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 "program/prog_parameter.h"
35 #include "program/prog_print.h"
36 #include "program/prog_optimize.h"
37 #include "program/sampler.h"
38 #include "program/hash_table.h"
39 #include "brw_context.h"
44 #include "../glsl/glsl_types.h"
45 #include "../glsl/ir_optimization.h"
46 #include "../glsl/ir_print_visitor.h"
49 ARF
= BRW_ARCHITECTURE_REGISTER_FILE
,
50 GRF
= BRW_GENERAL_REGISTER_FILE
,
51 MRF
= BRW_MESSAGE_REGISTER_FILE
,
52 IMM
= BRW_IMMEDIATE_VALUE
,
53 FIXED_HW_REG
, /* a struct brw_reg */
54 UNIFORM
, /* prog_data->params[hw_reg] */
59 FS_OPCODE_FB_WRITE
= 256,
77 static int using_new_fs
= -1;
78 static struct brw_reg
brw_reg_from_fs_reg(class fs_reg
*reg
);
81 brw_new_shader(GLcontext
*ctx
, GLuint name
, GLuint type
)
83 struct brw_shader
*shader
;
85 shader
= talloc_zero(NULL
, struct brw_shader
);
87 shader
->base
.Type
= type
;
88 shader
->base
.Name
= name
;
89 _mesa_init_shader(ctx
, &shader
->base
);
95 struct gl_shader_program
*
96 brw_new_shader_program(GLcontext
*ctx
, GLuint name
)
98 struct brw_shader_program
*prog
;
99 prog
= talloc_zero(NULL
, struct brw_shader_program
);
101 prog
->base
.Name
= name
;
102 _mesa_init_shader_program(ctx
, &prog
->base
);
108 brw_compile_shader(GLcontext
*ctx
, struct gl_shader
*shader
)
110 if (!_mesa_ir_compile_shader(ctx
, shader
))
117 brw_link_shader(GLcontext
*ctx
, struct gl_shader_program
*prog
)
119 if (using_new_fs
== -1)
120 using_new_fs
= getenv("INTEL_NEW_FS") != NULL
;
122 for (unsigned i
= 0; i
< prog
->_NumLinkedShaders
; i
++) {
123 struct brw_shader
*shader
= (struct brw_shader
*)prog
->_LinkedShaders
[i
];
125 if (using_new_fs
&& shader
->base
.Type
== GL_FRAGMENT_SHADER
) {
126 void *mem_ctx
= talloc_new(NULL
);
130 talloc_free(shader
->ir
);
131 shader
->ir
= new(shader
) exec_list
;
132 clone_ir_list(mem_ctx
, shader
->ir
, shader
->base
.ir
);
134 do_mat_op_to_vec(shader
->ir
);
135 do_mod_to_fract(shader
->ir
);
136 do_div_to_mul_rcp(shader
->ir
);
137 do_sub_to_add_neg(shader
->ir
);
138 do_explog_to_explog2(shader
->ir
);
143 brw_do_channel_expressions(shader
->ir
);
144 brw_do_vector_splitting(shader
->ir
);
146 progress
= do_lower_jumps(shader
->ir
, true, true,
147 true, /* main return */
148 false, /* continue */
152 progress
= do_common_optimization(shader
->ir
, true, 32) || progress
;
154 progress
= lower_noise(shader
->ir
) || progress
;
156 lower_variable_index_to_cond_assign(shader
->ir
,
158 GL_TRUE
, /* output */
160 GL_TRUE
/* uniform */
164 validate_ir_tree(shader
->ir
);
166 reparent_ir(shader
->ir
, shader
->ir
);
167 talloc_free(mem_ctx
);
171 if (!_mesa_ir_link_shader(ctx
, prog
))
178 type_size(const struct glsl_type
*type
)
180 unsigned int size
, i
;
182 switch (type
->base_type
) {
185 case GLSL_TYPE_FLOAT
:
187 return type
->components();
188 case GLSL_TYPE_ARRAY
:
189 /* FINISHME: uniform/varying arrays. */
190 return type_size(type
->fields
.array
) * type
->length
;
191 case GLSL_TYPE_STRUCT
:
193 for (i
= 0; i
< type
->length
; i
++) {
194 size
+= type_size(type
->fields
.structure
[i
].type
);
197 case GLSL_TYPE_SAMPLER
:
198 /* Samplers take up no register space, since they're baked in at
203 assert(!"not reached");
210 /* Callers of this talloc-based new need not call delete. It's
211 * easier to just talloc_free 'ctx' (or any of its ancestors). */
212 static void* operator new(size_t size
, void *ctx
)
216 node
= talloc_size(ctx
, size
);
217 assert(node
!= NULL
);
225 this->reg_offset
= 0;
231 /** Generic unset register constructor. */
235 this->file
= BAD_FILE
;
238 /** Immediate value constructor. */
243 this->type
= BRW_REGISTER_TYPE_F
;
247 /** Immediate value constructor. */
252 this->type
= BRW_REGISTER_TYPE_D
;
256 /** Immediate value constructor. */
261 this->type
= BRW_REGISTER_TYPE_UD
;
265 /** Fixed brw_reg Immediate value constructor. */
266 fs_reg(struct brw_reg fixed_hw_reg
)
269 this->file
= FIXED_HW_REG
;
270 this->fixed_hw_reg
= fixed_hw_reg
;
271 this->type
= fixed_hw_reg
.type
;
274 fs_reg(enum register_file file
, int hw_reg
);
275 fs_reg(class fs_visitor
*v
, const struct glsl_type
*type
);
277 /** Register file: ARF, GRF, MRF, IMM. */
278 enum register_file file
;
279 /** Abstract register number. 0 = fixed hw reg */
281 /** Offset within the abstract register. */
283 /** HW register number. Generally unset until register allocation. */
285 /** Register type. BRW_REGISTER_TYPE_* */
289 struct brw_reg fixed_hw_reg
;
291 /** Value for file == BRW_IMMMEDIATE_FILE */
299 static const fs_reg reg_undef
;
300 static const fs_reg
reg_null(ARF
, BRW_ARF_NULL
);
302 class fs_inst
: public exec_node
{
304 /* Callers of this talloc-based new need not call delete. It's
305 * easier to just talloc_free 'ctx' (or any of its ancestors). */
306 static void* operator new(size_t size
, void *ctx
)
310 node
= talloc_zero_size(ctx
, size
);
311 assert(node
!= NULL
);
318 this->opcode
= BRW_OPCODE_NOP
;
319 this->saturate
= false;
320 this->conditional_mod
= BRW_CONDITIONAL_NONE
;
321 this->predicated
= false;
325 this->shadow_compare
= false;
336 this->opcode
= opcode
;
339 fs_inst(int opcode
, fs_reg dst
, fs_reg src0
)
342 this->opcode
= opcode
;
347 fs_inst(int opcode
, fs_reg dst
, fs_reg src0
, fs_reg src1
)
350 this->opcode
= opcode
;
356 fs_inst(int opcode
, fs_reg dst
, fs_reg src0
, fs_reg src1
, fs_reg src2
)
359 this->opcode
= opcode
;
366 int opcode
; /* BRW_OPCODE_* or FS_OPCODE_* */
371 int conditional_mod
; /**< BRW_CONDITIONAL_* */
373 int mlen
; /**< SEND message length */
375 int target
; /**< MRT target. */
380 * Annotation for the generated IR. One of the two can be set.
