2 * Copyright (C) 2008 Nicolai Haehnle.
6 * Permission is hereby granted, free of charge, to any person obtaining
7 * a copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sublicense, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial
16 * portions of the Software.
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
19 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
21 * IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
22 * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
23 * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
24 * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
31 * Perform temporary register allocation and attempt to pair off instructions
32 * in RGB and Alpha pairs. Also attempts to optimize the TEX instruction
33 * vs. ALU instruction scheduling.
36 #include "radeon_program_pair.h"
38 #include "radeon_common.h"
40 #include "shader/prog_print.h"
42 #define error(fmt, args...) do { \
43 _mesa_problem(s->Ctx, "%s::%s(): " fmt "\n", \
44 __FILE__, __FUNCTION__, ##args); \
48 struct pair_state_instruction
{
49 GLuint IsTex
:1; /**< Is a texture instruction */
50 GLuint NeedRGB
:1; /**< Needs the RGB ALU */
51 GLuint NeedAlpha
:1; /**< Needs the Alpha ALU */
52 GLuint IsTranscendent
:1; /**< Is a special transcendent instruction */
55 * Number of (read and write) dependencies that must be resolved before
56 * this instruction can be scheduled.
58 GLuint NumDependencies
:5;
61 * Next instruction in the linked list of ready instructions.
63 struct pair_state_instruction
*NextReady
;
66 * Values that this instruction writes
68 struct reg_value
*Values
[4];
73 * Used to keep track of which instructions read a value.
75 struct reg_value_reader
{
76 GLuint IP
; /**< IP of the instruction that performs this access */
77 struct reg_value_reader
*Next
;
81 * Used to keep track which values are stored in each component of a
85 GLuint IP
; /**< IP of the instruction that writes this value */
86 struct reg_value
*Next
; /**< Pointer to the next value to be written to the same PROGRAM_TEMPORARY component */
89 * Unordered linked list of instructions that read from this value.
91 struct reg_value_reader
*Readers
;
94 * Number of readers of this value. This is calculated during @ref scan_instructions
95 * and continually decremented during code emission.
96 * When this count reaches zero, the instruction that writes the @ref Next value
103 * Used to translate a PROGRAM_INPUT or PROGRAM_TEMPORARY Mesa register
104 * to the proper hardware temporary.
106 struct pair_register_translation
{
109 GLuint RefCount
:23; /**< # of times this occurs in an unscheduled instruction SrcReg or DstReg */
112 * Notes the value that is currently contained in each component
113 * (only used for PROGRAM_TEMPORARY registers).
115 struct reg_value
*Value
[4];
120 struct gl_program
*Program
;
121 const struct radeon_pair_handler
*Handler
;
128 * Translate Mesa registers to hardware registers
130 struct pair_register_translation Inputs
[FRAG_ATTRIB_MAX
];
131 struct pair_register_translation Temps
[MAX_PROGRAM_TEMPS
];
134 * Derived information about program instructions.
136 struct pair_state_instruction
*Instructions
;
139 GLuint RefCount
; /**< # of times this occurs in an unscheduled SrcReg or DstReg */
143 * Linked list of instructions that can be scheduled right now,
144 * based on which ALU/TEX resources they require.
146 struct pair_state_instruction
*ReadyFullALU
;
147 struct pair_state_instruction
*ReadyRGB
;
148 struct pair_state_instruction
*ReadyAlpha
;
149 struct pair_state_instruction
*ReadyTEX
;
152 * Pool of @ref reg_value structures for fast allocation.
