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/hash_table.h"
38 #include "brw_context.h"
43 #include "../glsl/glsl_types.h"
44 #include "../glsl/ir_optimization.h"
45 #include "../glsl/ir_print_visitor.h"
48 ARF
= BRW_ARCHITECTURE_REGISTER_FILE
,
49 GRF
= BRW_GENERAL_REGISTER_FILE
,
50 MRF
= BRW_MESSAGE_REGISTER_FILE
,
51 IMM
= BRW_IMMEDIATE_VALUE
,
52 FIXED_HW_REG
, /* a struct brw_reg */
53 UNIFORM
, /* prog_data->params[hw_reg] */
58 FS_OPCODE_FB_WRITE
= 256,
76 static int using_new_fs
= -1;
79 brw_new_shader(GLcontext
*ctx
, GLuint name
, GLuint type
)
81 struct brw_shader
*shader
;
83 shader
= talloc_zero(NULL
, struct brw_shader
);
85 shader
->base
.Type
= type
;
86 shader
->base
.Name
= name
;
87 _mesa_init_shader(ctx
, &shader
->base
);
93 struct gl_shader_program
*
94 brw_new_shader_program(GLcontext
*ctx
, GLuint name
)
96 struct brw_shader_program
*prog
;
97 prog
= talloc_zero(NULL
, struct brw_shader_program
);
99 prog
->base
.Name
= name
;
100 _mesa_init_shader_program(ctx
, &prog
->base
);
106 brw_compile_shader(GLcontext
*ctx
, struct gl_shader
*shader
)
108 if (!_mesa_ir_compile_shader(ctx
, shader
))
115 brw_link_shader(GLcontext
*ctx
, struct gl_shader_program
*prog
)
117 if (using_new_fs
== -1)
118 using_new_fs
= getenv("INTEL_NEW_FS") != NULL
;
120 for (unsigned i
= 0; i
< prog
->_NumLinkedShaders
; i
++) {
121 struct brw_shader
*shader
= (struct brw_shader
*)prog
->_LinkedShaders
[i
];
123 if (using_new_fs
&& shader
->base
.Type
== GL_FRAGMENT_SHADER
) {
124 void *mem_ctx
= talloc_new(NULL
);
128 talloc_free(shader
->ir
);
129 shader
->ir
= new(shader
) exec_list
;
130 clone_ir_list(mem_ctx
, shader
->ir
, shader
->base
.ir
);
132 do_mat_op_to_vec(shader
->ir
);
133 do_mod_to_fract(shader
->ir
);
134 do_div_to_mul_rcp(shader
->ir
);
135 do_sub_to_add_neg(shader
->ir
);
136 do_explog_to_explog2(shader
->ir
);
138 brw_do_channel_expressions(shader
->ir
);
139 brw_do_vector_splitting(shader
->ir
);
144 progress
= do_common_optimization(shader
->ir
, true, 32) || progress
;
147 validate_ir_tree(shader
->ir
);
149 reparent_ir(shader
->ir
, shader
->ir
);
150 talloc_free(mem_ctx
);
154 if (!_mesa_ir_link_shader(ctx
, prog
))
161 type_size(const struct glsl_type
*type
)
163 unsigned int size
, i
;
165 switch (type
->base_type
) {
168 case GLSL_TYPE_FLOAT
:
170 return type
->components();
171 case GLSL_TYPE_ARRAY
:
172 /* FINISHME: uniform/varying arrays. */
173 return type_size(type
->fields
.array
) * type
->length
;
174 case GLSL_TYPE_STRUCT
:
176 for (i
= 0; i
< type
->length
; i
++) {
177 size
+= type_size(type
->fields
.structure
[i
].type
);
180 case GLSL_TYPE_SAMPLER
:
181 /* Samplers take up no register space, since they're baked in at
186 assert(!"not reached");
193 /* Callers of this talloc-based new need not call delete. It's
194 * easier to just talloc_free 'ctx' (or any of its ancestors). */
195 static void* operator new(size_t size
, void *ctx
)
199 node
= talloc_size(ctx
, size
);
200 assert(node
!= NULL
);
208 this->reg_offset
= 0;
214 /** Generic unset register constructor. */
218 this->file
= BAD_FILE
;
221 /** Immediate value constructor. */
226 this->type
= BRW_REGISTER_TYPE_F
;
230 /** Immediate value constructor. */
235 this->type
= BRW_REGISTER_TYPE_D
;
239 /** Immediate value constructor. */
244 this->type
= BRW_REGISTER_TYPE_UD
;
248 /** Fixed brw_reg Immediate value constructor. */
249 fs_reg(struct brw_reg fixed_hw_reg
)
252 this->file
= FIXED_HW_REG
;
253 this->fixed_hw_reg
= fixed_hw_reg
;
254 this->type
= fixed_hw_reg
.type
;
257 fs_reg(enum register_file file
, int hw_reg
);
258 fs_reg(class fs_visitor
*v
, const struct glsl_type
*type
);
260 /** Register file: ARF, GRF, MRF, IMM. */
261 enum register_file file
;
262 /** Abstract register number. 0 = fixed hw reg */
264 /** Offset within the abstract register. */
266 /** HW register number. Generally unset until register allocation. */
268 /** Register type. BRW_REGISTER_TYPE_* */
272 struct brw_reg fixed_hw_reg
;
274 /** Value for file == BRW_IMMMEDIATE_FILE */
282 static const fs_reg reg_undef
;
283 static const fs_reg
reg_null(ARF
, BRW_ARF_NULL
);
285 class fs_inst
: public exec_node
{
287 /* Callers of this talloc-based new need not call delete. It's
288 * easier to just talloc_free 'ctx' (or any of its ancestors). */
289 static void* operator new(size_t size
, void *ctx
)
293 node
= talloc_zero_size(ctx
, size
);
294 assert(node
!= NULL
);
301 this->opcode
= BRW_OPCODE_NOP
;
302 this->saturate
= false;
303 this->conditional_mod
= BRW_CONDITIONAL_NONE
;
304 this->predicated
= false;
306 this->shadow_compare
= false;
317 this->opcode
= opcode
;
320 fs_inst(int opcode
, fs_reg dst
, fs_reg src0
)
323 this->opcode
= opcode
;
328 fs_inst(int opcode
, fs_reg dst
, fs_reg src0
, fs_reg src1
)
331 this->opcode
= opcode
;
337 fs_inst(int opcode
, fs_reg dst
, fs_reg src0
, fs_reg src1
, fs_reg src2
)
340 this->opcode
= opcode
;
347 int opcode
; /* BRW_OPCODE_* or FS_OPCODE_* */
352 int conditional_mod
; /**< BRW_CONDITIONAL_* */
354 int mlen
; /** SEND message length */
359 * Annotation for the generated IR. One of the two can be set.
