2 #include "pipe/p_context.h"
3 #include "pipe/p_defines.h"
4 #include "pipe/p_state.h"
5 #include "util/u_inlines.h"
6 #include "util/u_debug.h"
8 #include "pipe/p_shader_tokens.h"
9 #include "tgsi/tgsi_parse.h"
10 #include "tgsi/tgsi_util.h"
11 #include "tgsi/tgsi_dump.h"
12 #include "tgsi/tgsi_ureg.h"
14 #include "nvfx_context.h"
15 #include "nvfx_shader.h"
16 #include "nvfx_resource.h"
18 #define MAX_CONSTS 128
22 struct nvfx_pipe_fragment_program
* pfp
;
23 struct nvfx_fragment_program
*fp
;
26 unsigned long long r_temps
;
27 unsigned long long r_temps_discard
;
28 struct nvfx_reg r_result
[PIPE_MAX_SHADER_OUTPUTS
];
29 struct nvfx_reg
*r_temp
;
30 unsigned sprite_coord_temp
;
43 struct nvfx_reg imm
[MAX_IMM
];
46 unsigned char generic_to_slot
[256]; /* semantic idx for each input semantic */
48 struct util_dynarray if_stack
;
49 //struct util_dynarray loop_stack;
50 struct util_dynarray label_relocs
;
53 static INLINE
struct nvfx_reg
54 temp(struct nvfx_fpc
*fpc
)
56 int idx
= __builtin_ctzll(~fpc
->r_temps
);
58 if (idx
>= fpc
->max_temps
) {
59 NOUVEAU_ERR("out of temps!!\n");
61 return nvfx_reg(NVFXSR_TEMP
, 0);
64 fpc
->r_temps
|= (1ULL << idx
);
65 fpc
->r_temps_discard
|= (1ULL << idx
);
66 return nvfx_reg(NVFXSR_TEMP
, idx
);
70 release_temps(struct nvfx_fpc
*fpc
)
72 fpc
->r_temps
&= ~fpc
->r_temps_discard
;
73 fpc
->r_temps_discard
= 0ULL;
76 static INLINE
struct nvfx_reg
77 constant(struct nvfx_fpc
*fpc
, int pipe
, float vals
[4])
81 if (fpc
->nr_consts
== MAX_CONSTS
)
83 idx
= fpc
->nr_consts
++;
85 fpc
->consts
[idx
].pipe
= pipe
;
87 memcpy(fpc
->consts
[idx
].vals
, vals
, 4 * sizeof(float));
88 return nvfx_reg(NVFXSR_CONST
, idx
);
92 grow_insns(struct nvfx_fpc
*fpc
, int size
)
94 struct nvfx_fragment_program
*fp
= fpc
->fp
;
97 fp
->insn
= realloc(fp
->insn
, sizeof(uint32_t) * fp
->insn_len
);
101 emit_src(struct nvfx_fpc
*fpc
, int pos
, struct nvfx_src src
)
103 struct nvfx_fragment_program
*fp
= fpc
->fp
;
104 uint32_t *hw
= &fp
->insn
[fpc
->inst_offset
];
107 switch (src
.reg
.type
) {
109 sr
|= (NVFX_FP_REG_TYPE_INPUT
<< NVFX_FP_REG_TYPE_SHIFT
);
110 hw
[0] |= (src
.reg
.index
<< NVFX_FP_OP_INPUT_SRC_SHIFT
);
113 sr
|= NVFX_FP_REG_SRC_HALF
;
116 sr
|= (NVFX_FP_REG_TYPE_TEMP
<< NVFX_FP_REG_TYPE_SHIFT
);
117 sr
|= (src
.reg
.index
<< NVFX_FP_REG_SRC_SHIFT
);
119 case NVFXSR_RELOCATED
:
120 sr
|= (NVFX_FP_REG_TYPE_TEMP
<< NVFX_FP_REG_TYPE_SHIFT
);
121 sr
|= (fpc
->sprite_coord_temp
<< NVFX_FP_REG_SRC_SHIFT
);
122 //printf("adding relocation at %x for %x\n", fpc->inst_offset, src.index);
123 util_dynarray_append(&fpc
->fp
->slot_relocations
[src
.reg
.index
], unsigned, fpc
->inst_offset
+ pos
+ 1);
126 if (!fpc
->have_const
) {
131 hw
= &fp
->insn
[fpc
->inst_offset
];
132 if (fpc
->consts
[src
.reg
.index
].pipe
>= 0) {
133 struct nvfx_fragment_program_data
*fpd
;
135 fp
->consts
= realloc(fp
->consts
, ++fp
->nr_consts
*
137 fpd
= &fp
->consts
[fp
->nr_consts
- 1];
138 fpd
->offset
= fpc
->inst_offset
+ 4;
139 fpd
->index
= fpc
->consts
[src
.reg
.index
].pipe
;
140 memset(&fp
->insn
[fpd
->offset
], 0, sizeof(uint32_t) * 4);
142 memcpy(&fp
->insn
[fpc
->inst_offset
+ 4],
143 fpc
->consts
[src
.reg
.index
].vals
,
144 sizeof(uint32_t) * 4);
147 sr
|= (NVFX_FP_REG_TYPE_CONST
<< NVFX_FP_REG_TYPE_SHIFT
);
150 sr
|= (NVFX_FP_REG_TYPE_INPUT
<< NVFX_FP_REG_TYPE_SHIFT
);
157 sr
|= NVFX_FP_REG_NEGATE
;
160 hw
[1] |= (1 << (29 + pos
));
162 sr
|= ((src
.swz
[0] << NVFX_FP_REG_SWZ_X_SHIFT
) |
163 (src
.swz
[1] << NVFX_FP_REG_SWZ_Y_SHIFT
) |
164 (src
.swz
[2] << NVFX_FP_REG_SWZ_Z_SHIFT
) |
165 (src
.swz
[3] << NVFX_FP_REG_SWZ_W_SHIFT
));
171 emit_dst(struct nvfx_fpc
*fpc
, struct nvfx_reg dst
)
173 struct nvfx_fragment_program
*fp
= fpc
->fp
;
174 uint32_t *hw
= &fp
->insn
[fpc
->inst_offset
];
178 if (fpc
->num_regs
< (dst
.index
+ 1))
179 fpc
->num_regs
= dst
.index
+ 1;
182 if (dst
.index
== 1) {
183 fp
->fp_control
|= 0xe;
185 hw
[0] |= NVFX_FP_OP_OUT_REG_HALF
;
195 hw
[0] |= (dst
.index
<< NVFX_FP_OP_OUT_REG_SHIFT
);
199 nvfx_fp_emit(struct nvfx_fpc
*fpc
, struct nvfx_insn insn
)
201 struct nvfx_fragment_program
*fp
= fpc
->fp
;
204 fpc
->inst_offset
= fp
->insn_len
;
207 hw
= &fp
->insn
[fpc
->inst_offset
];
208 memset(hw
, 0, sizeof(uint32_t) * 4);
210 if (insn
.op
== NVFX_FP_OP_OPCODE_KIL
)
211 fp
->fp_control
|= NV34TCL_FP_CONTROL_USES_KIL
;
212 hw
[0] |= (insn
.op
<< NVFX_FP_OP_OPCODE_SHIFT
);
213 hw
[0] |= (insn
.mask
<< NVFX_FP_OP_OUTMASK_SHIFT
);
214 hw
[2] |= (insn
.scale
<< NVFX_FP_OP_DST_SCALE_SHIFT
);
217 hw
[0] |= NVFX_FP_OP_OUT_SAT
;
220 hw
[0] |= NVFX_FP_OP_COND_WRITE_ENABLE
;
221 hw
[1] |= (insn
.cc_test
<< NVFX_FP_OP_COND_SHIFT
);
222 hw
[1] |= ((insn
.cc_swz
[0] << NVFX_FP_OP_COND_SWZ_X_SHIFT
) |
223 (insn
.cc_swz
[1] << NVFX_FP_OP_COND_SWZ_Y_SHIFT
) |
224 (insn
.cc_swz
[2] << NVFX_FP_OP_COND_SWZ_Z_SHIFT
) |
225 (insn
.cc_swz
[3] << NVFX_FP_OP_COND_SWZ_W_SHIFT
));
229 hw
[0] |= (insn
.unit
<< NVFX_FP_OP_TEX_UNIT_SHIFT
);
230 fp
->samplers
|= (1 << insn
.unit
);
233 emit_dst(fpc
, insn
.dst
);
234 emit_src(fpc
, 0, insn
.src
[0]);
235 emit_src(fpc
, 1, insn
.src
[1]);
236 emit_src(fpc
, 2, insn
.src
[2]);
239 #define arith(s,o,d,m,s0,s1,s2) \
240 nvfx_insn((s), NVFX_FP_OP_OPCODE_##o, -1, \
241 (d), (m), (s0), (s1), (s2))
243 #define tex(s,o,u,d,m,s0,s1,s2) \
244 nvfx_insn((s), NVFX_FP_OP_OPCODE_##o, (u), \
245 (d), (m), (s0), none, none)
247 /* IF src.x != 0, as TGSI specifies */
249 nv40_fp_if(struct nvfx_fpc
*fpc
, struct nvfx_src src
)
251 const struct nvfx_src none
= nvfx_src(nvfx_reg(NVFXSR_NONE
, 0));
252 struct nvfx_insn insn
= arith(0, MOV
, none
.reg
, NVFX_FP_MASK_X
, src
, none
, none
);
255 nvfx_fp_emit(fpc
, insn
);
257 fpc
->inst_offset
= fpc
->fp
->insn_len
;
259 hw
= &fpc
->fp
->insn
[fpc
->inst_offset
];
260 /* I really wonder why fp16 precision is used. Presumably the hardware ignores it? */
261 hw
[0] = (NV40_FP_OP_BRA_OPCODE_IF
<< NVFX_FP_OP_OPCODE_SHIFT
