projects / mesa.git / blob
? search:


6d05df6ae285dbc1a38da19bf8428ffa8f62e47b
[mesa.git] / src / gallium / drivers / r600 / r600_shader.c
1 /*
2 * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
3 *
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 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
10 *
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
13 * Software.
14 *
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 NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 */
23 #include "pipe/p_shader_tokens.h"
24 #include "tgsi/tgsi_info.h"
25 #include "tgsi/tgsi_parse.h"
26 #include "tgsi/tgsi_scan.h"
27 #include "tgsi/tgsi_dump.h"
28 #include "util/u_format.h"
29 #include "r600_pipe.h"
30 #include "r600_asm.h"
31 #include "r600_sq.h"
32 #include "r600_formats.h"
33 #include "r600_opcodes.h"
34 #include "r600d.h"
35 #include <stdio.h>
36 #include <errno.h>
37 #include <byteswap.h>
38
39 /* CAYMAN notes
40 Why CAYMAN got loops for lots of instructions is explained here.
41
42 -These 8xx t-slot only ops are implemented in all vector slots.
43 MUL_LIT, FLT_TO_UINT, INT_TO_FLT, UINT_TO_FLT
44 These 8xx t-slot only opcodes become vector ops, with all four
45 slots expecting the arguments on sources a and b. Result is
46 broadcast to all channels.
47 MULLO_INT, MULHI_INT, MULLO_UINT, MULHI_UINT
48 These 8xx t-slot only opcodes become vector ops in the z, y, and
49 x slots.
50 EXP_IEEE, LOG_IEEE/CLAMPED, RECIP_IEEE/CLAMPED/FF/INT/UINT/_64/CLAMPED_64
51 RECIPSQRT_IEEE/CLAMPED/FF/_64/CLAMPED_64
52 SQRT_IEEE/_64
53 SIN/COS
54 The w slot may have an independent co-issued operation, or if the
55 result is required to be in the w slot, the opcode above may be
56 issued in the w slot as well.
57 The compiler must issue the source argument to slots z, y, and x
58 */
59
60 static int r600_pipe_shader(struct pipe_context *ctx, struct r600_pipe_shader *shader)
61 {
62 struct r600_context *rctx = (struct r600_context *)ctx;
63 struct r600_shader *rshader = &shader->shader;
64 uint32_t *ptr;
65 int i;
66
67 /* copy new shader */
68 if (shader->bo == NULL) {
69 shader->bo = (struct r600_resource*)
70 pipe_buffer_create(ctx->screen, PIPE_BIND_CUSTOM, PIPE_USAGE_IMMUTABLE, rshader->bc.ndw * 4);
71 if (shader->bo == NULL) {
72 return -ENOMEM;
73 }
74 ptr = (uint32_t*)rctx->ws->buffer_map(shader->bo->buf, rctx->cs, PIPE_TRANSFER_WRITE);
75 if (R600_BIG_ENDIAN) {
76 for (i = 0; i < rshader->bc.ndw; ++i) {
77 ptr[i] = bswap_32(rshader->bc.bytecode[i]);
78 }
79 } else {
80 memcpy(ptr, rshader->bc.bytecode, rshader->bc.ndw * sizeof(*ptr));
81 }
82 rctx->ws->buffer_unmap(shader->bo->buf);
83 }
84 /* build state */
85 switch (rshader->processor_type) {
86 case TGSI_PROCESSOR_VERTEX:
87 if (rctx->chip_class >= EVERGREEN) {
88 evergreen_pipe_shader_vs(ctx, shader);
89 } else {
90 r600_pipe_shader_vs(ctx, shader);
91 }
92 break;
93 case TGSI_PROCESSOR_FRAGMENT:
94 if (rctx->chip_class >= EVERGREEN) {
95 evergreen_pipe_shader_ps(ctx, shader);
96 } else {
97 r600_pipe_shader_ps(ctx, shader);
98 }
99 break;
100 default:
101 return -EINVAL;
102 }
103 return 0;
104 }
105
106 static int r600_shader_from_tgsi(struct r600_context * rctx, struct r600_pipe_shader *pipeshader);
107
108 int r600_pipe_shader_create(struct pipe_context *ctx, struct r600_pipe_shader *shader)
109 {
110 static int dump_shaders = -1;
111 struct r600_context *rctx = (struct r600_context *)ctx;
112 int r;
113
114 /* Would like some magic "get_bool_option_once" routine.
115 */
116 if (dump_shaders == -1)
117 dump_shaders = debug_get_bool_option("R600_DUMP_SHADERS", FALSE);
118
119 if (dump_shaders) {
120 fprintf(stderr, "--------------------------------------------------------------\n");
121 tgsi_dump(shader->tokens, 0);
122
123 if (shader->so.num_outputs) {
124 unsigned i;
125 fprintf(stderr, "STREAMOUT\n");
126 for (i = 0; i < shader->so.num_outputs; i++) {
127 unsigned mask = ((1 << shader->so.output[i].num_components) - 1) <<
128 shader->so.output[i].start_component;
129 fprintf(stderr, " %i: MEM_STREAM0_BUF%i OUT[%i].%s%s%s%s\n", i,
130 shader->so.output[i].output_buffer, shader->so.output[i].register_index,
131 mask & 1 ? "x" : "_",
132 (mask >> 1) & 1 ? "y" : "_",
133 (mask >> 2) & 1 ? "z" : "_",
134 (mask >> 3) & 1 ? "w" : "_");
135 }
136 }
137 }
138 r = r600_shader_from_tgsi(rctx, shader);
139 if (r) {
140 R600_ERR("translation from TGSI failed !\n");
141 return r;
142 }
143 r = r600_bytecode_build(&shader->shader.bc);
144 if (r) {
145 R600_ERR("building bytecode failed !\n");
146 return r;
147 }
148 if (dump_shaders) {
149 r600_bytecode_dump(&shader->shader.bc);
150 fprintf(stderr, "______________________________________________________________\n");
151 }
152 return r600_pipe_shader(ctx, shader);
153 }
154
155 void r600_pipe_shader_destroy(struct pipe_context *ctx, struct r600_pipe_shader *shader)
156 {
157 pipe_resource_reference((struct pipe_resource**)&shader->bo, NULL);
158 r600_bytecode_clear(&shader->shader.bc);
159
160 memset(&shader->shader,0,sizeof(struct r600_shader));
161 }
162
163 /*
164 * tgsi -> r600 shader
165 */
166 struct r600_shader_tgsi_instruction;
167
168 struct r600_shader_src {
169 unsigned sel;
170 unsigned swizzle[4];
171 unsigned neg;
172 unsigned abs;
173 unsigned rel;
174 uint32_t value[4];
175 };
176
177 struct r600_shader_ctx {
178 struct tgsi_shader_info info;
179 struct tgsi_parse_context parse;
180 const struct tgsi_token *tokens;
181 unsigned type;
182 unsigned file_offset[TGSI_FILE_COUNT];
183 unsigned temp_reg;
184 struct r600_shader_tgsi_instruction *inst_info;
185 struct r600_bytecode *bc;
186 struct r600_shader *shader;
187 struct r600_shader_src src[4];
188 uint32_t *literals;
189 uint32_t nliterals;
190 uint32_t max_driver_temp_used;
191 /* needed for evergreen interpolation */
192 boolean input_centroid;
193 boolean input_linear;
194 boolean input_perspective;
195 int num_interp_gpr;
196 int face_gpr;
197 int colors_used;
198 boolean clip_vertex_write;
199 unsigned cv_output;
200 int fragcoord_input;
201 };
202
203 struct r600_shader_tgsi_instruction {
204 unsigned tgsi_opcode;
205 unsigned is_op3;
206 unsigned r600_opcode;
207 int (*process)(struct r600_shader_ctx *ctx);
208 };
209
210 static struct r600_shader_tgsi_instruction r600_shader_tgsi_instruction[], eg_shader_tgsi_instruction[], cm_shader_tgsi_instruction[];
211 static int tgsi_helper_tempx_replicate(struct r600_shader_ctx *ctx);
212
213 static int tgsi_is_supported(struct r600_shader_ctx *ctx)
214 {
215 struct tgsi_full_instruction *i = &ctx->parse.FullToken.FullInstruction;
216 int j;
217
218 if (i->Instruction.NumDstRegs > 1) {
219 R600_ERR("too many dst (%d)\n", i->Instruction.NumDstRegs);
220 return -EINVAL;
221 }
222 if (i->Instruction.Predicate) {
223 R600_ERR("predicate unsupported\n");
224 return -EINVAL;
225 }
226 #if 0
227 if (i->Instruction.Label) {
228 R600_ERR("label unsupported\n");
229 return -EINVAL;
230 }
231 #endif
232 for (j = 0; j < i->Instruction.NumSrcRegs; j++) {
233 if (i->Src[j].Register.Dimension) {
234 R600_ERR("unsupported src %d (dimension %d)\n", j,
235 i->Src[j].Register.Dimension);
236 return -EINVAL;
237 }
238 }
239 for (j = 0; j < i->Instruction.NumDstRegs; j++) {
240 if (i->Dst[j].Register.Dimension) {
241 R600_ERR("unsupported dst (dimension)\n");
242 return -EINVAL;
243 }
244 }
245 return 0;
246 }
247
248 static int evergreen_interp_alu(struct r600_shader_ctx *ctx, int input)
249 {
250 int i, r;
251 struct r600_bytecode_alu alu;
252 int gpr = 0, base_chan = 0;
253 int ij_index = 0;
254
255 if (ctx->shader->input[input].interpolate == TGSI_INTERPOLATE_PERSPECTIVE) {
256 ij_index = 0;
257 if (ctx->shader->input[input].centroid)
258 ij_index++;
259 } else if (ctx->shader->input[input].interpolate == TGSI_INTERPOLATE_LINEAR) {
260 ij_index = 0;
261 /* if we have perspective add one */
262 if (ctx->input_perspective) {
263 ij_index++;
264 /* if we have perspective centroid */
265 if (ctx->input_centroid)
266 ij_index++;
267 }
268 if (ctx->shader->input[input].centroid)
269 ij_index++;
270 }
271
272 /* work out gpr and base_chan from index */
273 gpr = ij_index / 2;
274 base_chan = (2 * (ij_index % 2)) + 1;
275
276 for (i = 0; i < 8; i++) {
277 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
278
279 if (i < 4)
280 alu.inst = EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_INTERP_ZW;
281 else
282 alu.inst = EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_INTERP_XY;
283
284 if ((i > 1) && (i < 6)) {
285 alu.dst.sel = ctx->shader->input[input].gpr;
286 alu.dst.write = 1;
287 }
288
289 alu.dst.chan = i % 4;
290
291 alu.src[0].sel = gpr;
292 alu.src[0].chan = (base_chan - (i % 2));
293
294 alu.src[1].sel = V_SQ_ALU_SRC_PARAM_BASE + ctx->shader->input[input].lds_pos;
295
296 alu.bank_swizzle_force = SQ_ALU_VEC_210;
297 if ((i % 4) == 3)
298 alu.last = 1;
299 r = r600_bytecode_add_alu(ctx->bc, &alu);
300 if (r)
301 return r;
302 }
303 return 0;
304 }
305
306 static int evergreen_interp_flat(struct r600_shader_ctx *ctx, int input)
307 {
308 int i, r;
309 struct r600_bytecode_alu alu;
310
311 for (i = 0; i < 4; i++) {
312 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
313
314 alu.inst = EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_INTERP_LOAD_P0;
315
316 alu.dst.sel = ctx->shader->input[input].gpr;
317 alu.dst.write = 1;
318
319 alu.dst.chan = i;
320
321 alu.src[0].sel = V_SQ_ALU_SRC_PARAM_BASE + ctx->shader->input[input].lds_pos;
322 alu.src[0].chan = i;
323
324 if (i == 3)
325 alu.last = 1;
326 r = r600_bytecode_add_alu(ctx->bc, &alu);
327 if (r)
328 return r;
329 }
330 return 0;
331 }
332
333 /*
334 * Special export handling in shaders
335 *
336 * shader export ARRAY_BASE for EXPORT_POS:
337 * 60 is position
338 * 61 is misc vector
339 * 62, 63 are clip distance vectors
340 *
341 * The use of the values exported in 61-63 are controlled by PA_CL_VS_OUT_CNTL:
342 * VS_OUT_MISC_VEC_ENA - enables the use of all fields in export 61
343 * USE_VTX_POINT_SIZE - point size in the X channel of export 61
344 * USE_VTX_EDGE_FLAG - edge flag in the Y channel of export 61
345 * USE_VTX_RENDER_TARGET_INDX - render target index in the Z channel of export 61
346 * USE_VTX_VIEWPORT_INDX - viewport index in the W channel of export 61
347 * USE_VTX_KILL_FLAG - kill flag in the Z channel of export 61 (mutually
348 * exclusive from render target index)
349 * VS_OUT_CCDIST0_VEC_ENA/VS_OUT_CCDIST1_VEC_ENA - enable clip distance vectors
350 *
351 *
352 * shader export ARRAY_BASE for EXPORT_PIXEL:
353 * 0-7 CB targets
354 * 61 computed Z vector
355 *
356 * The use of the values exported in the computed Z vector are controlled
357 * by DB_SHADER_CONTROL:
358 * Z_EXPORT_ENABLE - Z as a float in RED
359 * STENCIL_REF_EXPORT_ENABLE - stencil ref as int in GREEN
360 * COVERAGE_TO_MASK_ENABLE - alpha to mask in ALPHA
361 * MASK_EXPORT_ENABLE - pixel sample mask in BLUE
362 * DB_SOURCE_FORMAT - export control restrictions
363 *
364 */
365
366
367 /* Map name/sid pair from tgsi to the 8-bit semantic index for SPI setup */
368 static int r600_spi_sid(struct r600_shader_io * io)
369 {
370 int index, name = io->name;
371
372 /* These params are handled differently, they don't need
373 * semantic indices, so we'll use 0 for them.
374 */
375 if (name == TGSI_SEMANTIC_POSITION ||
376 name == TGSI_SEMANTIC_PSIZE ||
377 name == TGSI_SEMANTIC_FACE)
378 index = 0;
379 else {
380 if (name == TGSI_SEMANTIC_GENERIC) {
381 /* For generic params simply use sid from tgsi */
382 index = io->sid;
383 } else {
384 /* For non-generic params - pack name and sid into 8 bits */
385 index = 0x80 | (name<<3) | (io->sid);
386 }
387
388 /* Make sure that all really used indices have nonzero value, so
389 * we can just compare it to 0 later instead of comparing the name
390 * with different values to detect special cases. */
391 index++;
392 }
393
394 return index;
395 };
396
397 /* turn input into interpolate on EG */
398 static int evergreen_interp_input(struct r600_shader_ctx *ctx, int index)
399 {
400 int r = 0;
401
402 if (ctx->shader->input[index].spi_sid) {
403 ctx->shader->input[index].lds_pos = ctx->shader->nlds++;
404 if (ctx->shader->input[index].interpolate > 0) {
405 r = evergreen_interp_alu(ctx, index);
406 } else {
407 r = evergreen_interp_flat(ctx, index);
408 }
409 }
410 return r;
411 }
412
413 static int select_twoside_color(struct r600_shader_ctx *ctx, int front, int back)
414 {
415 struct r600_bytecode_alu alu;
416 int i, r;
417 int gpr_front = ctx->shader->input[front].gpr;
418 int gpr_back = ctx->shader->input[back].gpr;
419
420 for (i = 0; i < 4; i++) {
421 memset(&alu, 0, sizeof(alu));
422 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_CNDGT);
423 alu.is_op3 = 1;
424 alu.dst.write = 1;
425 alu.dst.sel = gpr_front;
426 alu.src[0].sel = ctx->face_gpr;
427 alu.src[1].sel = gpr_front;
428 alu.src[2].sel = gpr_back;
429
430 alu.dst.chan = i;
431 alu.src[1].chan = i;
432 alu.src[2].chan = i;
433 alu.last = (i==3);
434
435 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
436 return r;
437 }
438
439 return 0;
440 }
441
442 static int tgsi_declaration(struct r600_shader_ctx *ctx)
443 {
444 struct tgsi_full_declaration *d = &ctx->parse.FullToken.FullDeclaration;
445 unsigned i;
446 int r;
447
448 switch (d->Declaration.File) {
449 case TGSI_FILE_INPUT:
450 i = ctx->shader->ninput++;
451 ctx->shader->input[i].name = d->Semantic.Name;
452 ctx->shader->input[i].sid = d->Semantic.Index;
453 ctx->shader->input[i].spi_sid = r600_spi_sid(&ctx->shader->input[i]);
454 ctx->shader->input[i].interpolate = d->Declaration.Interpolate;
455 ctx->shader->input[i].centroid = d->Declaration.Centroid;
456 ctx->shader->input[i].gpr = ctx->file_offset[TGSI_FILE_INPUT] + d->Range.First;
457 if (ctx->type == TGSI_PROCESSOR_FRAGMENT) {
458 switch (ctx->shader->input[i].name) {
459 case TGSI_SEMANTIC_FACE:
460 ctx->face_gpr = ctx->shader->input[i].gpr;
461 break;
462 case TGSI_SEMANTIC_COLOR:
463 ctx->colors_used++;
464 break;
465 case TGSI_SEMANTIC_POSITION:
466 ctx->fragcoord_input = i;
467 break;
468 }
469 if (ctx->bc->chip_class >= EVERGREEN) {
470 if ((r = evergreen_interp_input(ctx, i)))
471 return r;
472 }
473 }
474 break;
475 case TGSI_FILE_OUTPUT:
476 i = ctx->shader->noutput++;
477 ctx->shader->output[i].name = d->Semantic.Name;
478 ctx->shader->output[i].sid = d->Semantic.Index;
479 ctx->shader->output[i].spi_sid = r600_spi_sid(&ctx->shader->output[i]);
480 ctx->shader->output[i].gpr = ctx->file_offset[TGSI_FILE_OUTPUT] + d->Range.First;
481 ctx->shader->output[i].interpolate = d->Declaration.Interpolate;
482 ctx->shader->output[i].write_mask = d->Declaration.UsageMask;
483 if (ctx->type == TGSI_PROCESSOR_VERTEX) {
484 switch (d->Semantic.Name) {
485 case TGSI_SEMANTIC_CLIPDIST:
486 ctx->shader->clip_dist_write |= d->Declaration.UsageMask << (d->Semantic.Index << 2);
487 break;
488 case TGSI_SEMANTIC_PSIZE:
489 ctx->shader->vs_out_misc_write = 1;
490 ctx->shader->vs_out_point_size = 1;
491 break;
492 case TGSI_SEMANTIC_CLIPVERTEX:
493 ctx->clip_vertex_write = TRUE;
494 ctx->cv_output = i;
495 break;
496 }
497 }
498 break;
499 case TGSI_FILE_CONSTANT:
500 case TGSI_FILE_TEMPORARY:
501 case TGSI_FILE_SAMPLER:
502 case TGSI_FILE_ADDRESS:
503 break;
504
505 case TGSI_FILE_SYSTEM_VALUE:
506 if (d->Semantic.Name == TGSI_SEMANTIC_INSTANCEID) {
507 struct r600_bytecode_alu alu;
508 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
509
510 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_INT_TO_FLT);
511 alu.src[0].sel = 0;
512 alu.src[0].chan = 3;
513
514 alu.dst.sel = 0;
515 alu.dst.chan = 3;
516 alu.dst.write = 1;
517 alu.last = 1;
518
519 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
520 return r;
521 break;
522 } else if (d->Semantic.Name == TGSI_SEMANTIC_VERTEXID)
523 break;
524 default:
525 R600_ERR("unsupported file %d declaration\n", d->Declaration.File);
526 return -EINVAL;
527 }
528 return 0;
529 }
530
531 static int r600_get_temp(struct r600_shader_ctx *ctx)
532 {
533 return ctx->temp_reg + ctx->max_driver_temp_used++;
534 }
535
536 /*
537 * for evergreen we need to scan the shader to find the number of GPRs we need to
538 * reserve for interpolation.