362 const char *annotation
;
366 class fs_visitor
: public ir_visitor
370 fs_visitor(struct brw_wm_compile
*c
, struct brw_shader
*shader
)
375 this->intel
= &brw
->intel
;
376 this->ctx
= &intel
->ctx
;
377 this->mem_ctx
= talloc_new(NULL
);
378 this->shader
= shader
;
380 this->next_abstract_grf
= 1;
381 this->variable_ht
= hash_table_ctor(0,
382 hash_table_pointer_hash
,
383 hash_table_pointer_compare
);
385 this->frag_color
= NULL
;
386 this->frag_data
= NULL
;
387 this->frag_depth
= NULL
;
388 this->first_non_payload_grf
= 0;
390 this->current_annotation
= NULL
;
391 this->annotation_string
= NULL
;
392 this->annotation_ir
= NULL
;
396 talloc_free(this->mem_ctx
);
397 hash_table_dtor(this->variable_ht
);
400 fs_reg
*variable_storage(ir_variable
*var
);
402 void visit(ir_variable
*ir
);
403 void visit(ir_assignment
*ir
);
404 void visit(ir_dereference_variable
*ir
);
405 void visit(ir_dereference_record
*ir
);
406 void visit(ir_dereference_array
*ir
);
407 void visit(ir_expression
*ir
);
408 void visit(ir_texture
*ir
);
409 void visit(ir_if
*ir
);
410 void visit(ir_constant
*ir
);
411 void visit(ir_swizzle
*ir
);
412 void visit(ir_return
*ir
);
413 void visit(ir_loop
*ir
);
414 void visit(ir_loop_jump
*ir
);
415 void visit(ir_discard
*ir
);
416 void visit(ir_call
*ir
);
417 void visit(ir_function
*ir
);
418 void visit(ir_function_signature
*ir
);
420 fs_inst
*emit(fs_inst inst
);
421 void assign_curb_setup();
422 void assign_urb_setup();
424 void generate_code();
425 void generate_fb_write(fs_inst
*inst
);
426 void generate_linterp(fs_inst
*inst
, struct brw_reg dst
,
427 struct brw_reg
*src
);
428 void generate_tex(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
);
429 void generate_math(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg
*src
);
430 void generate_discard(fs_inst
*inst
);
432 void emit_dummy_fs();
433 void emit_interpolation();
434 void emit_pinterp(int location
);
435 void emit_fb_writes();
437 struct brw_reg
interp_reg(int location
, int channel
);
439 struct brw_context
*brw
;
440 struct intel_context
*intel
;
442 struct brw_wm_compile
*c
;
443 struct brw_compile
*p
;
444 struct brw_shader
*shader
;
446 exec_list instructions
;
447 int next_abstract_grf
;
448 struct hash_table
*variable_ht
;
449 ir_variable
*frag_color
, *frag_data
, *frag_depth
;
450 int first_non_payload_grf
;
452 /** @{ debug annotation info */
453 const char *current_annotation
;
454 ir_instruction
*base_ir
;
455 const char **annotation_string
;
456 ir_instruction
**annotation_ir
;
461 /* Result of last visit() method. */
469 fs_reg interp_attrs
[64];
475 /** Fixed HW reg constructor. */
476 fs_reg::fs_reg(enum register_file file
, int hw_reg
)
480 this->hw_reg
= hw_reg
;
481 this->type
= BRW_REGISTER_TYPE_F
;
484 /** Automatic reg constructor. */
485 fs_reg::fs_reg(class fs_visitor
*v
, const struct glsl_type
*type
)
490 this->reg
= v
->next_abstract_grf
;
491 this->reg_offset
= 0;
492 v
->next_abstract_grf
+= type_size(type
);
494 switch (type
->base_type
) {
495 case GLSL_TYPE_FLOAT
:
496 this->type
= BRW_REGISTER_TYPE_F
;
500 this->type
= BRW_REGISTER_TYPE_D
;
503 this->type
= BRW_REGISTER_TYPE_UD
;
506 assert(!"not reached");
507 this->type
= BRW_REGISTER_TYPE_F
;
513 fs_visitor::variable_storage(ir_variable
*var
)
515 return (fs_reg
*)hash_table_find(this->variable_ht
, var
);
519 fs_visitor::visit(ir_variable
*ir
)
523 if (strcmp(ir
->name
, "gl_FragColor") == 0) {
524 this->frag_color
= ir
;
525 } else if (strcmp(ir
->name
, "gl_FragData") == 0) {
526 this->frag_data
= ir
;
527 } else if (strcmp(ir
->name
, "gl_FragDepth") == 0) {
528 this->frag_depth
= ir
;
529 assert(!"FINISHME: this hangs currently.");
532 if (ir
->mode
== ir_var_in
) {
533 reg
= &this->interp_attrs
[ir
->location
];
536 if (ir
->mode
== ir_var_uniform
) {
537 const float *vec_values
;
538 int param_index
= c
->prog_data
.nr_params
;
540 /* FINISHME: This is wildly incomplete. */
541 assert(ir
->type
->is_scalar() || ir
->type
->is_vector() ||
542 ir
->type
->is_sampler());
544 const struct gl_program
*fp
= &this->brw
->fragment_program
->Base
;
545 /* Our support for uniforms is piggy-backed on the struct
546 * gl_fragment_program, because that's where the values actually
547 * get stored, rather than in some global gl_shader_program uniform
550 vec_values
= fp
->Parameters
->ParameterValues
[ir
->location
];
551 for (unsigned int i
= 0; i
< ir
->type
->vector_elements
; i
++) {
552 c
->prog_data
.param
[c
->prog_data
.nr_params
++] = &vec_values
[i
];
555 reg
= new(this->mem_ctx
) fs_reg(UNIFORM
, param_index
);
559 reg
= new(this->mem_ctx
) fs_reg(this, ir
->type
);
561 hash_table_insert(this->variable_ht
, reg
, ir
);
565 fs_visitor::visit(ir_dereference_variable
*ir
)
567 fs_reg
*reg
= variable_storage(ir
->var
);
572 fs_visitor::visit(ir_dereference_record
*ir
)
578 fs_visitor::visit(ir_dereference_array
*ir
)
583 ir
->array
->accept(this);
584 index
= ir
->array_index
->as_constant();
586 if (ir
->type
->is_matrix()) {
587 element_size
= ir
->type
->vector_elements
;
589 element_size
= type_size(ir
->type
);
593 assert(this->result
.file
== UNIFORM
||
594 (this->result
.file
== GRF
&&
595 this->result
.reg
!= 0));
596 this->result
.reg_offset
+= index
->value
.i
[0] * element_size
;
598 assert(!"FINISHME: non-constant matrix column");
603 fs_visitor::visit(ir_expression
*ir
)
605 unsigned int operand
;
610 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
611 ir
->operands
[operand
]->accept(this);
612 if (this->result
.file
== BAD_FILE
) {
614 printf("Failed to get tree for expression operand:\n");
615 ir
->operands
[operand
]->accept(&v
);
618 op
[operand
] = this->result
;
620 /* Matrix expression operands should have been broken down to vector
621 * operations already.