383 const char *annotation
;
387 class fs_visitor
: public ir_visitor
391 fs_visitor(struct brw_wm_compile
*c
, struct brw_shader
*shader
)
396 this->fp
= brw
->fragment_program
;
397 this->intel
= &brw
->intel
;
398 this->ctx
= &intel
->ctx
;
399 this->mem_ctx
= talloc_new(NULL
);
400 this->shader
= shader
;
402 this->next_abstract_grf
= 1;
403 this->variable_ht
= hash_table_ctor(0,
404 hash_table_pointer_hash
,
405 hash_table_pointer_compare
);
407 this->frag_color
= NULL
;
408 this->frag_data
= NULL
;
409 this->frag_depth
= NULL
;
410 this->first_non_payload_grf
= 0;
412 this->current_annotation
= NULL
;
413 this->annotation_string
= NULL
;
414 this->annotation_ir
= NULL
;
415 this->base_ir
= NULL
;
419 talloc_free(this->mem_ctx
);
420 hash_table_dtor(this->variable_ht
);
423 fs_reg
*variable_storage(ir_variable
*var
);
425 void visit(ir_variable
*ir
);
426 void visit(ir_assignment
*ir
);
427 void visit(ir_dereference_variable
*ir
);
428 void visit(ir_dereference_record
*ir
);
429 void visit(ir_dereference_array
*ir
);
430 void visit(ir_expression
*ir
);
431 void visit(ir_texture
*ir
);
432 void visit(ir_if
*ir
);
433 void visit(ir_constant
*ir
);
434 void visit(ir_swizzle
*ir
);
435 void visit(ir_return
*ir
);
436 void visit(ir_loop
*ir
);
437 void visit(ir_loop_jump
*ir
);
438 void visit(ir_discard
*ir
);
439 void visit(ir_call
*ir
);
440 void visit(ir_function
*ir
);
441 void visit(ir_function_signature
*ir
);
443 fs_inst
*emit(fs_inst inst
);
444 void assign_curb_setup();
445 void assign_urb_setup();
447 void generate_code();
448 void generate_fb_write(fs_inst
*inst
);
449 void generate_linterp(fs_inst
*inst
, struct brw_reg dst
,
450 struct brw_reg
*src
);
451 void generate_tex(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
);
452 void generate_math(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg
*src
);
453 void generate_discard(fs_inst
*inst
);
454 void generate_ddx(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
);
455 void generate_ddy(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
);
457 void emit_dummy_fs();
458 void emit_fragcoord_interpolation(ir_variable
*ir
);
459 void emit_interpolation();
460 void emit_pinterp(int location
);
461 void emit_fb_writes();
463 struct brw_reg
interp_reg(int location
, int channel
);
464 int setup_uniform_values(int loc
, const glsl_type
*type
);
466 struct brw_context
*brw
;
467 const struct gl_fragment_program
*fp
;
468 struct intel_context
*intel
;
470 struct brw_wm_compile
*c
;
471 struct brw_compile
*p
;
472 struct brw_shader
*shader
;
474 exec_list instructions
;
475 int next_abstract_grf
;
476 struct hash_table
*variable_ht
;
477 ir_variable
*frag_color
, *frag_data
, *frag_depth
;
478 int first_non_payload_grf
;
480 /** @{ debug annotation info */
481 const char *current_annotation
;
482 ir_instruction
*base_ir
;
483 const char **annotation_string
;
484 ir_instruction
**annotation_ir
;
489 /* Result of last visit() method. */
498 fs_reg interp_attrs
[64];
504 /** Fixed HW reg constructor. */
505 fs_reg::fs_reg(enum register_file file
, int hw_reg
)
509 this->hw_reg
= hw_reg
;
510 this->type
= BRW_REGISTER_TYPE_F
;
514 brw_type_for_base_type(const struct glsl_type
*type
)
516 switch (type
->base_type
) {
517 case GLSL_TYPE_FLOAT
:
518 return BRW_REGISTER_TYPE_F
;
521 return BRW_REGISTER_TYPE_D
;
523 return BRW_REGISTER_TYPE_UD
;
524 case GLSL_TYPE_ARRAY
:
525 case GLSL_TYPE_STRUCT
:
526 /* These should be overridden with the type of the member when
527 * dereferenced into. BRW_REGISTER_TYPE_UD seems like a likely
528 * way to trip up if we don't.
530 return BRW_REGISTER_TYPE_UD
;
532 assert(!"not reached");
533 return BRW_REGISTER_TYPE_F
;
537 /** Automatic reg constructor. */
538 fs_reg::fs_reg(class fs_visitor
*v
, const struct glsl_type
*type
)
543 this->reg
= v
->next_abstract_grf
;
544 this->reg_offset
= 0;
545 v
->next_abstract_grf
+= type_size(type
);
546 this->type
= brw_type_for_base_type(type
);
550 fs_visitor::variable_storage(ir_variable
*var
)
552 return (fs_reg
*)hash_table_find(this->variable_ht
, var
);
555 /* Our support for uniforms is piggy-backed on the struct
556 * gl_fragment_program, because that's where the values actually
557 * get stored, rather than in some global gl_shader_program uniform
561 fs_visitor::setup_uniform_values(int loc
, const glsl_type
*type
)
563 unsigned int offset
= 0;
566 if (type
->is_matrix()) {
567 const glsl_type
*column
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
568 type
->vector_elements
,
571 for (unsigned int i
= 0; i
< type
->matrix_columns
; i
++) {
572 offset
+= setup_uniform_values(loc
+ offset
, column
);
578 switch (type
->base_type
) {
579 case GLSL_TYPE_FLOAT
:
583 vec_values
= fp
->Base
.Parameters
->ParameterValues
[loc
];
584 for (unsigned int i
= 0; i
< type
->vector_elements
; i
++) {
585 c
->prog_data
.param
[c
->prog_data
.nr_params
++] = &vec_values
[i
];
589 case GLSL_TYPE_STRUCT
:
590 for (unsigned int i
= 0; i
< type
->length
; i
++) {
591 offset
+= setup_uniform_values(loc
+ offset
,
592 type
->fields
.structure
[i
].type
);
596 case GLSL_TYPE_ARRAY
:
597 for (unsigned int i
= 0; i
< type
->length
; i
++) {
598 offset
+= setup_uniform_values(loc
+ offset
, type
->fields
.array
);
602 case GLSL_TYPE_SAMPLER
:
603 /* The sampler takes up a slot, but we don't use any values from it. */
607 assert(!"not reached");
613 fs_visitor::emit_fragcoord_interpolation(ir_variable
*ir
)
615 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(this, ir
->type
);
617 fs_reg neg_y
= this->pixel_y
;
621 if (ir
->pixel_center_integer
) {
622 emit(fs_inst(BRW_OPCODE_MOV
, wpos
, this->pixel_x
));
624 emit(fs_inst(BRW_OPCODE_ADD
, wpos
, this->pixel_x
, fs_reg(0.5f
)));
629 if (ir
->origin_upper_left
&& ir
->pixel_center_integer
) {
630 emit(fs_inst(BRW_OPCODE_MOV
, wpos
, this->pixel_y
));
632 fs_reg pixel_y
= this->pixel_y
;
633 float offset
= (ir
->pixel_center_integer
? 0.0 : 0.5);
635 if (!ir
->origin_upper_left
) {
636 pixel_y
.negate
= true;
637 offset
+= c
->key
.drawable_height
- 1.0;
640 emit(fs_inst(BRW_OPCODE_ADD
, wpos
, pixel_y
, fs_reg(offset
)));
645 emit(fs_inst(FS_OPCODE_LINTERP
, wpos
, this->delta_x
, this->delta_y
,
646 interp_reg(FRAG_ATTRIB_WPOS
, 2)));
649 /* gl_FragCoord.w: Already set up in emit_interpolation */
650 emit(fs_inst(BRW_OPCODE_MOV
, wpos
, this->wpos_w
));
652 hash_table_insert(this->variable_ht
, reg
, ir
);
656 fs_visitor::visit(ir_variable
*ir
)
660 if (variable_storage(ir
))
663 if (strcmp(ir
->name
, "gl_FragColor") == 0) {
664 this->frag_color
= ir
;
665 } else if (strcmp(ir
->name
, "gl_FragData") == 0) {
666 this->frag_data
= ir
;
667 } else if (strcmp(ir
->name
, "gl_FragDepth") == 0) {
668 this->frag_depth
= ir
;
671 if (ir
->mode
== ir_var_in
) {
672 if (!strcmp(ir
->name
, "gl_FragCoord")) {
673 emit_fragcoord_interpolation(ir
);
675 } else if (!strcmp(ir
->name
, "gl_FrontFacing")) {
676 reg
= new(this->mem_ctx
) fs_reg(this, ir
->type
);
677 struct brw_reg r1_6ud
= retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_UD
);
678 /* bit 31 is "primitive is back face", so checking < (1 << 31) gives
681 fs_inst
*inst
= emit(fs_inst(BRW_OPCODE_CMP
,
685 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
686 emit(fs_inst(BRW_OPCODE_AND
, *reg
, *reg
, fs_reg(1u)));
688 reg
= &this->interp_attrs
[ir
->location
];
692 if (ir
->mode