) |
262 NV40_FP_OP_OUT_NONE
|
263 (NVFX_FP_PRECISION_FP16
<< NVFX_FP_OP_PRECISION_SHIFT
);
264 /* Use .xxxx swizzle so that we check only src[0].x*/
265 hw
[1] = (0 << NVFX_FP_OP_COND_SWZ_X_SHIFT
) |
266 (0 << NVFX_FP_OP_COND_SWZ_Y_SHIFT
) |
267 (0 << NVFX_FP_OP_COND_SWZ_Z_SHIFT
) |
268 (0 << NVFX_FP_OP_COND_SWZ_W_SHIFT
) |
269 (NVFX_FP_OP_COND_NE
<< NVFX_FP_OP_COND_SHIFT
);
270 hw
[2] = 0; /* | NV40_FP_OP_OPCODE_IS_BRANCH | else_offset */
271 hw
[3] = 0; /* | endif_offset */
272 util_dynarray_append(&fpc
->if_stack
, unsigned, fpc
->inst_offset
);
275 /* IF src.x != 0, as TGSI specifies */
277 nv40_fp_cal(struct nvfx_fpc
*fpc
, unsigned target
)
279 struct nvfx_relocation reloc
;
281 fpc
->inst_offset
= fpc
->fp
->insn_len
;
283 hw
= &fpc
->fp
->insn
[fpc
->inst_offset
];
284 /* I really wonder why fp16 precision is used. Presumably the hardware ignores it? */
285 hw
[0] = (NV40_FP_OP_BRA_OPCODE_CAL
<< NVFX_FP_OP_OPCODE_SHIFT
);
286 /* Use .xxxx swizzle so that we check only src[0].x*/
287 hw
[1] = (NVFX_SWZ_IDENTITY
<< NVFX_FP_OP_COND_SWZ_ALL_SHIFT
) |
288 (NVFX_FP_OP_COND_TR
<< NVFX_FP_OP_COND_SHIFT
);
289 hw
[2] = NV40_FP_OP_OPCODE_IS_BRANCH
; /* | call_offset */
291 reloc
.target
= target
;
292 reloc
.location
= fpc
->inst_offset
+ 2;
293 util_dynarray_append(&fpc
->label_relocs
, struct nvfx_relocation
, reloc
);
297 nv40_fp_ret(struct nvfx_fpc
*fpc
)
300 fpc
->inst_offset
= fpc
->fp
->insn_len
;
302 hw
= &fpc
->fp
->insn
[fpc
->inst_offset
];
303 /* I really wonder why fp16 precision is used. Presumably the hardware ignores it? */
304 hw
[0] = (NV40_FP_OP_BRA_OPCODE_RET
<< NVFX_FP_OP_OPCODE_SHIFT
);
305 /* Use .xxxx swizzle so that we check only src[0].x*/
306 hw
[1] = (NVFX_SWZ_IDENTITY
<< NVFX_FP_OP_COND_SWZ_ALL_SHIFT
) |
307 (NVFX_FP_OP_COND_TR
<< NVFX_FP_OP_COND_SHIFT
);
308 hw
[2] = NV40_FP_OP_OPCODE_IS_BRANCH
; /* | call_offset */
313 nv40_fp_rep(struct nvfx_fpc
*fpc
, unsigned count
, unsigned target
)
315 struct nvfx_relocation reloc
;
317 fpc
->inst_offset
= fpc
->fp
->insn_len
;
319 hw
= &fpc
->fp
->insn
[fpc
->inst_offset
];
320 /* I really wonder why fp16 precision is used. Presumably the hardware ignores it? */
321 hw
[0] = (NV40_FP_OP_BRA_OPCODE_REP
<< NVFX_FP_OP_OPCODE_SHIFT
) |
322 NV40_FP_OP_OUT_NONE
|
323 (NVFX_FP_PRECISION_FP16
<< NVFX_FP_OP_PRECISION_SHIFT
);
324 /* Use .xxxx swizzle so that we check only src[0].x*/
325 hw
[1] = (NVFX_SWZ_IDENTITY
<< NVFX_FP_OP_COND_SWZ_ALL_SHIFT
) |
326 (NVFX_FP_OP_COND_TR
<< NVFX_FP_OP_COND_SHIFT
);
327 hw
[2] = NV40_FP_OP_OPCODE_IS_BRANCH
|
328 (count
<< NV40_FP_OP_REP_COUNT1_SHIFT
) |
329 (count
<< NV40_FP_OP_REP_COUNT2_SHIFT
) |
330 (count
<< NV40_FP_OP_REP_COUNT3_SHIFT
);
331 hw
[3] = 0; /* | end_offset */
332 reloc
.target
= target
;
333 reloc
.location
= fpc
->inst_offset
+ 3;
334 util_dynarray_append(&fpc
->label_relocs
, struct nvfx_relocation
, reloc
);
335 //util_dynarray_append(&fpc->loop_stack, unsigned, target);
338 /* warning: this only works forward, and probably only if not inside any IF */
340 nv40_fp_bra(struct nvfx_fpc
*fpc
, unsigned target
)
342 struct nvfx_relocation reloc
;
344 fpc
->inst_offset
= fpc
->fp
->insn_len
;
346 hw
= &fpc
->fp
->insn
[fpc
->inst_offset
];
347 /* I really wonder why fp16 precision is used. Presumably the hardware ignores it? */
348 hw
[0] = (NV40_FP_OP_BRA_OPCODE_IF
<< NVFX_FP_OP_OPCODE_SHIFT
) |
349 NV40_FP_OP_OUT_NONE
|
350 (NVFX_FP_PRECISION_FP16
<< NVFX_FP_OP_PRECISION_SHIFT
);
351 /* Use .xxxx swizzle so that we check only src[0].x*/
352 hw
[1] = (NVFX_SWZ_IDENTITY
<< NVFX_FP_OP_COND_SWZ_X_SHIFT
) |
353 (NVFX_FP_OP_COND_FL
<< NVFX_FP_OP_COND_SHIFT
);
354 hw
[2] = NV40_FP_OP_OPCODE_IS_BRANCH
; /* | else_offset */
355 hw
[3] = 0; /* | endif_offset */
356 reloc
.target
= target
;
357 reloc
.location
= fpc
->inst_offset
+ 2;
358 util_dynarray_append(&fpc
->label_relocs
, struct nvfx_relocation
, reloc
);
359 reloc
.target
= target
;
360 reloc
.location
= fpc
->inst_offset
+ 3;
361 util_dynarray_append(&fpc
->label_relocs
, struct nvfx_relocation
, reloc
);
365 nv40_fp_brk(struct nvfx_fpc
*fpc
)
368 fpc
->inst_offset
= fpc
->fp
->insn_len
;
370 hw
= &fpc
->fp
->insn
[fpc
->inst_offset
];
371 /* I really wonder why fp16 precision is used. Presumably the hardware ignores it? */
372 hw
[0] = (NV40_FP_OP_BRA_OPCODE_BRK
<< NVFX_FP_OP_OPCODE_SHIFT
) |
374 /* Use .xxxx swizzle so that we check only src[0].x*/
375 hw
[1] = (NVFX_SWZ_IDENTITY
<< NVFX_FP_OP_COND_SWZ_X_SHIFT
) |
376 (NVFX_FP_OP_COND_TR
<< NVFX_FP_OP_COND_SHIFT
);
377 hw
[2] = NV40_FP_OP_OPCODE_IS_BRANCH
;
381 static INLINE
struct nvfx_src
382 tgsi_src(struct nvfx_fpc
*fpc
, const struct tgsi_full_src_register
*fsrc
)
386 switch (fsrc
->Register
.File
) {
387 case TGSI_FILE_INPUT
:
388 if(fpc
->pfp
->info
.input_semantic_name
[fsrc
->Register
.Index
] == TGSI_SEMANTIC_POSITION
) {
389 assert(fpc
->pfp
->info
.input_semantic_index
[fsrc
->Register
.Index
] == 0);
390 src
.reg
= nvfx_reg(NVFXSR_INPUT
, NVFX_FP_OP_INPUT_SRC_POSITION
);
391 } else if(fpc
->pfp
->info
.input_semantic_name
[fsrc
->Register
.Index
] == TGSI_SEMANTIC_COLOR
) {
392 if(fpc
->pfp
->info
.input_semantic_index
[fsrc
->Register
.Index
] == 0)
393 src
.reg
= nvfx_reg(NVFXSR_INPUT
, NVFX_FP_OP_INPUT_SRC_COL0
);
394 else if(fpc
->pfp
->info
.input_semantic_index
[fsrc
->Register
.Index
] == 1)
395 src
.reg
= nvfx_reg(NVFXSR_INPUT
, NVFX_FP_OP_INPUT_SRC_COL1
);
398 } else if(fpc
->pfp
->info
.input_semantic_name
[fsrc
->Register
.Index
] == TGSI_SEMANTIC_FOG
) {
399 assert(fpc
->pfp
->info
.input_semantic_index
[fsrc
->Register
.Index
] == 0);
400 src
.reg
= nvfx_reg(NVFXSR_INPUT
, NVFX_FP_OP_INPUT_SRC_FOGC
);
401 } else if(fpc
->pfp
->info
.input_semantic_name
[fsrc
->Register
.Index
] == TGSI_SEMANTIC_FACE
) {
402 /* TODO: check this has the correct values */
403 /* XXX: what do we do for nv30 here (assuming it lacks facing)?! */
404 assert(fpc
->pfp
->info