623 assert(!ir
->operands
[operand
]->type
->is_matrix());
624 /* And then those vector operands should have been broken down to scalar.
626 assert(!ir
->operands
[operand
]->type
->is_vector());
629 /* Storage for our result. If our result goes into an assignment, it will
630 * just get copy-propagated out, so no worries.
632 this->result
= fs_reg(this, ir
->type
);
634 switch (ir
->operation
) {
635 case ir_unop_logic_not
:
636 emit(fs_inst(BRW_OPCODE_ADD
, this->result
, op
[0], fs_reg(-1)));
639 op
[0].negate
= ~op
[0].negate
;
640 this->result
= op
[0];
644 this->result
= op
[0];
647 temp
= fs_reg(this, ir
->type
);
649 emit(fs_inst(BRW_OPCODE_MOV
, this->result
, fs_reg(0.0f
)));
651 inst
= emit(fs_inst(BRW_OPCODE_CMP
, reg_null
, op
[0], fs_reg(0.0f
)));
652 inst
->conditional_mod
= BRW_CONDITIONAL_G
;
653 inst
= emit(fs_inst(BRW_OPCODE_MOV
, this->result
, fs_reg(1.0f
)));
654 inst
->predicated
= true;
656 inst
= emit(fs_inst(BRW_OPCODE_CMP
, reg_null
, op
[0], fs_reg(0.0f
)));
657 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
658 inst
= emit(fs_inst(BRW_OPCODE_MOV
, this->result
, fs_reg(-1.0f
)));
659 inst
->predicated
= true;
663 emit(fs_inst(FS_OPCODE_RCP
, this->result
, op
[0]));
667 emit(fs_inst(FS_OPCODE_EXP2
, this->result
, op
[0]));
670 emit(fs_inst(FS_OPCODE_LOG2
, this->result
, op
[0]));
674 assert(!"not reached: should be handled by ir_explog_to_explog2");
677 emit(fs_inst(FS_OPCODE_SIN
, this->result
, op
[0]));
680 emit(fs_inst(FS_OPCODE_COS
, this->result
, op
[0]));
684 emit(fs_inst(FS_OPCODE_DDX
, this->result
, op
[0]));
687 emit(fs_inst(FS_OPCODE_DDY
, this->result
, op
[0]));
691 emit(fs_inst(BRW_OPCODE_ADD
, this->result
, op
[0], op
[1]));
694 assert(!"not reached: should be handled by ir_sub_to_add_neg");
698 emit(fs_inst(BRW_OPCODE_MUL
, this->result
, op
[0], op
[1]));
701 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
704 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
708 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
709 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
710 emit(fs_inst(BRW_OPCODE_AND
, this->result
, this->result
, fs_reg(0x1)));
712 case ir_binop_greater
:
713 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
714 inst
->conditional_mod
= BRW_CONDITIONAL_G
;
715 emit(fs_inst(BRW_OPCODE_AND
, this->result
, this->result
, fs_reg(0x1)));
717 case ir_binop_lequal
:
718 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
719 inst
->conditional_mod
= BRW_CONDITIONAL_LE
;
720 emit(fs_inst(BRW_OPCODE_AND
, this->result
, this->result
, fs_reg(0x1)));
722 case ir_binop_gequal
:
723 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
724 inst
->conditional_mod
= BRW_CONDITIONAL_GE
;
725 emit(fs_inst(BRW_OPCODE_AND
, this->result
, this->result
, fs_reg(0x1)));
728 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
729 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
730 emit(fs_inst(BRW_OPCODE_AND
, this->result
, this->result
, fs_reg(0x1)));
732 case ir_binop_nequal
:
733 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
734 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
735 emit(fs_inst(BRW_OPCODE_AND
, this->result
, this->result
, fs_reg(0x1)));
738 case ir_binop_logic_xor
:
739 emit(fs_inst(BRW_OPCODE_XOR
, this->result
, op
[0], op
[1]));
742 case ir_binop_logic_or
:
743 emit(fs_inst(BRW_OPCODE_OR
, this->result
, op
[0], op
[1]));
746 case ir_binop_logic_and
:
747 emit(fs_inst(BRW_OPCODE_AND
, this->result
, op
[0], op
[1]));
753 assert(!"not reached: should be handled by brw_channel_expressions");
757 assert(!"not reached: should be handled by lower_noise");
761 emit(fs_inst(FS_OPCODE_SQRT
, this->result
, op
[0]));
765 emit(fs_inst(FS_OPCODE_RSQ
, this->result
, op
[0]));
771 emit(fs_inst(BRW_OPCODE_MOV
, this->result
, op
[0]));
774 emit(fs_inst(BRW_OPCODE_MOV
, this->result
, op
[0]));
778 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], fs_reg(0.0f
)));
779 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
782 emit(fs_inst(BRW_OPCODE_RNDD
, this->result
, op
[0]));
785 op
[0].negate
= ~op
[0].negate
;
786 inst
= emit(fs_inst(BRW_OPCODE_RNDD
, this->result
, op
[0]));
787 this->result
.negate
= true;
790 inst
= emit(fs_inst(BRW_OPCODE_RNDD
, this->result
, op
[0]));
793 inst
= emit(fs_inst(BRW_OPCODE_FRC
, this->result
, op
[0]));
797 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
798 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
800 inst
= emit(fs_inst(BRW_OPCODE_SEL
, this->result
, op
[0], op
[1]));
801 inst
->predicated
= true;
804 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, op
[0], op
[1]));
805 inst
->conditional_mod
= BRW_CONDITIONAL_G
;
807 inst
= emit(fs_inst(BRW_OPCODE_SEL
, this->result
, op
[0], op
[1]));
808 inst
->predicated
= true;
812 inst
= emit(fs_inst(FS_OPCODE_POW
, this->result
, op
[0], op
[1]));
815 case ir_unop_bit_not
:
817 case ir_binop_lshift
:
818 case ir_binop_rshift
:
819 case ir_binop_bit_and
:
820 case ir_binop_bit_xor
:
821 case ir_binop_bit_or
:
822 assert(!"GLSL 1.30 features unsupported");
828 fs_visitor::visit(ir_assignment
*ir
)
835 /* FINISHME: arrays on the lhs */
836 ir
->lhs
->accept(this);
839 ir
->rhs
->accept(this);
842 /* FINISHME: This should really set to the correct maximal writemask for each
843 * FINISHME: component written (in the loops below). This case can only
844 * FINISHME: occur for matrices, arrays, and structures.