154 struct reg_value
*ValuePool
;
155 GLuint ValuePoolUsed
;
156 struct reg_value_reader
*ReaderPool
;
157 GLuint ReaderPoolUsed
;
161 static struct pair_register_translation
*get_register(struct pair_state
*s
, GLuint file
, GLuint index
)
164 case PROGRAM_TEMPORARY
: return &s
->Temps
[index
];
165 case PROGRAM_INPUT
: return &s
->Inputs
[index
];
170 static void alloc_hw_reg(struct pair_state
*s
, GLuint file
, GLuint index
, GLuint hwindex
)
172 struct pair_register_translation
*t
= get_register(s
, file
, index
);
173 ASSERT(!s
->HwTemps
[hwindex
].RefCount
);
174 ASSERT(!t
->Allocated
);
175 s
->HwTemps
[hwindex
].RefCount
= t
->RefCount
;
177 t
->HwIndex
= hwindex
;
180 static GLuint
get_hw_reg(struct pair_state
*s
, GLuint file
, GLuint index
)
184 struct pair_register_translation
*t
= get_register(s
, file
, index
);
186 _mesa_problem(s
->Ctx
, "get_hw_reg: %i[%i]\n", file
, index
);
193 for(hwindex
= 0; hwindex
< s
->Handler
->MaxHwTemps
; ++hwindex
)
194 if (!s
->HwTemps
[hwindex
].RefCount
)
197 if (hwindex
>= s
->Handler
->MaxHwTemps
) {
198 error("Ran out of hardware temporaries");
202 alloc_hw_reg(s
, file
, index
, hwindex
);
207 static void deref_hw_reg(struct pair_state
*s
, GLuint hwindex
)
209 if (!s
->HwTemps
[hwindex
].RefCount
) {
210 error("Hwindex %i refcount error", hwindex
);
214 s
->HwTemps
[hwindex
].RefCount
--;
217 static void add_pairinst_to_list(struct pair_state_instruction
**list
, struct pair_state_instruction
*pairinst
)
219 pairinst
->NextReady
= *list
;
224 * The instruction at the given IP has become ready. Link it into the ready
227 static void instruction_ready(struct pair_state
*s
, int ip
)
229 struct pair_state_instruction
*pairinst
= s
->Instructions
+ ip
;
232 _mesa_printf("instruction_ready(%i)\n", ip
);
235 add_pairinst_to_list(&s
->ReadyTEX
, pairinst
);
236 else if (!pairinst
->NeedAlpha
)
237 add_pairinst_to_list(&s
->ReadyRGB
, pairinst
);
238 else if (!pairinst
->NeedRGB
)
239 add_pairinst_to_list(&s
->ReadyAlpha
, pairinst
);
241 add_pairinst_to_list(&s
->ReadyFullALU
, pairinst
);
246 * Finally rewrite ADD, MOV, MUL as the appropriate native instruction
247 * and reverse the order of arguments for CMP.
249 static void final_rewrite(struct pair_state
*s
, struct prog_instruction
*inst
)
251 struct prog_src_register tmp
;
253 switch(inst
->Opcode
) {
255 inst
->SrcReg
[2] = inst
->SrcReg
[1];
256 inst
->SrcReg
[1].File
= PROGRAM_BUILTIN
;
257 inst
->SrcReg
[1].Swizzle
= SWIZZLE_1111
;
258 inst
->SrcReg
[1].NegateBase
= 0;
259 inst
->SrcReg
[1].NegateAbs
= 0;
260 inst
->Opcode
= OPCODE_MAD
;
263 tmp
= inst
->SrcReg
[2];
264 inst
->SrcReg
[2] = inst
->SrcReg
[0];
265 inst
->SrcReg
[0] = tmp
;
268 /* AMD say we should use CMP.
269 * However, when we transform
272 * CMP tmp, -r0, -r0, 0;
274 * we get incorrect behaviour on R500 when r0 == 0.0.
275 * It appears that the R500 KIL hardware treats -0.0 as less
278 inst
->SrcReg
[1].File
= PROGRAM_BUILTIN
;
279 inst
->SrcReg
[1].Swizzle
= SWIZZLE_1111
;
280 inst
->SrcReg
[2].File
= PROGRAM_BUILTIN
;
281 inst
->SrcReg
[2].Swizzle
= SWIZZLE_0000
;
282 inst
->Opcode
= OPCODE_MAD
;
285 inst
->SrcReg
[2].File
= PROGRAM_BUILTIN
;
286 inst
->SrcReg
[2].Swizzle
= SWIZZLE_0000
;
287 inst
->Opcode
= OPCODE_MAD
;
297 * Classify an instruction according to which ALUs etc. it needs
299 static void classify_instruction(struct pair_state
*s
,
300 struct prog_instruction
*inst
, struct pair_state_instruction
*pairinst
)
302 pairinst
->NeedRGB
= (inst
->DstReg
.WriteMask
& WRITEMASK_XYZ
) ? 1 : 0;
303 pairinst
->NeedAlpha
= (inst
->DstReg
.WriteMask
& WRITEMASK_W
) ? 1 : 0;
305 switch(inst
->Opcode
) {
323 pairinst
->IsTranscendent
= 1;
324 pairinst
->NeedAlpha
= 1;
327 pairinst
->NeedAlpha
= 1;
330 pairinst
->NeedRGB
= 1;
340 error("Unknown opcode %d\n", inst
->Opcode
);
347 * Count which (input, temporary) register is read and written how often,
348 * and scan the instruction stream to find dependencies.