== ir_var_uniform
) {
693 int param_index
= c
->prog_data
.nr_params
;
695 setup_uniform_values(ir
->location
, ir
->type
);
697 reg
= new(this->mem_ctx
) fs_reg(UNIFORM
, param_index
);
701 reg
= new(this->mem_ctx
) fs_reg(this, ir
->type
);
703 hash_table_insert(this->variable_ht
, reg
, ir
);
707 fs_visitor::visit(ir_dereference_variable
*ir
)
709 fs_reg
*reg
= variable_storage(ir
->var
);
714 fs_visitor::visit(ir_dereference_record
*ir
)
716 const glsl_type
*struct_type
= ir
->record
->type
;
718 ir
->record
->accept(this);
720 unsigned int offset
= 0;
721 for (unsigned int i
= 0; i
< struct_type
->length
; i
++) {
722 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
724 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
726 this->result
.reg_offset
+= offset
;
727 this->result
.type
= brw_type_for_base_type(ir
->type
);
731 fs_visitor::visit(ir_dereference_array
*ir
)
736 ir
->array
->accept(this);
737 index
= ir
->array_index
->as_constant();
739 if (ir
->type
->is_matrix()) {
740 element_size
= ir
->type
->vector_elements
;
742 element_size
= type_size(ir
->type
);
743 this->result
.type
= brw_type_for_base_type(ir
->type
);
747 assert(this->result
.file
== UNIFORM
||
748 (this->result
.file
== GRF
&&
749 this->result
.reg
!= 0));
750 this->result
.reg_offset
+= index
->value
.i
[0] * element_size
;
752 assert(!"FINISHME: non-constant matrix column");
757 fs_visitor::visit(ir_expression
*ir
)
759 unsigned int operand
;
764 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
765 ir
->operands
[operand
]->accept(this);
766 if (this->result
.file
== BAD_FILE
) {
768 printf("Failed to get tree for expression operand:\n");
769 ir
->operands
[operand
]->accept(&v
);
772 op
[operand
] = this->result
;
774 /* Matrix expression operands should have been broken down to vector
775 * operations already.
777 assert(!ir
->operands
[operand
]->type
->is_matrix());
778 /* And then those vector operands should have been broken down to scalar.
780 assert(!ir
->operands
[operand
]->type
->is_vector());
783 /* Storage for our result. If our result goes into an assignment, it will
784 * just get copy-propagated out, so no worries.
786 this->result
= fs_reg(this, ir
->type
);
788 switch (ir
->operation
) {
789 case ir_unop_logic_not
:
790 emit(fs_inst(BRW_OPCODE_ADD
, this->result
, op
[0], fs_reg(-1)));
793 op
[0].negate
= !op
[0].negate
;
794 this->result
= op
[0];
798 this->result
= op
[0];
801 temp
= fs_reg(this, ir
->type
);
803 emit(fs_inst(BRW_OPCODE_MOV
, this->result
, fs_reg(0.0f
)));
805 inst
= emit(fs_inst(BRW_OPCODE_CMP
, reg_null
, op
[0], fs_reg(0.0f
)));
806 inst
->conditional_mod
= BRW_CONDITIONAL_G
;
807 inst
= emit(fs_inst(BRW_OPCODE_MOV
, this->result
, fs_reg(1.0f
)));
808 inst
->predicated
= true;
810 inst
= emit(fs_inst(BRW_OPCODE_CMP
, reg_null
, op
[0], fs_reg(0.0f
)));
811 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
812 inst
= emit(fs_inst(BRW_OPCODE_MOV
, this->result
, fs_reg(-1.0f
)));
813 inst
->predicated
= true;
817 emit(fs_inst(FS_OPCODE_RCP
, this->result
, op
[0]));
821 emit(fs_inst(FS_OPCODE_EXP2
, this->result
, op
[0]));
824 emit(fs_inst(FS_OPCODE_LOG2
, this->result
, op
[0]));
828 assert(!"not reached: should be handled by ir_explog_to_explog2");
831 emit(fs_inst(FS_OPCODE_SIN
, this->result
, op
[0]));
834 emit(fs_inst(FS_OPCODE_COS
, this->result
, op
[0]));
838 emit(fs_inst(FS_OPCODE_DDX
, this->result
, op
[0]));
841 emit(fs_inst(FS_OPCODE_DDY
, this->result
, op
[0]));
845 emit(fs_inst(BRW_OPCODE_ADD
, this->result
, op
[0], op
[1]));
848 assert(!"not reached: should be handled by ir_sub_to_add_neg");
852 emit(fs_inst(BRW_OPCODE_MUL
, this->result
, op
[0], op
[1]));
855 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
858 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
862 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
863 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
864 emit(fs_inst(BRW_OPCODE_AND
, this->result
, this->result
, fs_reg(0x1)));
866 case ir_binop_greater
:
867 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
868 inst
->conditional_mod
= BRW_CONDITIONAL_G
;
869 emit(fs_inst(BRW_OPCODE_AND
, this->result
, this->result
, fs_reg(0x1)));
871 case ir_binop_lequal
:
872 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
873 inst
->conditional_mod
= BRW_CONDITIONAL_LE
;
874 emit(fs_inst(BRW_OPCODE_AND
, this->result
, this->result
, fs_reg(0x1)));
876 case ir_binop_gequal
:
877 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
878 inst
->conditional_mod
= BRW_CONDITIONAL_GE
;
879 emit(fs_inst(BRW_OPCODE_AND
, this->result
, this->result
, fs_reg(0x1)));
882 case ir_binop_all_equal
: /* same as nequal for scalars */
883 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
884 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
885 emit(fs_inst(BRW_OPCODE_AND
, this->result
, this->result
, fs_reg(0x1)));
887 case ir_binop_nequal
:
888 case ir_binop_any_nequal
: /* same as nequal for scalars */
889 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
890 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
891 emit(fs_inst(BRW_OPCODE_AND
, this->result
, this->result
, fs_reg(0x1)));
894 case ir_binop_logic_xor
:
895 emit(fs_inst(BRW_OPCODE_XOR
, this->result
, op
[0], op
[1]));
898 case ir_binop_logic_or
:
899 emit(fs_inst(BRW_OPCODE_OR
, this->result
, op
[0], op
[1]));
902 case ir_binop_logic_and
:
903 emit(fs_inst(BRW_OPCODE_AND
, this->result
, op
[0], op
[1]));
909 assert(!"not reached: should be handled by brw_fs_channel_expressions");
913 assert(!"not reached: should be handled by lower_noise");
917 emit(fs_inst(FS_OPCODE_SQRT
, this->result
, op
[0]));
921 emit(fs_inst(FS_OPCODE_RSQ
, this->result
, op
[0]));
927 emit(fs_inst(BRW_OPCODE_MOV
, this->result
, op
[0]));
930 emit(fs_inst(BRW_OPCODE_MOV
, this->result
, op
[0]));
934 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], fs_reg(0.0f
)));
935 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
938 emit(fs_inst(BRW_OPCODE_RNDD
, this->result
, op
[0]));
941 op
[0].negate
= ~op
[0].negate
;
942 inst
= emit(fs_inst(BRW_OPCODE_RNDD
, this->result
, op
[0]));
943 this->result
.negate
= true;
946 inst
= emit(fs_inst(BRW_OPCODE_RNDD
, this->result
, op
[0]));
949 inst
= emit(fs_inst(BRW_OPCODE_FRC
, this->result
, op
[0]));
953 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
954 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
956 inst
= emit(fs_inst(BRW_OPCODE_SEL
, this->result
, op
[0], op
[1]));
957 inst
->predicated
= true;
960 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
961 inst
->conditional_mod
= BRW_CONDITIONAL_G
;
963 inst
= emit(fs_inst(BRW_OPCODE_SEL
, this->result
, op
[0], op
[1]));
964 inst
->predicated
= true;
968 inst
= emit(fs_inst(FS_OPCODE_POW
, this->result
, op
[0], op
[1]));
971 case ir_unop_bit_not
:
973 case ir_binop_lshift
:
974 case ir_binop_rshift
:
975 case ir_binop_bit_and
:
976 case ir_binop_bit_xor
:
977 case ir_binop_bit_or
:
978 assert(!"GLSL 1.30 features unsupported");
984 fs_visitor::visit(ir_assignment
*ir
)
991 /* FINISHME: arrays on the lhs */
992 ir
->lhs
->accept(this);
995 ir
->rhs
->accept(this);
998 /* FINISHME: This should really set to the correct maximal writemask for each
999 * FINISHME: component written (in the loops below). This case can only
1000 * FINISHME: occur for matrices, arrays, and structures.