846 if (ir
->write_mask
== 0) {
847 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
848 write_mask
= WRITEMASK_XYZW
;
850 assert(ir
->lhs
->type
->is_vector() || ir
->lhs
->type
->is_scalar());
851 write_mask
= ir
->write_mask
;
854 assert(l
.file
!= BAD_FILE
);
855 assert(r
.file
!= BAD_FILE
);
858 /* Get the condition bool into the predicate. */
859 ir
->condition
->accept(this);
860 inst
= emit(fs_inst(BRW_OPCODE_CMP
, this->result
, fs_reg(0)));
861 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
864 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
865 if (i
>= 4 || (write_mask
& (1 << i
))) {
866 inst
= emit(fs_inst(BRW_OPCODE_MOV
, l
, r
));
868 inst
->predicated
= true;
876 fs_visitor::visit(ir_texture
*ir
)
879 fs_inst
*inst
= NULL
;
880 unsigned int mlen
= 0;
882 ir
->coordinate
->accept(this);
883 fs_reg coordinate
= this->result
;
886 fs_reg inv_proj
= fs_reg(this, glsl_type::float_type
);
888 ir
->projector
->accept(this);
889 emit(fs_inst(FS_OPCODE_RCP
, inv_proj
, this->result
));
891 fs_reg proj_coordinate
= fs_reg(this, ir
->coordinate
->type
);
892 for (unsigned int i
= 0; i
< ir
->coordinate
->type
->vector_elements
; i
++) {
893 emit(fs_inst(BRW_OPCODE_MUL
, proj_coordinate
, coordinate
, inv_proj
));
894 coordinate
.reg_offset
++;
895 proj_coordinate
.reg_offset
++;
897 proj_coordinate
.reg_offset
= 0;
899 coordinate
= proj_coordinate
;
902 for (mlen
= 0; mlen
< ir
->coordinate
->type
->vector_elements
; mlen
++) {
903 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
), coordinate
));
904 coordinate
.reg_offset
++;
907 /* Pre-Ironlake, the 8-wide sampler always took u,v,r. */
911 if (ir
->shadow_comparitor
) {
912 /* For shadow comparisons, we have to supply u,v,r. */
915 ir
->shadow_comparitor
->accept(this);
916 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
), this->result
));
920 /* Do we ever want to handle writemasking on texture samples? Is it
921 * performance relevant?
923 fs_reg dst
= fs_reg(this, glsl_type::vec4_type
);
927 inst
= emit(fs_inst(FS_OPCODE_TEX
, dst
, fs_reg(MRF
, base_mrf
)));
930 ir
->lod_info
.bias
->accept(this);
931 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
), this->result
));
934 inst
= emit(fs_inst(FS_OPCODE_TXB
, dst
, fs_reg(MRF
, base_mrf
)));
937 ir
->lod_info
.lod
->accept(this);
938 emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ mlen
), this->result
));
941 inst
= emit(fs_inst(FS_OPCODE_TXL
, dst
, fs_reg(MRF
, base_mrf
)));
945 assert(!"GLSL 1.30 features unsupported");
951 if (ir
->shadow_comparitor
)
952 inst
->shadow_compare
= true;
957 fs_visitor::visit(ir_swizzle
*ir
)
959 ir
->val
->accept(this);
960 fs_reg val
= this->result
;
962 fs_reg result
= fs_reg(this, ir
->type
);
963 this->result
= result
;
965 for (unsigned int i
= 0; i
< ir
->type
->vector_elements
; i
++) {
966 fs_reg channel
= val
;
984 channel
.reg_offset
+= swiz
;
985 emit(fs_inst(BRW_OPCODE_MOV
, result
, channel
));
991 fs_visitor::visit(ir_discard
*ir
)
993 assert(ir
->condition
== NULL
); /* FINISHME */
995 emit(fs_inst(FS_OPCODE_DISCARD
));
999 fs_visitor::visit(ir_constant
*ir
)
1001 fs_reg
reg(this, ir
->type
);
1004 for (unsigned int i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1005 switch (ir
->type
->base_type
) {
1006 case GLSL_TYPE_FLOAT
:
1007 emit(fs_inst(BRW_OPCODE_MOV
, reg
, fs_reg(ir
->value
.f
[i
])));
1009 case GLSL_TYPE_UINT
:
1010 emit(fs_inst(BRW_OPCODE_MOV
, reg
, fs_reg(ir
->value
.u
[i
])));
1013 emit(fs_inst(BRW_OPCODE_MOV
, reg
, fs_reg(ir
->value
.i
[i
])));
1015 case GLSL_TYPE_BOOL
:
1016 emit(fs_inst(BRW_OPCODE_MOV
, reg
, fs_reg((int)ir
->value
.b
[i
])));
1019 assert(!"Non-float/uint/int/bool constant");
1026 fs_visitor::visit(ir_if
*ir
)
1030 /* Don't point the annotation at the if statement, because then it plus
1031 * the then and else blocks get printed.
1033 this->base_ir
= ir
->condition
;
1035 /* Generate the condition into the condition code. */
1036 ir
->condition
->accept(this);
1037 inst
= emit(fs_inst(BRW_OPCODE_MOV
, fs_reg(brw_null_reg()), this->result
));
1038 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
1040 inst
= emit(fs_inst(BRW_OPCODE_IF
));
1041 inst
->predicated
= true;
1043 foreach_iter(exec_list_iterator
, iter
, ir
->then_instructions
) {
1044 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1050 if (!ir
->else_instructions
.is_empty()) {
1051 emit(fs_inst(BRW_OPCODE_ELSE
));
1053 foreach_iter(exec_list_iterator
, iter
, ir
->else_instructions
) {
1054 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1061 emit(fs_inst(BRW_OPCODE_ENDIF
));
1065 fs_visitor::visit(ir_loop
*ir
)
1069 assert(!ir
->increment
);
1070 assert(!ir
->counter
);
1072 emit(fs_inst(BRW_OPCODE_DO
));
1074 /* Start a safety counter. If the user messed up their loop
1075 * counting, we don't want to hang the GPU.