350 static void scan_instructions(struct pair_state
*s
)
352 struct prog_instruction
*inst
;
353 struct pair_state_instruction
*pairinst
;
356 for(inst
= s
->Program
->Instructions
, pairinst
= s
->Instructions
, ip
= 0;
357 inst
->Opcode
!= OPCODE_END
;
358 ++inst
, ++pairinst
, ++ip
) {
359 final_rewrite(s
, inst
);
360 classify_instruction(s
, inst
, pairinst
);
362 int nsrc
= _mesa_num_inst_src_regs(inst
->Opcode
);
364 for(j
= 0; j
< nsrc
; j
++) {
365 struct pair_register_translation
*t
=
366 get_register(s
, inst
->SrcReg
[j
].File
, inst
->SrcReg
[j
].Index
);
372 if (inst
->SrcReg
[j
].File
== PROGRAM_TEMPORARY
) {
374 for(i
= 0; i
< 4; ++i
) {
375 GLuint swz
= GET_SWZ(inst
->SrcReg
[j
].Swizzle
, i
);
377 continue; /* constant or NIL swizzle */
379 continue; /* this is an undefined read */
381 /* Do not add a dependency if this instruction
382 * also rewrites the value. The code below adds
383 * a dependency for the DstReg, which is a superset
384 * of the SrcReg dependency. */
385 if (inst
->DstReg
.File
== PROGRAM_TEMPORARY
&&
386 inst
->DstReg
.Index
== inst
->SrcReg
[j
].Index
&&
387 GET_BIT(inst
->DstReg
.WriteMask
, swz
))
390 struct reg_value_reader
* r
= &s
->ReaderPool
[s
->ReaderPoolUsed
++];
391 pairinst
->NumDependencies
++;
392 t
->Value
[swz
]->NumReaders
++;
394 r
->Next
= t
->Value
[swz
]->Readers
;
395 t
->Value
[swz
]->Readers
= r
;
400 int ndst
= _mesa_num_inst_dst_regs(inst
->Opcode
);
402 struct pair_register_translation
*t
=
403 get_register(s
, inst
->DstReg
.File
, inst
->DstReg
.Index
);
407 if (inst
->DstReg
.File
== PROGRAM_TEMPORARY
) {
409 for(j
= 0; j
< 4; ++j
) {
410 if (!GET_BIT(inst
->DstReg
.WriteMask
, j
))
413 struct reg_value
* v
= &s
->ValuePool
[s
->ValuePoolUsed
++];
416 pairinst
->NumDependencies
++;
417 t
->Value
[j
]->Next
= v
;
420 pairinst
->Values
[j
] = v
;
427 _mesa_printf("scan(%i): NumDeps = %i\n", ip
, pairinst
->NumDependencies
);
429 if (!pairinst
->NumDependencies
)
430 instruction_ready(s
, ip
);
433 /* Clear the PROGRAM_TEMPORARY state */
435 for(i
= 0; i
< MAX_PROGRAM_TEMPS
; ++i
) {
436 for(j
= 0; j
< 4; ++j
)
437 s
->Temps
[i
].Value
[j
] = 0;
443 * Reserve hardware temporary registers for the program inputs.
445 * @note This allocation is performed explicitly, because the order of inputs
446 * is determined by the RS hardware.
448 static void allocate_input_registers(struct pair_state
*s
)
450 GLuint InputsRead
= s
->Program
->InputsRead
;
455 if (InputsRead
& FRAG_BIT_COL0
)
456 alloc_hw_reg(s
, PROGRAM_INPUT
, FRAG_ATTRIB_COL0
, hwindex
++);
457 InputsRead
&= ~FRAG_BIT_COL0
;
459 /* Secondary color */
460 if (InputsRead
& FRAG_BIT_COL1
)
461 alloc_hw_reg(s
, PROGRAM_INPUT
, FRAG_ATTRIB_COL1
, hwindex
++);
462 InputsRead
&= ~FRAG_BIT_COL1
;
465 for (i
= 0; i
< s
->Ctx
->Const
.MaxTextureUnits
; i
++) {
466 if (InputsRead
& (FRAG_BIT_TEX0
<< i
))
467 alloc_hw_reg(s
, PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
+i
, hwindex
++);
469 InputsRead
&= ~FRAG_BITS_TEX_ANY
;
471 /* Fogcoords treated as a texcoord */
472 if (InputsRead
& FRAG_BIT_FOGC
)
473 alloc_hw_reg(s
, PROGRAM_INPUT
, FRAG_ATTRIB_FOGC
, hwindex
++);
474 InputsRead
&= ~FRAG_BIT_FOGC
;
476 /* fragment position treated as a texcoord */
477 if (InputsRead
& FRAG_BIT_WPOS
)
478 alloc_hw_reg(s
, PROGRAM_INPUT
, FRAG_ATTRIB_WPOS
, hwindex
++);
479 InputsRead
&= ~FRAG_BIT_WPOS
;
483 error("Don't know how to handle inputs 0x%x\n", InputsRead
);
487 static void decrement_dependencies(struct pair_state
*s
, int ip
)
489 struct pair_state_instruction
*pairinst
= s
->Instructions
+ ip
;
490 ASSERT(pairinst
->NumDependencies
> 0);
491 if (!--pairinst
->NumDependencies
)
492 instruction_ready(s
, ip
);
496 * Update the dependency tracking state based on what the instruction
497 * at the given IP does.