1002 if (ir
->write_mask
== 0) {
1003 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
1004 write_mask
= WRITEMASK_XYZW
;
1006 assert(ir
->lhs
->type
->is_vector() || ir
->lhs
->type
->is_scalar());
1007 write_mask
= ir
->write_mask
;
1010 assert(l
.file
!= BAD_FILE
);
1011 assert(r
.file
!= BAD_FILE
);
1013 if (ir
->condition
) {
1014 /* Get the condition bool into the predicate. */
1015 ir
->condition
->accept(this);
1016 inst
= emit(fs_inst(BRW_OPCODE_CMP
, reg_null
, this->result
, fs_reg(0)));
1017 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
1020 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1021 if (i
>= 4 || (write_mask
& (1 << i
))) {
1022 inst
= emit(fs_inst(BRW_OPCODE_MOV
, l
, r
));
1024 inst
->predicated
= true;
1032 fs_visitor::visit(ir_texture
*ir
)
1035 fs_inst
*inst
= NULL
;
1036 unsigned int mlen
= 0;
1038 ir
->coordinate
->accept(this);
1039 fs_reg coordinate
= this->result
;
1041 if (ir
->projector
) {
1042 fs_reg inv_proj
= fs_reg(this, glsl_type::float_type
);
1044 ir
->projector
->accept(this);
1045 emit(fs_inst(FS_OPCODE_RCP
, inv_proj
, this->result
));
1047 fs_reg proj_coordinate
= fs_reg(this, ir
->coordinate
->type
);
1048 for (unsigned int i
= 0; i
< ir
->coordinate
->type
->vector_elements
; i
++) {
1049 emit(fs_inst(BRW_OPCODE_MUL
, proj_coordinate
, coordinate
, inv_proj
));
1050 coordinate
.reg_offset
++;
1051 proj_coordinate
.reg_offset
++;
1053 proj_coordinate
.reg_offset
= 0;
1055 coordinate
= proj_coordinate
;
1058 for (mlen
= 0; mlen
< ir
->coordinate
->type
->vector_elements
; mlen
++) {
1059 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
), coordinate
));
1060 coordinate
.reg_offset
++;
1063 /* Pre-Ironlake, the 8-wide sampler always took u,v,r. */
1067 if (ir
->shadow_comparitor
) {
1068 /* For shadow comparisons, we have to supply u,v,r. */
1071 ir
->shadow_comparitor
->accept(this);
1072 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
), this->result
));
1076 /* Do we ever want to handle writemasking on texture samples? Is it
1077 * performance relevant?
1079 fs_reg dst
= fs_reg(this, glsl_type::vec4_type
);
1083 inst
= emit(fs_inst(FS_OPCODE_TEX
, dst
, fs_reg(MRF
, base_mrf
)));
1086 ir
->lod_info
.bias
->accept(this);
1087 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
), this->result
));
1090 inst
= emit(fs_inst(FS_OPCODE_TXB
, dst
, fs_reg(MRF
, base_mrf
)));
1093 ir
->lod_info
.lod
->accept(this);
1094 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
), this->result
));
1097 inst
= emit(fs_inst(FS_OPCODE_TXL
, dst
, fs_reg(MRF
, base_mrf
)));
1101 assert(!"GLSL 1.30 features unsupported");
1106 _mesa_get_sampler_uniform_value(ir
->sampler
,
1107 ctx
->Shader
.CurrentProgram
,
1108 &brw
->fragment_program
->Base
);
1109 inst
->sampler
= c
->fp
->program
.Base
.SamplerUnits
[inst
->sampler
];
1113 if (ir
->shadow_comparitor
)
1114 inst
->shadow_compare
= true;
1119 fs_visitor::visit(ir_swizzle
*ir
)
1121 ir
->val
->accept(this);
1122 fs_reg val
= this->result
;
1124 fs_reg result
= fs_reg(this, ir
->type
);
1125 this->result
= result
;
1127 for (unsigned int i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1128 fs_reg channel
= val
;
1146 channel
.reg_offset
+= swiz
;
1147 emit(fs_inst(BRW_OPCODE_MOV
, result
, channel
));
1148 result
.reg_offset
++;
1153 fs_visitor::visit(ir_discard
*ir
)
1155 assert(ir
->condition
== NULL
); /* FINISHME */
1157 emit(fs_inst(FS_OPCODE_DISCARD
));
1161 fs_visitor::visit(ir_constant
*ir
)
1163 fs_reg
reg(this, ir
->type
);
1166 for (unsigned int i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1167 switch (ir
->type
->base_type
) {
1168 case GLSL_TYPE_FLOAT
:
1169 emit(fs_inst(BRW_OPCODE_MOV
, reg
, fs_reg(ir
->value
.f
[i
])));
1171 case GLSL_TYPE_UINT
:
1172 emit(fs_inst(BRW_OPCODE_MOV
, reg
, fs_reg(ir
->value
.u
[i
])));
1175 emit(fs_inst(BRW_OPCODE_MOV
, reg
, fs_reg(ir
->value
.i
[i
])));
1177 case GLSL_TYPE_BOOL
:
1178 emit(fs_inst(BRW_OPCODE_MOV
, reg
, fs_reg((int)ir
->value
.b
[i
])));
1181 assert(!"Non-float/uint/int/bool constant");
1188 fs_visitor::visit(ir_if
*ir
)
1192 /* Don't point the annotation at the if statement, because then it plus
1193 * the then and else blocks get printed.
1195 this->base_ir
= ir
->condition
;
1197 /* Generate the condition into the condition code. */
1198 ir
->condition
->accept(this);
1199 inst
= emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(brw_null_reg()), this->result
));
1200 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
1202 inst
= emit(fs_inst(BRW_OPCODE_IF
));
1203 inst
->predicated
= true;
1205 foreach_iter(exec_list_iterator
, iter
, ir
->then_instructions
) {
1206 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1212 if (!ir
->else_instructions
.is_empty()) {
1213 emit(fs_inst(BRW_OPCODE_ELSE
));
1215 foreach_iter(exec_list_iterator
, iter
, ir
->else_instructions
) {
1216 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1223 emit(fs_inst(BRW_OPCODE_ENDIF
));
1227 fs_visitor::visit(ir_loop
*ir
)
1229 fs_reg counter
= reg_undef
;
1232 this->base_ir
= ir
->counter
;
1233 ir
->counter
->accept(this);
1234 counter
= *(variable_storage(ir
->counter
));
1237 this->base_ir
= ir
->from
;
1238 ir
->from
->accept(this);
1240 emit(fs_inst(BRW_OPCODE_MOV
, counter
, this->result
));
1244 /* Start a safety counter. If the user messed up their loop
1245 * counting, we don't want to hang the GPU.