1077 fs_reg max_iter
= fs_reg(this, glsl_type::int_type
);
1078 emit(fs_inst(BRW_OPCODE_MOV
, max_iter
, fs_reg(10000)));
1080 foreach_iter(exec_list_iterator
, iter
, ir
->body_instructions
) {
1081 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1087 /* Check the maximum loop iters counter. */
1088 inst
= emit(fs_inst(BRW_OPCODE_ADD
, max_iter
, max_iter
, fs_reg(-1)));
1089 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
1091 inst
= emit(fs_inst(BRW_OPCODE_BREAK
));
1092 inst
->predicated
= true;
1095 emit(fs_inst(BRW_OPCODE_WHILE
));
1099 fs_visitor::visit(ir_loop_jump
*ir
)
1102 case ir_loop_jump::jump_break
:
1103 emit(fs_inst(BRW_OPCODE_BREAK
));
1105 case ir_loop_jump::jump_continue
:
1106 emit(fs_inst(BRW_OPCODE_CONTINUE
));
1112 fs_visitor::visit(ir_call
*ir
)
1114 assert(!"FINISHME");
1118 fs_visitor::visit(ir_return
*ir
)
1120 assert(!"FINISHME");
1124 fs_visitor::visit(ir_function
*ir
)
1126 /* Ignore function bodies other than main() -- we shouldn't see calls to
1127 * them since they should all be inlined before we get to ir_to_mesa.
1129 if (strcmp(ir
->name
, "main") == 0) {
1130 const ir_function_signature
*sig
;
1133 sig
= ir
->matching_signature(&empty
);
1137 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
1138 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1147 fs_visitor::visit(ir_function_signature
*ir
)
1149 assert(!"not reached");
1154 fs_visitor::emit(fs_inst inst
)
1156 fs_inst
*list_inst
= new(mem_ctx
) fs_inst
;
1159 list_inst
->annotation
= this->current_annotation
;
1160 list_inst
->ir
= this->base_ir
;
1162 this->instructions
.push_tail(list_inst
);
1167 /** Emits a dummy fragment shader consisting of magenta for bringup purposes. */
1169 fs_visitor::emit_dummy_fs()
1171 /* Everyone's favorite color. */
1172 emit(fs_inst(BRW_OPCODE_MOV
,
1175 emit(fs_inst(BRW_OPCODE_MOV
,
1178 emit(fs_inst(BRW_OPCODE_MOV
,
1181 emit(fs_inst(BRW_OPCODE_MOV
,
1186 write
= emit(fs_inst(FS_OPCODE_FB_WRITE
,
1191 /* The register location here is relative to the start of the URB
1192 * data. It will get adjusted to be a real location before
1193 * generate_code() time.
1196 fs_visitor::interp_reg(int location
, int channel
)
1198 int regnr
= location
* 2 + channel
/ 2;
1199 int stride
= (channel
& 1) * 4;
1201 return brw_vec1_grf(regnr
, stride
);
1204 /** Emits the interpolation for the varying inputs. */
1206 fs_visitor::emit_interpolation()
1208 struct brw_reg g1_uw
= retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW
);
1209 /* For now, the source regs for the setup URB data will be unset,
1210 * since we don't know until codegen how many push constants we'll
1211 * use, and therefore what the setup URB offset is.
1213 fs_reg src_reg
= reg_undef
;
1215 this->current_annotation
= "compute pixel centers";
1216 this->pixel_x
= fs_reg(this, glsl_type::uint_type
);
1217 this->pixel_y
= fs_reg(this, glsl_type::uint_type
);
1218 this->pixel_x
.type
= BRW_REGISTER_TYPE_UW
;
1219 this->pixel_y
.type
= BRW_REGISTER_TYPE_UW
;
1220 emit(fs_inst(BRW_OPCODE_ADD
,
1222 fs_reg(stride(suboffset(g1_uw
, 4), 2, 4, 0)),
1223 fs_reg(brw_imm_v(0x10101010))));
1224 emit(fs_inst(BRW_OPCODE_ADD
,
1226 fs_reg(stride(suboffset(g1_uw
, 5), 2, 4, 0)),
1227 fs_reg(brw_imm_v(0x11001100))));
1229 this->current_annotation
= "compute pixel deltas from v0";
1230 this->delta_x
= fs_reg(this, glsl_type::float_type
);
1231 this->delta_y
= fs_reg(this, glsl_type::float_type
);
1232 emit(fs_inst(BRW_OPCODE_ADD
,
1235 fs_reg(negate(brw_vec1_grf(1, 0)))));
1236 emit(fs_inst(BRW_OPCODE_ADD
,
1239 fs_reg(brw_vec1_grf(1, 1))));
1241 this->current_annotation
= "compute pos.w and 1/pos.w";
1242 /* Compute wpos. Unlike many other varying inputs, we usually need it
1243 * to produce 1/w, and the varying variable wouldn't show up.
