[i
])));
1523 case GLSL_TYPE_UINT
:
1524 emit(fs_inst(BRW_OPCODE_MOV
, reg
, fs_reg(ir
->value
.u
[i
])));
1527 emit(fs_inst(BRW_OPCODE_MOV
, reg
, fs_reg(ir
->value
.i
[i
])));
1529 case GLSL_TYPE_BOOL
:
1530 emit(fs_inst(BRW_OPCODE_MOV
, reg
, fs_reg((int)ir
->value
.b
[i
])));
1533 assert(!"Non-float/uint/int/bool constant");
1540 fs_visitor::visit(ir_if
*ir
)
1544 /* Don't point the annotation at the if statement, because then it plus
1545 * the then and else blocks get printed.
1547 this->base_ir
= ir
->condition
;
1549 /* Generate the condition into the condition code. */
1550 ir
->condition
->accept(this);
1551 inst
= emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(brw_null_reg()), this->result
));
1552 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
1554 inst
= emit(fs_inst(BRW_OPCODE_IF
));
1555 inst
->predicated
= true;
1557 foreach_iter(exec_list_iterator
, iter
, ir
->then_instructions
) {
1558 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1564 if (!ir
->else_instructions
.is_empty()) {
1565 emit(fs_inst(BRW_OPCODE_ELSE
));
1567 foreach_iter(exec_list_iterator
, iter
, ir
->else_instructions
) {
1568 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1575 emit(fs_inst(BRW_OPCODE_ENDIF
));
1579 fs_visitor::visit(ir_loop
*ir
)
1581 fs_reg counter
= reg_undef
;
1584 this->base_ir
= ir
->counter
;
1585 ir
->counter
->accept(this);
1586 counter
= *(variable_storage(ir
->counter
));
1589 this->base_ir
= ir
->from
;
1590 ir
->from
->accept(this);
1592 emit(fs_inst(BRW_OPCODE_MOV
, counter
, this->result
));
1596 emit(fs_inst(BRW_OPCODE_DO
));
1599 this->base_ir
= ir
->to
;
1600 ir
->to
->accept(this);
1602 fs_inst
*inst
= emit(fs_inst(BRW_OPCODE_CMP
, reg_null
,
1603 counter
, this->result
));
1605 case ir_binop_equal
:
1606 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
1608 case ir_binop_nequal
:
1609 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
1611 case ir_binop_gequal
:
1612 inst
->conditional_mod
= BRW_CONDITIONAL_GE
;
1614 case ir_binop_lequal
:
1615 inst
->conditional_mod
= BRW_CONDITIONAL_LE
;
1617 case ir_binop_greater
:
1618 inst
->conditional_mod
= BRW_CONDITIONAL_G
;
1621 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
1624 assert(!"not reached: unknown loop condition");
1629 inst
= emit(fs_inst(BRW_OPCODE_BREAK
));
1630 inst
->predicated
= true;
1633 foreach_iter(exec_list_iterator
, iter
, ir
->body_instructions
) {
1634 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1640 if (ir
->increment
) {
1641 this->base_ir
= ir
->increment
;
1642 ir
->increment
->accept(this);
1643 emit(fs_inst(BRW_OPCODE_ADD
, counter
, counter
, this->result
));
1646 emit(fs_inst(BRW_OPCODE_WHILE
));
1650 fs_visitor::visit(ir_loop_jump
*ir
)
1653 case ir_loop_jump::jump_break
:
1654 emit(fs_inst(BRW_OPCODE_BREAK
));
1656 case ir_loop_jump::jump_continue
:
1657 emit(fs_inst(BRW_OPCODE_CONTINUE
));
1663 fs_visitor::visit(ir_call
*ir
)
1665 assert(!"FINISHME");
1669 fs_visitor::visit(ir_return
*ir
)
1671 assert(!"FINISHME");
1675 fs_visitor::visit(ir_function
*ir
)
1677 /* Ignore function bodies other than main() -- we shouldn't see calls to
1678 * them since they should all be inlined before we get to ir_to_mesa.
1680 if (strcmp(ir
->name
, "main") == 0) {
1681 const ir_function_signature
*sig
;
1684 sig
= ir
->matching_signature(&empty
);
1688 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
1689 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1698 fs_visitor::visit(ir_function_signature
*ir
)
1700 assert(!"not reached");
1705 fs_visitor::emit(fs_inst inst
)
1707 fs_inst
*list_inst
= new(mem_ctx
) fs_inst
;
1710 list_inst
->annotation
= this->current_annotation
;
1711 list_inst
->ir
= this->base_ir
;
1713 this->instructions
.push_tail(list_inst
);
1718 /** Emits a dummy fragment shader consisting of magenta for bringup purposes. */
1720 fs_visitor::emit_dummy_fs()
1722 /* Everyone's favorite color. */
1723 emit(fs_inst(BRW_OPCODE_MOV
,
1726 emit(fs_inst(BRW_OPCODE_MOV
,
1729 emit(fs_inst(BRW_OPCODE_MOV
,
1732 emit(fs_inst(BRW_OPCODE_MOV
,
1737 write
= emit(fs_inst(FS_OPCODE_FB_WRITE
,
1742 /* The register location here is relative to the start of the URB
1743 * data. It will get adjusted to be a real location before
1744 * generate_code() time.