.input_semantic_index
[fsrc
->Register
.Index
] == 0);
405 src
.reg
= nvfx_reg(NVFXSR_INPUT
, NV40_FP_OP_INPUT_SRC_FACING
);
407 assert(fpc
->pfp
->info
.input_semantic_name
[fsrc
->Register
.Index
] == TGSI_SEMANTIC_GENERIC
);
408 src
.reg
= nvfx_reg(NVFXSR_RELOCATED
, fpc
->generic_to_slot
[fpc
->pfp
->info
.input_semantic_index
[fsrc
->Register
.Index
]]);
411 case TGSI_FILE_CONSTANT
:
412 src
.reg
= constant(fpc
, fsrc
->Register
.Index
, NULL
);
414 case TGSI_FILE_IMMEDIATE
:
415 assert(fsrc
->Register
.Index
< fpc
->nr_imm
);
416 src
.reg
= fpc
->imm
[fsrc
->Register
.Index
];
418 case TGSI_FILE_TEMPORARY
:
419 src
.reg
= fpc
->r_temp
[fsrc
->Register
.Index
];
421 /* NV40 fragprog result regs are just temps, so this is simple */
422 case TGSI_FILE_OUTPUT
:
423 src
.reg
= fpc
->r_result
[fsrc
->Register
.Index
];
426 NOUVEAU_ERR("bad src file\n");
432 src
.abs
= fsrc
->Register
.Absolute
;
433 src
.negate
= fsrc
->Register
.Negate
;
434 src
.swz
[0] = fsrc
->Register
.SwizzleX
;
435 src
.swz
[1] = fsrc
->Register
.SwizzleY
;
436 src
.swz
[2] = fsrc
->Register
.SwizzleZ
;
437 src
.swz
[3] = fsrc
->Register
.SwizzleW
;
441 static INLINE
struct nvfx_reg
442 tgsi_dst(struct nvfx_fpc
*fpc
, const struct tgsi_full_dst_register
*fdst
) {
443 switch (fdst
->Register
.File
) {
444 case TGSI_FILE_OUTPUT
:
445 return fpc
->r_result
[fdst
->Register
.Index
];
446 case TGSI_FILE_TEMPORARY
:
447 return fpc
->r_temp
[fdst
->Register
.Index
];
449 return nvfx_reg(NVFXSR_NONE
, 0);
451 NOUVEAU_ERR("bad dst file %d\n", fdst
->Register
.File
);
452 return nvfx_reg(NVFXSR_NONE
, 0);
461 if (tgsi
& TGSI_WRITEMASK_X
) mask
|= NVFX_FP_MASK_X
;
462 if (tgsi
& TGSI_WRITEMASK_Y
) mask
|= NVFX_FP_MASK_Y
;
463 if (tgsi
& TGSI_WRITEMASK_Z
) mask
|= NVFX_FP_MASK_Z
;
464 if (tgsi
& TGSI_WRITEMASK_W
) mask
|= NVFX_FP_MASK_W
;
469 nvfx_fragprog_parse_instruction(struct nvfx_context
* nvfx
, struct nvfx_fpc
*fpc
,
470 const struct tgsi_full_instruction
*finst
)
472 const struct nvfx_src none
= nvfx_src(nvfx_reg(NVFXSR_NONE
, 0));
473 struct nvfx_insn insn
;
474 struct nvfx_src src
[3], tmp
, tmp2
;
476 int mask
, sat
, unit
= 0;
477 int ai
= -1, ci
= -1, ii
= -1;
480 if (finst
->Instruction
.Opcode
== TGSI_OPCODE_END
)
483 for (i
= 0; i
< finst
->Instruction
.NumSrcRegs
; i
++) {
484 const struct tgsi_full_src_register
*fsrc
;
486 fsrc
= &finst
->Src
[i
];
487 if (fsrc
->Register
.File
== TGSI_FILE_TEMPORARY
) {
488 src
[i
] = tgsi_src(fpc
, fsrc
);
492 for (i
= 0; i
< finst
->Instruction
.NumSrcRegs
; i
++) {
493 const struct tgsi_full_src_register
*fsrc
;
495 fsrc
= &finst
->Src
[i
];
497 switch (fsrc
->Register
.File
) {
498 case TGSI_FILE_INPUT
:
499 if (ai
== -1 || ai
== fsrc
->Register
.Index
) {
500 ai
= fsrc
->Register
.Index
;
501 src
[i
] = tgsi_src(fpc
, fsrc
);
503 src
[i
] = nvfx_src(temp(fpc
));
504 nvfx_fp_emit(fpc
, arith(0, MOV
, src
[i
].reg
, NVFX_FP_MASK_ALL
, tgsi_src(fpc
, fsrc
), none
, none
));
507 case TGSI_FILE_CONSTANT
:
508 if ((ci
== -1 && ii
== -1) ||
509 ci
== fsrc
->Register
.Index
) {
510 ci
= fsrc
->Register
.Index
;
511 src
[i
] = tgsi_src(fpc
, fsrc
);
513 src
[i
] = nvfx_src(temp(fpc
));
514 nvfx_fp_emit(fpc
, arith(0, MOV
, src
[i
].reg
, NVFX_FP_MASK_ALL
, tgsi_src(fpc
, fsrc
), none
, none
));
517 case TGSI_FILE_IMMEDIATE
:
518 if ((ci
== -1 && ii
== -1) ||
519 ii
== fsrc
->Register
.Index
) {
520 ii
= fsrc
->Register
.Index
;
521 src
[i
] = tgsi_src(fpc
, fsrc
);
523 src
[i
] = nvfx_src(temp(fpc
));
524 nvfx_fp_emit(fpc
, arith(0, MOV
, src
[i
].reg
, NVFX_FP_MASK_ALL
, tgsi_src(fpc
, fsrc
), none
, none
));
527 case TGSI_FILE_TEMPORARY
:
530 case TGSI_FILE_SAMPLER
:
531 unit
= fsrc
->Register
.Index
;
533 case TGSI_FILE_OUTPUT
:
536 NOUVEAU_ERR("bad src file\n");
541 dst
= tgsi_dst(fpc
, &finst
->Dst
[0]);
542 mask
= tgsi_mask(finst
->Dst
[0].Register
.WriteMask
);
543 sat
= (finst
->Instruction
.Saturate
== TGSI_SAT_ZERO_ONE
);
545 switch (finst
->Instruction
.Opcode
) {
546 case TGSI_OPCODE_ABS
:
547 nvfx_fp_emit(fpc
, arith(sat
, MOV
, dst
, mask
, abs(src
[0]), none
, none
));
549 case TGSI_OPCODE_ADD
:
550 nvfx_fp_emit(fpc
, arith(sat
, ADD
, dst
, mask
, src
[0], src
[1], none
));
552 case TGSI_OPCODE_CMP
:
553 insn
= arith(0, MOV
, none
.reg
, mask
, src
[0], none
, none
);
555 nvfx_fp_emit(fpc
, insn
);
557 insn
= arith(sat
, MOV
, dst
, mask
, src
[2], none
, none
);
558 insn
.cc_test
= NVFX_COND_GE
;
559 nvfx_fp_emit(fpc
, insn
);
561 insn
= arith(sat
, MOV
, dst
, mask
, src
[1], none
, none
);
562 insn
.cc_test
= NVFX_COND_LT
;
563 nvfx_fp_emit(fpc
, insn
);
565 case TGSI_OPCODE_COS
:
566 nvfx_fp_emit(fpc
, arith(sat
, COS
, dst
, mask
, src
[0], none
, none
));
568 case TGSI_OPCODE_DDX
:
569 if (mask
& (NVFX_FP_MASK_Z
| NVFX_FP_MASK_W
)) {
570 tmp
= nvfx_src(temp(fpc
));
571 nvfx_fp_emit(fpc
, arith(sat
, DDX
, tmp
.reg
, NVFX_FP_MASK_X
| NVFX_FP_MASK_Y
, swz(src
[0], Z
, W
, Z
, W
), none
, none
));
572 nvfx_fp_emit(fpc
, arith(0, MOV
, tmp
.reg
, NVFX_FP_MASK_Z
| NVFX_FP_MASK_W
, swz(tmp
, X
, Y
, X
, Y
), none
, none
));
573 nvfx_fp_emit(fpc
, arith(sat
, DDX
, tmp
.reg
, NVFX_FP_MASK_X
| NVFX_FP_MASK_Y
, src
[0], none
, none
));
574 nvfx_fp_emit(fpc
, arith(0, MOV
, dst
, mask
, tmp
, none
, none
));
576 nvfx_fp_emit(fpc
, arith(sat
, DDX
, dst
, mask
, src
[0], none
, none
));
579 case TGSI_OPCODE_DDY
:
580 if (mask
& (NVFX_FP_MASK_Z
| NVFX_FP_MASK_W
)) {
581 tmp
= nvfx_src(temp(fpc
));
582 nvfx_fp_emit(fpc
, arith(sat
, DDY
, tmp
.reg
, NVFX_FP_MASK_X
| NVFX_FP_MASK_Y
, swz(src
[0], Z
, W
, Z
, W
), none
, none
));
583 nvfx_fp_emit(fpc
, arith(0, MOV
, tmp
.reg
, NVFX_FP_MASK_Z
| NVFX_FP_MASK_W
, swz(tmp
, X
, Y
, X
, Y
), none
, none
));
584 nvfx_fp_emit(fpc
, arith(sat
, DDY
, tmp
.reg
, NVFX_FP_MASK_X
| NVFX_FP_MASK_Y
, src
[0], none
, none
));
585 nvfx_fp_emit(fpc
, arith(0, MOV
, dst
, mask
, tmp
, none
, none
));
587 nvfx_fp_emit(fpc
, arith(sat
, DDY
, dst