539 *
540 * we need to know if we are going to emit
541 * any centroid inputs
542 * if perspective and linear are required
543 */
544 static int evergreen_gpr_count(struct r600_shader_ctx *ctx)
545 {
546 int i;
547 int num_baryc;
548
549 ctx->input_linear = FALSE;
550 ctx->input_perspective = FALSE;
551 ctx->input_centroid = FALSE;
552 ctx->num_interp_gpr = 1;
553
554 /* any centroid inputs */
555 for (i = 0; i < ctx->info.num_inputs; i++) {
556 /* skip position/face */
557 if (ctx->info.input_semantic_name[i] == TGSI_SEMANTIC_POSITION ||
558 ctx->info.input_semantic_name[i] == TGSI_SEMANTIC_FACE)
559 continue;
560 if (ctx->info.input_interpolate[i] == TGSI_INTERPOLATE_LINEAR)
561 ctx->input_linear = TRUE;
562 if (ctx->info.input_interpolate[i] == TGSI_INTERPOLATE_PERSPECTIVE)
563 ctx->input_perspective = TRUE;
564 if (ctx->info.input_centroid[i])
565 ctx->input_centroid = TRUE;
566 }
567
568 num_baryc = 0;
569 /* ignoring sample for now */
570 if (ctx->input_perspective)
571 num_baryc++;
572 if (ctx->input_linear)
573 num_baryc++;
574 if (ctx->input_centroid)
575 num_baryc *= 2;
576
577 ctx->num_interp_gpr += (num_baryc + 1) >> 1;
578
579 /* TODO PULL MODEL and LINE STIPPLE, FIXED PT POS */
580 return ctx->num_interp_gpr;
581 }
582
583 static void tgsi_src(struct r600_shader_ctx *ctx,
584 const struct tgsi_full_src_register *tgsi_src,
585 struct r600_shader_src *r600_src)
586 {
587 memset(r600_src, 0, sizeof(*r600_src));
588 r600_src->swizzle[0] = tgsi_src->Register.SwizzleX;
589 r600_src->swizzle[1] = tgsi_src->Register.SwizzleY;
590 r600_src->swizzle[2] = tgsi_src->Register.SwizzleZ;
591 r600_src->swizzle[3] = tgsi_src->Register.SwizzleW;
592 r600_src->neg = tgsi_src->Register.Negate;
593 r600_src->abs = tgsi_src->Register.Absolute;
594
595 if (tgsi_src->Register.File == TGSI_FILE_IMMEDIATE) {
596 int index;
597 if ((tgsi_src->Register.SwizzleX == tgsi_src->Register.SwizzleY) &&
598 (tgsi_src->Register.SwizzleX == tgsi_src->Register.SwizzleZ) &&
599 (tgsi_src->Register.SwizzleX == tgsi_src->Register.SwizzleW)) {
600
601 index = tgsi_src->Register.Index * 4 + tgsi_src->Register.SwizzleX;
602 r600_bytecode_special_constants(ctx->literals[index], &r600_src->sel, &r600_src->neg);
603 if (r600_src->sel != V_SQ_ALU_SRC_LITERAL)
604 return;
605 }
606 index = tgsi_src->Register.Index;
607 r600_src->sel = V_SQ_ALU_SRC_LITERAL;
608 memcpy(r600_src->value, ctx->literals + index * 4, sizeof(r600_src->value));
609 } else if (tgsi_src->Register.File == TGSI_FILE_SYSTEM_VALUE) {
610 if (ctx->info.system_value_semantic_name[tgsi_src->Register.Index] == TGSI_SEMANTIC_INSTANCEID) {
611 r600_src->swizzle[0] = 3;
612 r600_src->swizzle[1] = 3;
613 r600_src->swizzle[2] = 3;
614 r600_src->swizzle[3] = 3;
615 r600_src->sel = 0;
616 } else if (ctx->info.system_value_semantic_name[tgsi_src->Register.Index] == TGSI_SEMANTIC_VERTEXID) {
617 r600_src->swizzle[0] = 0;
618 r600_src->swizzle[1] = 0;
619 r600_src->swizzle[2] = 0;
620 r600_src->swizzle[3] = 0;
621 r600_src->sel = 0;
622 }
623 } else {
624 if (tgsi_src->Register.Indirect)
625 r600_src->rel = V_SQ_REL_RELATIVE;
626 r600_src->sel = tgsi_src->Register.Index;
627 r600_src->sel += ctx->file_offset[tgsi_src->Register.File];
628 }
629 }
630
631 static int tgsi_fetch_rel_const(struct r600_shader_ctx *ctx, unsigned int offset, unsigned int dst_reg)
632 {
633 struct r600_bytecode_vtx vtx;
634 unsigned int ar_reg;
635 int r;
636
637 if (offset) {
638 struct r600_bytecode_alu alu;
639
640 memset(&alu, 0, sizeof(alu));
641
642 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ADD_INT);
643 alu.src[0].sel = ctx->bc->ar_reg;
644
645 alu.src[1].sel = V_SQ_ALU_SRC_LITERAL;
646 alu.src[1].value = offset;
647
648 alu.dst.sel = dst_reg;
649 alu.dst.write = 1;
650 alu.last = 1;
651
652 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
653 return r;
654
655 ar_reg = dst_reg;
656 } else {
657 ar_reg = ctx->bc->ar_reg;
658 }
659
660 memset(&vtx, 0, sizeof(vtx));
661 vtx.fetch_type = 2; /* VTX_FETCH_NO_INDEX_OFFSET */
662 vtx.src_gpr = ar_reg;
663 vtx.mega_fetch_count = 16;
664 vtx.dst_gpr = dst_reg;
665 vtx.dst_sel_x = 0; /* SEL_X */
666 vtx.dst_sel_y = 1; /* SEL_Y */
667 vtx.dst_sel_z = 2; /* SEL_Z */
668 vtx.dst_sel_w = 3; /* SEL_W */
669 vtx.data_format = FMT_32_32_32_32_FLOAT;
670 vtx.num_format_all = 2; /* NUM_FORMAT_SCALED */
671 vtx.format_comp_all = 1; /* FORMAT_COMP_SIGNED */
672 vtx.srf_mode_all = 1; /* SRF_MODE_NO_ZERO */
673 vtx.endian = r600_endian_swap(32);
674
675 if ((r = r600_bytecode_add_vtx(ctx->bc, &vtx)))
676 return r;
677
678 return 0;
679 }
680
681 static int tgsi_split_constant(struct r600_shader_ctx *ctx)
682 {
683 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
684 struct r600_bytecode_alu alu;
685 int i, j, k, nconst, r;
686
687 for (i = 0, nconst = 0; i < inst->Instruction.NumSrcRegs; i++) {
688 if (inst->Src[i].Register.File == TGSI_FILE_CONSTANT) {
689 nconst++;
690 }
691 tgsi_src(ctx, &inst->Src[i], &ctx->src[i]);
692 }
693 for (i = 0, j = nconst - 1; i < inst->Instruction.NumSrcRegs; i++) {
694 if (inst->Src[i].Register.File != TGSI_FILE_CONSTANT) {
695 continue;
696 }
697
698 if (ctx->src[i].rel) {
699 int treg = r600_get_temp(ctx);
700 if ((r = tgsi_fetch_rel_const(ctx, ctx->src[i].sel - 512, treg)))
701 return r;
702
703 ctx->src[i].sel = treg;
704 ctx->src[i].rel = 0;
705 j--;
706 } else if (j > 0) {
707 int treg = r600_get_temp(ctx);
708 for (k = 0; k < 4; k++) {
709 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
710 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV);
711 alu.src[0].sel = ctx->src[i].sel;
712 alu.src[0].chan = k;
713 alu.src[0].rel = ctx->src[i].rel;
714 alu.dst.sel = treg;
715 alu.dst.chan = k;
716 alu.dst.write = 1;
717 if (k == 3)
718 alu.last = 1;
719 r = r600_bytecode_add_alu(ctx->bc, &alu);
720 if (r)
721 return r;
722 }
723 ctx->src[i].sel = treg;
724 ctx->src[i].rel =0;
725 j--;
726 }
727 }
728 return 0;
729 }
730
731 /* need to move any immediate into a temp - for trig functions which use literal for PI stuff */
732 static int tgsi_split_literal_constant(struct r600_shader_ctx *ctx)
733 {
734 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
735 struct r600_bytecode_alu alu;
736 int i, j, k, nliteral, r;
737
738 for (i = 0, nliteral = 0; i < inst->Instruction.NumSrcRegs; i++) {
739 if (ctx->src[i].sel == V_SQ_ALU_SRC_LITERAL) {
740 nliteral++;
741 }
742 }
743 for (i = 0, j = nliteral - 1; i < inst->Instruction.NumSrcRegs; i++) {
744 if (j > 0 && ctx->src[i].sel == V_SQ_ALU_SRC_LITERAL) {
745 int treg = r600_get_temp(ctx);
746 for (k = 0; k < 4; k++) {
747 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
748 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV);
749 alu.src[0].sel = ctx->src[i].sel;
750 alu.src[0].chan = k;
751 alu.src[0].value = ctx->src[i].value[k];
752 alu.dst.sel = treg;
753 alu.dst.chan = k;
754 alu.dst.write = 1;
755 if (k == 3)
756 alu.last = 1;
757 r = r600_bytecode_add_alu(ctx->bc, &alu);
758 if (r)
759 return r;
760 }
761 ctx->src[i].sel = treg;
762 j--;
763 }
764 }
765 return 0;
766 }
767
768 static int process_twoside_color_inputs(struct r600_shader_ctx *ctx)
769 {
770 int i, r, count = ctx->shader->ninput;
771
772 /* additional inputs will be allocated right after the existing inputs,
773 * we won't need them after the color selection, so we don't need to
774 * reserve these gprs for the rest of the shader code and to adjust
775 * output offsets etc. */
776 int gpr = ctx->file_offset[TGSI_FILE_INPUT] +
777 ctx->info.file_max[TGSI_FILE_INPUT] + 1;
778
779 if (ctx->face_gpr == -1) {
780 i = ctx->shader->ninput++;
781 ctx->shader->input[i].name = TGSI_SEMANTIC_FACE;
782 ctx->shader->input[i].spi_sid = 0;
783 ctx->shader->input[i].gpr = gpr++;
784 ctx->face_gpr = ctx->shader->input[i].gpr;
785 }
786
787 for (i = 0; i < count; i++) {
788 if (ctx->shader->input[i].name == TGSI_SEMANTIC_COLOR) {
789 int ni = ctx->shader->ninput++;
790 memcpy(&ctx->shader->input[ni],&ctx->shader->input[i], sizeof(struct r600_shader_io));
791 ctx->shader->input[ni].name = TGSI_SEMANTIC_BCOLOR;
792 ctx->shader->input[ni].spi_sid = r600_spi_sid(&ctx->shader->input[ni]);
793 ctx->shader->input[ni].gpr = gpr++;
794
795 if (ctx->bc->chip_class >= EVERGREEN) {
796 r = evergreen_interp_input(ctx, ni);
797 if (r)
798 return r;
799 }
800
801 r = select_twoside_color(ctx, i, ni);
802 if (r)
803 return r;
804 }
805 }
806 return 0;
807 }
808
809 static int r600_shader_from_tgsi(struct r600_context * rctx, struct r600_pipe_shader *pipeshader)
810 {
811 struct r600_shader *shader = &pipeshader->shader;
812 struct tgsi_token *tokens = pipeshader->tokens;
813 struct pipe_stream_output_info so = pipeshader->so;
814 struct tgsi_full_immediate *immediate;
815 struct tgsi_full_property *property;
816 struct r600_shader_ctx ctx;
817 struct r600_bytecode_output output[32];
818 unsigned output_done, noutput;
819 unsigned opcode;
820 int i, j, k, r = 0;
821 int next_pixel_base = 0, next_pos_base = 60, next_param_base = 0;
822
823 ctx.bc = &shader->bc;
824 ctx.shader = shader;
825 r600_bytecode_init(ctx.bc, rctx->chip_class, rctx->family);
826 ctx.tokens = tokens;
827 tgsi_scan_shader(tokens, &ctx.info);
828 tgsi_parse_init(&ctx.parse, tokens);
829 ctx.type = ctx.parse.FullHeader.Processor.Processor;
830 shader->processor_type = ctx.type;
831 ctx.bc->type = shader->processor_type;
832
833 ctx.face_gpr = -1;
834 ctx.fragcoord_input = -1;
835 ctx.colors_used = 0;
836 ctx.clip_vertex_write = 0;
837
838 shader->two_side = (ctx.type == TGSI_PROCESSOR_FRAGMENT) && rctx->two_side;
839 shader->nr_cbufs = rctx->nr_cbufs;
840
841 /* register allocations */
842 /* Values [0,127] correspond to GPR[0..127].
843 * Values [128,159] correspond to constant buffer bank 0
844 * Values [160,191] correspond to constant buffer bank 1
845 * Values [256,511] correspond to cfile constants c[0..255]. (Gone on EG)
846 * Values [256,287] correspond to constant buffer bank 2 (EG)
847 * Values [288,319] correspond to constant buffer bank 3 (EG)
848 * Other special values are shown in the list below.
849 * 244 ALU_SRC_1_DBL_L: special constant 1.0 double-float, LSW. (RV670+)
850 * 245 ALU_SRC_1_DBL_M: special constant 1.0 double-float, MSW. (RV670+)
851 * 246 ALU_SRC_0_5_DBL_L: special constant 0.5 double-float, LSW. (RV670+)
852 * 247 ALU_SRC_0_5_DBL_M: special constant 0.5 double-float, MSW. (RV670+)
853 * 248 SQ_ALU_SRC_0: special constant 0.0.
854 * 249 SQ_ALU_SRC_1: special constant 1.0 float.
855 * 250 SQ_ALU_SRC_1_INT: special constant 1 integer.
856 * 251 SQ_ALU_SRC_M_1_INT: special constant -1 integer.
857 * 252 SQ_ALU_SRC_0_5: special constant 0.5 float.
858 * 253 SQ_ALU_SRC_LITERAL: literal constant.
859 * 254 SQ_ALU_SRC_PV: previous vector result.
860 * 255 SQ_ALU_SRC_PS: previous scalar result.
861 */
862 for (i = 0; i < TGSI_FILE_COUNT; i++) {
863 ctx.file_offset[i] = 0;
864 }
865 if (ctx.type == TGSI_PROCESSOR_VERTEX) {
866 ctx.file_offset[TGSI_FILE_INPUT] = 1;
867 if (ctx.bc->chip_class >= EVERGREEN) {
868 r600_bytecode_add_cfinst(ctx.bc, EG_V_SQ_CF_WORD1_SQ_CF_INST_CALL_FS);
869 } else {
870 r600_bytecode_add_cfinst(ctx.bc, V_SQ_CF_WORD1_SQ_CF_INST_CALL_FS);
871 }
872 }
873 if (ctx.type == TGSI_PROCESSOR_FRAGMENT && ctx.bc->chip_class >= EVERGREEN) {
874 ctx.file_offset[TGSI_FILE_INPUT] = evergreen_gpr_count(&ctx);
875 }
876 ctx.file_offset[TGSI_FILE_OUTPUT] = ctx.file_offset[TGSI_FILE_INPUT] +
877 ctx.info.file_max[TGSI_FILE_INPUT] + 1;
878 ctx.file_offset[TGSI_FILE_TEMPORARY] = ctx.file_offset[TGSI_FILE_OUTPUT] +
879 ctx.info.file_max[TGSI_FILE_OUTPUT] + 1;
880
881 /* Outside the GPR range. This will be translated to one of the
882 * kcache banks later. */
883 ctx.file_offset[TGSI_FILE_CONSTANT] = 512;
884
885 ctx.file_offset[TGSI_FILE_IMMEDIATE] = V_SQ_ALU_SRC_LITERAL;
886 ctx.bc->ar_reg = ctx.file_offset[TGSI_FILE_TEMPORARY] +
887 ctx.info.file_max[TGSI_FILE_TEMPORARY] + 1;
888 ctx.temp_reg = ctx.bc->ar_reg + 1;
889
890 ctx.nliterals = 0;
891 ctx.literals = NULL;
892 shader->fs_write_all = FALSE;
893 while (!tgsi_parse_end_of_tokens(&ctx.parse)) {
894 tgsi_parse_token(&ctx.parse);
895 switch (ctx.parse.FullToken.Token.Type) {
896 case TGSI_TOKEN_TYPE_IMMEDIATE:
897 immediate = &ctx.parse.FullToken.FullImmediate;
898 ctx.literals = realloc(ctx.literals, (ctx.nliterals + 1) * 16);
899 if(ctx.literals == NULL) {
900 r = -ENOMEM;
901 goto out_err;
902 }
903 ctx.literals[ctx.nliterals * 4 + 0] = immediate->u[0].Uint;
904 ctx.literals[ctx.nliterals * 4 + 1] = immediate->u[1].Uint;
905 ctx.literals[ctx.nliterals * 4 + 2] = immediate->u[2].Uint;
906 ctx.literals[ctx.nliterals * 4 + 3] = immediate->u[3].Uint;
907 ctx.nliterals++;
908 break;
909 case TGSI_TOKEN_TYPE_DECLARATION:
910 r = tgsi_declaration(&ctx);
911 if (r)
912 goto out_err;
913 break;
914 case TGSI_TOKEN_TYPE_INSTRUCTION:
915 break;
916 case TGSI_TOKEN_TYPE_PROPERTY:
917 property = &ctx.parse.FullToken.FullProperty;
918 switch (property->Property.PropertyName) {
919 case TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS:
920 if (property->u[0].Data == 1)
921 shader->fs_write_all = TRUE;
922 break;
923 case TGSI_PROPERTY_VS_PROHIBIT_UCPS:
924 if (property->u[0].Data == 1)
925 shader->vs_prohibit_ucps = TRUE;
926 break;
927 }
928 break;
929 default:
930 R600_ERR("unsupported token type %d\n", ctx.parse.FullToken.Token.Type);
931 r = -EINVAL;
932 goto out_err;
933 }
934 }
935
936 if (ctx.fragcoord_input >= 0) {
937 if (ctx.bc->chip_class == CAYMAN) {
938 for (j = 0 ; j < 4; j++) {
939 struct r600_bytecode_alu alu;
940 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
941 alu.inst = BC_INST(ctx.bc, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_IEEE);
942 alu.src[0].sel = shader->input[ctx.fragcoord_input].gpr;
943 alu.src[0].chan = 3;
944
945 alu.dst.sel = shader->input[ctx.fragcoord_input].gpr;
946 alu.dst.chan = j;
947 alu.dst.write = (j == 3);
948 alu.last = 1;
949 if ((r = r600_bytecode_add_alu(ctx.bc, &alu)))
950 return r;
951 }
952 } else {
953 struct r600_bytecode_alu alu;
954 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
955 alu.inst = BC_INST(ctx.bc, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_IEEE);
956 alu.src[0].sel = shader->input[ctx.fragcoord_input].gpr;
957 alu.src[0].chan = 3;
958
959 alu.dst.sel = shader->input[ctx.fragcoord_input].gpr;
960 alu.dst.chan = 3;
961 alu.dst.write = 1;
962 alu.last = 1;
963 if ((r = r600_bytecode_add_alu(ctx.bc, &alu)))
964 return r;
965 }
966 }
967
968 if (shader->two_side && ctx.colors_used) {
969 if ((r = process_twoside_color_inputs(&ctx)))
970 return r;
971 }
972
973 tgsi_parse_init(&ctx.parse, tokens);
974 while (!tgsi_parse_end_of_tokens(&ctx.parse)) {
975 tgsi_parse_token(&ctx.parse);
976 switch (ctx.parse.FullToken.Token.Type) {
977 case TGSI_TOKEN_TYPE_INSTRUCTION:
978 r = tgsi_is_supported(&ctx);
979 if (r)
980 goto out_err;
981 ctx.max_driver_temp_used = 0;
982 /* reserve first tmp for everyone */
983 r600_get_temp(&ctx);
984
985 opcode = ctx.parse.FullToken.FullInstruction.Instruction.Opcode;
986 if ((r = tgsi_split_constant(&ctx)))
987 goto out_err;
988 if ((r = tgsi_split_literal_constant(&ctx)))
989 goto out_err;
990 if (ctx.bc->chip_class == CAYMAN)
991 ctx.inst_info = &cm_shader_tgsi_instruction[opcode];
992 else if (ctx.bc->chip_class >= EVERGREEN)
993 ctx.inst_info = &eg_shader_tgsi_instruction[opcode];
994 else
995 ctx.inst_info = &r600_shader_tgsi_instruction[opcode];
996 r = ctx.inst_info->process(&ctx);
997 if (r)
998 goto out_err;
999 break;
1000 default:
1001 break;
1002 }
1003 }
1004
1005 noutput = shader->noutput;
1006
1007 if (ctx.clip_vertex_write) {
1008 /* need to convert a clipvertex write into clipdistance writes and not export
1009 the clip vertex anymore */
1010
1011 memset(&shader->output[noutput], 0, 2*sizeof(struct r600_shader_io));
1012 shader->output[noutput].name = TGSI_SEMANTIC_CLIPDIST;
1013 shader->output[noutput].gpr = ctx.temp_reg;
1014 noutput++;
1015 shader->output[noutput].name = TGSI_SEMANTIC_CLIPDIST;
1016 shader->output[noutput].gpr = ctx.temp_reg+1;
1017 noutput++;
1018
1019 /* reset spi_sid for clipvertex output to avoid confusing spi */
1020 shader->output[ctx.cv_output].spi_sid = 0;
1021
1022 shader->clip_dist_write = 0xFF;
1023
1024 for (i = 0; i < 8; i++) {
1025 int oreg = i >> 2;
1026 int ochan = i & 3;
1027
1028 for (j = 0; j < 4; j++) {
1029 struct r600_bytecode_alu alu;
1030 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1031 alu.inst = BC_INST(ctx.bc, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4);
1032 alu.src[0].sel = shader->output[ctx.cv_output].gpr;
1033 alu.src[0].chan = j;
1034
1035 alu.src[1].sel = 512 + i;
1036 alu.src[1].kc_bank = 1;
1037 alu.src[1].chan = j;
1038
1039 alu.dst.sel = ctx.temp_reg + oreg;
1040 alu.dst.chan = j;
1041 alu.dst.write = (j == ochan);
1042 if (j == 3)
1043 alu.last = 1;
1044 r = r600_bytecode_add_alu(ctx.bc, &alu);
1045 if (r)
1046 return r;
1047 }
1048 }
1049 }
1050
1051 /* Add stream outputs. */
1052 if (ctx.type == TGSI_PROCESSOR_VERTEX && so.num_outputs) {
1053 for (i = 0; i < so.num_outputs; i++) {
1054 struct r600_bytecode_output output;
1055
1056 if (so.output[i].output_buffer >= 4) {
1057 R600_ERR("exceeded the max number of stream output buffers, got: %d\n",
1058 so.output[i].output_buffer);
1059 r = -EINVAL;
1060 goto out_err;
1061 }
1062 if (so.output[i].dst_offset < so.output[i].start_component) {
1063 R600_ERR("stream_output - dst_offset cannot be less than start_component\n");
1064 r = -EINVAL;
1065 goto out_err;
1066 }
1067
1068 memset(&output, 0, sizeof(struct r600_bytecode_output));
1069 output.gpr = shader->output[so.output[i].register_index].gpr;
1070 output.elem_size = 0;
1071 output.array_base = so.output[i].dst_offset - so.output[i].start_component;
1072 output.type = V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_WRITE;
1073 output.burst_count = 1;
1074 output.barrier = 1;
1075 /* array_size is an upper limit for the burst_count
1076 * with MEM_STREAM instructions */
1077 output.array_size = 0xFFF;
1078 output.comp_mask = ((1 << so.output[i].num_components) - 1) << so.output[i].start_component;
1079 if (ctx.bc->chip_class >= EVERGREEN) {
1080 switch (so.output[i].output_buffer) {
1081 case 0:
1082 output.inst = EG_V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_MEM_STREAM0_BUF0;
1083 break;
1084 case 1:
1085 output.inst = EG_V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_MEM_STREAM0_BUF1;
1086 break;
1087 case 2:
1088 output.inst = EG_V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_MEM_STREAM0_BUF2;
1089 break;
1090 case 3:
1091 output.inst = EG_V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_MEM_STREAM0_BUF3;
1092 break;
1093 }
1094 } else {
1095 switch (so.output[i].output_buffer) {
1096 case 0:
1097 output.inst = V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_MEM_STREAM0;
1098 break;
1099 case 1:
1100 output.inst = V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_MEM_STREAM1;
1101 break;
1102 case 2:
1103 output.inst = V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_MEM_STREAM2;
1104 break;
1105 case 3:
1106 output.inst = V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_MEM_STREAM3;
1107 break;
1108 }
1109 }
1110 r = r600_bytecode_add_output(ctx.bc, &output);
1111 if (r)
1112 goto out_err;
1113 }
1114 }
1115
1116 /* export output */
1117 for (i = 0, j = 0; i < noutput; i++, j++) {
1118 memset(&output[j], 0, sizeof(struct r600_bytecode_output));
1119 output[j].gpr = shader->output[i].gpr;
1120 output[j].elem_size = 3;
1121 output[j].swizzle_x = 0;
1122 output[j].swizzle_y = 1;
1123 output[j].swizzle_z = 2;
1124 output[j].swizzle_w = 3;
1125 output[j].burst_count = 1;
1126 output[j].barrier = 1;
1127 output[j].type = -1;
1128 output[j].inst = BC_INST(ctx.bc, V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT);
1129 switch (ctx.type) {
1130 case TGSI_PROCESSOR_VERTEX:
1131 switch (shader->output[i].name) {
1132 case TGSI_SEMANTIC_POSITION:
1133 output[j].array_base = next_pos_base++;
1134 output[j].type = V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_POS;
1135 break;
1136
1137 case TGSI_SEMANTIC_PSIZE:
1138 output[j].array_base = next_pos_base++;
1139 output[j].type = V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_POS;
1140 break;
1141 case TGSI_SEMANTIC_CLIPVERTEX:
1142 j--;
1143 break;
1144 case TGSI_SEMANTIC_CLIPDIST:
1145 output[j].array_base = next_pos_base++;
1146 output[j].type = V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_POS;
1147 /* spi_sid is 0 for clipdistance outputs that were generated
1148 * for clipvertex - we don't need to pass them to PS */
1149 if (shader->output[i].spi_sid) {
1150 j++;
1151 /* duplicate it as PARAM to pass to the pixel shader */
1152 memcpy(&output[j], &output[j-1], sizeof(struct r600_bytecode_output));
1153 output[j].array_base = next_param_base++;
1154 output[j].type = V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PARAM;
1155 }
1156 break;
1157 case TGSI_SEMANTIC_FOG:
1158 output[j].swizzle_y = 4; /* 0 */
1159 output[j].swizzle_z = 4; /* 0 */
1160 output[j].swizzle_w = 5; /* 1 */
1161 break;
1162 }
1163 break;
1164 case TGSI_PROCESSOR_FRAGMENT:
1165 if (shader->output[i].name == TGSI_SEMANTIC_COLOR) {
1166 output[j].array_base = next_pixel_base++;
1167 output[j].type = V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PIXEL;
1168 if (shader->fs_write_all && (rctx->chip_class >= EVERGREEN)) {
1169 for (k = 1; k < shader->nr_cbufs; k++) {
1170 j++;
1171 memset(&output[j], 0, sizeof(struct r600_bytecode_output));
1172 output[j].gpr = shader->output[i].gpr;
1173 output[j].elem_size = 3;
1174 output[j].swizzle_x = 0;
1175 output[j].swizzle_y = 1;
1176 output[j].swizzle_z = 2;
1177 output[j].swizzle_w = 3;
1178 output[j].burst_count = 1;
1179 output[j].barrier = 1;
1180 output[j].array_base = next_pixel_base++;
1181 output[j].inst = BC_INST(ctx.bc, V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT);
1182 output[j].type = V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PIXEL;
1183 }
1184 }
1185 } else if (shader->output[i].name == TGSI_SEMANTIC_POSITION) {
1186 output[j].array_base = 61;
1187 output[j].swizzle_x = 2;
1188 output[j].swizzle_y = 7;