1247 fs_reg max_iter
= fs_reg(this, glsl_type::int_type
);
1248 emit(fs_inst(BRW_OPCODE_MOV
, max_iter
, fs_reg(10000)));
1250 emit(fs_inst(BRW_OPCODE_DO
));
1253 this->base_ir
= ir
->to
;
1254 ir
->to
->accept(this);
1256 fs_inst
*inst
= emit(fs_inst(BRW_OPCODE_CMP
, reg_null
,
1257 counter
, this->result
));
1259 case ir_binop_equal
:
1260 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
1262 case ir_binop_nequal
:
1263 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
1265 case ir_binop_gequal
:
1266 inst
->conditional_mod
= BRW_CONDITIONAL_GE
;
1268 case ir_binop_lequal
:
1269 inst
->conditional_mod
= BRW_CONDITIONAL_LE
;
1271 case ir_binop_greater
:
1272 inst
->conditional_mod
= BRW_CONDITIONAL_G
;
1275 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
1278 assert(!"not reached: unknown loop condition");
1283 inst
= emit(fs_inst(BRW_OPCODE_BREAK
));
1284 inst
->predicated
= true;
1287 foreach_iter(exec_list_iterator
, iter
, ir
->body_instructions
) {
1288 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1294 /* Check the maximum loop iters counter. */
1295 inst
= emit(fs_inst(BRW_OPCODE_ADD
, max_iter
, max_iter
, fs_reg(-1)));
1296 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
1298 inst
= emit(fs_inst(BRW_OPCODE_BREAK
));
1299 inst
->predicated
= true;
1302 if (ir
->increment
) {
1303 this->base_ir
= ir
->increment
;
1304 ir
->increment
->accept(this);
1305 emit(fs_inst(BRW_OPCODE_ADD
, counter
, counter
, this->result
));
1308 emit(fs_inst(BRW_OPCODE_WHILE
));
1312 fs_visitor::visit(ir_loop_jump
*ir
)
1315 case ir_loop_jump::jump_break
:
1316 emit(fs_inst(BRW_OPCODE_BREAK
));
1318 case ir_loop_jump::jump_continue
:
1319 emit(fs_inst(BRW_OPCODE_CONTINUE
));
1325 fs_visitor::visit(ir_call
*ir
)
1327 assert(!"FINISHME");
1331 fs_visitor::visit(ir_return
*ir
)
1333 assert(!"FINISHME");
1337 fs_visitor::visit(ir_function
*ir
)
1339 /* Ignore function bodies other than main() -- we shouldn't see calls to
1340 * them since they should all be inlined before we get to ir_to_mesa.
1342 if (strcmp(ir
->name
, "main") == 0) {
1343 const ir_function_signature
*sig
;
1346 sig
= ir
->matching_signature(&empty
);
1350 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
1351 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1360 fs_visitor::visit(ir_function_signature
*ir
)
1362 assert(!"not reached");
1367 fs_visitor::emit(fs_inst inst
)
1369 fs_inst
*list_inst
= new(mem_ctx
) fs_inst
;
1372 list_inst
->annotation
= this->current_annotation
;
1373 list_inst
->ir
= this->base_ir
;
1375 this->instructions
.push_tail(list_inst
);
1380 /** Emits a dummy fragment shader consisting of magenta for bringup purposes. */
1382 fs_visitor::emit_dummy_fs()
1384 /* Everyone's favorite color. */
1385 emit(fs_inst(BRW_OPCODE_MOV
,
1388 emit(fs_inst(BRW_OPCODE_MOV
,
1391 emit(fs_inst(BRW_OPCODE_MOV
,
1394 emit(fs_inst(BRW_OPCODE_MOV
,
1399 write
= emit(fs_inst(FS_OPCODE_FB_WRITE
,
1404 /* The register location here is relative to the start of the URB
1405 * data. It will get adjusted to be a real location before
1406 * generate_code() time.
1409 fs_visitor::interp_reg(int location
, int channel
)
1411 int regnr
= location
* 2 + channel
/ 2;
1412 int stride
= (channel
& 1) * 4;
1414 return brw_vec1_grf(regnr
, stride
);
1417 /** Emits the interpolation for the varying inputs. */
1419 fs_visitor::emit_interpolation()
1421 struct brw_reg g1_uw
= retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW
);
1422 /* For now, the source regs for the setup URB data will be unset,
1423 * since we don't know until codegen how many push constants we'll
1424 * use, and therefore what the setup URB offset is.
1426 fs_reg src_reg
= reg_undef
;
1428 this->current_annotation
= "compute pixel centers";
1429 this->pixel_x
= fs_reg(this, glsl_type::uint_type
);
1430 this->pixel_y
= fs_reg(this, glsl_type::uint_type
);
1431 this->pixel_x
.type
= BRW_REGISTER_TYPE_UW
;
1432 this->pixel_y
.type
= BRW_REGISTER_TYPE_UW
;
1433 emit(fs_inst(BRW_OPCODE_ADD
,
1435 fs_reg(stride(suboffset(g1_uw
, 4), 2, 4, 0)),
1436 fs_reg(brw_imm_v(0x10101010))));
1437 emit(fs_inst(BRW_OPCODE_ADD
,
1439 fs_reg(stride(suboffset(g1_uw
, 5), 2, 4, 0)),
1440 fs_reg(brw_imm_v(0x11001100))));
1442 this->current_annotation
= "compute pixel deltas from v0";
1443 this->delta_x
= fs_reg(this, glsl_type::float_type
);
1444 this->delta_y
= fs_reg(this, glsl_type::float_type
);
1445 emit(fs_inst(BRW_OPCODE_ADD
,
1448 fs_reg(negate(brw_vec1_grf(1, 0)))));
1449 emit(fs_inst(BRW_OPCODE_ADD
,
1452 fs_reg(negate(brw_vec1_grf(1, 1)))));
1454 this->current_annotation
= "compute pos.w and 1/pos.w";
1455 /* Compute wpos.w. It's always in our setup, since it's needed to
1456 * interpolate the other attributes.