1747 fs_visitor::interp_reg(int location
, int channel
)
1749 int regnr
= urb_setup
[location
] * 2 + channel
/ 2;
1750 int stride
= (channel
& 1) * 4;
1752 assert(urb_setup
[location
] != -1);
1754 return brw_vec1_grf(regnr
, stride
);
1757 /** Emits the interpolation for the varying inputs. */
1759 fs_visitor::emit_interpolation_setup_gen4()
1761 struct brw_reg g1_uw
= retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW
);
1763 this->current_annotation
= "compute pixel centers";
1764 this->pixel_x
= fs_reg(this, glsl_type::uint_type
);
1765 this->pixel_y
= fs_reg(this, glsl_type::uint_type
);
1766 this->pixel_x
.type
= BRW_REGISTER_TYPE_UW
;
1767 this->pixel_y
.type
= BRW_REGISTER_TYPE_UW
;
1768 emit(fs_inst(BRW_OPCODE_ADD
,
1770 fs_reg(stride(suboffset(g1_uw
, 4), 2, 4, 0)),
1771 fs_reg(brw_imm_v(0x10101010))));
1772 emit(fs_inst(BRW_OPCODE_ADD
,
1774 fs_reg(stride(suboffset(g1_uw
, 5), 2, 4, 0)),
1775 fs_reg(brw_imm_v(0x11001100))));
1777 this->current_annotation
= "compute pixel deltas from v0";
1779 this->delta_x
= fs_reg(this, glsl_type::vec2_type
);
1780 this->delta_y
= this->delta_x
;
1781 this->delta_y
.reg_offset
++;
1783 this->delta_x
= fs_reg(this, glsl_type::float_type
);
1784 this->delta_y
= fs_reg(this, glsl_type::float_type
);
1786 emit(fs_inst(BRW_OPCODE_ADD
,
1789 fs_reg(negate(brw_vec1_grf(1, 0)))));
1790 emit(fs_inst(BRW_OPCODE_ADD
,
1793 fs_reg(negate(brw_vec1_grf(1, 1)))));
1795 this->current_annotation
= "compute pos.w and 1/pos.w";
1796 /* Compute wpos.w. It's always in our setup, since it's needed to
1797 * interpolate the other attributes.
1799 this->wpos_w
= fs_reg(this, glsl_type::float_type
);
1800 emit(fs_inst(FS_OPCODE_LINTERP
, wpos_w
, this->delta_x
, this->delta_y
,
1801 interp_reg(FRAG_ATTRIB_WPOS
, 3)));
1802 /* Compute the pixel 1/W value from wpos.w. */
1803 this->pixel_w
= fs_reg(this, glsl_type::float_type
);
1804 emit(fs_inst(FS_OPCODE_RCP
, this->pixel_w
, wpos_w
));
1805 this->current_annotation
= NULL
;
1808 /** Emits the interpolation for the varying inputs. */
1810 fs_visitor::emit_interpolation_setup_gen6()
1812 struct brw_reg g1_uw
= retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW
);
1814 /* If the pixel centers end up used, the setup is the same as for gen4. */
1815 this->current_annotation
= "compute pixel centers";
1816 this->pixel_x
= fs_reg(this, glsl_type::uint_type
);
1817 this->pixel_y
= fs_reg(this, glsl_type::uint_type
);
1818 this->pixel_x
.type
= BRW_REGISTER_TYPE_UW
;
1819 this->pixel_y
.type
= BRW_REGISTER_TYPE_UW
;
1820 emit(fs_inst(BRW_OPCODE_ADD
,
1822 fs_reg(stride(suboffset(g1_uw
, 4), 2, 4, 0)),
1823 fs_reg(brw_imm_v(0x10101010))));
1824 emit(fs_inst(BRW_OPCODE_ADD
,
1826 fs_reg(stride(suboffset(g1_uw
, 5), 2, 4, 0)),
1827 fs_reg(brw_imm_v(0x11001100))));
1829 this->current_annotation
= "compute 1/pos.w";
1830 this->wpos_w
= fs_reg(brw_vec8_grf(c
->key
.source_w_reg
, 0));
1831 this->pixel_w
= fs_reg(this, glsl_type::float_type
);
1832 emit(fs_inst(FS_OPCODE_RCP
, this->pixel_w
, wpos_w
));
1834 this->delta_x
= fs_reg(brw_vec8_grf(2, 0));
1835 this->delta_y
= fs_reg(brw_vec8_grf(3, 0));
1837 this->current_annotation
= NULL
;
1841 fs_visitor::emit_fb_writes()
1843 this->current_annotation
= "FB write header";
1849 if (c
->key
.aa_dest_stencil_reg
) {
1850 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, nr
++),
1851 fs_reg(brw_vec8_grf(c
->key
.aa_dest_stencil_reg
, 0))));
1854 /* Reserve space for color. It'll be filled in per MRT below. */
1858 if (c
->key
.source_depth_to_render_target
) {
1859 if (c
->key
.computes_depth
) {
1860 /* Hand over gl_FragDepth. */
1861 assert(this->frag_depth
);
1862 fs_reg depth
= *(variable_storage(this->frag_depth
));