1245 fs_reg wpos
= fs_reg(this, glsl_type::vec4_type
);
1246 this->interp_attrs
[FRAG_ATTRIB_WPOS
] = wpos
;
1247 emit(fs_inst(BRW_OPCODE_MOV
, wpos
, this->pixel_x
)); /* FINISHME: ARB_fcc */
1249 emit(fs_inst(BRW_OPCODE_MOV
, wpos
, this->pixel_y
)); /* FINISHME: ARB_fcc */
1251 emit(fs_inst(FS_OPCODE_LINTERP
, wpos
, this->delta_x
, this->delta_y
,
1252 interp_reg(FRAG_ATTRIB_WPOS
, 2)));
1254 emit(fs_inst(FS_OPCODE_LINTERP
, wpos
, this->delta_x
, this->delta_y
,
1255 interp_reg(FRAG_ATTRIB_WPOS
, 3)));
1256 /* Compute the pixel W value from wpos.w. */
1257 this->pixel_w
= fs_reg(this, glsl_type::float_type
);
1258 emit(fs_inst(FS_OPCODE_RCP
, this->pixel_w
, wpos
));
1260 /* FINISHME: gl_FrontFacing */
1262 foreach_iter(exec_list_iterator
, iter
, *this->shader
->ir
) {
1263 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1264 ir_variable
*var
= ir
->as_variable();
1269 if (var
->mode
!= ir_var_in
)
1272 /* If it's already set up (WPOS), skip. */
1273 if (var
->location
== 0)
1276 this->current_annotation
= talloc_asprintf(this->mem_ctx
,
1278 "(FRAG_ATTRIB[%d])",
1281 emit_pinterp(var
->location
);
1283 this->current_annotation
= NULL
;
1287 fs_visitor::emit_pinterp(int location
)
1289 fs_reg interp_attr
= fs_reg(this, glsl_type::vec4_type
);
1290 this->interp_attrs
[location
] = interp_attr
;
1292 for (unsigned int i
= 0; i
< 4; i
++) {
1293 struct brw_reg interp
= interp_reg(location
, i
);
1294 emit(fs_inst(FS_OPCODE_LINTERP
,
1299 interp_attr
.reg_offset
++;
1301 interp_attr
.reg_offset
-= 4;
1303 for (unsigned int i
= 0; i
< 4; i
++) {
1304 emit(fs_inst(BRW_OPCODE_MUL
,
1308 interp_attr
.reg_offset
++;
1313 fs_visitor::emit_fb_writes()
1315 this->current_annotation
= "FB write";
1317 assert(this->frag_color
|| !"FINISHME: MRT");
1318 fs_reg color
= *(variable_storage(this->frag_color
));
1320 for (int i
= 0; i
< 4; i
++) {
1321 emit(fs_inst(BRW_OPCODE_MOV
,
1327 emit(fs_inst(FS_OPCODE_FB_WRITE
,
1331 this->current_annotation
= NULL
;
1335 fs_visitor::generate_fb_write(fs_inst
*inst
)
1337 GLboolean eot
= 1; /* FINISHME: MRT */
1338 /* FINISHME: AADS */
1340 /* Header is 2 regs, g0 and g1 are the contents. g0 will be implied
1343 brw_push_insn_state(p
);
1344 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
1345 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
1348 brw_vec8_grf(1, 0));
1349 brw_pop_insn_state(p
);
1354 8, /* dispatch_width */
1355 retype(vec8(brw_null_reg()), BRW_REGISTER_TYPE_UW
),
1357 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW
),
1358 0, /* FINISHME: MRT target */
1365 fs_visitor::generate_linterp(fs_inst
*inst
,
1366 struct brw_reg dst
, struct brw_reg
*src
)
1368 struct brw_reg delta_x
= src
[0];
1369 struct brw_reg delta_y
= src
[1];
1370 struct brw_reg interp
= src
[2];
1373 delta_y
.nr
== delta_x
.nr
+ 1 &&
1374 (intel
->gen
>= 6 || (delta_x
.nr
& 1) == 0)) {
1375 brw_PLN(p
, dst
, interp
, delta_x
);
1377 brw_LINE(p
, brw_null_reg(), interp
, delta_x
);
1378 brw_MAC(p
, dst
, suboffset(interp
, 1), delta_y
);
1383 fs_visitor::generate_math(fs_inst
*inst
,
1384 struct brw_reg dst
, struct brw_reg
*src
)
1388 switch (inst
->opcode
) {
1390 op
= BRW_MATH_FUNCTION_INV
;
1393 op
= BRW_MATH_FUNCTION_RSQ
;
1395 case FS_OPCODE_SQRT
:
1396 op
= BRW_MATH_FUNCTION_SQRT
;
1398 case FS_OPCODE_EXP2
:
1399 op
= BRW_MATH_FUNCTION_EXP
;
1401 case FS_OPCODE_LOG2
:
1402 op
= BRW_MATH_FUNCTION_LOG
;
1405 op
= BRW_MATH_FUNCTION_POW
;
1408 op
= BRW_MATH_FUNCTION_SIN
;
1411 op
= BRW_MATH_FUNCTION_COS
;
1414 assert(!"not reached: unknown math function");
1419 if (inst
->opcode
== FS_OPCODE_POW
) {
1420 brw_MOV(p
, brw_message_reg(3), src
[1]);
1425 inst
->saturate
? BRW_MATH_SATURATE_SATURATE
:
1426 BRW_MATH_SATURATE_NONE
,
1428 BRW_MATH_DATA_VECTOR
,
1429 BRW_MATH_PRECISION_FULL
);
1433 fs_visitor::generate_tex(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
)
1438 if (intel
->gen
== 5) {
1439 switch (inst
->opcode
) {
1441 if (inst
->shadow_compare
) {
1442 msg_type
= BRW_SAMPLER_MESSAGE_SAMPLE_COMPARE_GEN5
;
1444 msg_type
= BRW_SAMPLER_MESSAGE_SAMPLE_GEN5
;
1448 if (inst
->shadow_compare
) {
1449 msg_type
= BRW_SAMPLER_MESSAGE_SAMPLE_BIAS_COMPARE_GEN5
;
1451 msg_type
= BRW_SAMPLER_MESSAGE_SAMPLE_BIAS_GEN5
;
1456 switch (inst
->opcode
) {
1458 /* Note that G45 and older determines shadow compare and dispatch width
1459 * from message length for most messages.
1461 if (inst
->shadow_compare
) {
1462 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_COMPARE
;
1464 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE
;
1467 if (inst
->shadow_compare
) {
1468 assert(!"FINISHME: shadow compare with bias.");
1469 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_BIAS
;
1471 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_BIAS
;
1477 assert(msg_type
!= -1);
1483 retype(dst
, BRW_REGISTER_TYPE_UW
),
1485 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW
),
1486 SURF_INDEX_TEXTURE(inst
->sampler
),
1494 BRW_SAMPLER_SIMD_MODE_SIMD8
);
1498 fs_visitor::generate_discard(fs_inst
*inst
)
1500 struct brw_reg g0
= retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW
);
1501 brw_push_insn_state(p
);
1502 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
1503 brw_NOT(p
, c
->emit_mask_reg
, brw_mask_reg(1)); /* IMASK */
1504 brw_AND(p
, g0
, c
->emit_mask_reg
, g0
);
1505 brw_pop_insn_state(p
);
1509 trivial_assign_reg(int header_size
, fs_reg
*reg
)
1511 if (reg
->file
== GRF
&& reg
->reg
!= 0) {
1512 reg
->hw_reg
= header_size
+ reg
->reg
- 1 + reg
->reg_offset
;
1518 fs_visitor::assign_curb_setup()
1520 c
->prog_data
.first_curbe_grf
= c
->key
.nr_payload_regs
;
1521 c
->prog_data
.curb_read_length
= ALIGN(c
->prog_data
.nr_params
, 8) / 8;
1523 if (intel
->gen
== 5 && (c
->prog_data
.first_curbe_grf
+
1524 c
->prog_data
.curb_read_length
) & 1) {
1525 /* Align the start of the interpolation coefficients so that we can use
1526 * the PLN instruction.