499 static void commit_instruction(struct pair_state
*s
, int ip
)
501 struct prog_instruction
*inst
= s
->Program
->Instructions
+ ip
;
502 struct pair_state_instruction
*pairinst
= s
->Instructions
+ ip
;
505 _mesa_printf("commit_instruction(%i)\n", ip
);
507 if (inst
->DstReg
.File
== PROGRAM_TEMPORARY
) {
508 struct pair_register_translation
*t
= &s
->Temps
[inst
->DstReg
.Index
];
509 deref_hw_reg(s
, t
->HwIndex
);
512 for(i
= 0; i
< 4; ++i
) {
513 if (!GET_BIT(inst
->DstReg
.WriteMask
, i
))
516 t
->Value
[i
] = pairinst
->Values
[i
];
517 if (t
->Value
[i
]->NumReaders
) {
518 struct reg_value_reader
*r
;
519 for(r
= pairinst
->Values
[i
]->Readers
; r
; r
= r
->Next
)
520 decrement_dependencies(s
, r
->IP
);
521 } else if (t
->Value
[i
]->Next
) {
522 /* This happens when the only reader writes
523 * the register at the same time */
524 decrement_dependencies(s
, t
->Value
[i
]->Next
->IP
);
529 int nsrc
= _mesa_num_inst_src_regs(inst
->Opcode
);
531 for(i
= 0; i
< nsrc
; i
++) {
532 struct pair_register_translation
*t
= get_register(s
, inst
->SrcReg
[i
].File
, inst
->SrcReg
[i
].Index
);
536 deref_hw_reg(s
, get_hw_reg(s
, inst
->SrcReg
[i
].File
, inst
->SrcReg
[i
].Index
));
538 if (inst
->SrcReg
[i
].File
!= PROGRAM_TEMPORARY
)
542 for(j
= 0; j
< 4; ++j
) {
543 GLuint swz
= GET_SWZ(inst
->SrcReg
[i
].Swizzle
, j
);
549 /* Do not free a dependency if this instruction
550 * also rewrites the value. See scan_instructions. */
551 if (inst
->DstReg
.File
== PROGRAM_TEMPORARY
&&
552 inst
->DstReg
.Index
== inst
->SrcReg
[i
].Index
&&
553 GET_BIT(inst
->DstReg
.WriteMask
, swz
))
556 if (!--t
->Value
[swz
]->NumReaders
) {
557 if (t
->Value
[swz
]->Next
)
558 decrement_dependencies(s
, t
->Value
[swz
]->Next
->IP
);
566 * Emit all ready texture instructions in a single block.
568 * Emit as a single block to (hopefully) sample many textures in parallel,
569 * and to avoid hardware indirections on R300.
571 * In R500, we don't really know when the result of a texture instruction
572 * arrives. So allocate all destinations first, to make sure they do not
573 * arrive early and overwrite a texture coordinate we're going to use later
576 static void emit_all_tex(struct pair_state
*s
)
578 struct pair_state_instruction
*readytex
;
579 struct pair_state_instruction
*pairinst
;
583 // Don't let the ready list change under us!
584 readytex
= s
->ReadyTEX
;
587 // Allocate destination hardware registers in one block to avoid conflicts.