, mask
, src
[0], none
, none
));
590 case TGSI_OPCODE_DP2
:
591 tmp
= nvfx_src(temp(fpc
));
592 nvfx_fp_emit(fpc
, arith(0, MUL
, tmp
.reg
, NVFX_FP_MASK_X
| NVFX_FP_MASK_Y
, src
[0], src
[1], none
));
593 nvfx_fp_emit(fpc
, arith(0, ADD
, dst
, mask
, swz(tmp
, X
, X
, X
, X
), swz(tmp
, Y
, Y
, Y
, Y
), none
));
595 case TGSI_OPCODE_DP3
:
596 nvfx_fp_emit(fpc
, arith(sat
, DP3
, dst
, mask
, src
[0], src
[1], none
));
598 case TGSI_OPCODE_DP4
:
599 nvfx_fp_emit(fpc
, arith(sat
, DP4
, dst
, mask
, src
[0], src
[1], none
));
601 case TGSI_OPCODE_DPH
:
602 tmp
= nvfx_src(temp(fpc
));
603 nvfx_fp_emit(fpc
, arith(0, DP3
, tmp
.reg
, NVFX_FP_MASK_X
, src
[0], src
[1], none
));
604 nvfx_fp_emit(fpc
, arith(sat
, ADD
, dst
, mask
, swz(tmp
, X
, X
, X
, X
), swz(src
[1], W
, W
, W
, W
), none
));
606 case TGSI_OPCODE_DST
:
607 nvfx_fp_emit(fpc
, arith(sat
, DST
, dst
, mask
, src
[0], src
[1], none
));
609 case TGSI_OPCODE_EX2
:
610 nvfx_fp_emit(fpc
, arith(sat
, EX2
, dst
, mask
, src
[0], none
, none
));
612 case TGSI_OPCODE_FLR
:
613 nvfx_fp_emit(fpc
, arith(sat
, FLR
, dst
, mask
, src
[0], none
, none
));
615 case TGSI_OPCODE_FRC
:
616 nvfx_fp_emit(fpc
, arith(sat
, FRC
, dst
, mask
, src
[0], none
, none
));
618 case TGSI_OPCODE_KILP
:
619 nvfx_fp_emit(fpc
, arith(0, KIL
, none
.reg
, 0, none
, none
, none
));
621 case TGSI_OPCODE_KIL
:
622 insn
= arith(0, MOV
, none
.reg
, NVFX_FP_MASK_ALL
, src
[0], none
, none
);
624 nvfx_fp_emit(fpc
, insn
);
626 insn
= arith(0, KIL
, none
.reg
, 0, none
, none
, none
);
627 insn
.cc_test
= NVFX_COND_LT
;
628 nvfx_fp_emit(fpc
, insn
);
630 case TGSI_OPCODE_LG2
:
631 nvfx_fp_emit(fpc
, arith(sat
, LG2
, dst
, mask
, src
[0], none
, none
));
633 case TGSI_OPCODE_LIT
:
635 nvfx_fp_emit(fpc
, arith(sat
, LIT_NV30
, dst
, mask
, src
[0], src
[1], src
[2]));
637 /* we use FLT_MIN, so that log2 never gives -infinity, and thus multiplication by
638 * specular 0 always gives 0, so that ex2 gives 1, to satisfy the 0^0 = 1 requirement
640 * NOTE: if we start using half precision, we might need an fp16 FLT_MIN here instead
642 float maxv
[4] = {0, FLT_MIN
, 0, 0};
643 struct nvfx_src maxs
= nvfx_src(constant(fpc
, -1, maxv
));
644 tmp
= nvfx_src(temp(fpc
));
645 if (ci
>= 0 || ii
>= 0) {
646 nvfx_fp_emit(fpc
, arith(0, MOV
, tmp
.reg
, NVFX_FP_MASK_X
| NVFX_FP_MASK_Y
, maxs
, none
, none
));
649 nvfx_fp_emit(fpc
, arith(0, MAX
, tmp
.reg
, NVFX_FP_MASK_Y
| NVFX_FP_MASK_W
, swz(src
[0], X
, X
, X
, Y
), swz(maxs
, X
, X
, Y
, Y
), none
));
650 nvfx_fp_emit(fpc
, arith(0, LG2
, tmp
.reg
, NVFX_FP_MASK_W
, swz(tmp
, W
, W
, W
, W
), none
, none
));
651 nvfx_fp_emit(fpc
, arith(0, MUL
, tmp
.reg
, NVFX_FP_MASK_W
, swz(tmp
, W
, W
, W
, W
), swz(src
[0], W
, W
, W
, W
), none
));
652 nvfx_fp_emit(fpc
, arith(sat
, LITEX2_NV40
, dst
, mask
, swz(tmp
, Y
, Y
, W
, W
), none
, none
));
655 case TGSI_OPCODE_LRP
:
657 nvfx_fp_emit(fpc
, arith(sat
, LRP_NV30
, dst
, mask
, src
[0], src
[1], src
[2]));
659 tmp
= nvfx_src(temp(fpc
));
660 nvfx_fp_emit(fpc
, arith(0, MAD
, tmp
.reg
, mask
, neg(src
[0]), src
[2], src
[2]));
661 nvfx_fp_emit(fpc
, arith(sat
, MAD
, dst
, mask
, src
[0], src
[1], tmp
));
664 case TGSI_OPCODE_MAD
:
665 nvfx_fp_emit(fpc
, arith(sat
, MAD
, dst
, mask
, src
[0], src
[1], src
[2]));
667 case TGSI_OPCODE_MAX
:
668 nvfx_fp_emit(fpc
, arith(sat
, MAX
, dst
, mask
, src
[0], src
[1], none
));
670 case TGSI_OPCODE_MIN
:
671 nvfx_fp_emit(fpc
, arith(sat
, MIN
, dst
, mask
, src
[0], src
[1], none
));
673 case TGSI_OPCODE_MOV
:
674 nvfx_fp_emit(fpc
, arith(sat
, MOV
, dst
, mask
, src
[0], none
, none
));
676 case TGSI_OPCODE_MUL
:
677 nvfx_fp_emit(fpc
, arith(sat
, MUL
, dst
, mask
, src
[0], src
[1], none
));
679 case TGSI_OPCODE_NOP
:
681 case TGSI_OPCODE_POW
:
683 nvfx_fp_emit(fpc
, arith(sat
, POW_NV30
, dst
, mask
, src
[0], src
[1], none
));
685 tmp
= nvfx_src(temp(fpc
));
686 nvfx_fp_emit(fpc
, arith(0, LG2
, tmp
.reg
, NVFX_FP_MASK_X
, swz(src
[0], X
, X
, X
, X
), none
, none
));
687 nvfx_fp_emit(fpc
, arith(0, MUL
, tmp
.reg
, NVFX_FP_MASK_X
, swz(tmp
, X
, X
, X
, X
), swz(src
[1], X
, X
, X
, X
), none
));
688 nvfx_fp_emit(fpc
, arith(sat
, EX2
, dst
, mask
, swz(tmp
, X
, X
, X
, X
), none
, none
));
691 case TGSI_OPCODE_RCP
:
692 nvfx_fp_emit(fpc
, arith(sat
, RCP
, dst
, mask
, src
[0], none
, none
));
694 case TGSI_OPCODE_RFL
:
696 nvfx_fp_emit(fpc
, arith(0, RFL_NV30
, dst
, mask
, src
[0], src
[1], none
));
698 tmp
= nvfx_src(temp(fpc
));
699 nvfx_fp_emit(fpc
, arith(0, DP3
, tmp
.reg
, NVFX_FP_MASK_X
, src
[0], src
[0], none
));
700 nvfx_fp_emit(fpc
, arith(0, DP3
, tmp
.reg
, NVFX_FP_MASK_Y
, src
[0], src
[1], none
));
701 insn
= arith(0, DIV
, tmp
.reg
, NVFX_FP_MASK_Z
, swz(tmp
, Y
, Y
, Y
, Y
), swz(tmp
, X
, X
, X
, X
), none
);
702 insn
.scale
= NVFX_FP_OP_DST_SCALE_2X
;
703 nvfx_fp_emit(fpc
, insn
);
704 nvfx_fp_emit(fpc
, arith(sat
, MAD
, dst
, mask
, swz(tmp
, Z
, Z
, Z
, Z
), src
[0], neg(src
[1])));
707 case TGSI_OPCODE_RSQ
:
709 nvfx_fp_emit(fpc
, arith(sat
, RSQ_NV30
, dst
, mask
, abs(swz(src
[0], X
, X
, X
, X
)), none
, none
));
711 tmp
= nvfx_src(temp(fpc
));
712 insn
= arith(0, LG2
, tmp
.reg
, NVFX_FP_MASK_X
, abs(swz(src
[0], X
, X
, X
, X
)), none
, none
);
713 insn
.scale
= NVFX_FP_OP_DST_SCALE_INV_2X
;
714 nvfx_fp_emit(fpc
, insn
);
715 nvfx_fp_emit(fpc
, arith(sat
, EX2
, dst
, mask
, neg(swz(tmp
, X
, X
, X
, X
)), none
, none
));
718 case TGSI_OPCODE_SCS
:
719 /* avoid overwriting the source */
720 if(src
[0].swz
[NVFX_SWZ_X
] != NVFX_SWZ_X
)
722 if (mask
& NVFX_FP_MASK_X
)
723 nvfx_fp_emit(fpc
, arith(sat
, COS
, dst
, NVFX_FP_MASK_X
, swz(src
[0], X
, X
, X
, X
), none
, none
));
724 if (mask
& NVFX_FP_MASK_Y
)
725 nvfx_fp_emit(fpc
, arith(sat
, SIN
, dst
, NVFX_FP_MASK_Y
, swz(src
[0], X
, X
, X
, X
), none
, none
));
729 if (mask
& NVFX_FP_MASK_Y
)
730 nvfx_fp_emit(fpc
, arith(sat
, SIN
, dst
, NVFX_FP_MASK_Y
, swz(src