1528 c
->prog_data
.first_curbe_grf
++;
1531 /* Map the offsets in the UNIFORM file to fixed HW regs. */
1532 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
1533 fs_inst
*inst
= (fs_inst
*)iter
.get();
1535 for (unsigned int i
= 0; i
< 3; i
++) {
1536 if (inst
->src
[i
].file
== UNIFORM
) {
1537 int constant_nr
= inst
->src
[i
].hw_reg
+ inst
->src
[i
].reg_offset
;
1538 struct brw_reg brw_reg
= brw_vec1_grf(c
->prog_data
.first_curbe_grf
+
1542 inst
->src
[i
].file
= FIXED_HW_REG
;
1543 inst
->src
[i
].fixed_hw_reg
= brw_reg
;
1550 fs_visitor::assign_urb_setup()
1552 int urb_start
= c
->prog_data
.first_curbe_grf
+ c
->prog_data
.curb_read_length
;
1553 int interp_reg_nr
[FRAG_ATTRIB_MAX
];
1555 c
->prog_data
.urb_read_length
= 0;
1557 /* Figure out where each of the incoming setup attributes lands. */
1558 for (unsigned int i
= 0; i
< FRAG_ATTRIB_MAX
; i
++) {
1559 interp_reg_nr
[i
] = -1;
1561 if (i
!= FRAG_ATTRIB_WPOS
&&
1562 !(brw
->fragment_program
->Base
.InputsRead
& BITFIELD64_BIT(i
)))
1565 /* Each attribute is 4 setup channels, each of which is half a reg. */
1566 interp_reg_nr
[i
] = urb_start
+ c
->prog_data
.urb_read_length
;
1567 c
->prog_data
.urb_read_length
+= 2;
1570 /* Map the register numbers for FS_OPCODE_LINTERP so that it uses
1571 * the correct setup input.
1573 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
1574 fs_inst
*inst
= (fs_inst
*)iter
.get();
1576 if (inst
->opcode
!= FS_OPCODE_LINTERP
)
1579 assert(inst
->src
[2].file
== FIXED_HW_REG
);
1581 int location
= inst
->src
[2].fixed_hw_reg
.nr
/ 2;
1582 assert(interp_reg_nr
[location
] != -1);
1583 inst
->src
[2].fixed_hw_reg
.nr
= (interp_reg_nr
[location
] +
1584 (inst
->src
[2].fixed_hw_reg
.nr
& 1));
1587 this->first_non_payload_grf
= urb_start
+ c
->prog_data
.urb_read_length
;
1591 fs_visitor::assign_regs()
1593 int header_size
= this->first_non_payload_grf
;
1596 /* FINISHME: trivial assignment of register numbers */
1597 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
1598 fs_inst
*inst
= (fs_inst
*)iter
.get();
1600 trivial_assign_reg(header_size
, &inst
->dst
);
1601 trivial_assign_reg(header_size
, &inst
->src
[0]);
1602 trivial_assign_reg(header_size
, &inst
->src
[1]);
1604 last_grf
= MAX2(last_grf
, inst
->dst
.hw_reg
);
1605 last_grf
= MAX2(last_grf
, inst
->src
[0].hw_reg
);
1606 last_grf
= MAX2(last_grf
, inst
->src
[1].hw_reg
);
1609 this->grf_used
= last_grf
+ 1;
1612 static struct brw_reg
brw_reg_from_fs_reg(fs_reg
*reg
)
1614 struct brw_reg brw_reg
;
1616 switch (reg
->file
) {
1620 brw_reg
= brw_vec8_reg(reg
->file
,
1622 brw_reg
= retype(brw_reg
, reg
->type
);
1625 switch (reg
->type
) {
1626 case BRW_REGISTER_TYPE_F
:
1627 brw_reg
= brw_imm_f(reg
->imm
.f
);
1629 case BRW_REGISTER_TYPE_D
:
1630 brw_reg
= brw_imm_d(reg
->imm
.i
);
1632 case BRW_REGISTER_TYPE_UD
:
1633 brw_reg
= brw_imm_ud(reg
->imm
.u
);
1636 assert(!"not reached");
1641 brw_reg
= reg
->fixed_hw_reg
;
1644 /* Probably unused. */
1645 brw_reg
= brw_null_reg();
1648 assert(!"not reached");
1649 brw_reg
= brw_null_reg();
1653 brw_reg
= brw_abs(brw_reg
);
1655 brw_reg
= negate(brw_reg
);
1661 fs_visitor::generate_code()
1663 unsigned int annotation_len
= 0;
1664 int last_native_inst
= 0;
1665 struct brw_instruction
*if_stack
[16], *loop_stack
[16];
1666 int if_stack_depth
= 0, loop_stack_depth
= 0;
1667 int if_depth_in_loop
[16];
1669 if_depth_in_loop
[loop_stack_depth
] = 0;
1671 memset(&if_stack
, 0, sizeof(if_stack
));
1672 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
1673 fs_inst
*inst
= (fs_inst
*)iter
.get();
1674 struct brw_reg src
[3], dst
;
1676 for (unsigned int i
= 0; i
< 3; i
++) {
1677 src
[i
] = brw_reg_from_fs_reg(&inst
->src
[i
]);
1679 dst
= brw_reg_from_fs_reg(&inst
->dst
);
1681 brw_set_conditionalmod(p
, inst
->conditional_mod
);
1682 brw_set_predicate_control(p
, inst
->predicated
);
1684 switch (inst
->opcode
) {
1685 case BRW_OPCODE_MOV
:
1686 brw_MOV(p
, dst
, src
[0]);
1688 case BRW_OPCODE_ADD
:
1689 brw_ADD(p
, dst
, src
[0], src
[1]);
1691 case BRW_OPCODE_MUL
:
1692 brw_MUL(p
, dst
, src
[0], src
[1]);
1695 case BRW_OPCODE_FRC
:
1696 brw_FRC(p
, dst
, src
[0]);
1698 case BRW_OPCODE_RNDD
:
1699 brw_RNDD(p
, dst
, src
[0]);
1701 case BRW_OPCODE_RNDZ
:
1702 brw_RNDZ(p
, dst
, src
[0]);
1705 case BRW_OPCODE_AND
:
1706 brw_AND(p
, dst
, src
[0], src
[1]);
1709 brw_OR(p
, dst
, src
[0], src
[1]);
1711 case BRW_OPCODE_XOR
:
1712 brw_XOR(p
, dst
, src
[0], src
[1]);
1715 case BRW_OPCODE_CMP
:
1716 brw_CMP(p
, dst
, inst
->conditional_mod
, src
[0], src
[1]);
1718 case BRW_OPCODE_SEL
:
1719 brw_SEL(p
, dst
, src
[0], src
[1]);