1458 this->wpos_w
= fs_reg(this, glsl_type::float_type
);
1459 emit(fs_inst(FS_OPCODE_LINTERP
, wpos_w
, this->delta_x
, this->delta_y
,
1460 interp_reg(FRAG_ATTRIB_WPOS
, 3)));
1461 /* Compute the pixel 1/W value from wpos.w. */
1462 this->pixel_w
= fs_reg(this, glsl_type::float_type
);
1463 emit(fs_inst(FS_OPCODE_RCP
, this->pixel_w
, wpos_w
));
1465 foreach_iter(exec_list_iterator
, iter
, *this->shader
->ir
) {
1466 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1467 ir_variable
*var
= ir
->as_variable();
1472 if (var
->mode
!= ir_var_in
)
1475 /* If it's already set up (WPOS), skip. */
1476 if (var
->location
== 0)
1479 this->current_annotation
= talloc_asprintf(this->mem_ctx
,
1481 "(FRAG_ATTRIB[%d])",
1484 emit_pinterp(var
->location
);
1486 this->current_annotation
= NULL
;
1490 fs_visitor::emit_pinterp(int location
)
1492 fs_reg interp_attr
= fs_reg(this, glsl_type::vec4_type
);
1493 this->interp_attrs
[location
] = interp_attr
;
1495 for (unsigned int i
= 0; i
< 4; i
++) {
1496 struct brw_reg interp
= interp_reg(location
, i
);
1497 emit(fs_inst(FS_OPCODE_LINTERP
,
1502 interp_attr
.reg_offset
++;
1504 interp_attr
.reg_offset
-= 4;
1506 for (unsigned int i
= 0; i
< 4; i
++) {
1507 emit(fs_inst(BRW_OPCODE_MUL
,
1511 interp_attr
.reg_offset
++;
1516 fs_visitor::emit_fb_writes()
1518 this->current_annotation
= "FB write header";
1524 if (c
->key
.aa_dest_stencil_reg
) {
1525 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, nr
++),
1526 fs_reg(brw_vec8_grf(c
->key
.aa_dest_stencil_reg
, 0))));
1529 /* Reserve space for color. It'll be filled in per MRT below. */
1533 if (c
->key
.source_depth_to_render_target
) {
1534 if (c
->key
.computes_depth
) {
1535 /* Hand over gl_FragDepth. */
1536 assert(this->frag_depth
);
1537 fs_reg depth
= *(variable_storage(this->frag_depth
));
1539 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, nr
++), depth
));
1541 /* Pass through the payload depth. */
1542 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, nr
++),
1543 fs_reg(brw_vec8_grf(c
->key
.source_depth_reg
, 0))));
1547 if (c
->key
.dest_depth_reg
) {
1548 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, nr
++),
1549 fs_reg(brw_vec8_grf(c
->key
.dest_depth_reg
, 0))));
1552 fs_reg color
= reg_undef
;
1553 if (this->frag_color
)
1554 color
= *(variable_storage(this->frag_color
));
1555 else if (this->frag_data
)
1556 color
= *(variable_storage(this->frag_data
));
1558 for (int target
= 0; target
< c
->key
.nr_color_regions
; target
++) {
1559 this->current_annotation
= talloc_asprintf(this->mem_ctx
,
1560 "FB write target %d",
1562 if (this->frag_color
|| this->frag_data
) {
1563 for (int i
= 0; i
< 4; i
++) {
1564 emit(fs_inst(BRW_OPCODE_MOV
,
1565 fs_reg(MRF
, color_mrf
+ i
),
1571 if (this->frag_color
)
1572 color
.reg_offset
-= 4;
1574 fs_inst
*inst
= emit(fs_inst(FS_OPCODE_FB_WRITE
,
1575 reg_undef
, reg_undef
));
1576 inst
->target
= target
;
1578 if (target
== c
->key
.nr_color_regions
- 1)
1582 if (c
->key
.nr_color_regions
== 0) {
1583 fs_inst
*inst
= emit(fs_inst(FS_OPCODE_FB_WRITE
,
1584 reg_undef
, reg_undef
));
1589 this->current_annotation
= NULL
;
1593 fs_visitor::generate_fb_write(fs_inst
*inst
)
1595 GLboolean eot
= inst
->eot
;
1597 /* Header is 2 regs, g0 and g1 are the contents. g0 will be implied
1600 brw_push_insn_state(p
);
1601 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
1602 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
1605 brw_vec8_grf(1, 0));
1606 brw_pop_insn_state(p
);
1609 8, /* dispatch_width */
1610 retype(vec8(brw_null_reg()), BRW_REGISTER_TYPE_UW
),
1612 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW
),
1620 fs_visitor::generate_linterp(fs_inst
*inst
,
1621 struct brw_reg dst
, struct brw_reg
*src
)
1623 struct brw_reg delta_x
= src
[0];
1624 struct brw_reg delta_y
= src
[1];
1625 struct brw_reg interp
= src
[2];
1628 delta_y
.nr
== delta_x
.nr
+ 1 &&
1629 (intel
->gen
>= 6 || (delta_x
.nr
& 1) == 0)) {
1630 brw_PLN(p
, dst
, interp
, delta_x
);
1632 brw_LINE(p
, brw_null_reg(), interp
, delta_x
);
1633 brw_MAC(p
, dst
, suboffset(interp
, 1), delta_y
);
1638 fs_visitor::generate_math(fs_inst
*inst
,
1639 struct brw_reg dst
, struct brw_reg
*src
)
1643 switch (inst
->opcode
) {
1645 op
= BRW_MATH_FUNCTION_INV
;
1648 op
= BRW_MATH_FUNCTION_RSQ
;
1650 case FS_OPCODE_SQRT
:
1651 op
= BRW_MATH_FUNCTION_SQRT
;
1653 case FS_OPCODE_EXP2
:
1654 op
= BRW_MATH_FUNCTION_EXP
;
1656 case FS_OPCODE_LOG2
:
1657 op
= BRW_MATH_FUNCTION_LOG
;
1660 op
= BRW_MATH_FUNCTION_POW
;
1663 op
= BRW_MATH_FUNCTION_SIN
;
1666 op
= BRW_MATH_FUNCTION_COS
;
1669 assert(!"not reached: unknown math function");
1674 if (inst
->opcode
== FS_OPCODE_POW
) {
1675 brw_MOV(p
, brw_message_reg(3), src
[1]);
1680 inst
->saturate
? BRW_MATH_SATURATE_SATURATE
:
1681 BRW_MATH_SATURATE_NONE
,
1683 BRW_MATH_DATA_VECTOR
,
1684 BRW_MATH_PRECISION_FULL
);
1688 fs_visitor::generate_tex(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
)
1693 if (intel
->gen
== 5) {
1694 switch (inst
->opcode
) {
1696 if (inst
->shadow_compare
) {
1697 msg_type
= BRW_SAMPLER_MESSAGE_SAMPLE_COMPARE_GEN5
;
1699 msg_type
= BRW_SAMPLER_MESSAGE_SAMPLE_GEN5
;
1703 if (inst
->shadow_compare
) {
1704 msg_type
= BRW_SAMPLER_MESSAGE_SAMPLE_BIAS_COMPARE_GEN5
;
1706 msg_type
= BRW_SAMPLER_MESSAGE_SAMPLE_BIAS_GEN5
;
1711 switch (inst
->opcode
) {
1713 /* Note that G45 and older determines shadow compare and dispatch width
1714 * from message length for most messages.
1716 if (inst
->shadow_compare
) {
1717 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_COMPARE
;
1719 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE
;
1722 if (inst
->shadow_compare
) {
1723 assert(!"FINISHME: shadow compare with bias.");
1724 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_BIAS
;
1726 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_BIAS
;
1732 assert(msg_type
!= -1);
1738 retype(dst
, BRW_REGISTER_TYPE_UW
),
1740 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW
),
1741 SURF_INDEX_TEXTURE(inst
->sampler
),
1749 BRW_SAMPLER_SIMD_MODE_SIMD8
);
1753 /* For OPCODE_DDX and OPCODE_DDY, per channel of output we've got input
1756 * arg0: ss0.tl ss0.tr ss0.bl ss0.br ss1.tl ss1.tr ss1.bl ss1.br
1758 * and we're trying to produce:
1761 * dst: (ss0.tr - ss0.tl) (ss0.tl - ss0.bl)
1762 * (ss0.tr - ss0.tl) (ss0.tr - ss0.br)
1763 * (ss0.br - ss0.bl) (ss0.tl - ss0.bl)
1764 * (ss0.br - ss0.bl) (ss0.tr - ss0.br)
1765 * (ss1.tr - ss1.tl) (ss1.tl - ss1.bl)
1766 * (ss1.tr - ss1.tl) (ss1.tr - ss1.br)
1767 * (ss1.br - ss1.bl) (ss1.tl - ss1.bl)
1768 * (ss1.br - ss1.bl) (ss1.tr - ss1.br)
1770 * and add another set of two more subspans if in 16-pixel dispatch mode.