[0], X
, X
, X
, X
), none
, none
));
731 if (mask
& NVFX_FP_MASK_X
)
732 nvfx_fp_emit(fpc
, arith(sat
, COS
, dst
, NVFX_FP_MASK_X
, swz(src
[0], X
, X
, X
, X
), none
, none
));
735 case TGSI_OPCODE_SEQ
:
736 nvfx_fp_emit(fpc
, arith(sat
, SEQ
, dst
, mask
, src
[0], src
[1], none
));
738 case TGSI_OPCODE_SFL
:
739 nvfx_fp_emit(fpc
, arith(sat
, SFL
, dst
, mask
, src
[0], src
[1], none
));
741 case TGSI_OPCODE_SGE
:
742 nvfx_fp_emit(fpc
, arith(sat
, SGE
, dst
, mask
, src
[0], src
[1], none
));
744 case TGSI_OPCODE_SGT
:
745 nvfx_fp_emit(fpc
, arith(sat
, SGT
, dst
, mask
, src
[0], src
[1], none
));
747 case TGSI_OPCODE_SIN
:
748 nvfx_fp_emit(fpc
, arith(sat
, SIN
, dst
, mask
, src
[0], none
, none
));
750 case TGSI_OPCODE_SLE
:
751 nvfx_fp_emit(fpc
, arith(sat
, SLE
, dst
, mask
, src
[0], src
[1], none
));
753 case TGSI_OPCODE_SLT
:
754 nvfx_fp_emit(fpc
, arith(sat
, SLT
, dst
, mask
, src
[0], src
[1], none
));
756 case TGSI_OPCODE_SNE
:
757 nvfx_fp_emit(fpc
, arith(sat
, SNE
, dst
, mask
, src
[0], src
[1], none
));
759 case TGSI_OPCODE_SSG
:
761 float minonesv
[4] = {-1.0, -1.0, -1.0, -1.0};
762 struct nvfx_src minones
= swz(nvfx_src(constant(fpc
, -1, minonesv
)), X
, X
, X
, X
);
764 insn
= arith(sat
, MOV
, dst
, mask
, src
[0], none
, none
);
766 nvfx_fp_emit(fpc
, insn
);
768 insn
= arith(0, STR
, dst
, mask
, none
, none
, none
);
769 insn
.cc_test
= NVFX_COND_GT
;
770 nvfx_fp_emit(fpc
, insn
);
773 insn
= arith(0, MOV
, dst
, mask
, minones
, none
, none
);
774 insn
.cc_test
= NVFX_COND_LT
;
775 nvfx_fp_emit(fpc
, insn
);
779 case TGSI_OPCODE_STR
:
780 nvfx_fp_emit(fpc
, arith(sat
, STR
, dst
, mask
, src
[0], src
[1], none
));
782 case TGSI_OPCODE_SUB
:
783 nvfx_fp_emit(fpc
, arith(sat
, ADD
, dst
, mask
, src
[0], neg(src
[1]), none
));
785 case TGSI_OPCODE_TEX
:
786 nvfx_fp_emit(fpc
, tex(sat
, TEX
, unit
, dst
, mask
, src
[0], none
, none
));
788 case TGSI_OPCODE_TRUNC
:
789 tmp
= nvfx_src(temp(fpc
));
790 insn
= arith(0, MOV
, none
.reg
, mask
, src
[0], none
, none
);
792 nvfx_fp_emit(fpc
, insn
);
794 nvfx_fp_emit(fpc
, arith(0, FLR
, tmp
.reg
, mask
, abs(src
[0]), none
, none
));
795 nvfx_fp_emit(fpc
, arith(sat
, MOV
, dst
, mask
, tmp
, none
, none
));
797 insn
= arith(sat
, MOV
, dst
, mask
, neg(tmp
), none
, none
);
798 insn
.cc_test
= NVFX_COND_LT
;
799 nvfx_fp_emit(fpc
, insn
);
801 case TGSI_OPCODE_TXB
:
802 nvfx_fp_emit(fpc
, tex(sat
, TXB
, unit
, dst
, mask
, src
[0], none
, none
));
804 case TGSI_OPCODE_TXL
:
806 nvfx_fp_emit(fpc
, tex(sat
, TXL_NV40
, unit
, dst
, mask
, src
[0], none
, none
));
807 else /* unsupported on nv30, use TEX and hope they like it */
808 nvfx_fp_emit(fpc
, tex(sat
, TEX
, unit
, dst
, mask
, src
[0], none
, none
));
810 case TGSI_OPCODE_TXP
:
811 nvfx_fp_emit(fpc
, tex(sat
, TXP
, unit
, dst
, mask
, src
[0], none
, none
));
813 case TGSI_OPCODE_XPD
:
814 tmp
= nvfx_src(temp(fpc
));
815 nvfx_fp_emit(fpc
, arith(0, MUL
, tmp
.reg
, mask
, swz(src
[0], Z
, X
, Y
, Y
), swz(src
[1], Y
, Z
, X
, X
), none
));
816 nvfx_fp_emit(fpc
, arith(sat
, MAD
, dst
, (mask
& ~NVFX_FP_MASK_W
), swz(src
[0], Y
, Z
, X
, X
), swz(src
[1], Z
, X
, Y
, Y
), neg(tmp
)));
820 // MOVRC0 R31 (TR0.xyzw), R<src>:
821 // IF (NE.xxxx) ELSE <else> END <end>
824 nv40_fp_if(fpc
, src
[0]);
827 case TGSI_OPCODE_ELSE
:
832 assert(util_dynarray_contains(&fpc
->if_stack
, unsigned));
833 hw
= &fpc
->fp
->insn
[util_dynarray_top(&fpc
->if_stack
, unsigned)];
834 hw
[2] = NV40_FP_OP_OPCODE_IS_BRANCH
| fpc
->fp
->insn_len
;
838 case TGSI_OPCODE_ENDIF
:
843 assert(util_dynarray_contains(&fpc
->if_stack
, unsigned));
844 hw
= &fpc
->fp
->insn
[util_dynarray_pop(&fpc
->if_stack
, unsigned)];
846 hw
[2] = NV40_FP_OP_OPCODE_IS_BRANCH
| fpc
->fp
->insn_len
;
847 hw
[3] = fpc
->fp
->insn_len
;
851 case TGSI_OPCODE_BRA
:
852 /* This can in limited cases be implemented with an IF with the else and endif labels pointing to the target */
853 /* no state tracker uses this, so don't implement this for now */
855 nv40_fp_bra(fpc
, finst
->Label
.Label
);
858 case TGSI_OPCODE_BGNSUB
:
859 case TGSI_OPCODE_ENDSUB
:
860 /* nothing to do here */
863 case TGSI_OPCODE_CAL
:
866 nv40_fp_cal(fpc
, finst
->Label
.Label
);
869 case TGSI_OPCODE_RET
:
875 case TGSI_OPCODE_BGNLOOP
:
878 /* TODO: we should support using two nested REPs to allow a > 255 iteration count */
879 nv40_fp_rep(fpc
, 255, finst
->Label
.Label
);
882 case TGSI_OPCODE_ENDLOOP
:
885 case TGSI_OPCODE_BRK
:
891 case TGSI_OPCODE_CONT
:
893 static int warned
= 0;
895 NOUVEAU_ERR("Sorry, the continue keyword is not implemented: ignoring it.\n");
902 NOUVEAU_ERR("invalid opcode %d\n", finst
->Instruction
.Opcode
);
911 static int warned
= 0;
914 "Sorry, control flow instructions are not supported in hardware on nv3x: ignoring them\n"
915 "If rendering is incorrect, try to disable GLSL support in the application.\n");
923 nvfx_fragprog_parse_decl_output(struct nvfx_context
* nvfx
, struct nvfx_fpc
*fpc
,
924 const struct tgsi_full_declaration
*fdec
)
926 unsigned idx
= fdec
->Range
.First
;
929 switch (fdec
->Semantic
.Name
) {
930 case TGSI_SEMANTIC_POSITION
:
933 case TGSI_SEMANTIC_COLOR
:
935 switch (fdec
->Semantic
.Index
) {
936 case 0: hw
= 0; break;
937 case 1: hw
= 2; break;
938 case 2: hw
= 3; break;
939 case 3: hw
= 4; break;
941 if(hw
> ((nvfx
->is_nv4x
) ? 4 : 2)) {
942 NOUVEAU_ERR("bad rcol index\n");
947 NOUVEAU_ERR("bad output semantic\n");
951 fpc
->r_result
[idx
] = nvfx_reg(NVFXSR_OUTPUT
, hw
);
952 fpc
->r_temps
|= (1ULL << hw
);
957 nvfx_fragprog_prepare(struct nvfx_context
* nvfx
, struct nvfx_fpc
*fpc
)
959 struct tgsi_parse_context p
;
960 int high_temp
= -1, i
;
961 struct util_semantic_set set
;
962 float const0v
[4] = {0, 0, 0, 0};
963 struct nvfx_reg const0
;
964 unsigned num_texcoords
= nvfx
->is_nv4x
? 10 : 8;
966 fpc
->fp
->num_slots
= util_semantic_set_from_program_file(&set
, fpc
->pfp
->pipe
.tokens