588 for(pairinst
= readytex
; pairinst
; pairinst
= pairinst
->NextReady
) {
589 int ip
= pairinst
- s
->Instructions
;
590 struct prog_instruction
*inst
= s
->Program
->Instructions
+ ip
;
591 if (inst
->Opcode
!= OPCODE_KIL
)
592 get_hw_reg(s
, inst
->DstReg
.File
, inst
->DstReg
.Index
);
596 _mesa_printf(" BEGIN_TEX\n");
598 if (s
->Handler
->BeginTexBlock
)
599 s
->Error
= s
->Error
|| !s
->Handler
->BeginTexBlock(s
->UserData
);
601 for(pairinst
= readytex
; pairinst
; pairinst
= pairinst
->NextReady
) {
602 int ip
= pairinst
- s
->Instructions
;
603 struct prog_instruction
*inst
= s
->Program
->Instructions
+ ip
;
604 commit_instruction(s
, ip
);
606 if (inst
->Opcode
!= OPCODE_KIL
)
607 inst
->DstReg
.Index
= get_hw_reg(s
, inst
->DstReg
.File
, inst
->DstReg
.Index
);
608 inst
->SrcReg
[0].Index
= get_hw_reg(s
, inst
->SrcReg
[0].File
, inst
->SrcReg
[0].Index
);
612 _mesa_print_instruction(inst
);
614 s
->Error
= s
->Error
|| !s
->Handler
->EmitTex(s
->UserData
, inst
);
618 _mesa_printf(" END_TEX\n");
622 static int alloc_pair_source(struct pair_state
*s
, struct radeon_pair_instruction
*pair
,
623 struct prog_src_register src
, GLboolean rgb
, GLboolean alpha
)
626 int candidate_quality
= -1;
635 if (src
.File
== PROGRAM_TEMPORARY
|| src
.File
== PROGRAM_INPUT
) {
637 index
= get_hw_reg(s
, src
.File
, src
.Index
);
640 s
->Error
|= !s
->Handler
->EmitConst(s
->UserData
, src
.File
, src
.Index
, &index
);
643 for(i
= 0; i
< 3; ++i
) {
646 if (pair
->RGB
.Src
[i
].Used
) {
647 if (pair
->RGB
.Src
[i
].Constant
!= constant
||
648 pair
->RGB
.Src
[i
].Index
!= index
)
654 if (pair
->Alpha
.Src
[i
].Used
) {
655 if (pair
->Alpha
.Src
[i
].Constant
!= constant
||
656 pair
->Alpha
.Src
[i
].Index
!= index
)
661 if (q
> candidate_quality
) {
662 candidate_quality
= q
;
667 if (candidate
>= 0) {
669 pair
->RGB
.Src
[candidate
].Used
= 1;
670 pair
->RGB
.Src
[candidate
].Constant
= constant
;
671 pair
->RGB
.Src
[candidate
].Index
= index
;
674 pair
->Alpha
.Src
[candidate
].Used
= 1;
675 pair
->Alpha
.Src
[candidate
].Constant
= constant
;
676 pair
->Alpha
.Src
[candidate
].Index
= index
;
684 * Fill the given ALU instruction's opcodes and source operands into the given pair,
687 static GLboolean
fill_instruction_into_pair(struct pair_state
*s
, struct radeon_pair_instruction
*pair
, int ip
)
689 struct pair_state_instruction
*pairinst
= s
->Instructions
+ ip
;
690 struct prog_instruction
*inst
= s
->Program
->Instructions
+ ip
;
692 ASSERT(!pairinst
->NeedRGB
|| pair
->RGB
.Opcode
== OPCODE_NOP
);
693 ASSERT(!pairinst
->NeedAlpha
|| pair
->Alpha
.Opcode
== OPCODE_NOP
);
695 if (pairinst
->NeedRGB
) {
696 if (pairinst
->IsTranscendent
)
697 pair
->RGB
.Opcode
= OPCODE_REPL_ALPHA
;
699 pair
->RGB
.Opcode
= inst
->Opcode
;
700 if (inst
->SaturateMode
== SATURATE_ZERO_ONE
)
701 pair
->RGB
.Saturate
= 1;
703 if (pairinst
->NeedAlpha
) {
704 pair
->Alpha
.Opcode
= inst
->Opcode
;
705 if (inst
->SaturateMode
== SATURATE_ZERO_ONE
)
706 pair
->Alpha
.Saturate
= 1;
709 int nargs
= _mesa_num_inst_src_regs(inst
->Opcode
);