1772 * For DDX, it ends up being easy: width = 2, horiz=0 gets us the same result
1773 * for each pair, and vertstride = 2 jumps us 2 elements after processing a
1774 * pair. But for DDY, it's harder, as we want to produce the pairs swizzled
1775 * between each other. We could probably do it like ddx and swizzle the right
1776 * order later, but bail for now and just produce
1777 * ((ss0.tl - ss0.bl)x4 (ss1.tl - ss1.bl)x4)
1780 fs_visitor::generate_ddx(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
)
1782 struct brw_reg src0
= brw_reg(src
.file
, src
.nr
, 1,
1783 BRW_REGISTER_TYPE_F
,
1784 BRW_VERTICAL_STRIDE_2
,
1786 BRW_HORIZONTAL_STRIDE_0
,
1787 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
1788 struct brw_reg src1
= brw_reg(src
.file
, src
.nr
, 0,
1789 BRW_REGISTER_TYPE_F
,
1790 BRW_VERTICAL_STRIDE_2
,
1792 BRW_HORIZONTAL_STRIDE_0
,
1793 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
1794 brw_ADD(p
, dst
, src0
, negate(src1
));
1798 fs_visitor::generate_ddy(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
)
1800 struct brw_reg src0
= brw_reg(src
.file
, src
.nr
, 0,
1801 BRW_REGISTER_TYPE_F
,
1802 BRW_VERTICAL_STRIDE_4
,
1804 BRW_HORIZONTAL_STRIDE_0
,
1805 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
1806 struct brw_reg src1
= brw_reg(src
.file
, src
.nr
, 2,
1807 BRW_REGISTER_TYPE_F
,
1808 BRW_VERTICAL_STRIDE_4
,
1810 BRW_HORIZONTAL_STRIDE_0
,
1811 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
1812 brw_ADD(p
, dst
, src0
, negate(src1
));
1816 fs_visitor::generate_discard(fs_inst
*inst
)
1818 struct brw_reg g0
= retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW
);
1819 brw_push_insn_state(p
);
1820 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
1821 brw_NOT(p
, c
->emit_mask_reg
, brw_mask_reg(1)); /* IMASK */
1822 brw_AND(p
, g0
, c
->emit_mask_reg
, g0
);
1823 brw_pop_insn_state(p
);
1827 trivial_assign_reg(int header_size
, fs_reg
*reg
)
1829 if (reg
->file
== GRF
&& reg
->reg
!= 0) {
1830 reg
->hw_reg
= header_size
+ reg
->reg
- 1 + reg
->reg_offset
;
1836 fs_visitor::assign_curb_setup()
1838 c
->prog_data
.first_curbe_grf
= c
->key
.nr_payload_regs
;
1839 c
->prog_data
.curb_read_length
= ALIGN(c
->prog_data
.nr_params
, 8) / 8;
1841 if (intel
->gen
== 5 && (c
->prog_data
.first_curbe_grf
+
1842 c
->prog_data
.curb_read_length
) & 1) {
1843 /* Align the start of the interpolation coefficients so that we can use
1844 * the PLN instruction.
1846 c
->prog_data
.first_curbe_grf
++;
1849 /* Map the offsets in the UNIFORM file to fixed HW regs. */
1850 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
1851 fs_inst
*inst
= (fs_inst
*)iter
.get();
1853 for (unsigned int i
= 0; i
< 3; i
++) {
1854 if (inst
->src
[i
].file
== UNIFORM
) {
1855 int constant_nr
= inst
->src
[i
].hw_reg
+ inst
->src
[i
].reg_offset
;
1856 struct brw_reg brw_reg
= brw_vec1_grf(c
->prog_data
.first_curbe_grf
+
1860 inst
->src
[i
].file
= FIXED_HW_REG
;
1861 inst
->src
[i
].fixed_hw_reg
= brw_reg
;
1868 fs_visitor::assign_urb_setup()
1870 int urb_start
= c
->prog_data
.first_curbe_grf
+ c
->prog_data
.curb_read_length
;
1871 int interp_reg_nr
[FRAG_ATTRIB_MAX
];
1873 c
->prog_data
.urb_read_length
= 0;
1875 /* Figure out where each of the incoming setup attributes lands. */
1876 for (unsigned int i
= 0; i
< FRAG_ATTRIB_MAX
; i
++) {
1877 interp_reg_nr
[i
] = -1;
1879 if (i
!= FRAG_ATTRIB_WPOS
&&
1880 !(brw
->fragment_program
->Base
.InputsRead
& BITFIELD64_BIT(i
)))
1883 /* Each attribute is 4 setup channels, each of which is half a reg. */
1884 interp_reg_nr
[i
] = urb_start
+ c
->prog_data
.urb_read_length
;
1885 c
->prog_data
.urb_read_length
+= 2;
1888 /* Map the register numbers for FS_OPCODE_LINTERP so that it uses
1889 * the correct setup input.
1891 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
1892 fs_inst
*inst
= (fs_inst
*)iter
.get();
1894 if (inst
->opcode
!= FS_OPCODE_LINTERP
)
1897 assert(inst
->src
[2].file
== FIXED_HW_REG
);
1899 int location
= inst
->src
[2].fixed_hw_reg
.nr
/ 2;
1900 assert(interp_reg_nr
[location
] != -1);
1901 inst
->src
[2].fixed_hw_reg
.nr
= (interp_reg_nr
[location
] +
1902 (inst
->src
[2].fixed_hw_reg
.nr
& 1));
1905 this->first_non_payload_grf
= urb_start
+ c
->prog_data
.urb_read_length
;
1909 fs_visitor::assign_regs()
1911 int header_size
= this->first_non_payload_grf
;
1914 /* FINISHME: trivial assignment of register numbers */
1915 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
1916 fs_inst
*inst
= (fs_inst
*)iter
.get();
1918 trivial_assign_reg(header_size
, &inst
->dst
);
1919 trivial_assign_reg(header_size
, &inst
->src
[0]);
1920 trivial_assign_reg(header_size
, &inst
->src
[1]);
1922 last_grf
= MAX2(last_grf
, inst
->dst
.hw_reg
);
1923 last_grf
= MAX2(last_grf
, inst
->src
[0].hw_reg
);
1924 last_grf
= MAX2(last_grf
, inst
->src
[1].hw_reg
);
1927 this->grf_used
= last_grf
+ 1;
1930 static struct brw_reg
brw_reg_from_fs_reg(fs_reg
*reg
)
1932 struct brw_reg brw_reg
;
1934 switch (reg
->file
) {
1938 brw_reg
= brw_vec8_reg(reg
->file
,
1940 brw_reg
= retype(brw_reg
, reg
->type
);
1943 switch (reg
->type
) {
1944 case BRW_REGISTER_TYPE_F
:
1945 brw_reg
= brw_imm_f(reg
->imm
.f
);
1947 case BRW_REGISTER_TYPE_D
:
1948 brw_reg
= brw_imm_d(reg
->imm
.i
);
1950 case BRW_REGISTER_TYPE_UD
:
1951 brw_reg
= brw_imm_ud(reg
->imm
.u
);
1954 assert(!"not reached");
1959 brw_reg
= reg
->fixed_hw_reg
;
1962 /* Probably unused. */
1963 brw_reg
= brw_null_reg();
1966 assert(!"not reached");
1967 brw_reg
= brw_null_reg();
1971 brw_reg
= brw_abs(brw_reg
);
1973 brw_reg
= negate(brw_reg
);
1979 fs_visitor::generate_code()
1981 unsigned int annotation_len
= 0;
1982 int last_native_inst
= 0;
1983 struct brw_instruction
*if_stack
[16], *loop_stack
[16];
1984 int if_stack_depth
= 0, loop_stack_depth
= 0;
1985 int if_depth_in_loop
[16];
1987 if_depth_in_loop
[loop_stack_depth
] = 0;
1989 memset(&if_stack
, 0, sizeof(if_stack
));