1864 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, nr
++), depth
));
1866 /* Pass through the payload depth. */
1867 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, nr
++),
1868 fs_reg(brw_vec8_grf(c
->key
.source_depth_reg
, 0))));
1872 if (c
->key
.dest_depth_reg
) {
1873 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, nr
++),
1874 fs_reg(brw_vec8_grf(c
->key
.dest_depth_reg
, 0))));
1877 fs_reg color
= reg_undef
;
1878 if (this->frag_color
)
1879 color
= *(variable_storage(this->frag_color
));
1880 else if (this->frag_data
)
1881 color
= *(variable_storage(this->frag_data
));
1883 for (int target
= 0; target
< c
->key
.nr_color_regions
; target
++) {
1884 this->current_annotation
= talloc_asprintf(this->mem_ctx
,
1885 "FB write target %d",
1887 if (this->frag_color
|| this->frag_data
) {
1888 for (int i
= 0; i
< 4; i
++) {
1889 emit(fs_inst(BRW_OPCODE_MOV
,
1890 fs_reg(MRF
, color_mrf
+ i
),
1896 if (this->frag_color
)
1897 color
.reg_offset
-= 4;
1899 fs_inst
*inst
= emit(fs_inst(FS_OPCODE_FB_WRITE
,
1900 reg_undef
, reg_undef
));
1901 inst
->target
= target
;
1903 if (target
== c
->key
.nr_color_regions
- 1)
1907 if (c
->key
.nr_color_regions
== 0) {
1908 fs_inst
*inst
= emit(fs_inst(FS_OPCODE_FB_WRITE
,
1909 reg_undef
, reg_undef
));
1914 this->current_annotation
= NULL
;
1918 fs_visitor::generate_fb_write(fs_inst
*inst
)
1920 GLboolean eot
= inst
->eot
;
1922 /* Header is 2 regs, g0 and g1 are the contents. g0 will be implied
1925 brw_push_insn_state(p
);
1926 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
1927 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
1930 brw_vec8_grf(1, 0));
1931 brw_pop_insn_state(p
);
1934 8, /* dispatch_width */
1935 retype(vec8(brw_null_reg()), BRW_REGISTER_TYPE_UW
),
1937 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW
),
1945 fs_visitor::generate_linterp(fs_inst
*inst
,
1946 struct brw_reg dst
, struct brw_reg
*src
)
1948 struct brw_reg delta_x
= src
[0];
1949 struct brw_reg delta_y
= src
[1];
1950 struct brw_reg interp
= src
[2];
1953 delta_y
.nr
== delta_x
.nr
+ 1 &&
1954 (intel
->gen
>= 6 || (delta_x
.nr
& 1) == 0)) {
1955 brw_PLN(p
, dst
, interp
, delta_x
);
1957 brw_LINE(p
, brw_null_reg(), interp
, delta_x
);
1958 brw_MAC(p
, dst
, suboffset(interp
, 1), delta_y
);
1963 fs_visitor::generate_math(fs_inst
*inst
,
1964 struct brw_reg dst
, struct brw_reg
*src
)
1968 switch (inst
->opcode
) {
1970 op
= BRW_MATH_FUNCTION_INV
;
1973 op
= BRW_MATH_FUNCTION_RSQ
;
1975 case FS_OPCODE_SQRT
:
1976 op
= BRW_MATH_FUNCTION_SQRT
;
1978 case FS_OPCODE_EXP2
:
1979 op
= BRW_MATH_FUNCTION_EXP
;
1981 case FS_OPCODE_LOG2
:
1982 op
= BRW_MATH_FUNCTION_LOG
;
1985 op
= BRW_MATH_FUNCTION_POW
;
1988 op
= BRW_MATH_FUNCTION_SIN
;
1991 op
= BRW_MATH_FUNCTION_COS
;
1994 assert(!"not reached: unknown math function");
1999 if (inst
->opcode
== FS_OPCODE_POW
) {
2000 brw_MOV(p
, brw_message_reg(3), src
[1]);
2005 inst
->saturate
? BRW_MATH_SATURATE_SATURATE
:
2006 BRW_MATH_SATURATE_NONE
,
2008 BRW_MATH_DATA_VECTOR
,
2009 BRW_MATH_PRECISION_FULL
);
2013 fs_visitor::generate_tex(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
)
2017 uint32_t simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD8
;
2019 if (intel
->gen
== 5) {
2020 switch (inst
->opcode
) {
2022 if (inst
->shadow_compare
) {
2023 msg_type
= BRW_SAMPLER_MESSAGE_SAMPLE_COMPARE_GEN5
;
2025 msg_type
= BRW_SAMPLER_MESSAGE_SAMPLE_GEN5
;
2029 if (inst
->shadow_compare
) {
2030 msg_type
= BRW_SAMPLER_MESSAGE_SAMPLE_BIAS_COMPARE_GEN5
;
2032 msg_type
= BRW_SAMPLER_MESSAGE_SAMPLE_BIAS_GEN5
;
2037 switch (inst
->opcode
) {
2039 /* Note that G45 and older determines shadow compare and dispatch width
2040 * from message length for most messages.