1189 output[j].swizzle_z = output[j].swizzle_w = 7;
1190 output[j].type = V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PIXEL;
1191 } else if (shader->output[i].name == TGSI_SEMANTIC_STENCIL) {
1192 output[j].array_base = 61;
1193 output[j].swizzle_x = 7;
1194 output[j].swizzle_y = 1;
1195 output[j].swizzle_z = output[j].swizzle_w = 7;
1196 output[j].type = V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PIXEL;
1197 } else {
1198 R600_ERR("unsupported fragment output name %d\n", shader->output[i].name);
1199 r = -EINVAL;
1200 goto out_err;
1201 }
1202 break;
1203 default:
1204 R600_ERR("unsupported processor type %d\n", ctx.type);
1205 r = -EINVAL;
1206 goto out_err;
1207 }
1208
1209 if (output[j].type==-1) {
1210 output[j].type = V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PARAM;
1211 output[j].array_base = next_param_base++;
1212 }
1213 }
1214
1215 /* add fake param output for vertex shader if no param is exported */
1216 if (ctx.type == TGSI_PROCESSOR_VERTEX && next_param_base == 0) {
1217 memset(&output[j], 0, sizeof(struct r600_bytecode_output));
1218 output[j].gpr = 0;
1219 output[j].elem_size = 3;
1220 output[j].swizzle_x = 7;
1221 output[j].swizzle_y = 7;
1222 output[j].swizzle_z = 7;
1223 output[j].swizzle_w = 7;
1224 output[j].burst_count = 1;
1225 output[j].barrier = 1;
1226 output[j].type = V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PARAM;
1227 output[j].array_base = 0;
1228 output[j].inst = BC_INST(ctx.bc, V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT);
1229 j++;
1230 }
1231
1232 /* add fake pixel export */
1233 if (ctx.type == TGSI_PROCESSOR_FRAGMENT && j == 0) {
1234 memset(&output[j], 0, sizeof(struct r600_bytecode_output));
1235 output[j].gpr = 0;
1236 output[j].elem_size = 3;
1237 output[j].swizzle_x = 7;
1238 output[j].swizzle_y = 7;
1239 output[j].swizzle_z = 7;
1240 output[j].swizzle_w = 7;
1241 output[j].burst_count = 1;
1242 output[j].barrier = 1;
1243 output[j].type = V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PIXEL;
1244 output[j].array_base = 0;
1245 output[j].inst = BC_INST(ctx.bc, V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT);
1246 j++;
1247 }
1248
1249 noutput = j;
1250
1251 /* set export done on last export of each type */
1252 for (i = noutput - 1, output_done = 0; i >= 0; i--) {
1253 if (ctx.bc->chip_class < CAYMAN) {
1254 if (i == (noutput - 1)) {
1255 output[i].end_of_program = 1;
1256 }
1257 }
1258 if (!(output_done & (1 << output[i].type))) {
1259 output_done |= (1 << output[i].type);
1260 output[i].inst = BC_INST(ctx.bc, V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT_DONE);
1261 }
1262 }
1263 /* add output to bytecode */
1264 for (i = 0; i < noutput; i++) {
1265 r = r600_bytecode_add_output(ctx.bc, &output[i]);
1266 if (r)
1267 goto out_err;
1268 }
1269 /* add program end */
1270 if (ctx.bc->chip_class == CAYMAN)
1271 cm_bytecode_add_cf_end(ctx.bc);
1272
1273 free(ctx.literals);
1274 tgsi_parse_free(&ctx.parse);
1275 return 0;
1276 out_err:
1277 free(ctx.literals);
1278 tgsi_parse_free(&ctx.parse);
1279 return r;
1280 }
1281
1282 static int tgsi_unsupported(struct r600_shader_ctx *ctx)
1283 {
1284 R600_ERR("%s tgsi opcode unsupported\n",
1285 tgsi_get_opcode_name(ctx->inst_info->tgsi_opcode));
1286 return -EINVAL;
1287 }
1288
1289 static int tgsi_end(struct r600_shader_ctx *ctx)
1290 {
1291 return 0;
1292 }
1293
1294 static void r600_bytecode_src(struct r600_bytecode_alu_src *bc_src,
1295 const struct r600_shader_src *shader_src,
1296 unsigned chan)
1297 {
1298 bc_src->sel = shader_src->sel;
1299 bc_src->chan = shader_src->swizzle[chan];
1300 bc_src->neg = shader_src->neg;
1301 bc_src->abs = shader_src->abs;
1302 bc_src->rel = shader_src->rel;
1303 bc_src->value = shader_src->value[bc_src->chan];
1304 }
1305
1306 static void r600_bytecode_src_set_abs(struct r600_bytecode_alu_src *bc_src)
1307 {
1308 bc_src->abs = 1;
1309 bc_src->neg = 0;
1310 }
1311
1312 static void r600_bytecode_src_toggle_neg(struct r600_bytecode_alu_src *bc_src)
1313 {
1314 bc_src->neg = !bc_src->neg;
1315 }
1316
1317 static void tgsi_dst(struct r600_shader_ctx *ctx,
1318 const struct tgsi_full_dst_register *tgsi_dst,
1319 unsigned swizzle,
1320 struct r600_bytecode_alu_dst *r600_dst)
1321 {
1322 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
1323
1324 r600_dst->sel = tgsi_dst->Register.Index;
1325 r600_dst->sel += ctx->file_offset[tgsi_dst->Register.File];
1326 r600_dst->chan = swizzle;
1327 r600_dst->write = 1;
1328 if (tgsi_dst->Register.Indirect)
1329 r600_dst->rel = V_SQ_REL_RELATIVE;
1330 if (inst->Instruction.Saturate) {
1331 r600_dst->clamp = 1;
1332 }
1333 }
1334
1335 static int tgsi_last_instruction(unsigned writemask)
1336 {
1337 int i, lasti = 0;
1338
1339 for (i = 0; i < 4; i++) {
1340 if (writemask & (1 << i)) {
1341 lasti = i;
1342 }
1343 }
1344 return lasti;
1345 }
1346
1347 static int tgsi_op2_s(struct r600_shader_ctx *ctx, int swap, int trans_only)
1348 {
1349 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
1350 struct r600_bytecode_alu alu;
1351 int i, j, r;
1352 int lasti = tgsi_last_instruction(inst->Dst[0].Register.WriteMask);
1353
1354 for (i = 0; i < lasti + 1; i++) {
1355 if (!(inst->Dst[0].Register.WriteMask & (1 << i)))
1356 continue;
1357
1358 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1359 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
1360
1361 alu.inst = ctx->inst_info->r600_opcode;
1362 if (!swap) {
1363 for (j = 0; j < inst->Instruction.NumSrcRegs; j++) {
1364 r600_bytecode_src(&alu.src[j], &ctx->src[j], i);
1365 }
1366 } else {
1367 r600_bytecode_src(&alu.src[0], &ctx->src[1], i);
1368 r600_bytecode_src(&alu.src[1], &ctx->src[0], i);
1369 }
1370 /* handle some special cases */
1371 switch (ctx->inst_info->tgsi_opcode) {
1372 case TGSI_OPCODE_SUB:
1373 r600_bytecode_src_toggle_neg(&alu.src[1]);
1374 break;
1375 case TGSI_OPCODE_ABS:
1376 r600_bytecode_src_set_abs(&alu.src[0]);
1377 break;
1378 default:
1379 break;
1380 }
1381 if (i == lasti || trans_only) {
1382 alu.last = 1;
1383 }
1384 r = r600_bytecode_add_alu(ctx->bc, &alu);
1385 if (r)
1386 return r;
1387 }
1388 return 0;
1389 }
1390
1391 static int tgsi_op2(struct r600_shader_ctx *ctx)
1392 {
1393 return tgsi_op2_s(ctx, 0, 0);
1394 }
1395
1396 static int tgsi_op2_swap(struct r600_shader_ctx *ctx)
1397 {
1398 return tgsi_op2_s(ctx, 1, 0);
1399 }
1400
1401 static int tgsi_op2_trans(struct r600_shader_ctx *ctx)
1402 {
1403 return tgsi_op2_s(ctx, 0, 1);
1404 }
1405
1406 static int tgsi_ineg(struct r600_shader_ctx *ctx)
1407 {
1408 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
1409 struct r600_bytecode_alu alu;
1410 int i, r;
1411 int lasti = tgsi_last_instruction(inst->Dst[0].Register.WriteMask);
1412
1413 for (i = 0; i < lasti + 1; i++) {
1414
1415 if (!(inst->Dst[0].Register.WriteMask & (1 << i)))
1416 continue;
1417 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1418 alu.inst = ctx->inst_info->r600_opcode;
1419
1420 alu.src[0].sel = V_SQ_ALU_SRC_0;
1421
1422 r600_bytecode_src(&alu.src[1], &ctx->src[0], i);
1423
1424 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
1425
1426 if (i == lasti) {
1427 alu.last = 1;
1428 }
1429 r = r600_bytecode_add_alu(ctx->bc, &alu);
1430 if (r)
1431 return r;
1432 }
1433 return 0;
1434
1435 }
1436
1437 static int cayman_emit_float_instr(struct r600_shader_ctx *ctx)
1438 {
1439 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
1440 int i, j, r;
1441 struct r600_bytecode_alu alu;
1442 int last_slot = (inst->Dst[0].Register.WriteMask & 0x8) ? 4 : 3;
1443
1444 for (i = 0 ; i < last_slot; i++) {
1445 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1446 alu.inst = ctx->inst_info->r600_opcode;
1447 for (j = 0; j < inst->Instruction.NumSrcRegs; j++) {
1448 r600_bytecode_src(&alu.src[j], &ctx->src[j], 0);
1449 }
1450 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
1451 alu.dst.write = (inst->Dst[0].Register.WriteMask >> i) & 1;
1452
1453 if (i == last_slot - 1)
1454 alu.last = 1;
1455 r = r600_bytecode_add_alu(ctx->bc, &alu);
1456 if (r)
1457 return r;
1458 }
1459 return 0;
1460 }
1461
1462 static int cayman_mul_int_instr(struct r600_shader_ctx *ctx)
1463 {
1464 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
1465 int i, j, k, r;
1466 struct r600_bytecode_alu alu;
1467 int last_slot = (inst->Dst[0].Register.WriteMask & 0x8) ? 4 : 3;
1468 for (k = 0; k < last_slot; k++) {
1469 if (!(inst->Dst[0].Register.WriteMask & (1 << k)))
1470 continue;
1471
1472 for (i = 0 ; i < 4; i++) {
1473 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1474 alu.inst = ctx->inst_info->r600_opcode;
1475 for (j = 0; j < inst->Instruction.NumSrcRegs; j++) {
1476 r600_bytecode_src(&alu.src[j], &ctx->src[j], k);
1477 }
1478 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
1479 alu.dst.write = (i == k);
1480 if (i == 3)
1481 alu.last = 1;
1482 r = r600_bytecode_add_alu(ctx->bc, &alu);
1483 if (r)
1484 return r;
1485 }
1486 }
1487 return 0;
1488 }
1489
1490 /*
1491 * r600 - trunc to -PI..PI range
1492 * r700 - normalize by dividing by 2PI
1493 * see fdo bug 27901
1494 */
1495 static int tgsi_setup_trig(struct r600_shader_ctx *ctx)
1496 {
1497 static float half_inv_pi = 1.0 /(3.1415926535 * 2);
1498 static float double_pi = 3.1415926535 * 2;
1499 static float neg_pi = -3.1415926535;
1500
1501 int r;
1502 struct r600_bytecode_alu alu;
1503
1504 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1505 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_MULADD);
1506 alu.is_op3 = 1;
1507
1508 alu.dst.chan = 0;
1509 alu.dst.sel = ctx->temp_reg;
1510 alu.dst.write = 1;
1511
1512 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
1513
1514 alu.src[1].sel = V_SQ_ALU_SRC_LITERAL;
1515 alu.src[1].chan = 0;
1516 alu.src[1].value = *(uint32_t *)&half_inv_pi;
1517 alu.src[2].sel = V_SQ_ALU_SRC_0_5;
1518 alu.src[2].chan = 0;
1519 alu.last = 1;
1520 r = r600_bytecode_add_alu(ctx->bc, &alu);
1521 if (r)
1522 return r;
1523
1524 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1525 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FRACT);
1526
1527 alu.dst.chan = 0;
1528 alu.dst.sel = ctx->temp_reg;
1529 alu.dst.write = 1;
1530
1531 alu.src[0].sel = ctx->temp_reg;
1532 alu.src[0].chan = 0;
1533 alu.last = 1;
1534 r = r600_bytecode_add_alu(ctx->bc, &alu);
1535 if (r)
1536 return r;
1537
1538 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1539 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_MULADD);
1540 alu.is_op3 = 1;
1541
1542 alu.dst.chan = 0;
1543 alu.dst.sel = ctx->temp_reg;
1544 alu.dst.write = 1;
1545
1546 alu.src[0].sel = ctx->temp_reg;
1547 alu.src[0].chan = 0;
1548
1549 alu.src[1].sel = V_SQ_ALU_SRC_LITERAL;
1550 alu.src[1].chan = 0;
1551 alu.src[2].sel = V_SQ_ALU_SRC_LITERAL;
1552 alu.src[2].chan = 0;
1553
1554 if (ctx->bc->chip_class == R600) {
1555 alu.src[1].value = *(uint32_t *)&double_pi;
1556 alu.src[2].value = *(uint32_t *)&neg_pi;
1557 } else {
1558 alu.src[1].sel = V_SQ_ALU_SRC_1;
1559 alu.src[2].sel = V_SQ_ALU_SRC_0_5;
1560 alu.src[2].neg = 1;
1561 }
1562
1563 alu.last = 1;
1564 r = r600_bytecode_add_alu(ctx->bc, &alu);
1565 if (r)
1566 return r;
1567 return 0;
1568 }
1569
1570 static int cayman_trig(struct r600_shader_ctx *ctx)
1571 {
1572 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
1573 struct r600_bytecode_alu alu;
1574 int last_slot = (inst->Dst[0].Register.WriteMask & 0x8) ? 4 : 3;
1575 int i, r;
1576
1577 r = tgsi_setup_trig(ctx);
1578 if (r)
1579 return r;
1580
1581
1582 for (i = 0; i < last_slot; i++) {
1583 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1584 alu.inst = ctx->inst_info->r600_opcode;
1585 alu.dst.chan = i;
1586
1587 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
1588 alu.dst.write = (inst->Dst[0].Register.WriteMask >> i) & 1;
1589
1590 alu.src[0].sel = ctx->temp_reg;
1591 alu.src[0].chan = 0;
1592 if (i == last_slot - 1)
1593 alu.last = 1;
1594 r = r600_bytecode_add_alu(ctx->bc, &alu);
1595 if (r)
1596 return r;
1597 }
1598 return 0;
1599 }
1600
1601 static int tgsi_trig(struct r600_shader_ctx *ctx)
1602 {
1603 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
1604 struct r600_bytecode_alu alu;
1605 int i, r;
1606 int lasti = tgsi_last_instruction(inst->Dst[0].Register.WriteMask);
1607
1608 r = tgsi_setup_trig(ctx);
1609 if (r)
1610 return r;
1611
1612 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1613 alu.inst = ctx->inst_info->r600_opcode;
1614 alu.dst.chan = 0;
1615 alu.dst.sel = ctx->temp_reg;
1616 alu.dst.write = 1;
1617
1618 alu.src[0].sel = ctx->temp_reg;
1619 alu.src[0].chan = 0;
1620 alu.last = 1;
1621 r = r600_bytecode_add_alu(ctx->bc, &alu);
1622 if (r)
1623 return r;
1624
1625 /* replicate result */
1626 for (i = 0; i < lasti + 1; i++) {
1627 if (!(inst->Dst[0].Register.WriteMask & (1 << i)))
1628 continue;
1629
1630 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1631 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV);
1632
1633 alu.src[0].sel = ctx->temp_reg;
1634 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
1635 if (i == lasti)
1636 alu.last = 1;
1637 r = r600_bytecode_add_alu(ctx->bc, &alu);
1638 if (r)
1639 return r;
1640 }
1641 return 0;
1642 }
1643
1644 static int tgsi_scs(struct r600_shader_ctx *ctx)
1645 {
1646 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
1647 struct r600_bytecode_alu alu;
1648 int i, r;
1649
1650 /* We'll only need the trig stuff if we are going to write to the
1651 * X or Y components of the destination vector.
1652 */
1653 if (likely(inst->Dst[0].Register.WriteMask & TGSI_WRITEMASK_XY)) {
1654 r = tgsi_setup_trig(ctx);
1655 if (r)
1656 return r;
1657 }
1658
1659 /* dst.x = COS */
1660 if (inst->Dst[0].Register.WriteMask & TGSI_WRITEMASK_X) {
1661 if (ctx->bc->chip_class == CAYMAN) {
1662 for (i = 0 ; i < 3; i++) {
1663 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1664 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_COS);
1665 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
1666
1667 if (i == 0)
1668 alu.dst.write = 1;
1669 else
1670 alu.dst.write = 0;
1671 alu.src[0].sel = ctx->temp_reg;
1672 alu.src[0].chan = 0;
1673 if (i == 2)
1674 alu.last = 1;
1675 r = r600_bytecode_add_alu(ctx->bc, &alu);
1676 if (r)
1677 return r;
1678 }
1679 } else {
1680 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1681 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_COS);
1682 tgsi_dst(ctx, &inst->Dst[0], 0, &alu.dst);
1683
1684 alu.src[0].sel = ctx->temp_reg;
1685 alu.src[0].chan = 0;
1686 alu.last = 1;
1687 r = r600_bytecode_add_alu(ctx->bc, &alu);
1688 if (r)
1689 return r;
1690 }
1691 }
1692
1693 /* dst.y = SIN */
1694 if (inst->Dst[0].Register.WriteMask & TGSI_WRITEMASK_Y) {
1695 if (ctx->bc->chip_class == CAYMAN) {
1696 for (i = 0 ; i < 3; i++) {
1697 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1698 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SIN);
1699 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
1700 if (i == 1)
1701 alu.dst.write = 1;
1702 else
1703 alu.dst.write = 0;
1704 alu.src[0].sel = ctx->temp_reg;
1705 alu.src[0].chan = 0;
1706 if (i == 2)
1707 alu.last = 1;
1708 r = r600_bytecode_add_alu(ctx->bc, &alu);
1709 if (r)
1710 return r;
1711 }
1712 } else {
1713 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1714 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SIN);
1715 tgsi_dst(ctx, &inst->Dst[0], 1, &alu.dst);
1716
1717 alu.src[0].sel = ctx->temp_reg;
1718 alu.src[0].chan = 0;
1719 alu.last = 1;
1720 r = r600_bytecode_add_alu(ctx->bc, &alu);
1721 if (r)
1722 return r;
1723 }
1724 }
1725
1726 /* dst.z = 0.0; */
1727 if (inst->Dst[0].Register.WriteMask & TGSI_WRITEMASK_Z) {
1728 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1729
1730 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV);
1731
1732 tgsi_dst(ctx, &inst->Dst[0], 2, &alu.dst);
1733
1734 alu.src[0].sel = V_SQ_ALU_SRC_0;
1735 alu.src[0].chan = 0;
1736
1737 alu.last = 1;
1738
1739 r = r600_bytecode_add_alu(ctx->bc, &alu);
1740 if (r)
1741 return r;
1742 }
1743
1744 /* dst.w = 1.0; */
1745 if (inst->Dst[0].Register.WriteMask & TGSI_WRITEMASK_W) {
1746 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1747
1748 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV);
1749
1750 tgsi_dst(ctx, &inst->Dst[0], 3, &alu.dst);
1751
1752 alu.src[0].sel = V_SQ_ALU_SRC_1;
1753 alu.src[0].chan = 0;
1754
1755 alu.last = 1;
1756
1757 r = r600_bytecode_add_alu(ctx->bc, &alu);
1758 if (r)
1759 return r;
1760 }
1761
1762 return 0;
1763 }
1764
1765 static int tgsi_kill(struct r600_shader_ctx *ctx)
1766 {
1767 struct r600_bytecode_alu alu;
1768 int i, r;
1769
1770 for (i = 0; i < 4; i++) {
1771 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1772 alu.inst = ctx->inst_info->r600_opcode;
1773
1774 alu.dst.chan = i;
1775
1776 alu.src[0].sel = V_SQ_ALU_SRC_0;
1777
1778 if (ctx->inst_info->tgsi_opcode == TGSI_OPCODE_KILP) {
1779 alu.src[1].sel = V_SQ_ALU_SRC_1;
1780 alu.src[1].neg = 1;
1781 } else {
1782 r600_bytecode_src(&alu.src[1], &ctx->src[0], i);
1783 }
1784 if (i == 3) {
1785 alu.last = 1;
1786 }
1787 r = r600_bytecode_add_alu(ctx->bc, &alu);
1788 if (r)
1789 return r;
1790 }
1791
1792 /* kill must be last in ALU */
1793 ctx->bc->force_add_cf = 1;
1794 ctx->shader->uses_kill = TRUE;
1795 return 0;
1796 }
1797
1798 static int tgsi_lit(struct r600_shader_ctx *ctx)
1799 {
1800 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
1801 struct r600_bytecode_alu alu;
1802 int r;
1803
1804 /* tmp.x = max(src.y, 0.0) */
1805 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1806 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MAX);
1807 r600_bytecode_src(&alu.src[0], &ctx->src[0], 1);
1808 alu.src[1].sel = V_SQ_ALU_SRC_0; /*0.0*/
1809 alu.src[1].chan = 1;
1810
1811 alu.dst.sel = ctx->temp_reg;
1812 alu.dst.chan = 0;
1813 alu.dst.write = 1;
1814
1815 alu.last = 1;
1816 r = r600_bytecode_add_alu(ctx->bc, &alu);
1817 if (r)
1818 return r;
1819
1820 if (inst->Dst[0].Register.WriteMask & (1 << 2))
1821 {
1822 int chan;
1823 int sel;
1824 int i;
1825
1826 if (ctx->bc->chip_class == CAYMAN) {
1827 for (i = 0; i < 3; i++) {
1828 /* tmp.z = log(tmp.x) */
1829 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1830 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LOG_CLAMPED);
1831 alu.src[0].sel = ctx->temp_reg;
1832 alu.src[0].chan = 0;
1833 alu.dst.sel = ctx->temp_reg;
1834 alu.dst.chan = i;
1835 if (i == 2) {
1836 alu.dst.write = 1;
1837 alu.last = 1;
1838 } else
1839 alu.dst.write = 0;
1840
1841 r = r600_bytecode_add_alu(ctx->bc, &alu);
1842 if (r)
1843 return r;
1844 }
1845 } else {
1846 /* tmp.z = log(tmp.x) */
1847 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1848 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LOG_CLAMPED);
1849 alu.src[0].sel = ctx->temp_reg;
1850 alu.src[0].chan = 0;
1851 alu.dst.sel = ctx->temp_reg;
1852 alu.dst.chan = 2;
1853 alu.dst.write = 1;
1854 alu.last = 1;
1855 r = r600_bytecode_add_alu(ctx->bc, &alu);
1856 if (r)
1857 return r;
1858 }
1859
1860 chan = alu.dst.chan;
1861 sel = alu.dst.sel;
1862
1863 /* tmp.x = amd MUL_LIT(tmp.z, src.w, src.x ) */
1864 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1865 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_MUL_LIT);
1866 alu.src[0].sel = sel;
1867 alu.src[0].chan = chan;
1868 r600_bytecode_src(&alu.src[1], &ctx->src[0], 3);
1869 r600_bytecode_src(&alu.src[2], &ctx->src[0], 0);
1870 alu.dst.sel = ctx->temp_reg;
1871 alu.dst.chan = 0;
1872 alu.dst.write = 1;
1873 alu.is_op3 = 1;
1874 alu.last = 1;
1875 r = r600_bytecode_add_alu(ctx->bc, &alu);
1876 if (r)
1877 return r;
1878
1879 if (ctx->bc->chip_class == CAYMAN) {
1880 for (i = 0; i < 3; i++) {
1881 /* dst.z = exp(tmp.x) */
1882 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1883 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_EXP_IEEE);
1884 alu.src[0].sel = ctx->temp_reg;
1885 alu.src[0].chan = 0;
1886 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
1887 if (i == 2) {
1888 alu.dst.write = 1;
1889 alu.last = 1;
1890 } else
1891 alu.dst.write = 0;
1892 r = r600_bytecode_add_alu(ctx->bc, &alu);
1893 if (r)
1894 return r;
1895 }
1896 } else {
1897 /* dst.z = exp(tmp.x) */
1898 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1899 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_EXP_IEEE);
1900 alu.src[0].sel = ctx->temp_reg;
1901 alu.src[0].chan = 0;
1902 tgsi_dst(ctx, &inst->Dst[0], 2, &alu.dst);
1903 alu.last = 1;
1904 r = r600_bytecode_add_alu(ctx->bc, &alu);
1905 if (r)
1906 return r;
1907 }
1908 }
1909
1910 /* dst.x, <- 1.0 */
1911 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1912 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV);
1913 alu.src[0].sel = V_SQ_ALU_SRC_1; /*1.0*/
1914 alu.src[0].chan = 0;
1915 tgsi_dst(ctx, &inst->Dst[0], 0, &alu.dst);
1916 alu.dst.write = (inst->Dst[0].Register.WriteMask >> 0) & 1;
1917 r = r600_bytecode_add_alu(ctx->bc, &alu);
1918 if (r)
1919 return r;
1920
1921 /* dst.y = max(src.x, 0.0) */
1922 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1923 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MAX);
1924 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
1925 alu.src[1].sel = V_SQ_ALU_SRC_0; /*0.0*/
1926 alu.src[1].chan = 0;
1927 tgsi_dst(ctx, &inst->Dst[0], 1, &alu.dst);
1928 alu.dst.write = (inst->Dst[0].Register.WriteMask >> 1) & 1;
1929 r = r600_bytecode_add_alu(ctx->bc, &alu);
1930 if (r)
1931 return r;
1932
1933 /* dst.w, <- 1.0 */
1934 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1935 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV);
1936 alu.src[0].sel = V_SQ_ALU_SRC_1;
1937 alu.src[0].chan = 0;
1938 tgsi_dst(ctx, &inst->Dst[0], 3, &alu.dst);
1939 alu.dst.write = (inst->Dst[0].Register.WriteMask >> 3) & 1;
1940 alu.last = 1;
1941 r = r600_bytecode_add_alu(ctx->bc, &alu);
1942 if (r)
1943 return r;
1944
1945 return 0;
1946 }
1947
1948 static int tgsi_rsq(struct r600_shader_ctx *ctx)
1949 {
1950 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
1951 struct r600_bytecode_alu alu;
1952 int i, r;
1953
1954 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1955
1956 /* FIXME:
1957 * For state trackers other than OpenGL, we'll want to use
1958 * _RECIPSQRT_IEEE instead.