1723 assert(if_stack_depth
< 16);
1724 if_stack
[if_stack_depth
] = brw_IF(p
, BRW_EXECUTE_8
);
1727 case BRW_OPCODE_ELSE
:
1728 if_stack
[if_stack_depth
- 1] =
1729 brw_ELSE(p
, if_stack
[if_stack_depth
- 1]);
1731 case BRW_OPCODE_ENDIF
:
1733 brw_ENDIF(p
, if_stack
[if_stack_depth
]);
1737 loop_stack
[loop_stack_depth
++] = brw_DO(p
, BRW_EXECUTE_8
);
1738 if_depth_in_loop
[loop_stack_depth
] = 0;
1741 case BRW_OPCODE_BREAK
:
1742 brw_BREAK(p
, if_depth_in_loop
[loop_stack_depth
]);
1743 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
1745 case BRW_OPCODE_CONTINUE
:
1746 brw_CONT(p
, if_depth_in_loop
[loop_stack_depth
]);
1747 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
1750 case BRW_OPCODE_WHILE
: {
1751 struct brw_instruction
*inst0
, *inst1
;
1754 if (intel
->gen
== 5)
1757 assert(loop_stack_depth
> 0);
1759 inst0
= inst1
= brw_WHILE(p
, loop_stack
[loop_stack_depth
]);
1760 /* patch all the BREAK/CONT instructions from last BGNLOOP */
1761 while (inst0
> loop_stack
[loop_stack_depth
]) {
1763 if (inst0
->header
.opcode
== BRW_OPCODE_BREAK
&&
1764 inst0
->bits3
.if_else
.jump_count
== 0) {
1765 inst0
->bits3
.if_else
.jump_count
= br
* (inst1
- inst0
+ 1);
1767 else if (inst0
->header
.opcode
== BRW_OPCODE_CONTINUE
&&
1768 inst0
->bits3
.if_else
.jump_count
== 0) {
1769 inst0
->bits3
.if_else
.jump_count
= br
* (inst1
- inst0
);
1777 case FS_OPCODE_SQRT
:
1778 case FS_OPCODE_EXP2
:
1779 case FS_OPCODE_LOG2
:
1783 generate_math(inst
, dst
, src
);
1785 case FS_OPCODE_LINTERP
:
1786 generate_linterp(inst
, dst
, src
);
1791 generate_tex(inst
, dst
, src
[0]);
1793 case FS_OPCODE_DISCARD
:
1794 generate_discard(inst
);
1796 case FS_OPCODE_FB_WRITE
:
1797 generate_fb_write(inst
);
1800 if (inst
->opcode
< (int)ARRAY_SIZE(brw_opcodes
)) {
1801 _mesa_problem(ctx
, "Unsupported opcode `%s' in FS",
1802 brw_opcodes
[inst
->opcode
].name
);
1804 _mesa_problem(ctx
, "Unsupported opcode %d in FS", inst
->opcode
);
1809 if (annotation_len
< p
->nr_insn
) {
1810 annotation_len
*= 2;
1811 if (annotation_len
< 16)
1812 annotation_len
= 16;
1814 this->annotation_string
= talloc_realloc(this->mem_ctx
,
1818 this->annotation_ir
= talloc_realloc(this->mem_ctx
,
1824 for (unsigned int i
= last_native_inst
; i
< p
->nr_insn
; i
++) {
1825 this->annotation_string
[i
] = inst
->annotation
;
1826 this->annotation_ir
[i
] = inst
->ir
;
1828 last_native_inst
= p
->nr_insn
;
1833 brw_wm_fs_emit(struct brw_context
*brw
, struct brw_wm_compile
*c
)
1835 struct brw_compile
*p
= &c
->func
;
1836 struct intel_context
*intel
= &brw
->intel
;
1837 GLcontext
*ctx
= &intel
->ctx
;
1838 struct brw_shader
*shader
= NULL
;
1839 struct gl_shader_program
*prog
= ctx
->Shader
.CurrentProgram
;
1847 for (unsigned int i
= 0; i
< prog
->_NumLinkedShaders
; i
++) {
1848 if (prog
->_LinkedShaders
[i
]->Type
== GL_FRAGMENT_SHADER
) {
1849 shader
= (struct brw_shader
*)prog
->_LinkedShaders
[i
];
1856 /* We always use 8-wide mode, at least for now. For one, flow
1857 * control only works in 8-wide. Also, when we're fragment shader
1858 * bound, we're almost always under register pressure as well, so
1859 * 8-wide would save us from the performance cliff of spilling
1862 c
->dispatch_width
= 8;
1864 if (INTEL_DEBUG
& DEBUG_WM
) {
1865 printf("GLSL IR for native fragment shader %d:\n", prog
->Name
);
1866 _mesa_print_ir(shader
->ir
, NULL
);
1870 /* Now the main event: Visit the shader IR and generate our FS IR for it.
1872 fs_visitor
v(c
, shader
);
1877 v
.emit_interpolation();
1879 /* Generate FS IR for main(). (the visitor only descends into
1880 * functions called "main").
1882 foreach_iter(exec_list_iterator
, iter
, *shader
->ir
) {
1883 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1889 v
.assign_curb_setup();
1890 v
.assign_urb_setup();
1896 assert(!v
.fail
); /* FINISHME: Cleanly fail, tested at link time, etc. */
1901 if (INTEL_DEBUG
& DEBUG_WM
) {
1902 const char *last_annotation_string
= NULL
;
1903 ir_instruction
*last_annotation_ir
= NULL
;
1905 printf("Native code for fragment shader %d:\n", prog
->Name
);
1906 for (unsigned int i
= 0; i
< p
->nr_insn
; i
++) {
1907 if (last_annotation_ir
!= v
.annotation_ir
[i
]) {
1908 last_annotation_ir
= v
.annotation_ir
[i
];
1909 if (last_annotation_ir
) {
1911 last_annotation_ir
->print();
1915 if (last_annotation_string
!= v
.annotation_string
[i
]) {
1916 last_annotation_string
= v
.annotation_string
[i
];
1917 if (last_annotation_string
)
1918 printf(" %s\n", last_annotation_string
);
1920 brw_disasm(stdout
, &p
->store
[i
], intel
->gen
);
1925 c
->prog_data
.total_grf
= v
.grf_used
;
1926 c
->prog_data
.total_scratch
= 0;