2042 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE
;
2043 if (inst
->shadow_compare
) {
2044 assert(inst
->mlen
== 5);
2046 assert(inst
->mlen
<= 6);
2050 if (inst
->shadow_compare
) {
2051 assert(inst
->mlen
== 5);
2052 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE
;
2054 assert(inst
->mlen
== 8);
2055 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_BIAS
;
2056 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
2061 assert(msg_type
!= -1);
2063 if (simd_mode
== BRW_SAMPLER_SIMD_MODE_SIMD16
) {
2072 retype(dst
, BRW_REGISTER_TYPE_UW
),
2074 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW
),
2075 SURF_INDEX_TEXTURE(inst
->sampler
),
2087 /* For OPCODE_DDX and OPCODE_DDY, per channel of output we've got input
2090 * arg0: ss0.tl ss0.tr ss0.bl ss0.br ss1.tl ss1.tr ss1.bl ss1.br
2092 * and we're trying to produce:
2095 * dst: (ss0.tr - ss0.tl) (ss0.tl - ss0.bl)
2096 * (ss0.tr - ss0.tl) (ss0.tr - ss0.br)
2097 * (ss0.br - ss0.bl) (ss0.tl - ss0.bl)
2098 * (ss0.br - ss0.bl) (ss0.tr - ss0.br)
2099 * (ss1.tr - ss1.tl) (ss1.tl - ss1.bl)
2100 * (ss1.tr - ss1.tl) (ss1.tr - ss1.br)
2101 * (ss1.br - ss1.bl) (ss1.tl - ss1.bl)
2102 * (ss1.br - ss1.bl) (ss1.tr - ss1.br)
2104 * and add another set of two more subspans if in 16-pixel dispatch mode.
2106 * For DDX, it ends up being easy: width = 2, horiz=0 gets us the same result
2107 * for each pair, and vertstride = 2 jumps us 2 elements after processing a
2108 * pair. But for DDY, it's harder, as we want to produce the pairs swizzled
2109 * between each other. We could probably do it like ddx and swizzle the right
2110 * order later, but bail for now and just produce
2111 * ((ss0.tl - ss0.bl)x4 (ss1.tl - ss1.bl)x4)
2114 fs_visitor::generate_ddx(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
)
2116 struct brw_reg src0
= brw_reg(src
.file
, src
.nr
, 1,
2117 BRW_REGISTER_TYPE_F
,
2118 BRW_VERTICAL_STRIDE_2
,
2120 BRW_HORIZONTAL_STRIDE_0
,
2121 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
2122 struct brw_reg src1
= brw_reg(src
.file
, src
.nr
, 0,
2123 BRW_REGISTER_TYPE_F
,
2124 BRW_VERTICAL_STRIDE_2
,
2126 BRW_HORIZONTAL_STRIDE_0
,
2127 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
2128 brw_ADD(p
, dst
, src0
, negate(src1
));
2132 fs_visitor::generate_ddy(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
)
2134 struct brw_reg src0
= brw_reg(src
.file
, src
.nr
, 0,
2135 BRW_REGISTER_TYPE_F
,
2136 BRW_VERTICAL_STRIDE_4
,
2138 BRW_HORIZONTAL_STRIDE_0
,
2139 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
2140 struct brw_reg src1
= brw_reg(src
.file
, src
.nr
, 2,
2141 BRW_REGISTER_TYPE_F
,
2142 BRW_VERTICAL_STRIDE_4
,
2144 BRW_HORIZONTAL_STRIDE_0
,
2145 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
2146 brw_ADD(p
, dst
, src0
, negate(src1
));
2150 fs_visitor::generate_discard(fs_inst
*inst
, struct brw_reg temp
)
2152 struct brw_reg g0
= retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW
);
2153 temp
= brw_uw1_reg(temp
.file
, temp
.nr
, 0);
2155 brw_push_insn_state(p
);
2156 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
2157 brw_NOT(p
, temp
, brw_mask_reg(1)); /* IMASK */
2158 brw_AND(p
, g0
, temp
, g0
);
2159 brw_pop_insn_state(p
);
2163 fs_visitor::assign_curb_setup()
2165 c
->prog_data
.first_curbe_grf
= c
->key
.nr_payload_regs
;
2166 c
->prog_data
.curb_read_length
= ALIGN(c
->prog_data
.nr_params
, 8) / 8;
2168 /* Map the offsets in the UNIFORM file to fixed HW regs. */
2169 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2170 fs_inst
*inst
= (fs_inst
*)iter
.get();
2172 for (unsigned int i
= 0; i
< 3; i
++) {
2173 if (inst
->src
[i
].file
== UNIFORM
) {
2174 int constant_nr
= inst
->src
[i
].hw_reg
+ inst
->src
[i
].reg_offset
;
2175 struct brw_reg brw_reg
= brw_vec1_grf(c
->prog_data
.first_curbe_grf
+
2179 inst
->src
[i
].file
= FIXED_HW_REG
;
2180 inst
->src
[i
].fixed_hw_reg
= brw_reg
;
2187 fs_visitor::calculate_urb_setup()
2189 for (unsigned int i
= 0; i
< FRAG_ATTRIB_MAX
; i
++) {
2194 /* Figure out where each of the incoming setup attributes lands. */
2195 if (intel
->gen
>= 6) {
2196 for (unsigned int i
= 0; i
< FRAG_ATTRIB_MAX
; i
++) {
2197 if (i
== FRAG_ATTRIB_WPOS
||
2198 (brw
->fragment_program
->Base
.InputsRead
& BITFIELD64_BIT(i
))) {
2199 urb_setup
[i
] = urb_next
++;
2203 /* FINISHME: The sf doesn't map VS->FS inputs for us very well. */
2204 for (unsigned int i
= 0; i
< VERT_RESULT_MAX
; i
++) {
2205 if (c
->key
.vp_outputs_written
& BITFIELD64_BIT(i
)) {
2208 if (i
>= VERT_RESULT_VAR0
)
2209 fp_index
= i
- (VERT_RESULT_VAR0
- FRAG_ATTRIB_VAR0
);
2210 else if (i
<= VERT_RESULT_TEX7
)
2216 urb_setup
[fp_index
] = urb_next
++;
2221 /* Each attribute is 4 setup channels, each of which is half a reg. */
2222 c
->prog_data
.urb_read_length
= urb_next
* 2;
2226 fs_visitor::assign_urb_setup()
2228 int urb_start
= c
->prog_data
.first_curbe_grf
+ c
->prog_data
.curb_read_length
;
2230 /* Offset all the urb_setup[] index by the actual position of the
2231 * setup regs, now that the location of the constants has been chosen.