1959 */
1960 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIPSQRT_CLAMPED);
1961
1962 for (i = 0; i < inst->Instruction.NumSrcRegs; i++) {
1963 r600_bytecode_src(&alu.src[i], &ctx->src[i], 0);
1964 r600_bytecode_src_set_abs(&alu.src[i]);
1965 }
1966 alu.dst.sel = ctx->temp_reg;
1967 alu.dst.write = 1;
1968 alu.last = 1;
1969 r = r600_bytecode_add_alu(ctx->bc, &alu);
1970 if (r)
1971 return r;
1972 /* replicate result */
1973 return tgsi_helper_tempx_replicate(ctx);
1974 }
1975
1976 static int tgsi_helper_tempx_replicate(struct r600_shader_ctx *ctx)
1977 {
1978 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
1979 struct r600_bytecode_alu alu;
1980 int i, r;
1981
1982 for (i = 0; i < 4; i++) {
1983 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
1984 alu.src[0].sel = ctx->temp_reg;
1985 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV);
1986 alu.dst.chan = i;
1987 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
1988 alu.dst.write = (inst->Dst[0].Register.WriteMask >> i) & 1;
1989 if (i == 3)
1990 alu.last = 1;
1991 r = r600_bytecode_add_alu(ctx->bc, &alu);
1992 if (r)
1993 return r;
1994 }
1995 return 0;
1996 }
1997
1998 static int tgsi_trans_srcx_replicate(struct r600_shader_ctx *ctx)
1999 {
2000 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
2001 struct r600_bytecode_alu alu;
2002 int i, r;
2003
2004 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2005 alu.inst = ctx->inst_info->r600_opcode;
2006 for (i = 0; i < inst->Instruction.NumSrcRegs; i++) {
2007 r600_bytecode_src(&alu.src[i], &ctx->src[i], 0);
2008 }
2009 alu.dst.sel = ctx->temp_reg;
2010 alu.dst.write = 1;
2011 alu.last = 1;
2012 r = r600_bytecode_add_alu(ctx->bc, &alu);
2013 if (r)
2014 return r;
2015 /* replicate result */
2016 return tgsi_helper_tempx_replicate(ctx);
2017 }
2018
2019 static int cayman_pow(struct r600_shader_ctx *ctx)
2020 {
2021 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
2022 int i, r;
2023 struct r600_bytecode_alu alu;
2024 int last_slot = (inst->Dst[0].Register.WriteMask & 0x8) ? 4 : 3;
2025
2026 for (i = 0; i < 3; i++) {
2027 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2028 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LOG_IEEE);
2029 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
2030 alu.dst.sel = ctx->temp_reg;
2031 alu.dst.chan = i;
2032 alu.dst.write = 1;
2033 if (i == 2)
2034 alu.last = 1;
2035 r = r600_bytecode_add_alu(ctx->bc, &alu);
2036 if (r)
2037 return r;
2038 }
2039
2040 /* b * LOG2(a) */
2041 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2042 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MUL);
2043 r600_bytecode_src(&alu.src[0], &ctx->src[1], 0);
2044 alu.src[1].sel = ctx->temp_reg;
2045 alu.dst.sel = ctx->temp_reg;
2046 alu.dst.write = 1;
2047 alu.last = 1;
2048 r = r600_bytecode_add_alu(ctx->bc, &alu);
2049 if (r)
2050 return r;
2051
2052 for (i = 0; i < last_slot; i++) {
2053 /* POW(a,b) = EXP2(b * LOG2(a))*/
2054 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2055 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_EXP_IEEE);
2056 alu.src[0].sel = ctx->temp_reg;
2057
2058 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
2059 alu.dst.write = (inst->Dst[0].Register.WriteMask >> i) & 1;
2060 if (i == last_slot - 1)
2061 alu.last = 1;
2062 r = r600_bytecode_add_alu(ctx->bc, &alu);
2063 if (r)
2064 return r;
2065 }
2066 return 0;
2067 }
2068
2069 static int tgsi_pow(struct r600_shader_ctx *ctx)
2070 {
2071 struct r600_bytecode_alu alu;
2072 int r;
2073
2074 /* LOG2(a) */
2075 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2076 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LOG_IEEE);
2077 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
2078 alu.dst.sel = ctx->temp_reg;
2079 alu.dst.write = 1;
2080 alu.last = 1;
2081 r = r600_bytecode_add_alu(ctx->bc, &alu);
2082 if (r)
2083 return r;
2084 /* b * LOG2(a) */
2085 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2086 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MUL);
2087 r600_bytecode_src(&alu.src[0], &ctx->src[1], 0);
2088 alu.src[1].sel = ctx->temp_reg;
2089 alu.dst.sel = ctx->temp_reg;
2090 alu.dst.write = 1;
2091 alu.last = 1;
2092 r = r600_bytecode_add_alu(ctx->bc, &alu);
2093 if (r)
2094 return r;
2095 /* POW(a,b) = EXP2(b * LOG2(a))*/
2096 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2097 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_EXP_IEEE);
2098 alu.src[0].sel = ctx->temp_reg;
2099 alu.dst.sel = ctx->temp_reg;
2100 alu.dst.write = 1;
2101 alu.last = 1;
2102 r = r600_bytecode_add_alu(ctx->bc, &alu);
2103 if (r)
2104 return r;
2105 return tgsi_helper_tempx_replicate(ctx);
2106 }
2107
2108 static int tgsi_divmod(struct r600_shader_ctx *ctx, int mod, int signed_op)
2109 {
2110 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
2111 struct r600_bytecode_alu alu;
2112 int i, r, j;
2113 unsigned write_mask = inst->Dst[0].Register.WriteMask;
2114 int tmp0 = ctx->temp_reg;
2115 int tmp1 = r600_get_temp(ctx);
2116 int tmp2 = r600_get_temp(ctx);
2117 int tmp3 = r600_get_temp(ctx);
2118 /* Unsigned path:
2119 *
2120 * we need to represent src1 as src2*q + r, where q - quotient, r - remainder
2121 *
2122 * 1. tmp0.x = rcp (src2) = 2^32/src2 + e, where e is rounding error
2123 * 2. tmp0.z = lo (tmp0.x * src2)
2124 * 3. tmp0.w = -tmp0.z
2125 * 4. tmp0.y = hi (tmp0.x * src2)
2126 * 5. tmp0.z = (tmp0.y == 0 ? tmp0.w : tmp0.z) = abs(lo(rcp*src2))
2127 * 6. tmp0.w = hi (tmp0.z * tmp0.x) = e, rounding error
2128 * 7. tmp1.x = tmp0.x - tmp0.w
2129 * 8. tmp1.y = tmp0.x + tmp0.w
2130 * 9. tmp0.x = (tmp0.y == 0 ? tmp1.y : tmp1.x)
2131 * 10. tmp0.z = hi(tmp0.x * src1) = q
2132 * 11. tmp0.y = lo (tmp0.z * src2) = src2*q = src1 - r
2133 *
2134 * 12. tmp0.w = src1 - tmp0.y = r
2135 * 13. tmp1.x = tmp0.w >= src2 = r >= src2 (uint comparison)
2136 * 14. tmp1.y = src1 >= tmp0.y = r >= 0 (uint comparison)
2137 *
2138 * if DIV
2139 *
2140 * 15. tmp1.z = tmp0.z + 1 = q + 1
2141 * 16. tmp1.w = tmp0.z - 1 = q - 1
2142 *
2143 * else MOD
2144 *
2145 * 15. tmp1.z = tmp0.w - src2 = r - src2
2146 * 16. tmp1.w = tmp0.w + src2 = r + src2
2147 *
2148 * endif
2149 *
2150 * 17. tmp1.x = tmp1.x & tmp1.y
2151 *
2152 * DIV: 18. tmp0.z = tmp1.x==0 ? tmp0.z : tmp1.z
2153 * MOD: 18. tmp0.z = tmp1.x==0 ? tmp0.w : tmp1.z
2154 *
2155 * 19. tmp0.z = tmp1.y==0 ? tmp1.w : tmp0.z
2156 * 20. dst = src2==0 ? MAX_UINT : tmp0.z
2157 *
2158 * Signed path:
2159 *
2160 * Same as unsigned, using abs values of the operands,
2161 * and fixing the sign of the result in the end.
2162 */
2163
2164 for (i = 0; i < 4; i++) {
2165 if (!(write_mask & (1<<i)))
2166 continue;
2167
2168 if (signed_op) {
2169
2170 /* tmp2.x = -src0 */
2171 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2172 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SUB_INT);
2173
2174 alu.dst.sel = tmp2;
2175 alu.dst.chan = 0;
2176 alu.dst.write = 1;
2177
2178 alu.src[0].sel = V_SQ_ALU_SRC_0;
2179
2180 r600_bytecode_src(&alu.src[1], &ctx->src[0], i);
2181
2182 alu.last = 1;
2183 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2184 return r;
2185
2186 /* tmp2.y = -src1 */
2187 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2188 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SUB_INT);
2189
2190 alu.dst.sel = tmp2;
2191 alu.dst.chan = 1;
2192 alu.dst.write = 1;
2193
2194 alu.src[0].sel = V_SQ_ALU_SRC_0;
2195
2196 r600_bytecode_src(&alu.src[1], &ctx->src[1], i);
2197
2198 alu.last = 1;
2199 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2200 return r;
2201
2202 /* tmp2.z sign bit is set if src0 and src2 signs are different */
2203 /* it will be a sign of the quotient */
2204 if (!mod) {
2205
2206 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2207 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_XOR_INT);
2208
2209 alu.dst.sel = tmp2;
2210 alu.dst.chan = 2;
2211 alu.dst.write = 1;
2212
2213 r600_bytecode_src(&alu.src[0], &ctx->src[0], i);
2214 r600_bytecode_src(&alu.src[1], &ctx->src[1], i);
2215
2216 alu.last = 1;
2217 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2218 return r;
2219 }
2220
2221 /* tmp2.x = |src0| */
2222 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2223 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_CNDGE_INT);
2224 alu.is_op3 = 1;
2225
2226 alu.dst.sel = tmp2;
2227 alu.dst.chan = 0;
2228 alu.dst.write = 1;
2229
2230 r600_bytecode_src(&alu.src[0], &ctx->src[0], i);
2231 r600_bytecode_src(&alu.src[1], &ctx->src[0], i);
2232 alu.src[2].sel = tmp2;
2233 alu.src[2].chan = 0;
2234
2235 alu.last = 1;
2236 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2237 return r;
2238
2239 /* tmp2.y = |src1| */
2240 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2241 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_CNDGE_INT);
2242 alu.is_op3 = 1;
2243
2244 alu.dst.sel = tmp2;
2245 alu.dst.chan = 1;
2246 alu.dst.write = 1;
2247
2248 r600_bytecode_src(&alu.src[0], &ctx->src[1], i);
2249 r600_bytecode_src(&alu.src[1], &ctx->src[1], i);
2250 alu.src[2].sel = tmp2;
2251 alu.src[2].chan = 1;
2252
2253 alu.last = 1;
2254 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2255 return r;
2256
2257 }
2258
2259 /* 1. tmp0.x = rcp_u (src2) = 2^32/src2 + e, where e is rounding error */
2260 if (ctx->bc->chip_class == CAYMAN) {
2261 /* tmp3.x = u2f(src2) */
2262 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2263 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_UINT_TO_FLT);
2264
2265 alu.dst.sel = tmp3;
2266 alu.dst.chan = 0;
2267 alu.dst.write = 1;
2268
2269 if (signed_op) {
2270 alu.src[0].sel = tmp2;
2271 alu.src[0].chan = 1;
2272 } else {
2273 r600_bytecode_src(&alu.src[0], &ctx->src[1], i);
2274 }
2275
2276 alu.last = 1;
2277 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2278 return r;
2279
2280 /* tmp0.x = recip(tmp3.x) */
2281 for (j = 0 ; j < 3; j++) {
2282 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2283 alu.inst = EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_IEEE;
2284
2285 alu.dst.sel = tmp0;
2286 alu.dst.chan = j;
2287 alu.dst.write = (j == 0);
2288
2289 alu.src[0].sel = tmp3;
2290 alu.src[0].chan = 0;
2291
2292 if (j == 2)
2293 alu.last = 1;
2294 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2295 return r;
2296 }
2297
2298 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2299 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MUL);
2300
2301 alu.src[0].sel = tmp0;
2302 alu.src[0].chan = 0;
2303
2304 alu.src[1].sel = V_SQ_ALU_SRC_LITERAL;
2305 alu.src[1].value = 0x4f800000;
2306
2307 alu.dst.sel = tmp3;
2308 alu.dst.write = 1;
2309 alu.last = 1;
2310 r = r600_bytecode_add_alu(ctx->bc, &alu);
2311 if (r)
2312 return r;
2313
2314 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2315 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLT_TO_UINT);
2316
2317 alu.dst.sel = tmp0;
2318 alu.dst.chan = 0;
2319 alu.dst.write = 1;
2320
2321 alu.src[0].sel = tmp3;
2322 alu.src[0].chan = 0;
2323
2324 alu.last = 1;
2325 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2326 return r;
2327
2328 } else {
2329 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2330 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_UINT);
2331
2332 alu.dst.sel = tmp0;
2333 alu.dst.chan = 0;
2334 alu.dst.write = 1;
2335
2336 if (signed_op) {
2337 alu.src[0].sel = tmp2;
2338 alu.src[0].chan = 1;
2339 } else {
2340 r600_bytecode_src(&alu.src[0], &ctx->src[1], i);
2341 }
2342
2343 alu.last = 1;
2344 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2345 return r;
2346 }
2347
2348 /* 2. tmp0.z = lo (tmp0.x * src2) */
2349 if (ctx->bc->chip_class == CAYMAN) {
2350 for (j = 0 ; j < 4; j++) {
2351 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2352 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MULLO_UINT);
2353
2354 alu.dst.sel = tmp0;
2355 alu.dst.chan = j;
2356 alu.dst.write = (j == 2);
2357
2358 alu.src[0].sel = tmp0;
2359 alu.src[0].chan = 0;
2360 if (signed_op) {
2361 alu.src[1].sel = tmp2;
2362 alu.src[1].chan = 1;
2363 } else {
2364 r600_bytecode_src(&alu.src[1], &ctx->src[1], i);
2365 }
2366
2367 alu.last = (j == 3);
2368 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2369 return r;
2370 }
2371 } else {
2372 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2373 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MULLO_UINT);
2374
2375 alu.dst.sel = tmp0;
2376 alu.dst.chan = 2;
2377 alu.dst.write = 1;
2378
2379 alu.src[0].sel = tmp0;
2380 alu.src[0].chan = 0;
2381 if (signed_op) {
2382 alu.src[1].sel = tmp2;
2383 alu.src[1].chan = 1;
2384 } else {
2385 r600_bytecode_src(&alu.src[1], &ctx->src[1], i);
2386 }
2387
2388 alu.last = 1;
2389 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2390 return r;
2391 }
2392
2393 /* 3. tmp0.w = -tmp0.z */
2394 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2395 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SUB_INT);
2396
2397 alu.dst.sel = tmp0;
2398 alu.dst.chan = 3;
2399 alu.dst.write = 1;
2400
2401 alu.src[0].sel = V_SQ_ALU_SRC_0;
2402 alu.src[1].sel = tmp0;
2403 alu.src[1].chan = 2;
2404
2405 alu.last = 1;
2406 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2407 return r;
2408
2409 /* 4. tmp0.y = hi (tmp0.x * src2) */
2410 if (ctx->bc->chip_class == CAYMAN) {
2411 for (j = 0 ; j < 4; j++) {
2412 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2413 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MULHI_UINT);
2414
2415 alu.dst.sel = tmp0;
2416 alu.dst.chan = j;
2417 alu.dst.write = (j == 1);
2418
2419 alu.src[0].sel = tmp0;
2420 alu.src[0].chan = 0;
2421
2422 if (signed_op) {
2423 alu.src[1].sel = tmp2;
2424 alu.src[1].chan = 1;
2425 } else {
2426 r600_bytecode_src(&alu.src[1], &ctx->src[1], i);
2427 }
2428 alu.last = (j == 3);
2429 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2430 return r;
2431 }
2432 } else {
2433 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2434 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MULHI_UINT);
2435
2436 alu.dst.sel = tmp0;
2437 alu.dst.chan = 1;
2438 alu.dst.write = 1;
2439
2440 alu.src[0].sel = tmp0;
2441 alu.src[0].chan = 0;
2442
2443 if (signed_op) {
2444 alu.src[1].sel = tmp2;
2445 alu.src[1].chan = 1;
2446 } else {
2447 r600_bytecode_src(&alu.src[1], &ctx->src[1], i);
2448 }
2449
2450 alu.last = 1;
2451 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2452 return r;
2453 }
2454
2455 /* 5. tmp0.z = (tmp0.y == 0 ? tmp0.w : tmp0.z) = abs(lo(rcp*src)) */
2456 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2457 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_CNDE_INT);
2458 alu.is_op3 = 1;
2459
2460 alu.dst.sel = tmp0;
2461 alu.dst.chan = 2;
2462 alu.dst.write = 1;
2463
2464 alu.src[0].sel = tmp0;
2465 alu.src[0].chan = 1;
2466 alu.src[1].sel = tmp0;
2467 alu.src[1].chan = 3;
2468 alu.src[2].sel = tmp0;
2469 alu.src[2].chan = 2;
2470
2471 alu.last = 1;
2472 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2473 return r;
2474
2475 /* 6. tmp0.w = hi (tmp0.z * tmp0.x) = e, rounding error */
2476 if (ctx->bc->chip_class == CAYMAN) {
2477 for (j = 0 ; j < 4; j++) {
2478 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2479 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MULHI_UINT);
2480
2481 alu.dst.sel = tmp0;
2482 alu.dst.chan = j;
2483 alu.dst.write = (j == 3);
2484
2485 alu.src[0].sel = tmp0;
2486 alu.src[0].chan = 2;
2487
2488 alu.src[1].sel = tmp0;
2489 alu.src[1].chan = 0;
2490
2491 alu.last = (j == 3);
2492 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2493 return r;
2494 }
2495 } else {
2496 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2497 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MULHI_UINT);
2498
2499 alu.dst.sel = tmp0;
2500 alu.dst.chan = 3;
2501 alu.dst.write = 1;
2502
2503 alu.src[0].sel = tmp0;
2504 alu.src[0].chan = 2;
2505
2506 alu.src[1].sel = tmp0;
2507 alu.src[1].chan = 0;
2508
2509 alu.last = 1;
2510 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2511 return r;
2512 }
2513
2514 /* 7. tmp1.x = tmp0.x - tmp0.w */
2515 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2516 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SUB_INT);
2517
2518 alu.dst.sel = tmp1;
2519 alu.dst.chan = 0;
2520 alu.dst.write = 1;
2521
2522 alu.src[0].sel = tmp0;
2523 alu.src[0].chan = 0;
2524 alu.src[1].sel = tmp0;
2525 alu.src[1].chan = 3;
2526
2527 alu.last = 1;
2528 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2529 return r;
2530
2531 /* 8. tmp1.y = tmp0.x + tmp0.w */
2532 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2533 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ADD_INT);
2534
2535 alu.dst.sel = tmp1;
2536 alu.dst.chan = 1;
2537 alu.dst.write = 1;
2538
2539 alu.src[0].sel = tmp0;
2540 alu.src[0].chan = 0;
2541 alu.src[1].sel = tmp0;
2542 alu.src[1].chan = 3;
2543
2544 alu.last = 1;
2545 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2546 return r;
2547
2548 /* 9. tmp0.x = (tmp0.y == 0 ? tmp1.y : tmp1.x) */
2549 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2550 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_CNDE_INT);
2551 alu.is_op3 = 1;
2552
2553 alu.dst.sel = tmp0;
2554 alu.dst.chan = 0;
2555 alu.dst.write = 1;
2556
2557 alu.src[0].sel = tmp0;
2558 alu.src[0].chan = 1;
2559 alu.src[1].sel = tmp1;
2560 alu.src[1].chan = 1;
2561 alu.src[2].sel = tmp1;
2562 alu.src[2].chan = 0;
2563
2564 alu.last = 1;
2565 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2566 return r;
2567
2568 /* 10. tmp0.z = hi(tmp0.x * src1) = q */
2569 if (ctx->bc->chip_class == CAYMAN) {
2570 for (j = 0 ; j < 4; j++) {
2571 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2572 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MULHI_UINT);
2573
2574 alu.dst.sel = tmp0;
2575 alu.dst.chan = j;
2576 alu.dst.write = (j == 2);
2577
2578 alu.src[0].sel = tmp0;
2579 alu.src[0].chan = 0;
2580
2581 if (signed_op) {
2582 alu.src[1].sel = tmp2;
2583 alu.src[1].chan = 0;
2584 } else {
2585 r600_bytecode_src(&alu.src[1], &ctx->src[0], i);
2586 }
2587
2588 alu.last = (j == 3);
2589 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2590 return r;
2591 }
2592 } else {
2593 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2594 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MULHI_UINT);
2595
2596 alu.dst.sel = tmp0;
2597 alu.dst.chan = 2;
2598 alu.dst.write = 1;
2599
2600 alu.src[0].sel = tmp0;
2601 alu.src[0].chan = 0;
2602
2603 if (signed_op) {
2604 alu.src[1].sel = tmp2;
2605 alu.src[1].chan = 0;
2606 } else {
2607 r600_bytecode_src(&alu.src[1], &ctx->src[0], i);
2608 }
2609
2610 alu.last = 1;
2611 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2612 return r;
2613 }
2614
2615 /* 11. tmp0.y = lo (src2 * tmp0.z) = src2*q = src1 - r */
2616 if (ctx->bc->chip_class == CAYMAN) {
2617 for (j = 0 ; j < 4; j++) {
2618 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2619 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MULLO_UINT);
2620
2621 alu.dst.sel = tmp0;
2622 alu.dst.chan = j;
2623 alu.dst.write = (j == 1);
2624
2625 if (signed_op) {
2626 alu.src[0].sel = tmp2;
2627 alu.src[0].chan = 1;
2628 } else {
2629 r600_bytecode_src(&alu.src[0], &ctx->src[1], i);
2630 }
2631
2632 alu.src[1].sel = tmp0;
2633 alu.src[1].chan = 2;
2634
2635 alu.last = (j == 3);
2636 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2637 return r;
2638 }
2639 } else {
2640 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2641 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MULLO_UINT);
2642
2643 alu.dst.sel = tmp0;
2644 alu.dst.chan = 1;
2645 alu.dst.write = 1;
2646
2647 if (signed_op) {
2648 alu.src[0].sel = tmp2;
2649 alu.src[0].chan = 1;
2650 } else {
2651 r600_bytecode_src(&alu.src[0], &ctx->src[1], i);
2652 }
2653
2654 alu.src[1].sel = tmp0;
2655 alu.src[1].chan = 2;
2656
2657 alu.last = 1;
2658 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2659 return r;
2660 }
2661
2662 /* 12. tmp0.w = src1 - tmp0.y = r */
2663 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2664 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SUB_INT);
2665
2666 alu.dst.sel = tmp0;
2667 alu.dst.chan = 3;
2668 alu.dst.write = 1;
2669
2670 if (signed_op) {
2671 alu.src[0].sel = tmp2;
2672 alu.src[0].chan = 0;
2673 } else {
2674 r600_bytecode_src(&alu.src[0], &ctx->src[0], i);
2675 }
2676
2677 alu.src[1].sel = tmp0;
2678 alu.src[1].chan = 1;
2679
2680 alu.last = 1;
2681 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2682 return r;
2683
2684 /* 13. tmp1.x = tmp0.w >= src2 = r >= src2 */
2685 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2686 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGE_UINT);
2687
2688 alu.dst.sel = tmp1;
2689 alu.dst.chan = 0;
2690 alu.dst.write = 1;
2691
2692 alu.src[0].sel = tmp0;
2693 alu.src[0].chan = 3;
2694 if (signed_op) {
2695 alu.src[1].sel = tmp2;
2696 alu.src[1].chan = 1;
2697 } else {
2698 r600_bytecode_src(&alu.src[1], &ctx->src[1], i);
2699 }
2700
2701 alu.last = 1;
2702 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2703 return r;
2704
2705 /* 14. tmp1.y = src1 >= tmp0.y = r >= 0 */
2706 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2707 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGE_UINT);
2708
2709 alu.dst.sel = tmp1;
2710 alu.dst.chan = 1;
2711 alu.dst.write = 1;
2712
2713 if (signed_op) {
2714 alu.src[0].sel = tmp2;
2715 alu.src[0].chan = 0;
2716 } else {
2717 r600_bytecode_src(&alu.src[0], &ctx->src[0], i);
2718 }
2719
2720 alu.src[1].sel = tmp0;
2721 alu.src[1].chan = 1;
2722
2723 alu.last = 1;
2724 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2725 return r;
2726
2727 if (mod) { /* UMOD */
2728
2729 /* 15. tmp1.z = tmp0.w - src2 = r - src2 */
2730 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2731 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SUB_INT);
2732
2733 alu.dst.sel = tmp1;
2734 alu.dst.chan = 2;
2735 alu.dst.write = 1;
2736
2737 alu.src[0].sel = tmp0;
2738 alu.src[0].chan = 3;
2739
2740 if (signed_op) {
2741 alu.src[1].sel = tmp2;
2742 alu.src[1].chan = 1;
2743 } else {
2744 r600_bytecode_src(&alu.src[1], &ctx->src[1], i);
2745 }
2746
2747 alu.last = 1;
2748 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2749 return r;
2750
2751 /* 16. tmp1.w = tmp0.w + src2 = r + src2 */
2752 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2753 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ADD_INT);
2754
2755 alu.dst.sel = tmp1;
2756 alu.dst.chan = 3;
2757 alu.dst.write = 1;
2758
2759 alu.src[0].sel = tmp0;
2760 alu.src[0].chan = 3;
2761 if (signed_op) {
2762 alu.src[1].sel = tmp2;
2763 alu.src[1].chan = 1;
2764 } else {
2765 r600_bytecode_src(&alu.src[1], &ctx->src[1], i);
2766 }
2767
2768 alu.last = 1;
2769 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2770 return r;
2771
2772 } else { /* UDIV */
2773
2774 /* 15. tmp1.z = tmp0.z + 1 = q + 1 DIV */
2775 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2776 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ADD_INT);
2777
2778 alu.dst.sel = tmp1;
2779 alu.dst.chan = 2;
2780 alu.dst.write = 1;
2781
2782 alu.src[0].sel = tmp0;
2783 alu.src[0].chan = 2;
2784 alu.src[1].sel = V_SQ_ALU_SRC_1_INT;
2785
2786 alu.last = 1;
2787 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2788 return r;
2789
2790 /* 16. tmp1.w = tmp0.z - 1 = q - 1 */
2791 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2792 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ADD_INT);
2793
2794 alu.dst.sel = tmp1;
2795 alu.dst.chan = 3;
2796 alu.dst.write = 1;
2797
2798 alu.src[0].sel = tmp0;
2799 alu.src[0].chan = 2;
2800 alu.src[1].sel = V_SQ_ALU_SRC_M_1_INT;
2801
2802 alu.last = 1;
2803 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2804 return r;
2805
2806 }
2807
2808 /* 17. tmp1.x = tmp1.x & tmp1.y */
2809 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2810 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_AND_INT);
2811
2812 alu.dst.sel = tmp1;
2813 alu.dst.chan = 0;