712 /* Special case for DDX/DDY (MDH/MDV). */
713 if (inst
->Opcode
== OPCODE_DDX
|| inst
->Opcode
== OPCODE_DDY
) {
714 if (pair
->RGB
.Src
[0].Used
|| pair
->Alpha
.Src
[0].Used
)
720 for(i
= 0; i
< nargs
; ++i
) {
722 if (pairinst
->NeedRGB
&& !pairinst
->IsTranscendent
) {
723 GLboolean srcrgb
= GL_FALSE
;
724 GLboolean srcalpha
= GL_FALSE
;
725 GLuint negatebase
= 0;
727 for(j
= 0; j
< 3; ++j
) {
728 GLuint swz
= GET_SWZ(inst
->SrcReg
[i
].Swizzle
, j
);
733 if (swz
!= SWIZZLE_NIL
&& GET_BIT(inst
->SrcReg
[i
].NegateBase
, j
))
736 source
= alloc_pair_source(s
, pair
, inst
->SrcReg
[i
], srcrgb
, srcalpha
);
739 pair
->RGB
.Arg
[i
].Source
= source
;
740 pair
->RGB
.Arg
[i
].Swizzle
= inst
->SrcReg
[i
].Swizzle
& 0x1ff;
741 pair
->RGB
.Arg
[i
].Abs
= inst
->SrcReg
[i
].Abs
;
742 pair
->RGB
.Arg
[i
].Negate
= (negatebase
& ~pair
->RGB
.Arg
[i
].Abs
) ^ inst
->SrcReg
[i
].NegateAbs
;
744 if (pairinst
->NeedAlpha
) {
745 GLboolean srcrgb
= GL_FALSE
;
746 GLboolean srcalpha
= GL_FALSE
;
747 GLuint negatebase
= GET_BIT(inst
->SrcReg
[i
].NegateBase
, pairinst
->IsTranscendent
? 0 : 3);
748 GLuint swz
= GET_SWZ(inst
->SrcReg
[i
].Swizzle
, pairinst
->IsTranscendent
? 0 : 3);
753 source
= alloc_pair_source(s
, pair
, inst
->SrcReg
[i
], srcrgb
, srcalpha
);
756 pair
->Alpha
.Arg
[i
].Source
= source
;
757 pair
->Alpha
.Arg
[i
].Swizzle
= swz
;
758 pair
->Alpha
.Arg
[i
].Abs
= inst
->SrcReg
[i
].Abs
;
759 pair
->Alpha
.Arg
[i
].Negate
= (negatebase
& ~pair
->RGB
.Arg
[i
].Abs
) ^ inst
->SrcReg
[i
].NegateAbs
;
768 * Fill in the destination register information.
770 * This is split from filling in source registers because we want
771 * to avoid allocating hardware temporaries for destinations until
772 * we are absolutely certain that we're going to emit a certain
773 * instruction pairing.
775 static void fill_dest_into_pair(struct pair_state
*s
, struct radeon_pair_instruction
*pair
, int ip
)
777 struct pair_state_instruction
*pairinst
= s
->Instructions
+ ip
;
778 struct prog_instruction
*inst
= s
->Program
->Instructions
+ ip
;
780 if (inst
->DstReg
.File
== PROGRAM_OUTPUT
) {
781 if (inst
->DstReg
.Index
== FRAG_RESULT_COLOR
) {
782 pair
->RGB
.OutputWriteMask
|= inst
->DstReg
.WriteMask
& WRITEMASK_XYZ
;
783 pair
->Alpha
.OutputWriteMask
|= GET_BIT(inst
->DstReg
.WriteMask
, 3);
784 } else if (inst
->DstReg
.Index
== FRAG_RESULT_DEPTH
) {
785 pair
->Alpha
.DepthWriteMask
|= GET_BIT(inst
->DstReg
.WriteMask
, 3);
788 GLuint hwindex
= get_hw_reg(s
, inst
->DstReg
.File
, inst
->DstReg
.Index
);
789 if (pairinst
->NeedRGB
) {
790 pair
->RGB
.DestIndex
= hwindex
;
791 pair
->RGB
.WriteMask
|= inst
->DstReg
.WriteMask
& WRITEMASK_XYZ
;
793 if (pairinst
->NeedAlpha
) {
794 pair
->Alpha
.DestIndex
= hwindex
;
795 pair
->Alpha
.WriteMask
|= GET_BIT(inst
->DstReg
.WriteMask
, 3);
802 * Find a good ALU instruction or pair of ALU instruction and emit it.
804 * Prefer emitting full ALU instructions, so that when we reach a point
805 * where no full ALU instruction can be emitted, we have more candidates
806 * for RGB/Alpha pairing.