2233 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2234 fs_inst
*inst
= (fs_inst
*)iter
.get();
2236 if (inst
->opcode
!= FS_OPCODE_LINTERP
)
2239 assert(inst
->src
[2].file
== FIXED_HW_REG
);
2241 inst
->src
[2].fixed_hw_reg
.nr
+= urb_start
;
2244 this->first_non_payload_grf
= urb_start
+ c
->prog_data
.urb_read_length
;
2248 assign_reg(int *reg_hw_locations
, fs_reg
*reg
)
2250 if (reg
->file
== GRF
&& reg
->reg
!= 0) {
2251 reg
->hw_reg
= reg_hw_locations
[reg
->reg
] + reg
->reg_offset
;
2257 fs_visitor::assign_regs_trivial()
2260 int hw_reg_mapping
[this->virtual_grf_next
];
2263 hw_reg_mapping
[0] = 0;
2264 hw_reg_mapping
[1] = this->first_non_payload_grf
;
2265 for (i
= 2; i
< this->virtual_grf_next
; i
++) {
2266 hw_reg_mapping
[i
] = (hw_reg_mapping
[i
- 1] +
2267 this->virtual_grf_sizes
[i
- 1]);
2269 last_grf
= hw_reg_mapping
[i
- 1] + this->virtual_grf_sizes
[i
- 1];
2271 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2272 fs_inst
*inst
= (fs_inst
*)iter
.get();
2274 assign_reg(hw_reg_mapping
, &inst
->dst
);
2275 assign_reg(hw_reg_mapping
, &inst
->src
[0]);
2276 assign_reg(hw_reg_mapping
, &inst
->src
[1]);
2279 this->grf_used
= last_grf
+ 1;
2283 fs_visitor::assign_regs()
2286 int hw_reg_mapping
[this->virtual_grf_next
+ 1];
2287 int base_reg_count
= BRW_MAX_GRF
- this->first_non_payload_grf
;
2288 int class_sizes
[base_reg_count
];
2289 int class_count
= 0;
2290 int aligned_pair_class
= -1;
2292 calculate_live_intervals();
2294 /* Set up the register classes.
2296 * The base registers store a scalar value. For texture samples,
2297 * we get virtual GRFs composed of 4 contiguous hw register. For
2298 * structures and arrays, we store them as contiguous larger things
2299 * than that, though we should be able to do better most of the
2302 class_sizes
[class_count
++] = 1;
2303 if (brw
->has_pln
&& intel
->gen
< 6) {
2304 /* Always set up the (unaligned) pairs for gen5, so we can find
2305 * them for making the aligned pair class.
2307 class_sizes
[class_count
++] = 2;
2309 for (int r
= 1; r
< this->virtual_grf_next
; r
++) {
2312 for (i
= 0; i
< class_count
; i
++) {
2313 if (class_sizes
[i
] == this->virtual_grf_sizes
[r
])
2316 if (i
== class_count
) {
2317 if (this->virtual_grf_sizes
[r
] >= base_reg_count
) {
2318 fprintf(stderr
, "Object too large to register allocate.\n");
2322 class_sizes
[class_count
++] = this->virtual_grf_sizes
[r
];
2326 int ra_reg_count
= 0;
2327 int class_base_reg
[class_count
];
2328 int class_reg_count
[class_count
];
2329 int classes
[class_count
+ 1];
2331 for (int i
= 0; i
< class_count
; i
++) {
2332 class_base_reg
[i
] = ra_reg_count
;
2333 class_reg_count
[i
] = base_reg_count
- (class_sizes
[i
] - 1);
2334 ra_reg_count
+= class_reg_count
[i
];
2337 struct ra_regs
*regs
= ra_alloc_reg_set(ra_reg_count
);
2338 for (int i
= 0; i
< class_count
; i
++) {
2339 classes
[i
] = ra_alloc_reg_class(regs
);
2341 for (int i_r
= 0; i_r
< class_reg_count
[i
]; i_r
++) {
2342 ra_class_add_reg(regs
, classes
[i
], class_base_reg
[i
] + i_r
);
2345 /* Add conflicts between our contiguous registers aliasing
2346 * base regs and other register classes' contiguous registers
2347 * that alias base regs, or the base regs themselves for classes[0].
2349 for (int c
= 0; c
<= i
; c
++) {
2350 for (int i_r
= 0; i_r
< class_reg_count
[i
]; i_r
++) {
2351 for (int c_r
= MAX2(0, i_r
- (class_sizes
[c
] - 1));
2352 c_r
< MIN2(class_reg_count
[c
], i_r
+ class_sizes
[i
]);
2356 printf("%d/%d conflicts %d/%d\n",
2357 class_sizes
[i
], this->first_non_payload_grf
+ i_r
,
2358 class_sizes
[c
], this->first_non_payload_grf
+ c_r
);
2361 ra_add_reg_conflict(regs
,
2362 class_base_reg
[i
] + i_r
,
2363 class_base_reg
[c
] + c_r
);
2369 /* Add a special class for aligned pairs, which we'll put delta_x/y
2370 * in on gen5 so that we can do PLN.