2814 alu.dst.write = 1;
2815
2816 alu.src[0].sel = tmp1;
2817 alu.src[0].chan = 0;
2818 alu.src[1].sel = tmp1;
2819 alu.src[1].chan = 1;
2820
2821 alu.last = 1;
2822 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2823 return r;
2824
2825 /* 18. tmp0.z = tmp1.x==0 ? tmp0.z : tmp1.z DIV */
2826 /* 18. tmp0.z = tmp1.x==0 ? tmp0.w : tmp1.z MOD */
2827 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2828 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_CNDE_INT);
2829 alu.is_op3 = 1;
2830
2831 alu.dst.sel = tmp0;
2832 alu.dst.chan = 2;
2833 alu.dst.write = 1;
2834
2835 alu.src[0].sel = tmp1;
2836 alu.src[0].chan = 0;
2837 alu.src[1].sel = tmp0;
2838 alu.src[1].chan = mod ? 3 : 2;
2839 alu.src[2].sel = tmp1;
2840 alu.src[2].chan = 2;
2841
2842 alu.last = 1;
2843 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2844 return r;
2845
2846 /* 19. tmp0.z = tmp1.y==0 ? tmp1.w : tmp0.z */
2847 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2848 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_CNDE_INT);
2849 alu.is_op3 = 1;
2850
2851 if (signed_op) {
2852 alu.dst.sel = tmp0;
2853 alu.dst.chan = 2;
2854 alu.dst.write = 1;
2855 } else {
2856 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
2857 }
2858
2859 alu.src[0].sel = tmp1;
2860 alu.src[0].chan = 1;
2861 alu.src[1].sel = tmp1;
2862 alu.src[1].chan = 3;
2863 alu.src[2].sel = tmp0;
2864 alu.src[2].chan = 2;
2865
2866 alu.last = 1;
2867 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2868 return r;
2869
2870 if (signed_op) {
2871
2872 /* fix the sign of the result */
2873
2874 if (mod) {
2875
2876 /* tmp0.x = -tmp0.z */
2877 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2878 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SUB_INT);
2879
2880 alu.dst.sel = tmp0;
2881 alu.dst.chan = 0;
2882 alu.dst.write = 1;
2883
2884 alu.src[0].sel = V_SQ_ALU_SRC_0;
2885 alu.src[1].sel = tmp0;
2886 alu.src[1].chan = 2;
2887
2888 alu.last = 1;
2889 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2890 return r;
2891
2892 /* sign of the remainder is the same as the sign of src0 */
2893 /* tmp0.x = src0>=0 ? tmp0.z : tmp0.x */
2894 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2895 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_CNDGE_INT);
2896 alu.is_op3 = 1;
2897
2898 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
2899
2900 r600_bytecode_src(&alu.src[0], &ctx->src[0], i);
2901 alu.src[1].sel = tmp0;
2902 alu.src[1].chan = 2;
2903 alu.src[2].sel = tmp0;
2904 alu.src[2].chan = 0;
2905
2906 alu.last = 1;
2907 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2908 return r;
2909
2910 } else {
2911
2912 /* tmp0.x = -tmp0.z */
2913 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2914 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SUB_INT);
2915
2916 alu.dst.sel = tmp0;
2917 alu.dst.chan = 0;
2918 alu.dst.write = 1;
2919
2920 alu.src[0].sel = V_SQ_ALU_SRC_0;
2921 alu.src[1].sel = tmp0;
2922 alu.src[1].chan = 2;
2923
2924 alu.last = 1;
2925 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2926 return r;
2927
2928 /* fix the quotient sign (same as the sign of src0*src1) */
2929 /* tmp0.x = tmp2.z>=0 ? tmp0.z : tmp0.x */
2930 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2931 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_CNDGE_INT);
2932 alu.is_op3 = 1;
2933
2934 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
2935
2936 alu.src[0].sel = tmp2;
2937 alu.src[0].chan = 2;
2938 alu.src[1].sel = tmp0;
2939 alu.src[1].chan = 2;
2940 alu.src[2].sel = tmp0;
2941 alu.src[2].chan = 0;
2942
2943 alu.last = 1;
2944 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
2945 return r;
2946 }
2947 }
2948 }
2949 return 0;
2950 }
2951
2952 static int tgsi_udiv(struct r600_shader_ctx *ctx)
2953 {
2954 return tgsi_divmod(ctx, 0, 0);
2955 }
2956
2957 static int tgsi_umod(struct r600_shader_ctx *ctx)
2958 {
2959 return tgsi_divmod(ctx, 1, 0);
2960 }
2961
2962 static int tgsi_idiv(struct r600_shader_ctx *ctx)
2963 {
2964 return tgsi_divmod(ctx, 0, 1);
2965 }
2966
2967 static int tgsi_imod(struct r600_shader_ctx *ctx)
2968 {
2969 return tgsi_divmod(ctx, 1, 1);
2970 }
2971
2972 static int tgsi_iabs(struct r600_shader_ctx *ctx)
2973 {
2974 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
2975 struct r600_bytecode_alu alu;
2976 int i, r;
2977 unsigned write_mask = inst->Dst[0].Register.WriteMask;
2978 int last_inst = tgsi_last_instruction(write_mask);
2979
2980 /* tmp = -src */
2981 for (i = 0; i < 4; i++) {
2982 if (!(write_mask & (1<<i)))
2983 continue;
2984
2985 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
2986 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SUB_INT);
2987
2988 alu.dst.sel = ctx->temp_reg;
2989 alu.dst.chan = i;
2990 alu.dst.write = 1;
2991
2992 r600_bytecode_src(&alu.src[1], &ctx->src[0], i);
2993 alu.src[0].sel = V_SQ_ALU_SRC_0;
2994
2995 if (i == last_inst)
2996 alu.last = 1;
2997 r = r600_bytecode_add_alu(ctx->bc, &alu);
2998 if (r)
2999 return r;
3000 }
3001
3002 /* dst = (src >= 0 ? src : tmp) */
3003 for (i = 0; i < 4; i++) {
3004 if (!(write_mask & (1<<i)))
3005 continue;
3006
3007 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3008 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_CNDGE_INT);
3009 alu.is_op3 = 1;
3010 alu.dst.write = 1;
3011
3012 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
3013
3014 r600_bytecode_src(&alu.src[0], &ctx->src[0], i);
3015 r600_bytecode_src(&alu.src[1], &ctx->src[0], i);
3016 alu.src[2].sel = ctx->temp_reg;
3017 alu.src[2].chan = i;
3018
3019 if (i == last_inst)
3020 alu.last = 1;
3021 r = r600_bytecode_add_alu(ctx->bc, &alu);
3022 if (r)
3023 return r;
3024 }
3025 return 0;
3026 }
3027
3028 static int tgsi_issg(struct r600_shader_ctx *ctx)
3029 {
3030 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
3031 struct r600_bytecode_alu alu;
3032 int i, r;
3033 unsigned write_mask = inst->Dst[0].Register.WriteMask;
3034 int last_inst = tgsi_last_instruction(write_mask);
3035
3036 /* tmp = (src >= 0 ? src : -1) */
3037 for (i = 0; i < 4; i++) {
3038 if (!(write_mask & (1<<i)))
3039 continue;
3040
3041 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3042 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_CNDGE_INT);
3043 alu.is_op3 = 1;
3044
3045 alu.dst.sel = ctx->temp_reg;
3046 alu.dst.chan = i;
3047 alu.dst.write = 1;
3048
3049 r600_bytecode_src(&alu.src[0], &ctx->src[0], i);
3050 r600_bytecode_src(&alu.src[1], &ctx->src[0], i);
3051 alu.src[2].sel = V_SQ_ALU_SRC_M_1_INT;
3052
3053 if (i == last_inst)
3054 alu.last = 1;
3055 r = r600_bytecode_add_alu(ctx->bc, &alu);
3056 if (r)
3057 return r;
3058 }
3059
3060 /* dst = (tmp > 0 ? 1 : tmp) */
3061 for (i = 0; i < 4; i++) {
3062 if (!(write_mask & (1<<i)))
3063 continue;
3064
3065 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3066 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_CNDGT_INT);
3067 alu.is_op3 = 1;
3068 alu.dst.write = 1;
3069
3070 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
3071
3072 alu.src[0].sel = ctx->temp_reg;
3073 alu.src[0].chan = i;
3074
3075 alu.src[1].sel = V_SQ_ALU_SRC_1_INT;
3076
3077 alu.src[2].sel = ctx->temp_reg;
3078 alu.src[2].chan = i;
3079
3080 if (i == last_inst)
3081 alu.last = 1;
3082 r = r600_bytecode_add_alu(ctx->bc, &alu);
3083 if (r)
3084 return r;
3085 }
3086 return 0;
3087 }
3088
3089
3090
3091 static int tgsi_ssg(struct r600_shader_ctx *ctx)
3092 {
3093 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
3094 struct r600_bytecode_alu alu;
3095 int i, r;
3096
3097 /* tmp = (src > 0 ? 1 : src) */
3098 for (i = 0; i < 4; i++) {
3099 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3100 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_CNDGT);
3101 alu.is_op3 = 1;
3102
3103 alu.dst.sel = ctx->temp_reg;
3104 alu.dst.chan = i;
3105
3106 r600_bytecode_src(&alu.src[0], &ctx->src[0], i);
3107 alu.src[1].sel = V_SQ_ALU_SRC_1;
3108 r600_bytecode_src(&alu.src[2], &ctx->src[0], i);
3109
3110 if (i == 3)
3111 alu.last = 1;
3112 r = r600_bytecode_add_alu(ctx->bc, &alu);
3113 if (r)
3114 return r;
3115 }
3116
3117 /* dst = (-tmp > 0 ? -1 : tmp) */
3118 for (i = 0; i < 4; i++) {
3119 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3120 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_CNDGT);
3121 alu.is_op3 = 1;
3122 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
3123
3124 alu.src[0].sel = ctx->temp_reg;
3125 alu.src[0].chan = i;
3126 alu.src[0].neg = 1;
3127
3128 alu.src[1].sel = V_SQ_ALU_SRC_1;
3129 alu.src[1].neg = 1;
3130
3131 alu.src[2].sel = ctx->temp_reg;
3132 alu.src[2].chan = i;
3133
3134 if (i == 3)
3135 alu.last = 1;
3136 r = r600_bytecode_add_alu(ctx->bc, &alu);
3137 if (r)
3138 return r;
3139 }
3140 return 0;
3141 }
3142
3143 static int tgsi_helper_copy(struct r600_shader_ctx *ctx, struct tgsi_full_instruction *inst)
3144 {
3145 struct r600_bytecode_alu alu;
3146 int i, r;
3147
3148 for (i = 0; i < 4; i++) {
3149 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3150 if (!(inst->Dst[0].Register.WriteMask & (1 << i))) {
3151 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP);
3152 alu.dst.chan = i;
3153 } else {
3154 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV);
3155 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
3156 alu.src[0].sel = ctx->temp_reg;
3157 alu.src[0].chan = i;
3158 }
3159 if (i == 3) {
3160 alu.last = 1;
3161 }
3162 r = r600_bytecode_add_alu(ctx->bc, &alu);
3163 if (r)
3164 return r;
3165 }
3166 return 0;
3167 }
3168
3169 static int tgsi_op3(struct r600_shader_ctx *ctx)
3170 {
3171 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
3172 struct r600_bytecode_alu alu;
3173 int i, j, r;
3174 int lasti = tgsi_last_instruction(inst->Dst[0].Register.WriteMask);
3175
3176 for (i = 0; i < lasti + 1; i++) {
3177 if (!(inst->Dst[0].Register.WriteMask & (1 << i)))
3178 continue;
3179
3180 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3181 alu.inst = ctx->inst_info->r600_opcode;
3182 for (j = 0; j < inst->Instruction.NumSrcRegs; j++) {
3183 r600_bytecode_src(&alu.src[j], &ctx->src[j], i);
3184 }
3185
3186 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
3187 alu.dst.chan = i;
3188 alu.dst.write = 1;
3189 alu.is_op3 = 1;
3190 if (i == lasti) {
3191 alu.last = 1;
3192 }
3193 r = r600_bytecode_add_alu(ctx->bc, &alu);
3194 if (r)
3195 return r;
3196 }
3197 return 0;
3198 }
3199
3200 static int tgsi_dp(struct r600_shader_ctx *ctx)
3201 {
3202 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
3203 struct r600_bytecode_alu alu;
3204 int i, j, r;
3205
3206 for (i = 0; i < 4; i++) {
3207 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3208 alu.inst = ctx->inst_info->r600_opcode;
3209 for (j = 0; j < inst->Instruction.NumSrcRegs; j++) {
3210 r600_bytecode_src(&alu.src[j], &ctx->src[j], i);
3211 }
3212
3213 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
3214 alu.dst.chan = i;
3215 alu.dst.write = (inst->Dst[0].Register.WriteMask >> i) & 1;
3216 /* handle some special cases */
3217 switch (ctx->inst_info->tgsi_opcode) {
3218 case TGSI_OPCODE_DP2:
3219 if (i > 1) {
3220 alu.src[0].sel = alu.src[1].sel = V_SQ_ALU_SRC_0;
3221 alu.src[0].chan = alu.src[1].chan = 0;
3222 }
3223 break;
3224 case TGSI_OPCODE_DP3:
3225 if (i > 2) {
3226 alu.src[0].sel = alu.src[1].sel = V_SQ_ALU_SRC_0;
3227 alu.src[0].chan = alu.src[1].chan = 0;
3228 }
3229 break;
3230 case TGSI_OPCODE_DPH:
3231 if (i == 3) {
3232 alu.src[0].sel = V_SQ_ALU_SRC_1;
3233 alu.src[0].chan = 0;
3234 alu.src[0].neg = 0;
3235 }
3236 break;
3237 default:
3238 break;
3239 }
3240 if (i == 3) {
3241 alu.last = 1;
3242 }
3243 r = r600_bytecode_add_alu(ctx->bc, &alu);
3244 if (r)
3245 return r;
3246 }
3247 return 0;
3248 }
3249
3250 static inline boolean tgsi_tex_src_requires_loading(struct r600_shader_ctx *ctx,
3251 unsigned index)
3252 {
3253 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
3254 return (inst->Src[index].Register.File != TGSI_FILE_TEMPORARY &&
3255 inst->Src[index].Register.File != TGSI_FILE_INPUT &&
3256 inst->Src[index].Register.File != TGSI_FILE_OUTPUT) ||
3257 ctx->src[index].neg || ctx->src[index].abs;
3258 }
3259
3260 static inline unsigned tgsi_tex_get_src_gpr(struct r600_shader_ctx *ctx,
3261 unsigned index)
3262 {
3263 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
3264 return ctx->file_offset[inst->Src[index].Register.File] + inst->Src[index].Register.Index;
3265 }
3266
3267 static int tgsi_tex(struct r600_shader_ctx *ctx)
3268 {
3269 static float one_point_five = 1.5f;
3270 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
3271 struct r600_bytecode_tex tex;
3272 struct r600_bytecode_alu alu;
3273 unsigned src_gpr;
3274 int r, i, j;
3275 int opcode;
3276 /* Texture fetch instructions can only use gprs as source.
3277 * Also they cannot negate the source or take the absolute value */
3278 const boolean src_requires_loading = tgsi_tex_src_requires_loading(ctx, 0);
3279 boolean src_loaded = FALSE;
3280 unsigned sampler_src_reg = 1;
3281 uint8_t offset_x = 0, offset_y = 0, offset_z = 0;
3282
3283 src_gpr = tgsi_tex_get_src_gpr(ctx, 0);
3284
3285 if (inst->Instruction.Opcode == TGSI_OPCODE_TXF) {
3286 /* get offset values */
3287 if (inst->Texture.NumOffsets) {
3288 assert(inst->Texture.NumOffsets == 1);
3289
3290 offset_x = ctx->literals[inst->TexOffsets[0].Index + inst->TexOffsets[0].SwizzleX] << 1;
3291 offset_y = ctx->literals[inst->TexOffsets[0].Index + inst->TexOffsets[0].SwizzleY] << 1;
3292 offset_z = ctx->literals[inst->TexOffsets[0].Index + inst->TexOffsets[0].SwizzleZ] << 1;
3293 }
3294 } else if (inst->Instruction.Opcode == TGSI_OPCODE_TXD) {
3295 /* TGSI moves the sampler to src reg 3 for TXD */
3296 sampler_src_reg = 3;
3297
3298 for (i = 1; i < 3; i++) {
3299 /* set gradients h/v */
3300 memset(&tex, 0, sizeof(struct r600_bytecode_tex));
3301 tex.inst = (i == 1) ? SQ_TEX_INST_SET_GRADIENTS_H :
3302 SQ_TEX_INST_SET_GRADIENTS_V;
3303 tex.sampler_id = tgsi_tex_get_src_gpr(ctx, sampler_src_reg);
3304 tex.resource_id = tex.sampler_id + R600_MAX_CONST_BUFFERS;
3305
3306 if (tgsi_tex_src_requires_loading(ctx, i)) {
3307 tex.src_gpr = r600_get_temp(ctx);
3308 tex.src_sel_x = 0;
3309 tex.src_sel_y = 1;
3310 tex.src_sel_z = 2;
3311 tex.src_sel_w = 3;
3312
3313 for (j = 0; j < 4; j++) {
3314 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3315 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV);
3316 r600_bytecode_src(&alu.src[0], &ctx->src[i], j);
3317 alu.dst.sel = tex.src_gpr;
3318 alu.dst.chan = j;
3319 if (j == 3)
3320 alu.last = 1;
3321 alu.dst.write = 1;
3322 r = r600_bytecode_add_alu(ctx->bc, &alu);
3323 if (r)
3324 return r;
3325 }
3326
3327 } else {
3328 tex.src_gpr = tgsi_tex_get_src_gpr(ctx, i);
3329 tex.src_sel_x = ctx->src[i].swizzle[0];
3330 tex.src_sel_y = ctx->src[i].swizzle[1];
3331 tex.src_sel_z = ctx->src[i].swizzle[2];
3332 tex.src_sel_w = ctx->src[i].swizzle[3];
3333 tex.src_rel = ctx->src[i].rel;
3334 }
3335 tex.dst_gpr = ctx->temp_reg; /* just to avoid confusing the asm scheduler */
3336 tex.dst_sel_x = tex.dst_sel_y = tex.dst_sel_z = tex.dst_sel_w = 7;
3337 if (inst->Texture.Texture != TGSI_TEXTURE_RECT) {
3338 tex.coord_type_x = 1;
3339 tex.coord_type_y = 1;
3340 tex.coord_type_z = 1;
3341 tex.coord_type_w = 1;
3342 }
3343 r = r600_bytecode_add_tex(ctx->bc, &tex);
3344 if (r)
3345 return r;
3346 }
3347 } else if (inst->Instruction.Opcode == TGSI_OPCODE_TXP) {
3348 int out_chan;
3349 /* Add perspective divide */
3350 if (ctx->bc->chip_class == CAYMAN) {
3351 out_chan = 2;
3352 for (i = 0; i < 3; i++) {
3353 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3354 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_IEEE);
3355 r600_bytecode_src(&alu.src[0], &ctx->src[0], 3);
3356
3357 alu.dst.sel = ctx->temp_reg;
3358 alu.dst.chan = i;
3359 if (i == 2)
3360 alu.last = 1;
3361 if (out_chan == i)
3362 alu.dst.write = 1;
3363 r = r600_bytecode_add_alu(ctx->bc, &alu);
3364 if (r)
3365 return r;
3366 }
3367
3368 } else {
3369 out_chan = 3;
3370 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3371 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_IEEE);
3372 r600_bytecode_src(&alu.src[0], &ctx->src[0], 3);
3373
3374 alu.dst.sel = ctx->temp_reg;
3375 alu.dst.chan = out_chan;
3376 alu.last = 1;
3377 alu.dst.write = 1;
3378 r = r600_bytecode_add_alu(ctx->bc, &alu);
3379 if (r)
3380 return r;
3381 }
3382
3383 for (i = 0; i < 3; i++) {
3384 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3385 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MUL);
3386 alu.src[0].sel = ctx->temp_reg;
3387 alu.src[0].chan = out_chan;
3388 r600_bytecode_src(&alu.src[1], &ctx->src[0], i);
3389 alu.dst.sel = ctx->temp_reg;
3390 alu.dst.chan = i;
3391 alu.dst.write = 1;
3392 r = r600_bytecode_add_alu(ctx->bc, &alu);
3393 if (r)
3394 return r;
3395 }
3396 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3397 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV);
3398 alu.src[0].sel = V_SQ_ALU_SRC_1;
3399 alu.src[0].chan = 0;
3400 alu.dst.sel = ctx->temp_reg;
3401 alu.dst.chan = 3;
3402 alu.last = 1;
3403 alu.dst.write = 1;
3404 r = r600_bytecode_add_alu(ctx->bc, &alu);
3405 if (r)
3406 return r;
3407 src_loaded = TRUE;
3408 src_gpr = ctx->temp_reg;
3409 }
3410
3411 if ((inst->Texture.Texture == TGSI_TEXTURE_CUBE ||
3412 inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE) &&
3413 inst->Instruction.Opcode != TGSI_OPCODE_TXQ) {
3414
3415 static const unsigned src0_swizzle[] = {2, 2, 0, 1};
3416 static const unsigned src1_swizzle[] = {1, 0, 2, 2};
3417
3418 /* tmp1.xyzw = CUBE(R0.zzxy, R0.yxzz) */
3419 for (i = 0; i < 4; i++) {
3420 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3421 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_CUBE);
3422 r600_bytecode_src(&alu.src[0], &ctx->src[0], src0_swizzle[i]);
3423 r600_bytecode_src(&alu.src[1], &ctx->src[0], src1_swizzle[i]);
3424 alu.dst.sel = ctx->temp_reg;
3425 alu.dst.chan = i;
3426 if (i == 3)
3427 alu.last = 1;
3428 alu.dst.write = 1;
3429 r = r600_bytecode_add_alu(ctx->bc, &alu);
3430 if (r)
3431 return r;
3432 }
3433
3434 /* tmp1.z = RCP_e(|tmp1.z|) */
3435 if (ctx->bc->chip_class == CAYMAN) {
3436 for (i = 0; i < 3; i++) {
3437 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3438 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_IEEE);
3439 alu.src[0].sel = ctx->temp_reg;
3440 alu.src[0].chan = 2;
3441 alu.src[0].abs = 1;
3442 alu.dst.sel = ctx->temp_reg;
3443 alu.dst.chan = i;
3444 if (i == 2)
3445 alu.dst.write = 1;
3446 if (i == 2)
3447 alu.last = 1;
3448 r = r600_bytecode_add_alu(ctx->bc, &alu);
3449 if (r)
3450 return r;
3451 }
3452 } else {
3453 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3454 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_IEEE);
3455 alu.src[0].sel = ctx->temp_reg;
3456 alu.src[0].chan = 2;
3457 alu.src[0].abs = 1;
3458 alu.dst.sel = ctx->temp_reg;
3459 alu.dst.chan = 2;
3460 alu.dst.write = 1;
3461 alu.last = 1;
3462 r = r600_bytecode_add_alu(ctx->bc, &alu);
3463 if (r)
3464 return r;
3465 }
3466
3467 /* MULADD R0.x, R0.x, PS1, (0x3FC00000, 1.5f).x
3468 * MULADD R0.y, R0.y, PS1, (0x3FC00000, 1.5f).x
3469 * muladd has no writemask, have to use another temp
3470 */
3471 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3472 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_MULADD);
3473 alu.is_op3 = 1;
3474
3475 alu.src[0].sel = ctx->temp_reg;
3476 alu.src[0].chan = 0;
3477 alu.src[1].sel = ctx->temp_reg;
3478 alu.src[1].chan = 2;
3479
3480 alu.src[2].sel = V_SQ_ALU_SRC_LITERAL;
3481 alu.src[2].chan = 0;
3482 alu.src[2].value = *(uint32_t *)&one_point_five;
3483
3484 alu.dst.sel = ctx->temp_reg;
3485 alu.dst.chan = 0;
3486 alu.dst.write = 1;
3487
3488 r = r600_bytecode_add_alu(ctx->bc, &alu);
3489 if (r)
3490 return r;
3491
3492 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3493 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_MULADD);
3494 alu.is_op3 = 1;
3495
3496 alu.src[0].sel = ctx->temp_reg;
3497 alu.src[0].chan = 1;
3498 alu.src[1].sel = ctx->temp_reg;
3499 alu.src[1].chan = 2;
3500
3501 alu.src[2].sel = V_SQ_ALU_SRC_LITERAL;
3502 alu.src[2].chan = 0;
3503 alu.src[2].value = *(uint32_t *)&one_point_five;
3504
3505 alu.dst.sel = ctx->temp_reg;
3506 alu.dst.chan = 1;
3507 alu.dst.write = 1;
3508
3509 alu.last = 1;
3510 r = r600_bytecode_add_alu(ctx->bc, &alu);
3511 if (r)
3512 return r;
3513 /* write initial W value into Z component */
3514 if (inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE) {
3515 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3516 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV);
3517 r600_bytecode_src(&alu.src[0], &ctx->src[0], 3);
3518 alu.dst.sel = ctx->temp_reg;
3519 alu.dst.chan = 2;
3520 alu.dst.write = 1;
3521 alu.last = 1;
3522 r = r600_bytecode_add_alu(ctx->bc, &alu);
3523 if (r)
3524 return r;
3525 }
3526 src_loaded = TRUE;
3527 src_gpr = ctx->temp_reg;
3528 }
3529
3530 if (src_requires_loading && !src_loaded) {
3531 for (i = 0; i < 4; i++) {
3532 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3533 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV);
3534 r600_bytecode_src(&alu.src[0], &ctx->src[0], i);
3535 alu.dst.sel = ctx->temp_reg;
3536 alu.dst.chan = i;
3537 if (i == 3)
3538 alu.last = 1;
3539 alu.dst.write = 1;
3540 r = r600_bytecode_add_alu(ctx->bc, &alu);
3541 if (r)
3542 return r;
3543 }
3544 src_loaded = TRUE;
3545 src_gpr = ctx->temp_reg;
3546 }
3547
3548 opcode = ctx->inst_info->r600_opcode;
3549 if (inst->Texture.Texture == TGSI_TEXTURE_SHADOW1D ||
3550 inst->Texture.Texture == TGSI_TEXTURE_SHADOW2D ||
3551 inst->Texture.Texture == TGSI_TEXTURE_SHADOWRECT ||
3552 inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE ||
3553 inst->Texture.Texture == TGSI_TEXTURE_SHADOW1D_ARRAY ||
3554 inst->Texture.Texture == TGSI_TEXTURE_SHADOW2D_ARRAY) {
3555 switch (opcode) {
3556 case SQ_TEX_INST_SAMPLE:
3557 opcode = SQ_TEX_INST_SAMPLE_C;
3558 break;
3559 case SQ_TEX_INST_SAMPLE_L:
3560 opcode = SQ_TEX_INST_SAMPLE_C_L;
3561 break;
3562 case SQ_TEX_INST_SAMPLE_LB:
3563 opcode = SQ_TEX_INST_SAMPLE_C_LB;
3564 break;
3565 case SQ_TEX_INST_SAMPLE_G:
3566 opcode = SQ_TEX_INST_SAMPLE_C_G;
3567 break;
3568 }
3569 }
3570
3571 memset(&tex, 0, sizeof(struct r600_bytecode_tex));
3572 tex.inst = opcode;
3573
3574 tex.sampler_id = tgsi_tex_get_src_gpr(ctx, sampler_src_reg);
3575 tex.resource_id = tex.sampler_id + R600_MAX_CONST_BUFFERS;
3576 tex.src_gpr = src_gpr;
3577 tex.dst_gpr = ctx->file_offset[inst->Dst[0].Register.File] + inst->Dst[0].Register.Index;
3578 tex.dst_sel_x = (inst->Dst[0].Register.WriteMask & 1) ? 0 : 7;
3579 tex.dst_sel_y = (inst->Dst[0].Register.WriteMask & 2) ? 1 : 7;
3580 tex.dst_sel_z = (inst->Dst[0].Register.WriteMask & 4) ? 2 : 7;
3581 tex.dst_sel_w = (inst->Dst[0].Register.WriteMask & 8) ? 3 : 7;
3582 if (src_loaded) {
3583 tex.src_sel_x = 0;
3584 tex.src_sel_y = 1;
3585 tex.src_sel_z = 2;
3586 tex.src_sel_w = 3;
3587 } else {
3588 tex.src_sel_x = ctx->src[0].swizzle[0];
3589 tex.src_sel_y = ctx->src[0].swizzle[1];
3590 tex.src_sel_z = ctx->src[0].swizzle[2];
3591 tex.src_sel_w = ctx->src[0].swizzle[3];
3592 tex.src_rel = ctx->src[0].rel;
3593 }
3594
3595 if (inst->Texture.Texture == TGSI_TEXTURE_CUBE) {
3596 tex.src_sel_x = 1;
3597 tex.src_sel_y = 0;
3598 tex.src_sel_z = 3;
3599 tex.src_sel_w = 1;
3600 }
3601 if (inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE) {
3602 tex.src_sel_x = 1;
3603 tex.src_sel_y = 0;
3604 tex.src_sel_z = 3;
3605 tex.src_sel_w = 2; /* route Z compare value into W */
3606 }
3607
3608 if (inst->Texture.Texture != TGSI_TEXTURE_RECT &&
3609 inst->Texture.Texture != TGSI_TEXTURE_SHADOWRECT) {
3610 tex.coord_type_x = 1;
3611 tex.coord_type_y = 1;
3612 }
3613 tex.coord_type_z = 1;
3614 tex.coord_type_w = 1;
3615
3616 tex.offset_x = offset_x;
3617 tex.offset_y = offset_y;
3618 tex.offset_z = offset_z;
3619
3620 /* Put the depth for comparison in W.