, TGSI_FILE_INPUT
);
967 if(fpc
->fp
->num_slots
> num_texcoords
)
969 util_semantic_layout_from_set(fpc
->fp
->slot_to_generic
, &set
, 0, num_texcoords
);
970 util_semantic_table_from_layout(fpc
->generic_to_slot
, fpc
->fp
->slot_to_generic
, 0, num_texcoords
);
972 memset(fpc
->fp
->slot_to_fp_input
, 0xff, sizeof(fpc
->fp
->slot_to_fp_input
));
974 const0
= constant(fpc
, -1, const0v
);
975 assert(const0
.index
== 0);
977 tgsi_parse_init(&p
, fpc
->pfp
->pipe
.tokens
);
978 while (!tgsi_parse_end_of_tokens(&p
)) {
979 const union tgsi_full_token
*tok
= &p
.FullToken
;
981 tgsi_parse_token(&p
);
982 switch(tok
->Token
.Type
) {
983 case TGSI_TOKEN_TYPE_DECLARATION
:
985 const struct tgsi_full_declaration
*fdec
;
986 fdec
= &p
.FullToken
.FullDeclaration
;
987 switch (fdec
->Declaration
.File
) {
988 case TGSI_FILE_OUTPUT
:
989 if (!nvfx_fragprog_parse_decl_output(nvfx
, fpc
, fdec
))
992 case TGSI_FILE_TEMPORARY
:
993 if (fdec
->Range
.Last
> high_temp
) {
1003 case TGSI_TOKEN_TYPE_IMMEDIATE
:
1005 struct tgsi_full_immediate
*imm
;
1008 imm
= &p
.FullToken
.FullImmediate
;
1009 assert(imm
->Immediate
.DataType
== TGSI_IMM_FLOAT32
);
1010 assert(fpc
->nr_imm
< MAX_IMM
);
1012 vals
[0] = imm
->u
[0].Float
;
1013 vals
[1] = imm
->u
[1].Float
;
1014 vals
[2] = imm
->u
[2].Float
;
1015 vals
[3] = imm
->u
[3].Float
;
1016 fpc
->imm
[fpc
->nr_imm
++] = constant(fpc
, -1, vals
);
1023 tgsi_parse_free(&p
);
1026 fpc
->r_temp
= CALLOC(high_temp
, sizeof(struct nvfx_reg
));
1027 for (i
= 0; i
< high_temp
; i
++)
1028 fpc
->r_temp
[i
] = temp(fpc
);
1029 fpc
->r_temps_discard
= 0ULL;
1039 tgsi_parse_free(&p
);
1043 DEBUG_GET_ONCE_BOOL_OPTION(nvfx_dump_fp
, "NVFX_DUMP_FP", FALSE
)
1045 static struct nvfx_fragment_program
*
1046 nvfx_fragprog_translate(struct nvfx_context
*nvfx
,
1047 struct nvfx_pipe_fragment_program
*pfp
,
1048 boolean emulate_sprite_flipping
)
1050 struct tgsi_parse_context parse
;
1051 struct nvfx_fpc
*fpc
= NULL
;
1052 struct util_dynarray insns
;
1053 struct nvfx_fragment_program
* fp
= NULL
;
1054 const int min_size
= 4096;
1056 fp
= CALLOC_STRUCT(nvfx_fragment_program
);
1060 fpc
= CALLOC_STRUCT(nvfx_fpc
);
1064 fpc
->max_temps
= nvfx
->is_nv4x
? 48 : 32;
1069 for (unsigned i
= 0; i
< pfp
->info
.num_properties
; ++i
) {
1070 if (pfp
->info
.properties
[i
].name
== TGSI_PROPERTY_FS_COORD_ORIGIN
) {
1071 if(pfp
->info
.properties
[i
].data
[0])
1072 fp
->coord_conventions
|= NV34TCL_COORD_CONVENTIONS_ORIGIN_INVERTED
;
1073 } else if (pfp
->info
.properties
[i
].name
== TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
) {
1074 if(pfp
->info
.properties
[i
].data
[0])
1075 fp
->coord_conventions
|= NV34TCL_COORD_CONVENTIONS_CENTER_INTEGER
;
1079 if (!nvfx_fragprog_prepare(nvfx
, fpc
))
1082 tgsi_parse_init(&parse
, pfp
->pipe
.tokens
);
1083 util_dynarray_init(&insns
);
1085 if(emulate_sprite_flipping
)
1087 struct nvfx_reg reg
= temp(fpc
);
1088 struct nvfx_src sprite_input
= nvfx_src(nvfx_reg(NVFXSR_RELOCATED
, fp
->num_slots
));
1089 float v
[4] = {1, -1, 0, 0};
1090 struct nvfx_src imm
= nvfx_src(constant(fpc
, -1, v
));
1092 fpc
->sprite_coord_temp
= reg
.index
;
1093 fpc
->r_temps_discard
= 0ULL;
1094 nvfx_fp_emit(fpc
, arith(0, MAD
, reg
, NVFX_FP_MASK_ALL
, sprite_input
, swz(imm
, X
, Y
, X
, X
), swz(imm
, Z
, X
, Z
, Z
)));
1097 while (!tgsi_parse_end_of_tokens(&parse
)) {
1098 tgsi_parse_token(&parse
);
1100 switch (parse
.FullToken
.Token
.Type
) {
1101 case TGSI_TOKEN_TYPE_INSTRUCTION
:
1103 const struct tgsi_full_instruction
*finst
;
1105 util_dynarray_append(&insns
, unsigned, fp
->insn_len
);
1106 finst
= &parse
.FullToken
.FullInstruction
;
1107 if (!nvfx_fragprog_parse_instruction(nvfx
, fpc
, finst
))
1115 util_dynarray_append(&insns
, unsigned, fp
->insn_len
);
1117 for(unsigned i
= 0; i
< fpc
->label_relocs
.size
; i
+= sizeof(struct nvfx_relocation
))
1119 struct nvfx_relocation
* label_reloc
= (struct nvfx_relocation
*)((char*)fpc
->label_relocs
.data
+ i
);
1120 fp
->insn
[label_reloc
->location
] |= ((unsigned*)insns
.data
)[label_reloc
->target
];
1122 util_dynarray_fini(&insns
);
1125 fp
->fp_control
|= (fpc
->num_regs
-1)/2;
1127 fp
->fp_control
|= fpc
->num_regs
<< NV40TCL_FP_CONTROL_TEMP_COUNT_SHIFT
;
1129 /* Terminate final instruction */
1131 fp
->insn
[fpc
->inst_offset
] |= 0x00000001;
1133 /* Append NOP + END instruction for branches to the end of the program */
1134 fpc
->inst_offset
= fp
->insn_len
;
1136 fp
->insn
[fpc
->inst_offset
+ 0] = 0x00000001;
1137 fp
->insn
[fpc
->inst_offset
+ 1] = 0x00000000;
1138 fp
->insn
[fpc
->inst_offset
+ 2] = 0x00000000;
1139 fp
->insn
[fpc
->inst_offset
+ 3] = 0x00000000;
1141 if(debug_get_option_nvfx_dump_fp())
1144 tgsi_dump(pfp
->pipe
.tokens
, 0);
1146 debug_printf("\n%s fragment program:\n", nvfx
->is_nv4x
? "nv4x" : "nv3x");
1147 for (unsigned i
= 0; i
< fp
->insn_len
; i
+= 4)
1148 debug_printf("%3u: %08x %08x %08x %08x\n", i
>> 2, fp
->insn
[i
], fp
->insn
[i
+ 1], fp
->insn
[i
+ 2], fp
->insn
[i
+ 3]);
1152 fp
->prog_size
= (fp
->insn_len
* 4 + 63) & ~63;
1154 if(fp
->prog_size
>= min_size
)
1155 fp
->progs_per_bo
= 1;
1157 fp
->progs_per_bo
= min_size
/ fp
->prog_size
;
1158 fp
->bo_prog_idx
= fp
->progs_per_bo
- 1;
1161 tgsi_parse_free(&parse
);
1166 util_dynarray_fini(&fpc
->if_stack
);
1167 util_dynarray_fini(&fpc
->label_relocs
);
1168 //util_dynarray_fini(&fpc->loop_stack);
1174 _debug_printf("Error: failed to compile this fragment program:\n");
1175 tgsi_dump(pfp
->pipe
.tokens
, 0);
1186 nvfx_fp_memcpy(void* dst
, const void* src
, size_t len
)
1188 #ifndef WORDS_BIGENDIAN
1189 memcpy(dst
, src
, len
);
1192 for(i
= 0; i
< len
; i
+= 4) {
1193 uint32_t v
= (uint32_t*)((char*)src
+ i
);
1194 *(uint32_t*)((char*)dst
+ i
) = (v
>> 16) | (v
<< 16);
1199 /* The hardware only supports immediate constants inside the fragment program,
1200 * and at least on nv30 doesn't support an indirect linkage table.