2372 if (brw
->has_pln
&& intel
->gen
< 6) {
2373 int reg_count
= (base_reg_count
- 1) / 2;
2374 int unaligned_pair_class
= 1;
2375 assert(class_sizes
[unaligned_pair_class
] == 2);
2377 aligned_pair_class
= class_count
;
2378 classes
[aligned_pair_class
] = ra_alloc_reg_class(regs
);
2379 class_base_reg
[aligned_pair_class
] = 0;
2380 class_reg_count
[aligned_pair_class
] = 0;
2381 int start
= (this->first_non_payload_grf
& 1) ? 1 : 0;
2383 for (int i
= 0; i
< reg_count
; i
++) {
2384 ra_class_add_reg(regs
, classes
[aligned_pair_class
],
2385 class_base_reg
[unaligned_pair_class
] + i
* 2 + start
);
2390 ra_set_finalize(regs
);
2392 struct ra_graph
*g
= ra_alloc_interference_graph(regs
,
2393 this->virtual_grf_next
);
2394 /* Node 0 is just a placeholder to keep virtual_grf[] mapping 1:1
2397 ra_set_node_class(g
, 0, classes
[0]);
2399 for (int i
= 1; i
< this->virtual_grf_next
; i
++) {
2400 for (int c
= 0; c
< class_count
; c
++) {
2401 if (class_sizes
[c
] == this->virtual_grf_sizes
[i
]) {
2402 if (aligned_pair_class
>= 0 &&
2403 this->delta_x
.reg
== i
) {
2404 ra_set_node_class(g
, i
, classes
[aligned_pair_class
]);
2406 ra_set_node_class(g
, i
, classes
[c
]);
2412 for (int j
= 1; j
< i
; j
++) {
2413 if (virtual_grf_interferes(i
, j
)) {
2414 ra_add_node_interference(g
, i
, j
);
2419 /* FINISHME: Handle spilling */
2420 if (!ra_allocate_no_spills(g
)) {
2421 fprintf(stderr
, "Failed to allocate registers.\n");
2426 /* Get the chosen virtual registers for each node, and map virtual
2427 * regs in the register classes back down to real hardware reg
2430 hw_reg_mapping
[0] = 0; /* unused */
2431 for (int i
= 1; i
< this->virtual_grf_next
; i
++) {
2432 int reg
= ra_get_node_reg(g
, i
);
2435 for (int c
= 0; c
< class_count
; c
++) {
2436 if (reg
>= class_base_reg
[c
] &&
2437 reg
< class_base_reg
[c
] + class_reg_count
[c
]) {
2438 hw_reg
= reg
- class_base_reg
[c
];
2443 assert(hw_reg
!= -1);
2444 hw_reg_mapping
[i
] = this->first_non_payload_grf
+ hw_reg
;
2445 last_grf
= MAX2(last_grf
,
2446 hw_reg_mapping
[i
] + this->virtual_grf_sizes
[i
] - 1);
2449 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2450 fs_inst
*inst
= (fs_inst
*)iter
.get();
2452 assign_reg(hw_reg_mapping
, &inst
->dst
);
2453 assign_reg(hw_reg_mapping
, &inst
->src
[0]);
2454 assign_reg(hw_reg_mapping
, &inst
->src
[1]);
2457 this->grf_used
= last_grf
+ 1;
2464 fs_visitor::calculate_live_intervals()
2466 int num_vars
= this->virtual_grf_next
;
2467 int *def
= talloc_array(mem_ctx
, int, num_vars
);
2468 int *use
= talloc_array(mem_ctx
, int, num_vars
);
2472 for (int i
= 0; i
< num_vars
; i
++) {
2478 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2479 fs_inst
*inst
= (fs_inst
*)iter
.get();
2481 if (inst
->opcode
== BRW_OPCODE_DO
) {
2482 if (loop_depth
++ == 0)
2484 } else if (inst
->opcode
== BRW_OPCODE_WHILE
) {
2487 if (loop_depth
== 0) {
2490 * Patches up any vars marked for use within the loop as
2491 * live until the end. This is conservative, as there
2492 * will often be variables defined and used inside the
2493 * loop but dead at the end of the loop body.