3621 * TGSI_TEXTURE_SHADOW2D_ARRAY already has the depth in W.
3622 * Some instructions expect the depth in Z. */
3623 if ((inst->Texture.Texture == TGSI_TEXTURE_SHADOW1D ||
3624 inst->Texture.Texture == TGSI_TEXTURE_SHADOW2D ||
3625 inst->Texture.Texture == TGSI_TEXTURE_SHADOWRECT ||
3626 inst->Texture.Texture == TGSI_TEXTURE_SHADOW1D_ARRAY) &&
3627 opcode != SQ_TEX_INST_SAMPLE_C_L &&
3628 opcode != SQ_TEX_INST_SAMPLE_C_LB) {
3629 tex.src_sel_w = tex.src_sel_z;
3630 }
3631
3632 if (inst->Texture.Texture == TGSI_TEXTURE_1D_ARRAY ||
3633 inst->Texture.Texture == TGSI_TEXTURE_SHADOW1D_ARRAY) {
3634 if (opcode == SQ_TEX_INST_SAMPLE_C_L ||
3635 opcode == SQ_TEX_INST_SAMPLE_C_LB) {
3636 /* the array index is read from Y */
3637 tex.coord_type_y = 0;
3638 } else {
3639 /* the array index is read from Z */
3640 tex.coord_type_z = 0;
3641 tex.src_sel_z = tex.src_sel_y;
3642 }
3643 } else if (inst->Texture.Texture == TGSI_TEXTURE_2D_ARRAY ||
3644 inst->Texture.Texture == TGSI_TEXTURE_SHADOW2D_ARRAY)
3645 /* the array index is read from Z */
3646 tex.coord_type_z = 0;
3647
3648 r = r600_bytecode_add_tex(ctx->bc, &tex);
3649 if (r)
3650 return r;
3651
3652 /* add shadow ambient support - gallium doesn't do it yet */
3653 return 0;
3654 }
3655
3656 static int tgsi_lrp(struct r600_shader_ctx *ctx)
3657 {
3658 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
3659 struct r600_bytecode_alu alu;
3660 int lasti = tgsi_last_instruction(inst->Dst[0].Register.WriteMask);
3661 unsigned i;
3662 int r;
3663
3664 /* optimize if it's just an equal balance */
3665 if (ctx->src[0].sel == V_SQ_ALU_SRC_0_5) {
3666 for (i = 0; i < lasti + 1; i++) {
3667 if (!(inst->Dst[0].Register.WriteMask & (1 << i)))
3668 continue;
3669
3670 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3671 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ADD);
3672 r600_bytecode_src(&alu.src[0], &ctx->src[1], i);
3673 r600_bytecode_src(&alu.src[1], &ctx->src[2], i);
3674 alu.omod = 3;
3675 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
3676 alu.dst.chan = i;
3677 if (i == lasti) {
3678 alu.last = 1;
3679 }
3680 r = r600_bytecode_add_alu(ctx->bc, &alu);
3681 if (r)
3682 return r;
3683 }
3684 return 0;
3685 }
3686
3687 /* 1 - src0 */
3688 for (i = 0; i < lasti + 1; i++) {
3689 if (!(inst->Dst[0].Register.WriteMask & (1 << i)))
3690 continue;
3691
3692 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3693 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ADD);
3694 alu.src[0].sel = V_SQ_ALU_SRC_1;
3695 alu.src[0].chan = 0;
3696 r600_bytecode_src(&alu.src[1], &ctx->src[0], i);
3697 r600_bytecode_src_toggle_neg(&alu.src[1]);
3698 alu.dst.sel = ctx->temp_reg;
3699 alu.dst.chan = i;
3700 if (i == lasti) {
3701 alu.last = 1;
3702 }
3703 alu.dst.write = 1;
3704 r = r600_bytecode_add_alu(ctx->bc, &alu);
3705 if (r)
3706 return r;
3707 }
3708
3709 /* (1 - src0) * src2 */
3710 for (i = 0; i < lasti + 1; i++) {
3711 if (!(inst->Dst[0].Register.WriteMask & (1 << i)))
3712 continue;
3713
3714 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3715 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MUL);
3716 alu.src[0].sel = ctx->temp_reg;
3717 alu.src[0].chan = i;
3718 r600_bytecode_src(&alu.src[1], &ctx->src[2], i);
3719 alu.dst.sel = ctx->temp_reg;
3720 alu.dst.chan = i;
3721 if (i == lasti) {
3722 alu.last = 1;
3723 }
3724 alu.dst.write = 1;
3725 r = r600_bytecode_add_alu(ctx->bc, &alu);
3726 if (r)
3727 return r;
3728 }
3729
3730 /* src0 * src1 + (1 - src0) * src2 */
3731 for (i = 0; i < lasti + 1; i++) {
3732 if (!(inst->Dst[0].Register.WriteMask & (1 << i)))
3733 continue;
3734
3735 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3736 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_MULADD);
3737 alu.is_op3 = 1;
3738 r600_bytecode_src(&alu.src[0], &ctx->src[0], i);
3739 r600_bytecode_src(&alu.src[1], &ctx->src[1], i);
3740 alu.src[2].sel = ctx->temp_reg;
3741 alu.src[2].chan = i;
3742
3743 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
3744 alu.dst.chan = i;
3745 if (i == lasti) {
3746 alu.last = 1;
3747 }
3748 r = r600_bytecode_add_alu(ctx->bc, &alu);
3749 if (r)
3750 return r;
3751 }
3752 return 0;
3753 }
3754
3755 static int tgsi_cmp(struct r600_shader_ctx *ctx)
3756 {
3757 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
3758 struct r600_bytecode_alu alu;
3759 int i, r;
3760 int lasti = tgsi_last_instruction(inst->Dst[0].Register.WriteMask);
3761
3762 for (i = 0; i < lasti + 1; i++) {
3763 if (!(inst->Dst[0].Register.WriteMask & (1 << i)))
3764 continue;
3765
3766 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3767 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_CNDGE);
3768 r600_bytecode_src(&alu.src[0], &ctx->src[0], i);
3769 r600_bytecode_src(&alu.src[1], &ctx->src[2], i);
3770 r600_bytecode_src(&alu.src[2], &ctx->src[1], i);
3771 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
3772 alu.dst.chan = i;
3773 alu.dst.write = 1;
3774 alu.is_op3 = 1;
3775 if (i == lasti)
3776 alu.last = 1;
3777 r = r600_bytecode_add_alu(ctx->bc, &alu);
3778 if (r)
3779 return r;
3780 }
3781 return 0;
3782 }
3783
3784 static int tgsi_xpd(struct r600_shader_ctx *ctx)
3785 {
3786 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
3787 static const unsigned int src0_swizzle[] = {2, 0, 1};
3788 static const unsigned int src1_swizzle[] = {1, 2, 0};
3789 struct r600_bytecode_alu alu;
3790 uint32_t use_temp = 0;
3791 int i, r;
3792
3793 if (inst->Dst[0].Register.WriteMask != 0xf)
3794 use_temp = 1;
3795
3796 for (i = 0; i < 4; i++) {
3797 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3798 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MUL);
3799 if (i < 3) {
3800 r600_bytecode_src(&alu.src[0], &ctx->src[0], src0_swizzle[i]);
3801 r600_bytecode_src(&alu.src[1], &ctx->src[1], src1_swizzle[i]);
3802 } else {
3803 alu.src[0].sel = V_SQ_ALU_SRC_0;
3804 alu.src[0].chan = i;
3805 alu.src[1].sel = V_SQ_ALU_SRC_0;
3806 alu.src[1].chan = i;
3807 }
3808
3809 alu.dst.sel = ctx->temp_reg;
3810 alu.dst.chan = i;
3811 alu.dst.write = 1;
3812
3813 if (i == 3)
3814 alu.last = 1;
3815 r = r600_bytecode_add_alu(ctx->bc, &alu);
3816 if (r)
3817 return r;
3818 }
3819
3820 for (i = 0; i < 4; i++) {
3821 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3822 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_MULADD);
3823
3824 if (i < 3) {
3825 r600_bytecode_src(&alu.src[0], &ctx->src[0], src1_swizzle[i]);
3826 r600_bytecode_src(&alu.src[1], &ctx->src[1], src0_swizzle[i]);
3827 } else {
3828 alu.src[0].sel = V_SQ_ALU_SRC_0;
3829 alu.src[0].chan = i;
3830 alu.src[1].sel = V_SQ_ALU_SRC_0;
3831 alu.src[1].chan = i;
3832 }
3833
3834 alu.src[2].sel = ctx->temp_reg;
3835 alu.src[2].neg = 1;
3836 alu.src[2].chan = i;
3837
3838 if (use_temp)
3839 alu.dst.sel = ctx->temp_reg;
3840 else
3841 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
3842 alu.dst.chan = i;
3843 alu.dst.write = 1;
3844 alu.is_op3 = 1;
3845 if (i == 3)
3846 alu.last = 1;
3847 r = r600_bytecode_add_alu(ctx->bc, &alu);
3848 if (r)
3849 return r;
3850 }
3851 if (use_temp)
3852 return tgsi_helper_copy(ctx, inst);
3853 return 0;
3854 }
3855
3856 static int tgsi_exp(struct r600_shader_ctx *ctx)
3857 {
3858 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
3859 struct r600_bytecode_alu alu;
3860 int r;
3861 int i;
3862
3863 /* result.x = 2^floor(src); */
3864 if (inst->Dst[0].Register.WriteMask & 1) {
3865 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3866
3867 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLOOR);
3868 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
3869
3870 alu.dst.sel = ctx->temp_reg;
3871 alu.dst.chan = 0;
3872 alu.dst.write = 1;
3873 alu.last = 1;
3874 r = r600_bytecode_add_alu(ctx->bc, &alu);
3875 if (r)
3876 return r;
3877
3878 if (ctx->bc->chip_class == CAYMAN) {
3879 for (i = 0; i < 3; i++) {
3880 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_EXP_IEEE);
3881 alu.src[0].sel = ctx->temp_reg;
3882 alu.src[0].chan = 0;
3883
3884 alu.dst.sel = ctx->temp_reg;
3885 alu.dst.chan = i;
3886 if (i == 0)
3887 alu.dst.write = 1;
3888 if (i == 2)
3889 alu.last = 1;
3890 r = r600_bytecode_add_alu(ctx->bc, &alu);
3891 if (r)
3892 return r;
3893 }
3894 } else {
3895 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_EXP_IEEE);
3896 alu.src[0].sel = ctx->temp_reg;
3897 alu.src[0].chan = 0;
3898
3899 alu.dst.sel = ctx->temp_reg;
3900 alu.dst.chan = 0;
3901 alu.dst.write = 1;
3902 alu.last = 1;
3903 r = r600_bytecode_add_alu(ctx->bc, &alu);
3904 if (r)
3905 return r;
3906 }
3907 }
3908
3909 /* result.y = tmp - floor(tmp); */
3910 if ((inst->Dst[0].Register.WriteMask >> 1) & 1) {
3911 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3912
3913 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FRACT);
3914 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
3915
3916 alu.dst.sel = ctx->temp_reg;
3917 #if 0
3918 r = tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
3919 if (r)
3920 return r;
3921 #endif
3922 alu.dst.write = 1;
3923 alu.dst.chan = 1;
3924
3925 alu.last = 1;
3926
3927 r = r600_bytecode_add_alu(ctx->bc, &alu);
3928 if (r)
3929 return r;
3930 }
3931
3932 /* result.z = RoughApprox2ToX(tmp);*/
3933 if ((inst->Dst[0].Register.WriteMask >> 2) & 0x1) {
3934 if (ctx->bc->chip_class == CAYMAN) {
3935 for (i = 0; i < 3; i++) {
3936 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3937 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_EXP_IEEE);
3938 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
3939
3940 alu.dst.sel = ctx->temp_reg;
3941 alu.dst.chan = i;
3942 if (i == 2) {
3943 alu.dst.write = 1;
3944 alu.last = 1;
3945 }
3946
3947 r = r600_bytecode_add_alu(ctx->bc, &alu);
3948 if (r)
3949 return r;
3950 }
3951 } else {
3952 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3953 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_EXP_IEEE);
3954 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
3955
3956 alu.dst.sel = ctx->temp_reg;
3957 alu.dst.write = 1;
3958 alu.dst.chan = 2;
3959
3960 alu.last = 1;
3961
3962 r = r600_bytecode_add_alu(ctx->bc, &alu);
3963 if (r)
3964 return r;
3965 }
3966 }
3967
3968 /* result.w = 1.0;*/
3969 if ((inst->Dst[0].Register.WriteMask >> 3) & 0x1) {
3970 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3971
3972 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV);
3973 alu.src[0].sel = V_SQ_ALU_SRC_1;
3974 alu.src[0].chan = 0;
3975
3976 alu.dst.sel = ctx->temp_reg;
3977 alu.dst.chan = 3;
3978 alu.dst.write = 1;
3979 alu.last = 1;
3980 r = r600_bytecode_add_alu(ctx->bc, &alu);
3981 if (r)
3982 return r;
3983 }
3984 return tgsi_helper_copy(ctx, inst);
3985 }
3986
3987 static int tgsi_log(struct r600_shader_ctx *ctx)
3988 {
3989 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
3990 struct r600_bytecode_alu alu;
3991 int r;
3992 int i;
3993
3994 /* result.x = floor(log2(|src|)); */
3995 if (inst->Dst[0].Register.WriteMask & 1) {
3996 if (ctx->bc->chip_class == CAYMAN) {
3997 for (i = 0; i < 3; i++) {
3998 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
3999
4000 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LOG_IEEE);
4001 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
4002 r600_bytecode_src_set_abs(&alu.src[0]);
4003
4004 alu.dst.sel = ctx->temp_reg;
4005 alu.dst.chan = i;
4006 if (i == 0)
4007 alu.dst.write = 1;
4008 if (i == 2)
4009 alu.last = 1;
4010 r = r600_bytecode_add_alu(ctx->bc, &alu);
4011 if (r)
4012 return r;
4013 }
4014
4015 } else {
4016 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
4017
4018 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LOG_IEEE);
4019 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
4020 r600_bytecode_src_set_abs(&alu.src[0]);
4021
4022 alu.dst.sel = ctx->temp_reg;
4023 alu.dst.chan = 0;
4024 alu.dst.write = 1;
4025 alu.last = 1;
4026 r = r600_bytecode_add_alu(ctx->bc, &alu);
4027 if (r)
4028 return r;
4029 }
4030
4031 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLOOR);
4032 alu.src[0].sel = ctx->temp_reg;
4033 alu.src[0].chan = 0;
4034
4035 alu.dst.sel = ctx->temp_reg;
4036 alu.dst.chan = 0;
4037 alu.dst.write = 1;
4038 alu.last = 1;
4039
4040 r = r600_bytecode_add_alu(ctx->bc, &alu);
4041 if (r)
4042 return r;
4043 }
4044
4045 /* result.y = |src.x| / (2 ^ floor(log2(|src.x|))); */
4046 if ((inst->Dst[0].Register.WriteMask >> 1) & 1) {
4047
4048 if (ctx->bc->chip_class == CAYMAN) {
4049 for (i = 0; i < 3; i++) {
4050 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
4051
4052 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LOG_IEEE);
4053 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
4054 r600_bytecode_src_set_abs(&alu.src[0]);
4055
4056 alu.dst.sel = ctx->temp_reg;
4057 alu.dst.chan = i;
4058 if (i == 1)
4059 alu.dst.write = 1;
4060 if (i == 2)
4061 alu.last = 1;
4062
4063 r = r600_bytecode_add_alu(ctx->bc, &alu);
4064 if (r)
4065 return r;
4066 }
4067 } else {
4068 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
4069
4070 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LOG_IEEE);
4071 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
4072 r600_bytecode_src_set_abs(&alu.src[0]);
4073
4074 alu.dst.sel = ctx->temp_reg;
4075 alu.dst.chan = 1;
4076 alu.dst.write = 1;
4077 alu.last = 1;
4078
4079 r = r600_bytecode_add_alu(ctx->bc, &alu);
4080 if (r)
4081 return r;
4082 }
4083
4084 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
4085
4086 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLOOR);
4087 alu.src[0].sel = ctx->temp_reg;
4088 alu.src[0].chan = 1;
4089
4090 alu.dst.sel = ctx->temp_reg;
4091 alu.dst.chan = 1;
4092 alu.dst.write = 1;
4093 alu.last = 1;
4094
4095 r = r600_bytecode_add_alu(ctx->bc, &alu);
4096 if (r)
4097 return r;
4098
4099 if (ctx->bc->chip_class == CAYMAN) {
4100 for (i = 0; i < 3; i++) {
4101 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
4102 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_EXP_IEEE);
4103 alu.src[0].sel = ctx->temp_reg;
4104 alu.src[0].chan = 1;
4105
4106 alu.dst.sel = ctx->temp_reg;
4107 alu.dst.chan = i;
4108 if (i == 1)
4109 alu.dst.write = 1;
4110 if (i == 2)
4111 alu.last = 1;
4112
4113 r = r600_bytecode_add_alu(ctx->bc, &alu);
4114 if (r)
4115 return r;
4116 }
4117 } else {
4118 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
4119 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_EXP_IEEE);
4120 alu.src[0].sel = ctx->temp_reg;
4121 alu.src[0].chan = 1;
4122
4123 alu.dst.sel = ctx->temp_reg;
4124 alu.dst.chan = 1;
4125 alu.dst.write = 1;
4126 alu.last = 1;
4127
4128 r = r600_bytecode_add_alu(ctx->bc, &alu);
4129 if (r)
4130 return r;
4131 }
4132
4133 if (ctx->bc->chip_class == CAYMAN) {
4134 for (i = 0; i < 3; i++) {
4135 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
4136 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_IEEE);
4137 alu.src[0].sel = ctx->temp_reg;
4138 alu.src[0].chan = 1;
4139
4140 alu.dst.sel = ctx->temp_reg;
4141 alu.dst.chan = i;
4142 if (i == 1)
4143 alu.dst.write = 1;
4144 if (i == 2)
4145 alu.last = 1;
4146
4147 r = r600_bytecode_add_alu(ctx->bc, &alu);
4148 if (r)
4149 return r;
4150 }
4151 } else {
4152 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
4153 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_IEEE);
4154 alu.src[0].sel = ctx->temp_reg;
4155 alu.src[0].chan = 1;
4156
4157 alu.dst.sel = ctx->temp_reg;
4158 alu.dst.chan = 1;
4159 alu.dst.write = 1;
4160 alu.last = 1;
4161
4162 r = r600_bytecode_add_alu(ctx->bc, &alu);
4163 if (r)
4164 return r;
4165 }
4166
4167 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
4168
4169 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MUL);
4170
4171 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
4172 r600_bytecode_src_set_abs(&alu.src[0]);
4173
4174 alu.src[1].sel = ctx->temp_reg;
4175 alu.src[1].chan = 1;
4176
4177 alu.dst.sel = ctx->temp_reg;