808 static void emit_alu(struct pair_state
*s
)
810 struct radeon_pair_instruction pair
;
812 if (s
->ReadyFullALU
|| !(s
->ReadyRGB
&& s
->ReadyAlpha
)) {
814 if (s
->ReadyFullALU
) {
815 ip
= s
->ReadyFullALU
- s
->Instructions
;
816 s
->ReadyFullALU
= s
->ReadyFullALU
->NextReady
;
817 } else if (s
->ReadyRGB
) {
818 ip
= s
->ReadyRGB
- s
->Instructions
;
819 s
->ReadyRGB
= s
->ReadyRGB
->NextReady
;
821 ip
= s
->ReadyAlpha
- s
->Instructions
;
822 s
->ReadyAlpha
= s
->ReadyAlpha
->NextReady
;
825 _mesa_bzero(&pair
, sizeof(pair
));
826 fill_instruction_into_pair(s
, &pair
, ip
);
827 fill_dest_into_pair(s
, &pair
, ip
);
828 commit_instruction(s
, ip
);
830 struct pair_state_instruction
**prgb
;
831 struct pair_state_instruction
**palpha
;
833 /* Some pairings might fail because they require too
834 * many source slots; try all possible pairings if necessary */
835 for(prgb
= &s
->ReadyRGB
; *prgb
; prgb
= &(*prgb
)->NextReady
) {
836 for(palpha
= &s
->ReadyAlpha
; *palpha
; palpha
= &(*palpha
)->NextReady
) {
837 int rgbip
= *prgb
- s
->Instructions
;
838 int alphaip
= *palpha
- s
->Instructions
;
839 _mesa_bzero(&pair
, sizeof(pair
));
840 fill_instruction_into_pair(s
, &pair
, rgbip
);
841 if (!fill_instruction_into_pair(s
, &pair
, alphaip
))
843 *prgb
= (*prgb
)->NextReady
;
844 *palpha
= (*palpha
)->NextReady
;
845 fill_dest_into_pair(s
, &pair
, rgbip
);
846 fill_dest_into_pair(s
, &pair
, alphaip
);
847 commit_instruction(s
, rgbip
);
848 commit_instruction(s
, alphaip
);
853 /* No success in pairing; just take the first RGB instruction */
854 int ip
= s
->ReadyRGB
- s
->Instructions
;
855 s
->ReadyRGB
= s
->ReadyRGB
->NextReady
;
856 _mesa_bzero(&pair
, sizeof(pair
));
857 fill_instruction_into_pair(s
, &pair
, ip
);
858 fill_dest_into_pair(s
, &pair
, ip
);
859 commit_instruction(s
, ip
);
864 radeonPrintPairInstruction(&pair
);
866 s
->Error
= s
->Error
|| !s
->Handler
->EmitPaired(s
->UserData
, &pair
);
870 GLboolean
radeonPairProgram(GLcontext
*ctx
, struct gl_program
*program
,
871 const struct radeon_pair_handler
* handler
, void *userdata
)
875 _mesa_bzero(&s
, sizeof(s
));
879 s
.UserData
= userdata
;
880 s
.Debug
= (RADEON_DEBUG
& DEBUG_PIXEL
) ? GL_TRUE
: GL_FALSE
;
881 s
.Verbose
= GL_FALSE
&& s
.Debug
;
883 s
.Instructions
= (struct pair_state_instruction
*)_mesa_calloc(
884 sizeof(struct pair_state_instruction
)*s
.Program
->NumInstructions
);
885 s
.ValuePool
= (struct reg_value
*)_mesa_calloc(sizeof(struct reg_value
)*s
.Program
->NumInstructions
*4);
886 s
.ReaderPool
= (struct reg_value_reader
*)_mesa_calloc(
887 sizeof(struct reg_value_reader
)*s
.Program
->NumInstructions
*12);
890 _mesa_printf("Emit paired program\n");
892 scan_instructions(&s
);
893 allocate_input_registers(&s
);
896 (s
.ReadyTEX
|| s
.ReadyRGB
|| s
.ReadyAlpha
|| s
.ReadyFullALU
)) {
900 while(s
.ReadyFullALU
|| s
.ReadyRGB
|| s
.ReadyAlpha
)