1202 * Hence, we need to patch the fragment program itself both to update constants
1203 * and update linkage.
1205 * Using a single fragment program would entail unacceptable stalls if the GPU is
1206 * already rendering with that fragment program.
1207 * Thus, we instead use a "rotating queue" of buffer objects, each of which is
1208 * packed with multiple versions of the same program.
1210 * Whenever we need to patch something, we move to the next program and
1211 * patch it. If all buffer objects are in use by the GPU, we allocate another one,
1212 * expanding the queue.
1214 * As an additional optimization, we record when all the programs have the
1215 * current input slot configuration, and at that point we stop patching inputs.
1216 * This happens, for instance, if a given fragment program is always used with
1217 * the same vertex program (i.e. always with GLSL), or if the layouts match
1218 * enough (non-GLSL).
1220 * Note that instead of using multiple programs, we could push commands
1221 * on the FIFO to patch a single program: it's not fully clear which option is
1222 * faster, but my guess is that the current way is faster.
1224 * We also track the previous slot assignments for each version and don't
1225 * patch if they are the same (this could perhaps be removed).
1229 nvfx_fragprog_validate(struct nvfx_context
*nvfx
)
1231 struct nouveau_channel
* chan
= nvfx
->screen
->base
.channel
;
1232 struct nvfx_pipe_fragment_program
*pfp
= nvfx
->fragprog
;
1233 struct nvfx_vertex_program
* vp
;
1234 /* Gallium always puts the point coord in GENERIC[0]
1235 * TODO: this is wrong, Gallium needs to be fixed
1237 unsigned sprite_coord_enable
= nvfx
->rasterizer
->pipe
.point_quad_rasterization
* (nvfx
->rasterizer
->pipe
.sprite_coord_enable
| 1);
1239 boolean emulate_sprite_flipping
= sprite_coord_enable
&& nvfx
->rasterizer
->pipe
.sprite_coord_mode
;
1240 unsigned key
= emulate_sprite_flipping
;
1241 struct nvfx_fragment_program
* fp
;
1246 fp
= nvfx_fragprog_translate(nvfx
, pfp
, emulate_sprite_flipping
);
1252 struct ureg_program
*ureg
= ureg_create( TGSI_PROCESSOR_FRAGMENT
);
1256 nvfx
->dummy_fs
= ureg_create_shader_and_destroy( ureg
, &nvfx
->pipe
);
1261 _debug_printf("Error: unable to create a dummy fragment shader: aborting.");
1266 fp
= nvfx_fragprog_translate(nvfx
, nvfx
->dummy_fs
, FALSE
);
1267 emulate_sprite_flipping
= FALSE
;
1271 _debug_printf("Error: unable to compile even a dummy fragment shader: aborting.");
1279 vp
= nvfx
->render_mode
== HW
? nvfx
->vertprog
: nvfx
->swtnl
.vertprog
;
1281 if (fp
->last_vp_id
!= vp
->id
|| fp
->last_sprite_coord_enable
!= sprite_coord_enable
) {
1282 int sprite_real_input
= -1;
1283 int sprite_reloc_input
;
1285 fp
->last_vp_id
= vp
->id
;
1286 fp
->last_sprite_coord_enable
= sprite_coord_enable
;
1288 if(sprite_coord_enable
)
1290 sprite_real_input
= vp
->sprite_fp_input
;
1291 if(sprite_real_input
< 0)
1293 unsigned used_texcoords
= 0;
1294 for(unsigned i
= 0; i
< fp
->num_slots
; ++i
) {
1295 unsigned generic
= fp
->slot_to_generic
[i
];
1296 if(!((1 << generic
) & sprite_coord_enable
))
1298 unsigned char slot_mask
= vp
->generic_to_fp_input
[generic
];
1299 if(slot_mask
>= 0xf0)
1300 used_texcoords
|= 1 << ((slot_mask
& 0xf) - NVFX_FP_OP_INPUT_SRC_TC0
);
1304 sprite_real_input
= NVFX_FP_OP_INPUT_SRC_TC(__builtin_ctz(~used_texcoords
));
1307 fp
->point_sprite_control
|= (1 << (sprite_real_input
- NVFX_FP_OP_INPUT_SRC_TC0
+ 8));
1310 fp
->point_sprite_control
= 0;
1312 if(emulate_sprite_flipping
)
1313 sprite_reloc_input
= 0;
1315 sprite_reloc_input
= sprite_real_input
;
1317 for(i
= 0; i
< fp
->num_slots
; ++i
) {
1318 unsigned generic
= fp
->slot_to_generic
[i
];
1319 if((1 << generic
) & sprite_coord_enable
)
1321 if(fp
->slot_to_fp_input
[i
] != sprite_reloc_input
)
1326 unsigned char slot_mask
= vp
->generic_to_fp_input
[generic
];
1327 if((slot_mask
>> 4) & (slot_mask
^ fp
->slot_to_fp_input
[i
]))
1332 if(emulate_sprite_flipping
)
1334 if(fp
->slot_to_fp_input
[fp
->num_slots
] != sprite_real_input
)
1341 /* optimization: we start updating from the slot we found the first difference in */
1342 for(; i
< fp
->num_slots
; ++i
)
1344 unsigned generic
= fp
->slot_to_generic
[i
];
1345 if((1 << generic
) & sprite_coord_enable
)
1346 fp
->slot_to_fp_input
[i
] = sprite_reloc_input
;
1348 fp
->slot_to_fp_input
[i
] = vp
->generic_to_fp_input
[generic
] & 0xf;
1351 fp
->slot_to_fp_input
[fp
->num_slots
] = sprite_real_input
;
1356 for(i
= 0; i
<= fp
->num_slots
; ++i
) {
1357 unsigned fp_input
= fp
->slot_to_fp_input
[i
];
1358 if(fp_input
== NVFX_FP_OP_INPUT_SRC_TC(8))
1359 fp
->or |= (1 << 12);
1360 else if(fp_input
== NVFX_FP_OP_INPUT_SRC_TC(9))
1361 fp
->or |= (1 << 13);
1362 else if(fp_input
>= NVFX_FP_OP_INPUT_SRC_TC(0) && fp_input
<= NVFX_FP_OP_INPUT_SRC_TC(7))
1363 fp
->or |= (1 << (fp_input
- NVFX_FP_OP_INPUT_SRC_TC0
+ 14));
1367 fp
->progs_left_with_obsolete_slot_assignments
= fp
->progs
;
1372 /* We must update constants even on "just" fragprog changes, because
1373 * we don't check whether the current constant buffer matches the latest
1374 * one bound to this fragment program.
1375 * Doing such a check would likely be a pessimization.