4178 alu.dst.chan = 1;
4179 alu.dst.write = 1;
4180 alu.last = 1;
4181
4182 r = r600_bytecode_add_alu(ctx->bc, &alu);
4183 if (r)
4184 return r;
4185 }
4186
4187 /* result.z = log2(|src|);*/
4188 if ((inst->Dst[0].Register.WriteMask >> 2) & 1) {
4189 if (ctx->bc->chip_class == CAYMAN) {
4190 for (i = 0; i < 3; i++) {
4191 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
4192
4193 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LOG_IEEE);
4194 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
4195 r600_bytecode_src_set_abs(&alu.src[0]);
4196
4197 alu.dst.sel = ctx->temp_reg;
4198 if (i == 2)
4199 alu.dst.write = 1;
4200 alu.dst.chan = i;
4201 if (i == 2)
4202 alu.last = 1;
4203
4204 r = r600_bytecode_add_alu(ctx->bc, &alu);
4205 if (r)
4206 return r;
4207 }
4208 } else {
4209 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
4210
4211 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LOG_IEEE);
4212 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
4213 r600_bytecode_src_set_abs(&alu.src[0]);
4214
4215 alu.dst.sel = ctx->temp_reg;
4216 alu.dst.write = 1;
4217 alu.dst.chan = 2;
4218 alu.last = 1;
4219
4220 r = r600_bytecode_add_alu(ctx->bc, &alu);
4221 if (r)
4222 return r;
4223 }
4224 }
4225
4226 /* result.w = 1.0; */
4227 if ((inst->Dst[0].Register.WriteMask >> 3) & 1) {
4228 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
4229
4230 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV);
4231 alu.src[0].sel = V_SQ_ALU_SRC_1;
4232 alu.src[0].chan = 0;
4233
4234 alu.dst.sel = ctx->temp_reg;
4235 alu.dst.chan = 3;
4236 alu.dst.write = 1;
4237 alu.last = 1;
4238
4239 r = r600_bytecode_add_alu(ctx->bc, &alu);
4240 if (r)
4241 return r;
4242 }
4243
4244 return tgsi_helper_copy(ctx, inst);
4245 }
4246
4247 static int tgsi_eg_arl(struct r600_shader_ctx *ctx)
4248 {
4249 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
4250 struct r600_bytecode_alu alu;
4251 int r;
4252
4253 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
4254
4255 switch (inst->Instruction.Opcode) {
4256 case TGSI_OPCODE_ARL:
4257 alu.inst = EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLT_TO_INT_FLOOR;
4258 break;
4259 case TGSI_OPCODE_ARR:
4260 alu.inst = EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLT_TO_INT;
4261 break;
4262 case TGSI_OPCODE_UARL:
4263 alu.inst = EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV;
4264 break;
4265 default:
4266 assert(0);
4267 return -1;
4268 }
4269
4270 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
4271 alu.last = 1;
4272 alu.dst.sel = ctx->bc->ar_reg;
4273 alu.dst.write = 1;
4274 r = r600_bytecode_add_alu(ctx->bc, &alu);
4275 if (r)
4276 return r;
4277
4278 ctx->bc->ar_loaded = 0;
4279 return 0;
4280 }
4281 static int tgsi_r600_arl(struct r600_shader_ctx *ctx)
4282 {
4283 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
4284 struct r600_bytecode_alu alu;
4285 int r;
4286
4287 switch (inst->Instruction.Opcode) {
4288 case TGSI_OPCODE_ARL:
4289 memset(&alu, 0, sizeof(alu));
4290 alu.inst = V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLOOR;
4291 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
4292 alu.dst.sel = ctx->bc->ar_reg;
4293 alu.dst.write = 1;
4294 alu.last = 1;
4295
4296 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
4297 return r;
4298
4299 memset(&alu, 0, sizeof(alu));
4300 alu.inst = V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLT_TO_INT;
4301 alu.src[0].sel = ctx->bc->ar_reg;
4302 alu.dst.sel = ctx->bc->ar_reg;
4303 alu.dst.write = 1;
4304 alu.last = 1;
4305
4306 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
4307 return r;
4308 break;
4309 case TGSI_OPCODE_ARR:
4310 memset(&alu, 0, sizeof(alu));
4311 alu.inst = V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLT_TO_INT;
4312 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
4313 alu.dst.sel = ctx->bc->ar_reg;
4314 alu.dst.write = 1;
4315 alu.last = 1;
4316
4317 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
4318 return r;
4319 break;
4320 case TGSI_OPCODE_UARL:
4321 memset(&alu, 0, sizeof(alu));
4322 alu.inst = V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV;
4323 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
4324 alu.dst.sel = ctx->bc->ar_reg;
4325 alu.dst.write = 1;
4326 alu.last = 1;
4327
4328 if ((r = r600_bytecode_add_alu(ctx->bc, &alu)))
4329 return r;
4330 break;
4331 default:
4332 assert(0);
4333 return -1;
4334 }
4335
4336 ctx->bc->ar_loaded = 0;
4337 return 0;
4338 }
4339
4340 static int tgsi_opdst(struct r600_shader_ctx *ctx)
4341 {
4342 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
4343 struct r600_bytecode_alu alu;
4344 int i, r = 0;
4345
4346 for (i = 0; i < 4; i++) {
4347 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
4348
4349 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MUL);
4350 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
4351
4352 if (i == 0 || i == 3) {
4353 alu.src[0].sel = V_SQ_ALU_SRC_1;
4354 } else {
4355 r600_bytecode_src(&alu.src[0], &ctx->src[0], i);
4356 }
4357
4358 if (i == 0 || i == 2) {
4359 alu.src[1].sel = V_SQ_ALU_SRC_1;
4360 } else {
4361 r600_bytecode_src(&alu.src[1], &ctx->src[1], i);
4362 }
4363 if (i == 3)
4364 alu.last = 1;
4365 r = r600_bytecode_add_alu(ctx->bc, &alu);
4366 if (r)
4367 return r;
4368 }
4369 return 0;
4370 }
4371
4372 static int emit_logic_pred(struct r600_shader_ctx *ctx, int opcode)
4373 {
4374 struct r600_bytecode_alu alu;
4375 int r;
4376
4377 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
4378 alu.inst = opcode;
4379 alu.predicate = 1;
4380
4381 alu.dst.sel = ctx->temp_reg;
4382 alu.dst.write = 1;
4383 alu.dst.chan = 0;
4384
4385 r600_bytecode_src(&alu.src[0], &ctx->src[0], 0);
4386 alu.src[1].sel = V_SQ_ALU_SRC_0;
4387 alu.src[1].chan = 0;
4388
4389 alu.last = 1;
4390
4391 r = r600_bytecode_add_alu_type(ctx->bc, &alu, CTX_INST(V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_PUSH_BEFORE));
4392 if (r)
4393 return r;
4394 return 0;
4395 }
4396
4397 static int pops(struct r600_shader_ctx *ctx, int pops)
4398 {
4399 unsigned force_pop = ctx->bc->force_add_cf;
4400
4401 if (!force_pop) {
4402 int alu_pop = 3;
4403 if (ctx->bc->cf_last) {
4404 if (ctx->bc->cf_last->inst == CTX_INST(V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU))
4405 alu_pop = 0;
4406 else if (ctx->bc->cf_last->inst == CTX_INST(V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP_AFTER))
4407 alu_pop = 1;
4408 }
4409 alu_pop += pops;
4410 if (alu_pop == 1) {
4411 ctx->bc->cf_last->inst = CTX_INST(V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP_AFTER);
4412 ctx->bc->force_add_cf = 1;
4413 } else if (alu_pop == 2) {
4414 ctx->bc->cf_last->inst = CTX_INST(V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP2_AFTER);
4415 ctx->bc->force_add_cf = 1;
4416 } else {
4417 force_pop = 1;
4418 }
4419 }
4420
4421 if (force_pop) {
4422 r600_bytecode_add_cfinst(ctx->bc, CTX_INST(V_SQ_CF_WORD1_SQ_CF_INST_POP));
4423 ctx->bc->cf_last->pop_count = pops;
4424 ctx->bc->cf_last->cf_addr = ctx->bc->cf_last->id + 2;
4425 }
4426
4427 return 0;
4428 }
4429
4430 static inline void callstack_decrease_current(struct r600_shader_ctx *ctx, unsigned reason)
4431 {
4432 switch(reason) {
4433 case FC_PUSH_VPM:
4434 ctx->bc->callstack[ctx->bc->call_sp].current--;
4435 break;
4436 case FC_PUSH_WQM:
4437 case FC_LOOP:
4438 ctx->bc->callstack[ctx->bc->call_sp].current -= 4;
4439 break;
4440 case FC_REP:
4441 /* TOODO : for 16 vp asic should -= 2; */
4442 ctx->bc->callstack[ctx->bc->call_sp].current --;
4443 break;
4444 }
4445 }
4446
4447 static inline void callstack_check_depth(struct r600_shader_ctx *ctx, unsigned reason, unsigned check_max_only)
4448 {
4449 if (check_max_only) {
4450 int diff;
4451 switch (reason) {
4452 case FC_PUSH_VPM:
4453 diff = 1;
4454 break;
4455 case FC_PUSH_WQM:
4456 diff = 4;
4457 break;
4458 default:
4459 assert(0);
4460 diff = 0;
4461 }
4462 if ((ctx->bc->callstack[ctx->bc->call_sp].current + diff) >
4463 ctx->bc->callstack[ctx->bc->call_sp].max) {
4464 ctx->bc->callstack[ctx->bc->call_sp].max =
4465 ctx->bc->callstack[ctx->bc->call_sp].current + diff;
4466 }
4467 return;
4468 }
4469 switch (reason) {
4470 case FC_PUSH_VPM:
4471 ctx->bc->callstack[ctx->bc->call_sp].current++;
4472 break;
4473 case FC_PUSH_WQM:
4474 case FC_LOOP:
4475 ctx->bc->callstack[ctx->bc->call_sp].current += 4;
4476 break;
4477 case FC_REP:
4478 ctx->bc->callstack[ctx->bc->call_sp].current++;
4479 break;
4480 }
4481
4482 if ((ctx->bc->callstack[ctx->bc->call_sp].current) >
4483 ctx->bc->callstack[ctx->bc->call_sp].max) {
4484 ctx->bc->callstack[ctx->bc->call_sp].max =
4485 ctx->bc->callstack[ctx->bc->call_sp].current;
4486 }
4487 }
4488
4489 static void fc_set_mid(struct r600_shader_ctx *ctx, int fc_sp)
4490 {
4491 struct r600_cf_stack_entry *sp = &ctx->bc->fc_stack[fc_sp];
4492
4493 sp->mid = (struct r600_bytecode_cf **)realloc((void *)sp->mid,
4494 sizeof(struct r600_bytecode_cf *) * (sp->num_mid + 1));
4495 sp->mid[sp->num_mid] = ctx->bc->cf_last;
4496 sp->num_mid++;
4497 }
4498
4499 static void fc_pushlevel(struct r600_shader_ctx *ctx, int type)
4500 {
4501 ctx->bc->fc_sp++;
4502 ctx->bc->fc_stack[ctx->bc->fc_sp].type = type;
4503 ctx->bc->fc_stack[ctx->bc->fc_sp].start = ctx->bc->cf_last;
4504 }
4505
4506 static void fc_poplevel(struct r600_shader_ctx *ctx)
4507 {
4508 struct r600_cf_stack_entry *sp = &ctx->bc->fc_stack[ctx->bc->fc_sp];
4509 if (sp->mid) {
4510 free(sp->mid);
4511 sp->mid = NULL;
4512 }
4513 sp->num_mid = 0;
4514 sp->start = NULL;
4515 sp->type = 0;
4516 ctx->bc->fc_sp--;
4517 }
4518
4519 #if 0
4520 static int emit_return(struct r600_shader_ctx *ctx)
4521 {
4522 r600_bytecode_add_cfinst(ctx->bc, CTX_INST(V_SQ_CF_WORD1_SQ_CF_INST_RETURN));
4523 return 0;
4524 }
4525
4526 static int emit_jump_to_offset(struct r600_shader_ctx *ctx, int pops, int offset)
4527 {
4528
4529 r600_bytecode_add_cfinst(ctx->bc, CTX_INST(V_SQ_CF_WORD1_SQ_CF_INST_JUMP));
4530 ctx->bc->cf_last->pop_count = pops;
4531 /* TODO work out offset */
4532 return 0;
4533 }
4534
4535 static int emit_setret_in_loop_flag(struct r600_shader_ctx *ctx, unsigned flag_value)
4536 {
4537 return 0;
4538 }
4539
4540 static void emit_testflag(struct r600_shader_ctx *ctx)
4541 {
4542
4543 }
4544
4545 static void emit_return_on_flag(struct r600_shader_ctx *ctx, unsigned ifidx)
4546 {
4547 emit_testflag(ctx);
4548 emit_jump_to_offset(ctx, 1, 4);
4549 emit_setret_in_loop_flag(ctx, V_SQ_ALU_SRC_0);
4550 pops(ctx, ifidx + 1);
4551 emit_return(ctx);
4552 }
4553
4554 static void break_loop_on_flag(struct r600_shader_ctx *ctx, unsigned fc_sp)
4555 {
4556 emit_testflag(ctx);
4557
4558 r600_bytecode_add_cfinst(ctx->bc, ctx->inst_info->r600_opcode);
4559 ctx->bc->cf_last->pop_count = 1;
4560
4561 fc_set_mid(ctx, fc_sp);
4562
4563 pops(ctx, 1);
4564 }
4565 #endif
4566
4567 static int tgsi_if(struct r600_shader_ctx *ctx)
4568 {
4569 emit_logic_pred(ctx, CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETNE_INT));
4570
4571 r600_bytecode_add_cfinst(ctx->bc, CTX_INST(V_SQ_CF_WORD1_SQ_CF_INST_JUMP));
4572
4573 fc_pushlevel(ctx, FC_IF);
4574
4575 callstack_check_depth(ctx, FC_PUSH_VPM, 0);
4576 return 0;
4577 }
4578
4579 static int tgsi_else(struct r600_shader_ctx *ctx)
4580 {
4581 r600_bytecode_add_cfinst(ctx->bc, CTX_INST(V_SQ_CF_WORD1_SQ_CF_INST_ELSE));
4582 ctx->bc->cf_last->pop_count = 1;
4583
4584 fc_set_mid(ctx, ctx->bc->fc_sp);
4585 ctx->bc->fc_stack[ctx->bc->fc_sp].start->cf_addr = ctx->bc->cf_last->id;
4586 return 0;
4587 }
4588
4589 static int tgsi_endif(struct r600_shader_ctx *ctx)
4590 {
4591 pops(ctx, 1);
4592 if (ctx->bc->fc_stack[ctx->bc->fc_sp].type != FC_IF) {
4593 R600_ERR("if/endif unbalanced in shader\n");
4594 return -1;
4595 }
4596
4597 if (ctx->bc->fc_stack[ctx->bc->fc_sp].mid == NULL) {
4598 ctx->bc->fc_stack[ctx->bc->fc_sp].start->cf_addr = ctx->bc->cf_last->id + 2;
4599 ctx->bc->fc_stack[ctx->bc->fc_sp].start->pop_count = 1;
4600 } else {
4601 ctx->bc->fc_stack[ctx->bc->fc_sp].mid[0]->cf_addr = ctx->bc->cf_last->id + 2;
4602 }
4603 fc_poplevel(ctx);
4604
4605 callstack_decrease_current(ctx, FC_PUSH_VPM);
4606 return 0;
4607 }
4608
4609 static int tgsi_bgnloop(struct r600_shader_ctx *ctx)
4610 {
4611 r600_bytecode_add_cfinst(ctx->bc, CTX_INST(V_SQ_CF_WORD1_SQ_CF_INST_LOOP_START_NO_AL));
4612
4613 fc_pushlevel(ctx, FC_LOOP);
4614
4615 /* check stack depth */
4616 callstack_check_depth(ctx, FC_LOOP, 0);
4617 return 0;
4618 }
4619
4620 static int tgsi_endloop(struct r600_shader_ctx *ctx)
4621 {
4622 int i;
4623
4624 r600_bytecode_add_cfinst(ctx->bc, CTX_INST(V_SQ_CF_WORD1_SQ_CF_INST_LOOP_END));
4625
4626 if (ctx->bc->fc_stack[ctx->bc->fc_sp].type != FC_LOOP) {
4627 R600_ERR("loop/endloop in shader code are not paired.\n");
4628 return -EINVAL;
4629 }
4630
4631 /* fixup loop pointers - from r600isa
4632 LOOP END points to CF after LOOP START,
4633 LOOP START point to CF after LOOP END
4634 BRK/CONT point to LOOP END CF
4635 */
4636 ctx->bc->cf_last->cf_addr = ctx->bc->fc_stack[ctx->bc->fc_sp].start->id + 2;
4637
4638 ctx->bc->fc_stack[ctx->bc->fc_sp].start->cf_addr = ctx->bc->cf_last->id + 2;
4639
4640 for (i = 0; i < ctx->bc->fc_stack[ctx->bc->fc_sp].num_mid; i++) {
4641 ctx->bc->fc_stack[ctx->bc->fc_sp].mid[i]->cf_addr = ctx->bc->cf_last->id;
4642 }
4643 /* TODO add LOOPRET support */
4644 fc_poplevel(ctx);
4645 callstack_decrease_current(ctx, FC_LOOP);
4646 return 0;
4647 }
4648
4649 static int tgsi_loop_brk_cont(struct r600_shader_ctx *ctx)
4650 {
4651 unsigned int fscp;
4652
4653 for (fscp = ctx->bc->fc_sp; fscp > 0; fscp--)
4654 {
4655 if (FC_LOOP == ctx->bc->fc_stack[fscp].type)
4656 break;
4657 }
4658
4659 if (fscp == 0) {
4660 R600_ERR("Break not inside loop/endloop pair\n");
4661 return -EINVAL;
4662 }
4663
4664 r600_bytecode_add_cfinst(ctx->bc, ctx->inst_info->r600_opcode);
4665
4666 fc_set_mid(ctx, fscp);
4667
4668 callstack_check_depth(ctx, FC_PUSH_VPM, 1);
4669 return 0;
4670 }
4671
4672 static int tgsi_umad(struct r600_shader_ctx *ctx)
4673 {
4674 struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
4675 struct r600_bytecode_alu alu;
4676 int i, j, r;
4677 int lasti = tgsi_last_instruction(inst->Dst[0].Register.WriteMask);
4678
4679 /* src0 * src1 */
4680 for (i = 0; i < lasti + 1; i++) {
4681 if (!(inst->Dst[0].Register.WriteMask & (1 << i)))
4682 continue;
4683
4684 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
4685
4686 alu.dst.chan = i;
4687 alu.dst.sel = ctx->temp_reg;
4688 alu.dst.write = 1;
4689
4690 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MULLO_UINT);
4691 for (j = 0; j < 2; j++) {
4692 r600_bytecode_src(&alu.src[j], &ctx->src[j], i);
4693 }
4694
4695 alu.last = 1;
4696 r = r600_bytecode_add_alu(ctx->bc, &alu);
4697 if (r)
4698 return r;
4699 }
4700
4701
4702 for (i = 0; i < lasti + 1; i++) {
4703 if (!(inst->Dst[0].Register.WriteMask & (1 << i)))
4704 continue;
4705
4706 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
4707 tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
4708
4709 alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ADD_INT);
4710
4711 alu.src[0].sel = ctx->temp_reg;
4712 alu.src[0].chan = i;
4713
4714 r600_bytecode_src(&alu.src[1], &ctx->src[2], i);
4715 if (i == lasti) {
4716 alu.last = 1;
4717 }
4718 r = r600_bytecode_add_alu(ctx->bc, &alu);
4719 if (r)
4720 return r;
4721 }
4722 return 0;
4723 }
4724
4725 static struct r600_shader_tgsi_instruction r600_shader_tgsi_instruction[] = {
4726 {TGSI_OPCODE_ARL, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_r600_arl},
4727 {TGSI_OPCODE_MOV, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV, tgsi_op2},
4728 {TGSI_OPCODE_LIT, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_lit},
4729
4730 /* FIXME:
4731 * For state trackers other than OpenGL, we'll want to use
4732 * _RECIP_IEEE instead.