2495 for (int i
= 0; i
< num_vars
; i
++) {
2496 if (use
[i
] == loop_start
) {
2507 for (unsigned int i
= 0; i
< 3; i
++) {
2508 if (inst
->src
[i
].file
== GRF
&& inst
->src
[i
].reg
!= 0) {
2509 use
[inst
->src
[i
].reg
] = MAX2(use
[inst
->src
[i
].reg
], eip
);
2512 if (inst
->dst
.file
== GRF
&& inst
->dst
.reg
!= 0) {
2513 def
[inst
->dst
.reg
] = MIN2(def
[inst
->dst
.reg
], eip
);
2520 this->virtual_grf_def
= def
;
2521 this->virtual_grf_use
= use
;
2525 fs_visitor::virtual_grf_interferes(int a
, int b
)
2527 int start
= MAX2(this->virtual_grf_def
[a
], this->virtual_grf_def
[b
]);
2528 int end
= MIN2(this->virtual_grf_use
[a
], this->virtual_grf_use
[b
]);
2530 /* For dead code, just check if the def interferes with the other range. */
2531 if (this->virtual_grf_use
[a
] == -1) {
2532 return (this->virtual_grf_def
[a
] >= this->virtual_grf_def
[b
] &&
2533 this->virtual_grf_def
[a
] < this->virtual_grf_use
[b
]);
2535 if (this->virtual_grf_use
[b
] == -1) {
2536 return (this->virtual_grf_def
[b
] >= this->virtual_grf_def
[a
] &&
2537 this->virtual_grf_def
[b
] < this->virtual_grf_use
[a
]);
2540 return start
<= end
;
2543 static struct brw_reg
brw_reg_from_fs_reg(fs_reg
*reg
)
2545 struct brw_reg brw_reg
;
2547 switch (reg
->file
) {
2551 brw_reg
= brw_vec8_reg(reg
->file
,
2553 brw_reg
= retype(brw_reg
, reg
->type
);
2556 switch (reg
->type
) {
2557 case BRW_REGISTER_TYPE_F
:
2558 brw_reg
= brw_imm_f(reg
->imm
.f
);
2560 case BRW_REGISTER_TYPE_D
:
2561 brw_reg
= brw_imm_d(reg
->imm
.i
);
2563 case BRW_REGISTER_TYPE_UD
:
2564 brw_reg
= brw_imm_ud(reg
->imm
.u
);
2567 assert(!"not reached");
2572 brw_reg
= reg
->fixed_hw_reg
;
2575 /* Probably unused. */
2576 brw_reg
= brw_null_reg();
2579 assert(!"not reached");
2580 brw_reg
= brw_null_reg();
2584 brw_reg
= brw_abs(brw_reg
);
2586 brw_reg
= negate(brw_reg
);
2592 fs_visitor::generate_code()
2594 unsigned int annotation_len
= 0;
2595 int last_native_inst
= 0;
2596 struct brw_instruction
*if_stack
[16], *loop_stack
[16];
2597 int if_stack_depth
= 0, loop_stack_depth
= 0;
2598 int if_depth_in_loop
[16];
2600 if_depth_in_loop
[loop_stack_depth
] = 0;
2602 memset(&if_stack
, 0, sizeof(if_stack
));
2603 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2604 fs_inst
*inst
= (fs_inst
*)iter
.get();
2605 struct brw_reg src
[3], dst
;
2607 for (unsigned int i
= 0; i
< 3; i
++) {
2608 src
[i
] = brw_reg_from_fs_reg(&inst
->src
[i
]);
2610 dst
= brw_reg_from_fs_reg(&inst
->dst
);
2612 brw_set_conditionalmod(p
, inst
->conditional_mod
);
2613 brw_set_predicate_control(p
, inst
->predicated
);
2615 switch (inst
->opcode
) {
2616 case BRW_OPCODE_MOV
:
2617 brw_MOV(p
, dst
, src
[0]);
2619 case BRW_OPCODE_ADD
:
2620 brw_ADD(p
, dst
, src
[0], src
[1]);
2622 case BRW_OPCODE_MUL
:
2623 brw_MUL(p
, dst
, src
[0], src
[1]);
2626 case BRW_OPCODE_FRC
:
2627 brw_FRC(p
, dst
, src
[0]);
2629 case BRW_OPCODE_RNDD
:
2630 brw_RNDD(p
, dst
, src
[0]);
2632 case BRW_OPCODE_RNDZ
:
2633 brw_RNDZ(p
, dst
, src
[0]);
2636 case BRW_OPCODE_AND
:
2637 brw_AND(p
, dst
, src
[0], src
[1]);
2640 brw_OR(p
, dst
, src
[0], src
[1]);
2642 case BRW_OPCODE_XOR
:
2643 brw_XOR(p
, dst
, src
[0], src
[1]);
2646 case BRW_OPCODE_CMP
:
2647 brw_CMP(p
, dst
, inst
->conditional_mod
, src
[0], src
[1]);
2649 case BRW_OPCODE_SEL
:
2650 brw_SEL(p
, dst
, src
[0], src
[1]);
2654 assert(if_stack_depth
< 16);
2655 if_stack
[if_stack_depth
] = brw_IF(p
, BRW_EXECUTE_8
);
2656 if_depth_in_loop
[loop_stack_depth
]++;
2659 case BRW_OPCODE_ELSE
:
2660 if_stack
[if_stack_depth
- 1] =
2661 brw_ELSE(p
, if_stack
[if_stack_depth
- 1]);
2663 case BRW_OPCODE_ENDIF
:
2665 brw_ENDIF(p
, if_stack
[if_stack_depth
]);
2666 if_depth_in_loop
[loop_stack_depth
]--;
2670 loop_stack
[loop_stack_depth
++] = brw_DO(p
, BRW_EXECUTE_8
);
2671 if_depth_in_loop
[loop_stack_depth
] = 0;
2674 case BRW_OPCODE_BREAK
:
2675 brw_BREAK(p
, if_depth_in_loop
[loop_stack_depth
]);
2676 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
2678 case BRW_OPCODE_CONTINUE
:
2679 brw_CONT(p
, if_depth_in_loop
[loop_stack_depth
]);
2680 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
2683 case BRW_OPCODE_WHILE