905 _mesa_printf(" END\n");
907 _mesa_free(s
.Instructions
);
908 _mesa_free(s
.ValuePool
);
909 _mesa_free(s
.ReaderPool
);
915 static void print_pair_src(int i
, struct radeon_pair_instruction_source
* src
)
917 _mesa_printf(" Src%i = %s[%i]", i
, src
->Constant
? "CNST" : "TEMP", src
->Index
);
920 static const char* opcode_string(GLuint opcode
)
922 if (opcode
== OPCODE_REPL_ALPHA
)
925 return _mesa_opcode_string(opcode
);
928 static int num_pairinst_args(GLuint opcode
)
930 if (opcode
== OPCODE_REPL_ALPHA
)
933 return _mesa_num_inst_src_regs(opcode
);
936 static char swizzle_char(GLuint swz
)
939 case SWIZZLE_X
: return 'x';
940 case SWIZZLE_Y
: return 'y';
941 case SWIZZLE_Z
: return 'z';
942 case SWIZZLE_W
: return 'w';
943 case SWIZZLE_ZERO
: return '0';
944 case SWIZZLE_ONE
: return '1';
945 case SWIZZLE_NIL
: return '_';
950 void radeonPrintPairInstruction(struct radeon_pair_instruction
*inst
)
955 _mesa_printf(" RGB: ");
956 for(i
= 0; i
< 3; ++i
) {
957 if (inst
->RGB
.Src
[i
].Used
)
958 print_pair_src(i
, inst
->RGB
.Src
+ i
);
961 _mesa_printf(" Alpha:");
962 for(i
= 0; i
< 3; ++i
) {
963 if (inst
->Alpha
.Src
[i
].Used
)
964 print_pair_src(i
, inst
->Alpha
.Src
+ i
);
968 _mesa_printf(" %s%s", opcode_string(inst
->RGB
.Opcode
), inst
->RGB
.Saturate
? "_SAT" : "");
969 if (inst
->RGB
.WriteMask
)
970 _mesa_printf(" TEMP[%i].%s%s%s", inst
->RGB
.DestIndex
,
971 (inst
->RGB
.WriteMask
& 1) ? "x" : "",
972 (inst
->RGB
.WriteMask
& 2) ? "y" : "",
973 (inst
->RGB
.WriteMask
& 4) ? "z" : "");
974 if (inst
->RGB
.OutputWriteMask
)
975 _mesa_printf(" COLOR.%s%s%s",
976 (inst
->RGB
.OutputWriteMask
& 1) ? "x" : "",
977 (inst
->RGB
.OutputWriteMask
& 2) ? "y" : "",
978 (inst
->RGB
.OutputWriteMask
& 4) ? "z" : "");
979 nargs
= num_pairinst_args(inst
->RGB
.Opcode
);
980 for(i
= 0; i
< nargs
; ++i
) {
981 const char* abs
= inst
->RGB
.Arg
[i
].Abs
? "|" : "";
982 const char* neg
= inst
->RGB
.Arg
[i
].Negate
? "-" : "";
983 _mesa_printf(", %s%sSrc%i.%c%c%c%s", neg
, abs
, inst
->RGB
.Arg
[i
].Source
,
984 swizzle_char(GET_SWZ(inst
->RGB
.Arg
[i
].Swizzle
, 0)),
985 swizzle_char(GET_SWZ(inst
->RGB
.Arg
[i
].Swizzle
, 1)),
986 swizzle_char(GET_SWZ(inst
->RGB
.Arg
[i
].Swizzle
, 2)),
991 _mesa_printf(" %s%s", opcode_string(inst
->Alpha
.Opcode
), inst
->Alpha
.Saturate
? "_SAT" : "");
992 if (inst
->Alpha
.WriteMask
)
993 _mesa_printf(" TEMP[%i].w", inst
->Alpha
.DestIndex
);
994 if (inst
->Alpha
.OutputWriteMask
)
995 _mesa_printf(" COLOR.w");
996 if (inst
->Alpha
.DepthWriteMask
)
997 _mesa_printf(" DEPTH.w");
998 nargs
= num_pairinst_args(inst
->Alpha
.Opcode
);
999 for(i
= 0; i
< nargs
; ++i
) {
1000 const char* abs
= inst
->Alpha
.Arg
[i
].Abs
? "|" : "";
1001 const char* neg
= inst
->Alpha
.Arg
[i
].Negate
? "-" : "";
1002 _mesa_printf(", %s%sSrc%i.%c%s", neg
, abs
, inst
->Alpha
.Arg
[i
].Source
,
1003 swizzle_char(inst
->Alpha
.Arg
[i
].Swizzle
), abs
);