1990 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
1991 fs_inst
*inst
= (fs_inst
*)iter
.get();
1992 struct brw_reg src
[3], dst
;
1994 for (unsigned int i
= 0; i
< 3; i
++) {
1995 src
[i
] = brw_reg_from_fs_reg(&inst
->src
[i
]);
1997 dst
= brw_reg_from_fs_reg(&inst
->dst
);
1999 brw_set_conditionalmod(p
, inst
->conditional_mod
);
2000 brw_set_predicate_control(p
, inst
->predicated
);
2002 switch (inst
->opcode
) {
2003 case BRW_OPCODE_MOV
:
2004 brw_MOV(p
, dst
, src
[0]);
2006 case BRW_OPCODE_ADD
:
2007 brw_ADD(p
, dst
, src
[0], src
[1]);
2009 case BRW_OPCODE_MUL
:
2010 brw_MUL(p
, dst
, src
[0], src
[1]);
2013 case BRW_OPCODE_FRC
:
2014 brw_FRC(p
, dst
, src
[0]);
2016 case BRW_OPCODE_RNDD
:
2017 brw_RNDD(p
, dst
, src
[0]);
2019 case BRW_OPCODE_RNDZ
:
2020 brw_RNDZ(p
, dst
, src
[0]);
2023 case BRW_OPCODE_AND
:
2024 brw_AND(p
, dst
, src
[0], src
[1]);
2027 brw_OR(p
, dst
, src
[0], src
[1]);
2029 case BRW_OPCODE_XOR
:
2030 brw_XOR(p
, dst
, src
[0], src
[1]);
2033 case BRW_OPCODE_CMP
:
2034 brw_CMP(p
, dst
, inst
->conditional_mod
, src
[0], src
[1]);
2036 case BRW_OPCODE_SEL
:
2037 brw_SEL(p
, dst
, src
[0], src
[1]);
2041 assert(if_stack_depth
< 16);
2042 if_stack
[if_stack_depth
] = brw_IF(p
, BRW_EXECUTE_8
);
2043 if_depth_in_loop
[loop_stack_depth
]++;
2046 case BRW_OPCODE_ELSE
:
2047 if_stack
[if_stack_depth
- 1] =
2048 brw_ELSE(p
, if_stack
[if_stack_depth
- 1]);
2050 case BRW_OPCODE_ENDIF
:
2052 brw_ENDIF(p
, if_stack
[if_stack_depth
]);
2053 if_depth_in_loop
[loop_stack_depth
]--;
2057 loop_stack
[loop_stack_depth
++] = brw_DO(p
, BRW_EXECUTE_8
);
2058 if_depth_in_loop
[loop_stack_depth
] = 0;
2061 case BRW_OPCODE_BREAK
:
2062 brw_BREAK(p
, if_depth_in_loop
[loop_stack_depth
]);
2063 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
2065 case BRW_OPCODE_CONTINUE
:
2066 brw_CONT(p
, if_depth_in_loop
[loop_stack_depth
]);
2067 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
2070 case BRW_OPCODE_WHILE
: {
2071 struct brw_instruction
*inst0
, *inst1
;
2074 if (intel
->gen
== 5)
2077 assert(loop_stack_depth
> 0);
2079 inst0
= inst1
= brw_WHILE(p
, loop_stack
[loop_stack_depth
]);
2080 /* patch all the BREAK/CONT instructions from last BGNLOOP */
2081 while (inst0
> loop_stack
[loop_stack_depth
]) {
2083 if (inst0
->header
.opcode
== BRW_OPCODE_BREAK
&&
2084 inst0
->bits3
.if_else
.jump_count
== 0) {
2085 inst0
->bits3
.if_else
.jump_count
= br
* (inst1
- inst0
+ 1);
2087 else if (inst0
->header
.opcode
== BRW_OPCODE_CONTINUE
&&
2088 inst0
->bits3
.if_else
.jump_count
== 0) {
2089 inst0
->bits3
.if_else
.jump_count
= br
* (inst1
- inst0
);
2097 case FS_OPCODE_SQRT
:
2098 case FS_OPCODE_EXP2
:
2099 case FS_OPCODE_LOG2
:
2103 generate_math(inst
, dst
, src
);
2105 case FS_OPCODE_LINTERP
:
2106 generate_linterp(inst
, dst
, src
);
2111 generate_tex(inst
, dst
, src
[0]);
2113 case FS_OPCODE_DISCARD
:
2114 generate_discard(inst
);
2117 generate_ddx(inst
, dst
, src
[0]);
2120 generate_ddy(inst
, dst
, src
[0]);
2122 case FS_OPCODE_FB_WRITE
:
2123 generate_fb_write(inst
);
2126 if (inst
->opcode
< (int)ARRAY_SIZE(brw_opcodes
)) {
2127 _mesa_problem(ctx
, "Unsupported opcode `%s' in FS",
2128 brw_opcodes
[inst
->opcode
].name
);
2130 _mesa_problem(ctx
, "Unsupported opcode %d in FS", inst
->opcode
);
2135 if (annotation_len
< p
->nr_insn
) {
2136 annotation_len
*= 2;
2137 if (annotation_len
< 16)
2138 annotation_len
= 16;
2140 this->annotation_string
= talloc_realloc(this->mem_ctx
,
2144 this->annotation_ir
= talloc_realloc(this->mem_ctx
,
2150 for (unsigned int i
= last_native_inst
; i
< p
->nr_insn
; i
++) {
2151 this->annotation_string
[i
] = inst
->annotation
;
2152 this->annotation_ir
[i
] = inst
->ir
;
2154 last_native_inst
= p
->nr_insn
;
2159 brw_wm_fs_emit(struct brw_context
*brw
, struct brw_wm_compile
*c
)
2161 struct brw_compile
*p
= &c
->func
;
2162 struct intel_context
*intel
= &brw
->intel
;
2163 GLcontext
*ctx
= &intel
->ctx
;
2164 struct brw_shader
*shader
= NULL
;
2165 struct gl_shader_program
*prog
= ctx
->Shader
.CurrentProgram
;
2173 for (unsigned int i
= 0; i
< prog
->_NumLinkedShaders
; i
++) {
2174 if (prog
->_LinkedShaders
[i
]->Type
== GL_FRAGMENT_SHADER
) {
2175 shader
= (struct brw_shader
*)prog
->_LinkedShaders
[i
];
2182 /* We always use 8-wide mode, at least for now. For one, flow
2183 * control only works in 8-wide. Also, when we're fragment shader
2184 * bound, we're almost always under register pressure as well, so
2185 * 8-wide would save us from the performance cliff of spilling
2188 c
->dispatch_width
= 8;
2190 if (INTEL_DEBUG
& DEBUG_WM
) {
2191 printf("GLSL IR for native fragment shader %d:\n", prog
->Name
);
2192 _mesa_print_ir(shader
->ir
, NULL
);
2196 /* Now the main event: Visit the shader IR and generate our FS IR for it.
2198 fs_visitor
v(c
, shader
);
2203 v
.emit_interpolation();
2205 /* Generate FS IR for main(). (the visitor only descends into
2206 * functions called "main").
2208 foreach_iter(exec_list_iterator
, iter
, *shader
->ir
) {
2209 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
2215 v
.assign_curb_setup();
2216 v
.assign_urb_setup();
2222 assert(!v
.fail
); /* FINISHME: Cleanly fail, tested at link time, etc. */
2227 if (INTEL_DEBUG
& DEBUG_WM
) {
2228 const char *last_annotation_string
= NULL
;
2229 ir_instruction
*last_annotation_ir
= NULL
;
2231 printf("Native code for fragment shader %d:\n", prog
->Name
);
2232 for (unsigned int i
= 0; i
< p
->nr_insn
; i
++) {
2233 if (last_annotation_ir
!= v
.annotation_ir
[i
]) {
2234 last_annotation_ir
= v
.annotation_ir
[i
];
2235 if (last_annotation_ir
) {
2237 last_annotation_ir
->print();
2241 if (last_annotation_string
!= v
.annotation_string
[i
]) {
2242 last_annotation_string
= v
.annotation_string
[i
];
2243 if (last_annotation_string
)
2244 printf(" %s\n", last_annotation_string
);
2246 brw_disasm(stdout
, &p
->store
[i
], intel
->gen
);
2251 c
->prog_data
.total_grf
= v
.grf_used
;
2252 c
->prog_data
.total_scratch
= 0;