: {
2684 struct brw_instruction
*inst0
, *inst1
;
2687 if (intel
->gen
>= 5)
2690 assert(loop_stack_depth
> 0);
2692 inst0
= inst1
= brw_WHILE(p
, loop_stack
[loop_stack_depth
]);
2693 /* patch all the BREAK/CONT instructions from last BGNLOOP */
2694 while (inst0
> loop_stack
[loop_stack_depth
]) {
2696 if (inst0
->header
.opcode
== BRW_OPCODE_BREAK
&&
2697 inst0
->bits3
.if_else
.jump_count
== 0) {
2698 inst0
->bits3
.if_else
.jump_count
= br
* (inst1
- inst0
+ 1);
2700 else if (inst0
->header
.opcode
== BRW_OPCODE_CONTINUE
&&
2701 inst0
->bits3
.if_else
.jump_count
== 0) {
2702 inst0
->bits3
.if_else
.jump_count
= br
* (inst1
- inst0
);
2710 case FS_OPCODE_SQRT
:
2711 case FS_OPCODE_EXP2
:
2712 case FS_OPCODE_LOG2
:
2716 generate_math(inst
, dst
, src
);
2718 case FS_OPCODE_LINTERP
:
2719 generate_linterp(inst
, dst
, src
);
2724 generate_tex(inst
, dst
, src
[0]);
2726 case FS_OPCODE_DISCARD
:
2727 generate_discard(inst
, dst
/* src0 == dst */);
2730 generate_ddx(inst
, dst
, src
[0]);
2733 generate_ddy(inst
, dst
, src
[0]);
2735 case FS_OPCODE_FB_WRITE
:
2736 generate_fb_write(inst
);
2739 if (inst
->opcode
< (int)ARRAY_SIZE(brw_opcodes
)) {
2740 _mesa_problem(ctx
, "Unsupported opcode `%s' in FS",
2741 brw_opcodes
[inst
->opcode
].name
);
2743 _mesa_problem(ctx
, "Unsupported opcode %d in FS", inst
->opcode
);
2748 if (annotation_len
< p
->nr_insn
) {
2749 annotation_len
*= 2;
2750 if (annotation_len
< 16)
2751 annotation_len
= 16;
2753 this->annotation_string
= talloc_realloc(this->mem_ctx
,
2757 this->annotation_ir
= talloc_realloc(this->mem_ctx
,
2763 for (unsigned int i
= last_native_inst
; i
< p
->nr_insn
; i
++) {
2764 this->annotation_string
[i
] = inst
->annotation
;
2765 this->annotation_ir
[i
] = inst
->ir
;
2767 last_native_inst
= p
->nr_insn
;
2772 brw_wm_fs_emit(struct brw_context
*brw
, struct brw_wm_compile
*c
)
2774 struct brw_compile
*p
= &c
->func
;
2775 struct intel_context
*intel
= &brw
->intel
;
2776 GLcontext
*ctx
= &intel
->ctx
;
2777 struct brw_shader
*shader
= NULL
;
2778 struct gl_shader_program
*prog
= ctx
->Shader
.CurrentProgram
;
2786 for (unsigned int i
= 0; i
< prog
->_NumLinkedShaders
; i
++) {
2787 if (prog
->_LinkedShaders
[i
]->Type
== GL_FRAGMENT_SHADER
) {
2788 shader
= (struct brw_shader
*)prog
->_LinkedShaders
[i
];
2795 /* We always use 8-wide mode, at least for now. For one, flow
2796 * control only works in 8-wide. Also, when we're fragment shader
2797 * bound, we're almost always under register pressure as well, so
2798 * 8-wide would save us from the performance cliff of spilling
2801 c
->dispatch_width
= 8;
2803 if (INTEL_DEBUG
& DEBUG_WM
) {
2804 printf("GLSL IR for native fragment shader %d:\n", prog
->Name
);
2805 _mesa_print_ir(shader
->ir
, NULL
);
2809 /* Now the main event: Visit the shader IR and generate our FS IR for it.
2811 fs_visitor
v(c
, shader
);
2816 v
.calculate_urb_setup();
2818 v
.emit_interpolation_setup_gen4();
2820 v
.emit_interpolation_setup_gen6();
2822 /* Generate FS IR for main(). (the visitor only descends into
2823 * functions called "main").
2825 foreach_iter(exec_list_iterator
, iter
, *shader
->ir
) {
2826 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
2832 v
.assign_curb_setup();
2833 v
.assign_urb_setup();
2835 v
.assign_regs_trivial();
2843 assert(!v
.fail
); /* FINISHME: Cleanly fail, tested at link time, etc. */
2848 if (INTEL_DEBUG
& DEBUG_WM
) {
2849 const char *last_annotation_string
= NULL
;
2850 ir_instruction
*last_annotation_ir
= NULL
;
2852 printf("Native code for fragment shader %d:\n", prog
->Name
);
2853 for (unsigned int i
= 0; i
< p
->nr_insn
; i
++) {
2854 if (last_annotation_ir
!= v
.annotation_ir
[i
]) {
2855 last_annotation_ir
= v
.annotation_ir
[i
];
2856 if (last_annotation_ir
) {
2858 last_annotation_ir
->print();
2862 if (last_annotation_string
!= v
.annotation_string
[i
]) {
2863 last_annotation_string
= v
.annotation_string
[i
];
2864 if (last_annotation_string
)
2865 printf(" %s\n", last_annotation_string
);
2867 brw_disasm(stdout
, &p
->store
[i
], intel
->gen
);
2872 c
->prog_data
.total_grf
= v
.grf_used
;
2873 c
->prog_data
.total_scratch
= 0;