1377 if ((nvfx
->hw_fragprog
!= fp
) || (nvfx
->dirty
& (NVFX_NEW_FRAGPROG
| NVFX_NEW_FRAGCONST
))) {
1383 if(fp
->bo_prog_idx
>= fp
->progs_per_bo
)
1385 if(fp
->fpbo
&& !nouveau_bo_busy(fp
->fpbo
->next
->bo
, NOUVEAU_BO_WR
))
1387 fp
->fpbo
= fp
->fpbo
->next
;
1391 struct nvfx_fragment_program_bo
* fpbo
= os_malloc_aligned(sizeof(struct nvfx_fragment_program
) + (fp
->prog_size
+ 8) * fp
->progs_per_bo
, 16);
1395 fpbo
->slots
= (unsigned char*)&fpbo
->insn
[(fp
->prog_size
) * fp
->progs_per_bo
];
1396 memset(fpbo
->slots
, 0, 8 * fp
->progs_per_bo
);
1399 fpbo
->next
= fp
->fpbo
->next
;
1400 fp
->fpbo
->next
= fpbo
;
1406 fp
->progs
+= fp
->progs_per_bo
;
1407 fp
->progs_left_with_obsolete_slot_assignments
+= fp
->progs_per_bo
;
1408 nouveau_bo_new(nvfx
->screen
->base
.device
, NOUVEAU_BO_VRAM
| NOUVEAU_BO_MAP
, 64, fp
->prog_size
* fp
->progs_per_bo
, &fpbo
->bo
);
1409 nouveau_bo_map(fpbo
->bo
, NOUVEAU_BO_NOSYNC
);
1411 map
= fpbo
->bo
->map
;
1412 buf
= (uint8_t*)fpbo
->insn
;
1413 for(unsigned i
= 0; i
< fp
->progs_per_bo
; ++i
)
1415 memcpy(buf
, fp
->insn
, fp
->insn_len
* 4);
1416 nvfx_fp_memcpy(map
, fp
->insn
, fp
->insn_len
* 4);
1417 map
+= fp
->prog_size
;
1418 buf
+= fp
->prog_size
;
1421 fp
->bo_prog_idx
= 0;
1424 offset
= fp
->bo_prog_idx
* fp
->prog_size
;
1425 fpmap
= (uint32_t*)((char*)fp
->fpbo
->bo
->map
+ offset
);
1427 if(nvfx
->constbuf
[PIPE_SHADER_FRAGMENT
]) {
1428 struct pipe_resource
* constbuf
= nvfx
->constbuf
[PIPE_SHADER_FRAGMENT
];
1429 uint32_t* map
= (uint32_t*)nvfx_buffer(constbuf
)->data
;
1430 uint32_t* fpmap
= (uint32_t*)((char*)fp
->fpbo
->bo
->map
+ offset
);
1431 uint32_t* buf
= (uint32_t*)((char*)fp
->fpbo
->insn
+ offset
);
1433 for (i
= 0; i
< fp
->nr_consts
; ++i
) {
1434 unsigned off
= fp
->consts
[i
].offset
;
1435 unsigned idx
= fp
->consts
[i
].index
* 4;
1437 /* TODO: is checking a good idea? */
1438 if(memcmp(&buf
[off
], &map
[idx
], 4 * sizeof(uint32_t))) {
1439 memcpy(&buf
[off
], &map
[idx
], 4 * sizeof(uint32_t));
1440 nvfx_fp_memcpy(&fpmap
[off
], &map
[idx
], 4 * sizeof(uint32_t));
1445 /* we only do this if we aren't sure that all program versions have the
1446 * current slot assignments, otherwise we just update constants for speed
1448 if(fp
->progs_left_with_obsolete_slot_assignments
) {
1449 unsigned char* fpbo_slots
= &fp
->fpbo
->slots
[fp
->bo_prog_idx
* 8];
1450 /* also relocate sprite coord slot, if any */
1451 for(unsigned i
= 0; i
<= fp
->num_slots
; ++i
) {
1452 unsigned value
= fp
->slot_to_fp_input
[i
];;
1453 if(value
!= fpbo_slots
[i
]) {
1455 unsigned* begin
= (unsigned*)fp
->slot_relocations
[i
].data
;
1456 unsigned* end
= (unsigned*)((char*)fp
->slot_relocations
[i
].data
+ fp
->slot_relocations
[i
].size
);
1457 //printf("fp %p reloc slot %u/%u: %u -> %u\n", fp, i, fp->num_slots, fpbo_slots[i], value);
1460 /* was relocated to an input, switch type to temporary */
1461 for(p
= begin
; p
!= end
; ++p
) {
1463 unsigned dw
= fp
->insn
[off
];
1464 dw
&=~ NVFX_FP_REG_TYPE_MASK
;
1465 //printf("reloc_tmp at %x\n", off);
1466 nvfx_fp_memcpy(&fpmap
[off
], &dw
, sizeof(dw
));
1471 /* was relocated to a temporary, switch type to input */
1472 for(p
= begin
; p
!= end
; ++p
) {
1474 unsigned dw
= fp
->insn
[off
];
1475 //printf("reloc_in at %x\n", off);
1476 dw
|= NVFX_FP_REG_TYPE_INPUT
<< NVFX_FP_REG_TYPE_SHIFT
;
1477 nvfx_fp_memcpy(&fpmap
[off
], &dw
, sizeof(dw
));
1481 /* set the correct input index */
1482 for(p
= begin
; p
!= end
; ++p
) {
1483 unsigned off
= *p
& ~3;
1484 unsigned dw
= fp
->insn
[off
];
1485 //printf("reloc&~3 at %x\n", off);
1486 dw
= (dw
& ~NVFX_FP_OP_INPUT_SRC_MASK
) | (value
<< NVFX_FP_OP_INPUT_SRC_SHIFT
);
1487 nvfx_fp_memcpy(&fpmap
[off
], &dw
, sizeof(dw
));
1490 fpbo_slots
[i
] = value
;
1493 --fp
->progs_left_with_obsolete_slot_assignments
;
1496 nvfx
->hw_fragprog
= fp
;
1498 MARK_RING(chan
, 8, 1);
1499 OUT_RING(chan
, RING_3D(NV34TCL_FP_ACTIVE_PROGRAM
, 1));
1500 OUT_RELOC(chan
, fp
->fpbo
->bo
, offset
, NOUVEAU_BO_VRAM
|
1501 NOUVEAU_BO_GART
| NOUVEAU_BO_RD
| NOUVEAU_BO_LOW
|
1502 NOUVEAU_BO_OR
, NV34TCL_FP_ACTIVE_PROGRAM_DMA0
,
1503 NV34TCL_FP_ACTIVE_PROGRAM_DMA1
);
1504 OUT_RING(chan
, RING_3D(NV34TCL_FP_CONTROL
, 1));
1505 OUT_RING(chan
, fp
->fp_control
);
1506 if(!nvfx
->is_nv4x
) {
1507 OUT_RING(chan
, RING_3D(NV34TCL_FP_REG_CONTROL
, 1));
1508 OUT_RING(chan
, (1<<16)|0x4);
1509 OUT_RING(chan
, RING_3D(NV34TCL_TX_UNITS_ENABLE
, 1));
1510 OUT_RING(chan
, fp
->samplers
);
1515 unsigned pointsprite_control
= fp
->point_sprite_control
| nvfx
->rasterizer
->pipe
.point_quad_rasterization
;
1516 if(pointsprite_control
!= nvfx
->hw_pointsprite_control
)
1519 OUT_RING(chan
, RING_3D(NV34TCL_POINT_SPRITE
, 1));
1520 OUT_RING(chan
, pointsprite_control
);
1521 nvfx
->hw_pointsprite_control
= pointsprite_control
;
1525 nvfx
->relocs_needed
&=~ NVFX_RELOCATE_FRAGPROG
;
1529 nvfx_fragprog_relocate(struct nvfx_context
*nvfx
)
1531 struct nouveau_channel
* chan
= nvfx
->screen
->base
.channel
;
1532 struct nvfx_fragment_program
*fp
= nvfx
->hw_fragprog
;
1533 struct nouveau_bo
* bo
= fp
->fpbo
->bo
;
1534 int offset
= fp
->bo_prog_idx
* fp
->prog_size
;
1535 unsigned fp_flags
= NOUVEAU_BO_VRAM
| NOUVEAU_BO_RD
; // TODO: GART?
1536 fp_flags
|= NOUVEAU_BO_DUMMY
;
1537 MARK_RING(chan
, 2, 2);
1538 OUT_RELOC(chan
, bo
, RING_3D(NV34TCL_FP_ACTIVE_PROGRAM
, 1), fp_flags
, 0, 0);
1539 OUT_RELOC(chan
, bo
, offset
, fp_flags
| NOUVEAU_BO_LOW
|
1540 NOUVEAU_BO_OR
, NV34TCL_FP_ACTIVE_PROGRAM_DMA0
,
1541 NV34TCL_FP_ACTIVE_PROGRAM_DMA1
);
1542 nvfx
->relocs_needed
&=~ NVFX_RELOCATE_FRAGPROG
;
1546 nvfx_fragprog_destroy(struct nvfx_context
*nvfx
,
1547 struct nvfx_fragment_program
*fp
)
1550 struct nvfx_fragment_program_bo
* fpbo
= fp
->fpbo
;
1555 struct nvfx_fragment_program_bo
* next
= fpbo
->next
;
1556 nouveau_bo_unmap(fpbo
->bo
);
1557 nouveau_bo_ref(0, &fpbo
->bo
);
1561 while(fpbo
!= fp
->fpbo
);
1564 for(i
= 0; i
< Elements(fp
->slot_relocations
); ++i
)
1565 util_dynarray_fini(&fp
->slot_relocations
[i
]);
1572 nvfx_fp_state_create(struct pipe_context
*pipe
,
1573 const struct pipe_shader_state
*cso
)
1575 struct nvfx_pipe_fragment_program
*pfp
;
1577 pfp
= CALLOC(1, sizeof(struct nvfx_pipe_fragment_program
));
1578 pfp
->pipe
.tokens
= tgsi_dup_tokens(cso
->tokens
);
1580 tgsi_scan_shader(pfp
->pipe
.tokens
, &pfp
->info
);
1586 nvfx_fp_state_bind(struct pipe_context
*pipe
, void *hwcso
)
1588 struct nvfx_context
*nvfx
= nvfx_context(pipe
);
1590 nvfx
->fragprog
= hwcso
;
1591 nvfx
->dirty
|= NVFX_NEW_FRAGPROG
;
1595 nvfx_fp_state_delete(struct pipe_context
*pipe
, void *hwcso
)
1597 struct nvfx_context
*nvfx
= nvfx_context(pipe
);
1598 struct nvfx_pipe_fragment_program
*pfp
= hwcso
;
1601 for(i
= 0; i
< Elements(pfp
->fps
); ++i
)
1605 nvfx_fragprog_destroy(nvfx
, pfp
->fps
[i
]);
1610 FREE((void*)pfp
->pipe
.tokens
);
1615 nvfx_init_fragprog_functions(struct nvfx_context
*nvfx
)
1617 nvfx
->pipe
.create_fs_state
= nvfx_fp_state_create
;
1618 nvfx
->pipe
.bind_fs_state
= nvfx_fp_state_bind
;
1619 nvfx
->pipe
.delete_fs_state
= nvfx_fp_state_delete
;