4733 */
4734 {TGSI_OPCODE_RCP, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_CLAMPED, tgsi_trans_srcx_replicate},
4735
4736 {TGSI_OPCODE_RSQ, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_rsq},
4737 {TGSI_OPCODE_EXP, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_exp},
4738 {TGSI_OPCODE_LOG, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_log},
4739 {TGSI_OPCODE_MUL, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MUL, tgsi_op2},
4740 {TGSI_OPCODE_ADD, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ADD, tgsi_op2},
4741 {TGSI_OPCODE_DP3, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4, tgsi_dp},
4742 {TGSI_OPCODE_DP4, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4, tgsi_dp},
4743 {TGSI_OPCODE_DST, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_opdst},
4744 {TGSI_OPCODE_MIN, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MIN, tgsi_op2},
4745 {TGSI_OPCODE_MAX, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MAX, tgsi_op2},
4746 {TGSI_OPCODE_SLT, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGT, tgsi_op2_swap},
4747 {TGSI_OPCODE_SGE, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGE, tgsi_op2},
4748 {TGSI_OPCODE_MAD, 1, V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_MULADD, tgsi_op3},
4749 {TGSI_OPCODE_SUB, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ADD, tgsi_op2},
4750 {TGSI_OPCODE_LRP, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_lrp},
4751 {TGSI_OPCODE_CND, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4752 /* gap */
4753 {20, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4754 {TGSI_OPCODE_DP2A, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4755 /* gap */
4756 {22, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4757 {23, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4758 {TGSI_OPCODE_FRC, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FRACT, tgsi_op2},
4759 {TGSI_OPCODE_CLAMP, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4760 {TGSI_OPCODE_FLR, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLOOR, tgsi_op2},
4761 {TGSI_OPCODE_ROUND, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RNDNE, tgsi_op2},
4762 {TGSI_OPCODE_EX2, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_EXP_IEEE, tgsi_trans_srcx_replicate},
4763 {TGSI_OPCODE_LG2, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LOG_IEEE, tgsi_trans_srcx_replicate},
4764 {TGSI_OPCODE_POW, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_pow},
4765 {TGSI_OPCODE_XPD, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_xpd},
4766 /* gap */
4767 {32, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4768 {TGSI_OPCODE_ABS, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV, tgsi_op2},
4769 {TGSI_OPCODE_RCC, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4770 {TGSI_OPCODE_DPH, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4, tgsi_dp},
4771 {TGSI_OPCODE_COS, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_COS, tgsi_trig},
4772 {TGSI_OPCODE_DDX, 0, SQ_TEX_INST_GET_GRADIENTS_H, tgsi_tex},
4773 {TGSI_OPCODE_DDY, 0, SQ_TEX_INST_GET_GRADIENTS_V, tgsi_tex},
4774 {TGSI_OPCODE_KILP, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT, tgsi_kill}, /* predicated kill */
4775 {TGSI_OPCODE_PK2H, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4776 {TGSI_OPCODE_PK2US, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4777 {TGSI_OPCODE_PK4B, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4778 {TGSI_OPCODE_PK4UB, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4779 {TGSI_OPCODE_RFL, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4780 {TGSI_OPCODE_SEQ, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETE, tgsi_op2},
4781 {TGSI_OPCODE_SFL, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4782 {TGSI_OPCODE_SGT, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGT, tgsi_op2},
4783 {TGSI_OPCODE_SIN, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SIN, tgsi_trig},
4784 {TGSI_OPCODE_SLE, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGE, tgsi_op2_swap},
4785 {TGSI_OPCODE_SNE, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETNE, tgsi_op2},
4786 {TGSI_OPCODE_STR, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4787 {TGSI_OPCODE_TEX, 0, SQ_TEX_INST_SAMPLE, tgsi_tex},
4788 {TGSI_OPCODE_TXD, 0, SQ_TEX_INST_SAMPLE_G, tgsi_tex},
4789 {TGSI_OPCODE_TXP, 0, SQ_TEX_INST_SAMPLE, tgsi_tex},
4790 {TGSI_OPCODE_UP2H, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4791 {TGSI_OPCODE_UP2US, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4792 {TGSI_OPCODE_UP4B, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4793 {TGSI_OPCODE_UP4UB, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4794 {TGSI_OPCODE_X2D, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4795 {TGSI_OPCODE_ARA, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4796 {TGSI_OPCODE_ARR, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_r600_arl},
4797 {TGSI_OPCODE_BRA, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4798 {TGSI_OPCODE_CAL, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4799 {TGSI_OPCODE_RET, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4800 {TGSI_OPCODE_SSG, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_ssg},
4801 {TGSI_OPCODE_CMP, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_cmp},
4802 {TGSI_OPCODE_SCS, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_scs},
4803 {TGSI_OPCODE_TXB, 0, SQ_TEX_INST_SAMPLE_LB, tgsi_tex},
4804 {TGSI_OPCODE_NRM, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4805 {TGSI_OPCODE_DIV, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4806 {TGSI_OPCODE_DP2, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4, tgsi_dp},
4807 {TGSI_OPCODE_TXL, 0, SQ_TEX_INST_SAMPLE_L, tgsi_tex},
4808 {TGSI_OPCODE_BRK, 0, V_SQ_CF_WORD1_SQ_CF_INST_LOOP_BREAK, tgsi_loop_brk_cont},
4809 {TGSI_OPCODE_IF, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_if},
4810 /* gap */
4811 {75, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4812 {76, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4813 {TGSI_OPCODE_ELSE, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_else},
4814 {TGSI_OPCODE_ENDIF, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_endif},
4815 /* gap */
4816 {79, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4817 {80, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4818 {TGSI_OPCODE_PUSHA, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4819 {TGSI_OPCODE_POPA, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4820 {TGSI_OPCODE_CEIL, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4821 {TGSI_OPCODE_I2F, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_INT_TO_FLT, tgsi_op2_trans},
4822 {TGSI_OPCODE_NOT, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOT_INT, tgsi_op2},
4823 {TGSI_OPCODE_TRUNC, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_TRUNC, tgsi_op2},
4824 {TGSI_OPCODE_SHL, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LSHL_INT, tgsi_op2_trans},
4825 /* gap */
4826 {88, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4827 {TGSI_OPCODE_AND, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_AND_INT, tgsi_op2},
4828 {TGSI_OPCODE_OR, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_OR_INT, tgsi_op2},
4829 {TGSI_OPCODE_MOD, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_imod},
4830 {TGSI_OPCODE_XOR, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_XOR_INT, tgsi_op2},
4831 {TGSI_OPCODE_SAD, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4832 {TGSI_OPCODE_TXF, 0, SQ_TEX_INST_LD, tgsi_tex},
4833 {TGSI_OPCODE_TXQ, 0, SQ_TEX_INST_GET_TEXTURE_RESINFO, tgsi_tex},
4834 {TGSI_OPCODE_CONT, 0, V_SQ_CF_WORD1_SQ_CF_INST_LOOP_CONTINUE, tgsi_loop_brk_cont},
4835 {TGSI_OPCODE_EMIT, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4836 {TGSI_OPCODE_ENDPRIM, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4837 {TGSI_OPCODE_BGNLOOP, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_bgnloop},
4838 {TGSI_OPCODE_BGNSUB, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4839 {TGSI_OPCODE_ENDLOOP, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_endloop},
4840 {TGSI_OPCODE_ENDSUB, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4841 /* gap */
4842 {103, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4843 {104, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4844 {105, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4845 {106, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4846 {TGSI_OPCODE_NOP, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4847 /* gap */
4848 {108, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4849 {109, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4850 {110, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4851 {111, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4852 {TGSI_OPCODE_NRM4, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4853 {TGSI_OPCODE_CALLNZ, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4854 {TGSI_OPCODE_IFC, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4855 {TGSI_OPCODE_BREAKC, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4856 {TGSI_OPCODE_KIL, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT, tgsi_kill}, /* conditional kill */
4857 {TGSI_OPCODE_END, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_end}, /* aka HALT */
4858 /* gap */
4859 {118, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4860 {TGSI_OPCODE_F2I, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLT_TO_INT, tgsi_op2_trans},
4861 {TGSI_OPCODE_IDIV, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_idiv},
4862 {TGSI_OPCODE_IMAX, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MAX_INT, tgsi_op2},
4863 {TGSI_OPCODE_IMIN, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MIN_INT, tgsi_op2},
4864 {TGSI_OPCODE_INEG, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SUB_INT, tgsi_ineg},
4865 {TGSI_OPCODE_ISGE, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGE_INT, tgsi_op2},
4866 {TGSI_OPCODE_ISHR, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ASHR_INT, tgsi_op2_trans},
4867 {TGSI_OPCODE_ISLT, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGT_INT, tgsi_op2_swap},
4868 {TGSI_OPCODE_F2U, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLT_TO_UINT, tgsi_op2},
4869 {TGSI_OPCODE_U2F, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_UINT_TO_FLT, tgsi_op2_trans},
4870 {TGSI_OPCODE_UADD, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ADD_INT, tgsi_op2},
4871 {TGSI_OPCODE_UDIV, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_udiv},
4872 {TGSI_OPCODE_UMAD, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_umad},
4873 {TGSI_OPCODE_UMAX, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MAX_UINT, tgsi_op2},
4874 {TGSI_OPCODE_UMIN, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MIN_UINT, tgsi_op2},
4875 {TGSI_OPCODE_UMOD, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_umod},
4876 {TGSI_OPCODE_UMUL, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MULLO_UINT, tgsi_op2_trans},
4877 {TGSI_OPCODE_USEQ, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETE_INT, tgsi_op2},
4878 {TGSI_OPCODE_USGE, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGE_UINT, tgsi_op2},
4879 {TGSI_OPCODE_USHR, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LSHR_INT, tgsi_op2_trans},
4880 {TGSI_OPCODE_USLT, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGT_UINT, tgsi_op2_swap},
4881 {TGSI_OPCODE_USNE, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETNE_INT, tgsi_op2_swap},
4882 {TGSI_OPCODE_SWITCH, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4883 {TGSI_OPCODE_CASE, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4884 {TGSI_OPCODE_DEFAULT, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4885 {TGSI_OPCODE_ENDSWITCH, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4886 {TGSI_OPCODE_LOAD, 0, 0, tgsi_unsupported},
4887 {TGSI_OPCODE_LOAD_MS, 0, 0, tgsi_unsupported},
4888 {TGSI_OPCODE_SAMPLE, 0, 0, tgsi_unsupported},
4889 {TGSI_OPCODE_SAMPLE_B, 0, 0, tgsi_unsupported},
4890 {TGSI_OPCODE_SAMPLE_C, 0, 0, tgsi_unsupported},
4891 {TGSI_OPCODE_SAMPLE_C_LZ, 0, 0, tgsi_unsupported},
4892 {TGSI_OPCODE_SAMPLE_D, 0, 0, tgsi_unsupported},
4893 {TGSI_OPCODE_SAMPLE_L, 0, 0, tgsi_unsupported},
4894 {TGSI_OPCODE_GATHER4, 0, 0, tgsi_unsupported},
4895 {TGSI_OPCODE_RESINFO, 0, 0, tgsi_unsupported},
4896 {TGSI_OPCODE_SAMPLE_POS, 0, 0, tgsi_unsupported},
4897 {TGSI_OPCODE_SAMPLE_INFO, 0, 0, tgsi_unsupported},
4898 {TGSI_OPCODE_UARL, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOVA_INT, tgsi_r600_arl},
4899 {TGSI_OPCODE_UCMP, 0, 0, tgsi_unsupported},
4900 {TGSI_OPCODE_IABS, 0, 0, tgsi_iabs},
4901 {TGSI_OPCODE_ISSG, 0, 0, tgsi_issg},
4902 {TGSI_OPCODE_LAST, 0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4903 };
4904
4905 static struct r600_shader_tgsi_instruction eg_shader_tgsi_instruction[] = {
4906 {TGSI_OPCODE_ARL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_eg_arl},
4907 {TGSI_OPCODE_MOV, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV, tgsi_op2},
4908 {TGSI_OPCODE_LIT, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_lit},
4909 {TGSI_OPCODE_RCP, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_IEEE, tgsi_trans_srcx_replicate},
4910 {TGSI_OPCODE_RSQ, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIPSQRT_IEEE, tgsi_rsq},
4911 {TGSI_OPCODE_EXP, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_exp},
4912 {TGSI_OPCODE_LOG, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_log},
4913 {TGSI_OPCODE_MUL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MUL, tgsi_op2},
4914 {TGSI_OPCODE_ADD, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ADD, tgsi_op2},
4915 {TGSI_OPCODE_DP3, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4, tgsi_dp},
4916 {TGSI_OPCODE_DP4, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4, tgsi_dp},
4917 {TGSI_OPCODE_DST, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_opdst},
4918 {TGSI_OPCODE_MIN, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MIN, tgsi_op2},
4919 {TGSI_OPCODE_MAX, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MAX, tgsi_op2},
4920 {TGSI_OPCODE_SLT, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGT, tgsi_op2_swap},
4921 {TGSI_OPCODE_SGE, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGE, tgsi_op2},
4922 {TGSI_OPCODE_MAD, 1, EG_V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_MULADD, tgsi_op3},
4923 {TGSI_OPCODE_SUB, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ADD, tgsi_op2},
4924 {TGSI_OPCODE_LRP, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_lrp},
4925 {TGSI_OPCODE_CND, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4926 /* gap */
4927 {20, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4928 {TGSI_OPCODE_DP2A, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4929 /* gap */
4930 {22, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4931 {23, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4932 {TGSI_OPCODE_FRC, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FRACT, tgsi_op2},
4933 {TGSI_OPCODE_CLAMP, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4934 {TGSI_OPCODE_FLR, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLOOR, tgsi_op2},
4935 {TGSI_OPCODE_ROUND, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RNDNE, tgsi_op2},
4936 {TGSI_OPCODE_EX2, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_EXP_IEEE, tgsi_trans_srcx_replicate},
4937 {TGSI_OPCODE_LG2, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LOG_IEEE, tgsi_trans_srcx_replicate},
4938 {TGSI_OPCODE_POW, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_pow},
4939 {TGSI_OPCODE_XPD, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_xpd},
4940 /* gap */
4941 {32, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4942 {TGSI_OPCODE_ABS, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV, tgsi_op2},
4943 {TGSI_OPCODE_RCC, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4944 {TGSI_OPCODE_DPH, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4, tgsi_dp},
4945 {TGSI_OPCODE_COS, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_COS, tgsi_trig},
4946 {TGSI_OPCODE_DDX, 0, SQ_TEX_INST_GET_GRADIENTS_H, tgsi_tex},
4947 {TGSI_OPCODE_DDY, 0, SQ_TEX_INST_GET_GRADIENTS_V, tgsi_tex},
4948 {TGSI_OPCODE_KILP, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT, tgsi_kill}, /* predicated kill */
4949 {TGSI_OPCODE_PK2H, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4950 {TGSI_OPCODE_PK2US, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4951 {TGSI_OPCODE_PK4B, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4952 {TGSI_OPCODE_PK4UB, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4953 {TGSI_OPCODE_RFL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4954 {TGSI_OPCODE_SEQ, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETE, tgsi_op2},
4955 {TGSI_OPCODE_SFL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4956 {TGSI_OPCODE_SGT, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGT, tgsi_op2},
4957 {TGSI_OPCODE_SIN, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SIN, tgsi_trig},
4958 {TGSI_OPCODE_SLE, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGE, tgsi_op2_swap},
4959 {TGSI_OPCODE_SNE, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETNE, tgsi_op2},
4960 {TGSI_OPCODE_STR, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4961 {TGSI_OPCODE_TEX, 0, SQ_TEX_INST_SAMPLE, tgsi_tex},
4962 {TGSI_OPCODE_TXD, 0, SQ_TEX_INST_SAMPLE_G, tgsi_tex},
4963 {TGSI_OPCODE_TXP, 0, SQ_TEX_INST_SAMPLE, tgsi_tex},
4964 {TGSI_OPCODE_UP2H, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4965 {TGSI_OPCODE_UP2US, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4966 {TGSI_OPCODE_UP4B, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4967 {TGSI_OPCODE_UP4UB, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4968 {TGSI_OPCODE_X2D, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4969 {TGSI_OPCODE_ARA, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4970 {TGSI_OPCODE_ARR, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_eg_arl},
4971 {TGSI_OPCODE_BRA, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4972 {TGSI_OPCODE_CAL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4973 {TGSI_OPCODE_RET, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4974 {TGSI_OPCODE_SSG, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_ssg},
4975 {TGSI_OPCODE_CMP, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_cmp},
4976 {TGSI_OPCODE_SCS, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_scs},
4977 {TGSI_OPCODE_TXB, 0, SQ_TEX_INST_SAMPLE_LB, tgsi_tex},
4978 {TGSI_OPCODE_NRM, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4979 {TGSI_OPCODE_DIV, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4980 {TGSI_OPCODE_DP2, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4, tgsi_dp},
4981 {TGSI_OPCODE_TXL, 0, SQ_TEX_INST_SAMPLE_L, tgsi_tex},
4982 {TGSI_OPCODE_BRK, 0, EG_V_SQ_CF_WORD1_SQ_CF_INST_LOOP_BREAK, tgsi_loop_brk_cont},
4983 {TGSI_OPCODE_IF, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_if},
4984 /* gap */
4985 {75, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4986 {76, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4987 {TGSI_OPCODE_ELSE, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_else},
4988 {TGSI_OPCODE_ENDIF, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_endif},
4989 /* gap */
4990 {79, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4991 {80, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4992 {TGSI_OPCODE_PUSHA, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4993 {TGSI_OPCODE_POPA, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4994 {TGSI_OPCODE_CEIL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
4995 {TGSI_OPCODE_I2F, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_INT_TO_FLT, tgsi_op2_trans},
4996 {TGSI_OPCODE_NOT, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOT_INT, tgsi_op2},
4997 {TGSI_OPCODE_TRUNC, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_TRUNC, tgsi_op2},
4998 {TGSI_OPCODE_SHL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LSHL_INT, tgsi_op2},
4999 /* gap */
5000 {88, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5001 {TGSI_OPCODE_AND, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_AND_INT, tgsi_op2},
5002 {TGSI_OPCODE_OR, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_OR_INT, tgsi_op2},
5003 {TGSI_OPCODE_MOD, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_imod},
5004 {TGSI_OPCODE_XOR, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_XOR_INT, tgsi_op2},
5005 {TGSI_OPCODE_SAD, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5006 {TGSI_OPCODE_TXF, 0, SQ_TEX_INST_LD, tgsi_tex},
5007 {TGSI_OPCODE_TXQ, 0, SQ_TEX_INST_GET_TEXTURE_RESINFO, tgsi_tex},
5008 {TGSI_OPCODE_CONT, 0, EG_V_SQ_CF_WORD1_SQ_CF_INST_LOOP_CONTINUE, tgsi_loop_brk_cont},
5009 {TGSI_OPCODE_EMIT, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5010 {TGSI_OPCODE_ENDPRIM, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5011 {TGSI_OPCODE_BGNLOOP, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_bgnloop},
5012 {TGSI_OPCODE_BGNSUB, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5013 {TGSI_OPCODE_ENDLOOP, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_endloop},
5014 {TGSI_OPCODE_ENDSUB, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5015 /* gap */
5016 {103, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5017 {104, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5018 {105, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5019 {106, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5020 {TGSI_OPCODE_NOP, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5021 /* gap */
5022 {108, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5023 {109, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5024 {110, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5025 {111, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5026 {TGSI_OPCODE_NRM4, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5027 {TGSI_OPCODE_CALLNZ, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5028 {TGSI_OPCODE_IFC, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5029 {TGSI_OPCODE_BREAKC, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5030 {TGSI_OPCODE_KIL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT, tgsi_kill}, /* conditional kill */
5031 {TGSI_OPCODE_END, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_end}, /* aka HALT */
5032 /* gap */
5033 {118, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5034 {TGSI_OPCODE_F2I, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLT_TO_INT, tgsi_op2},
5035 {TGSI_OPCODE_IDIV, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_idiv},
5036 {TGSI_OPCODE_IMAX, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MAX_INT, tgsi_op2},
5037 {TGSI_OPCODE_IMIN, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MIN_INT, tgsi_op2},
5038 {TGSI_OPCODE_INEG, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SUB_INT, tgsi_ineg},
5039 {TGSI_OPCODE_ISGE, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGE_INT, tgsi_op2},
5040 {TGSI_OPCODE_ISHR, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ASHR_INT, tgsi_op2},
5041 {TGSI_OPCODE_ISLT, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGT_INT, tgsi_op2_swap},
5042 {TGSI_OPCODE_F2U, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLT_TO_UINT, tgsi_op2_trans},
5043 {TGSI_OPCODE_U2F, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_UINT_TO_FLT, tgsi_op2_trans},
5044 {TGSI_OPCODE_UADD, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ADD_INT, tgsi_op2},
5045 {TGSI_OPCODE_UDIV, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_udiv},
5046 {TGSI_OPCODE_UMAD, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_umad},
5047 {TGSI_OPCODE_UMAX, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MAX_UINT, tgsi_op2},
5048 {TGSI_OPCODE_UMIN, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MIN_UINT, tgsi_op2},
5049 {TGSI_OPCODE_UMOD, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_umod},
5050 {TGSI_OPCODE_UMUL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MULLO_UINT, tgsi_op2_trans},
5051 {TGSI_OPCODE_USEQ, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETE_INT, tgsi_op2},
5052 {TGSI_OPCODE_USGE, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGE_UINT, tgsi_op2},
5053 {TGSI_OPCODE_USHR, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LSHR_INT, tgsi_op2},
5054 {TGSI_OPCODE_USLT, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGT_UINT, tgsi_op2_swap},
5055 {TGSI_OPCODE_USNE, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETNE_INT, tgsi_op2},
5056 {TGSI_OPCODE_SWITCH, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5057 {TGSI_OPCODE_CASE, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5058 {TGSI_OPCODE_DEFAULT, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5059 {TGSI_OPCODE_ENDSWITCH, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5060 {TGSI_OPCODE_LOAD, 0, 0, tgsi_unsupported},
5061 {TGSI_OPCODE_LOAD_MS, 0, 0, tgsi_unsupported},
5062 {TGSI_OPCODE_SAMPLE, 0, 0, tgsi_unsupported},
5063 {TGSI_OPCODE_SAMPLE_B, 0, 0, tgsi_unsupported},
5064 {TGSI_OPCODE_SAMPLE_C, 0, 0, tgsi_unsupported},
5065 {TGSI_OPCODE_SAMPLE_C_LZ, 0, 0, tgsi_unsupported},
5066 {TGSI_OPCODE_SAMPLE_D, 0, 0, tgsi_unsupported},
5067 {TGSI_OPCODE_SAMPLE_L, 0, 0, tgsi_unsupported},
5068 {TGSI_OPCODE_GATHER4, 0, 0, tgsi_unsupported},
5069 {TGSI_OPCODE_RESINFO, 0, 0, tgsi_unsupported},
5070 {TGSI_OPCODE_SAMPLE_POS, 0, 0, tgsi_unsupported},
5071 {TGSI_OPCODE_SAMPLE_INFO, 0, 0, tgsi_unsupported},
5072 {TGSI_OPCODE_UARL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOVA_INT, tgsi_eg_arl},
5073 {TGSI_OPCODE_UCMP, 0, 0, tgsi_unsupported},
5074 {TGSI_OPCODE_IABS, 0, 0, tgsi_iabs},
5075 {TGSI_OPCODE_ISSG, 0, 0, tgsi_issg},
5076 {TGSI_OPCODE_LAST, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5077 };
5078
5079 static struct r600_shader_tgsi_instruction cm_shader_tgsi_instruction[] = {
5080 {TGSI_OPCODE_ARL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_eg_arl},
5081 {TGSI_OPCODE_MOV, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV, tgsi_op2},
5082 {TGSI_OPCODE_LIT, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_lit},
5083 {TGSI_OPCODE_RCP, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_IEEE, cayman_emit_float_instr},
5084 {TGSI_OPCODE_RSQ, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIPSQRT_IEEE, cayman_emit_float_instr},
5085 {TGSI_OPCODE_EXP, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_exp},
5086 {TGSI_OPCODE_LOG, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_log},
5087 {TGSI_OPCODE_MUL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MUL, tgsi_op2},
5088 {TGSI_OPCODE_ADD, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ADD, tgsi_op2},
5089 {TGSI_OPCODE_DP3, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4, tgsi_dp},
5090 {TGSI_OPCODE_DP4, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4, tgsi_dp},
5091 {TGSI_OPCODE_DST, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_opdst},
5092 {TGSI_OPCODE_MIN, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MIN, tgsi_op2},
5093 {TGSI_OPCODE_MAX, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MAX, tgsi_op2},
5094 {TGSI_OPCODE_SLT, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGT, tgsi_op2_swap},
5095 {TGSI_OPCODE_SGE, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGE, tgsi_op2},
5096 {TGSI_OPCODE_MAD, 1, EG_V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_MULADD, tgsi_op3},
5097 {TGSI_OPCODE_SUB, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ADD, tgsi_op2},
5098 {TGSI_OPCODE_LRP, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_lrp},
5099 {TGSI_OPCODE_CND, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5100 /* gap */
5101 {20, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5102 {TGSI_OPCODE_DP2A, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5103 /* gap */
5104 {22, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5105 {23, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5106 {TGSI_OPCODE_FRC, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FRACT, tgsi_op2},
5107 {TGSI_OPCODE_CLAMP, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5108 {TGSI_OPCODE_FLR, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLOOR, tgsi_op2},
5109 {TGSI_OPCODE_ROUND, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RNDNE, tgsi_op2},
5110 {TGSI_OPCODE_EX2, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_EXP_IEEE, cayman_emit_float_instr},
5111 {TGSI_OPCODE_LG2, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LOG_IEEE, cayman_emit_float_instr},
5112 {TGSI_OPCODE_POW, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, cayman_pow},
5113 {TGSI_OPCODE_XPD, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_xpd},
5114 /* gap */
5115 {32, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5116 {TGSI_OPCODE_ABS, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV, tgsi_op2},
5117 {TGSI_OPCODE_RCC, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5118 {TGSI_OPCODE_DPH, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4, tgsi_dp},
5119 {TGSI_OPCODE_COS, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_COS, cayman_trig},
5120 {TGSI_OPCODE_DDX, 0, SQ_TEX_INST_GET_GRADIENTS_H, tgsi_tex},
5121 {TGSI_OPCODE_DDY, 0, SQ_TEX_INST_GET_GRADIENTS_V, tgsi_tex},
5122 {TGSI_OPCODE_KILP, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT, tgsi_kill}, /* predicated kill */
5123 {TGSI_OPCODE_PK2H, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5124 {TGSI_OPCODE_PK2US, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5125 {TGSI_OPCODE_PK4B, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5126 {TGSI_OPCODE_PK4UB, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5127 {TGSI_OPCODE_RFL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5128 {TGSI_OPCODE_SEQ, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETE, tgsi_op2},
5129 {TGSI_OPCODE_SFL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5130 {TGSI_OPCODE_SGT, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGT, tgsi_op2},
5131 {TGSI_OPCODE_SIN, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SIN, cayman_trig},
5132 {TGSI_OPCODE_SLE, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGE, tgsi_op2_swap},
5133 {TGSI_OPCODE_SNE, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETNE, tgsi_op2},
5134 {TGSI_OPCODE_STR, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5135 {TGSI_OPCODE_TEX, 0, SQ_TEX_INST_SAMPLE, tgsi_tex},
5136 {TGSI_OPCODE_TXD, 0, SQ_TEX_INST_SAMPLE_G, tgsi_tex},
5137 {TGSI_OPCODE_TXP, 0, SQ_TEX_INST_SAMPLE, tgsi_tex},
5138 {TGSI_OPCODE_UP2H, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5139 {TGSI_OPCODE_UP2US, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5140 {TGSI_OPCODE_UP4B, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5141 {TGSI_OPCODE_UP4UB, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5142 {TGSI_OPCODE_X2D, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5143 {TGSI_OPCODE_ARA, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5144 {TGSI_OPCODE_ARR, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_eg_arl},
5145 {TGSI_OPCODE_BRA, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5146 {TGSI_OPCODE_CAL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5147 {TGSI_OPCODE_RET, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5148 {TGSI_OPCODE_SSG, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_ssg},
5149 {TGSI_OPCODE_CMP, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_cmp},
5150 {TGSI_OPCODE_SCS, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_scs},
5151 {TGSI_OPCODE_TXB, 0, SQ_TEX_INST_SAMPLE_LB, tgsi_tex},
5152 {TGSI_OPCODE_NRM, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5153 {TGSI_OPCODE_DIV, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5154 {TGSI_OPCODE_DP2, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4, tgsi_dp},
5155 {TGSI_OPCODE_TXL, 0, SQ_TEX_INST_SAMPLE_L, tgsi_tex},
5156 {TGSI_OPCODE_BRK, 0, EG_V_SQ_CF_WORD1_SQ_CF_INST_LOOP_BREAK, tgsi_loop_brk_cont},
5157 {TGSI_OPCODE_IF, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_if},
5158 /* gap */
5159 {75, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5160 {76, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5161 {TGSI_OPCODE_ELSE, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_else},
5162 {TGSI_OPCODE_ENDIF, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_endif},
5163 /* gap */
5164 {79, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5165 {80, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5166 {TGSI_OPCODE_PUSHA, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5167 {TGSI_OPCODE_POPA, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5168 {TGSI_OPCODE_CEIL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5169 {TGSI_OPCODE_I2F, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_INT_TO_FLT, tgsi_op2},
5170 {TGSI_OPCODE_NOT, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOT_INT, tgsi_op2},
5171 {TGSI_OPCODE_TRUNC, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_TRUNC, tgsi_op2},
5172 {TGSI_OPCODE_SHL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LSHL_INT, tgsi_op2},
5173 /* gap */
5174 {88, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5175 {TGSI_OPCODE_AND, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_AND_INT, tgsi_op2},
5176 {TGSI_OPCODE_OR, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_OR_INT, tgsi_op2},
5177 {TGSI_OPCODE_MOD, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_imod},
5178 {TGSI_OPCODE_XOR, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_XOR_INT, tgsi_op2},
5179 {TGSI_OPCODE_SAD, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5180 {TGSI_OPCODE_TXF, 0, SQ_TEX_INST_LD, tgsi_tex},
5181 {TGSI_OPCODE_TXQ, 0, SQ_TEX_INST_GET_TEXTURE_RESINFO, tgsi_tex},
5182 {TGSI_OPCODE_CONT, 0, EG_V_SQ_CF_WORD1_SQ_CF_INST_LOOP_CONTINUE, tgsi_loop_brk_cont},
5183 {TGSI_OPCODE_EMIT, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5184 {TGSI_OPCODE_ENDPRIM, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5185 {TGSI_OPCODE_BGNLOOP, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_bgnloop},
5186 {TGSI_OPCODE_BGNSUB, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5187 {TGSI_OPCODE_ENDLOOP, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_endloop},
5188 {TGSI_OPCODE_ENDSUB, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5189 /* gap */
5190 {103, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5191 {104, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5192 {105, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5193 {106, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5194 {TGSI_OPCODE_NOP, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5195 /* gap */
5196 {108, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5197 {109, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5198 {110, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5199 {111, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5200 {TGSI_OPCODE_NRM4, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5201 {TGSI_OPCODE_CALLNZ, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5202 {TGSI_OPCODE_IFC, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5203 {TGSI_OPCODE_BREAKC, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5204 {TGSI_OPCODE_KIL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT, tgsi_kill}, /* conditional kill */
5205 {TGSI_OPCODE_END, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_end}, /* aka HALT */
5206 /* gap */
5207 {118, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5208 {TGSI_OPCODE_F2I, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLT_TO_INT, tgsi_op2},
5209 {TGSI_OPCODE_IDIV, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_idiv},
5210 {TGSI_OPCODE_IMAX, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MAX_INT, tgsi_op2},
5211 {TGSI_OPCODE_IMIN, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MIN_INT, tgsi_op2},
5212 {TGSI_OPCODE_INEG, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SUB_INT, tgsi_ineg},
5213 {TGSI_OPCODE_ISGE, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGE_INT, tgsi_op2},
5214 {TGSI_OPCODE_ISHR, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ASHR_INT, tgsi_op2},
5215 {TGSI_OPCODE_ISLT, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGT_INT, tgsi_op2_swap},
5216 {TGSI_OPCODE_F2U, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLT_TO_UINT, tgsi_op2},
5217 {TGSI_OPCODE_U2F, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_UINT_TO_FLT, tgsi_op2},
5218 {TGSI_OPCODE_UADD, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ADD_INT, tgsi_op2},
5219 {TGSI_OPCODE_UDIV, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_udiv},
5220 {TGSI_OPCODE_UMAD, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_umad},
5221 {TGSI_OPCODE_UMAX, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MAX_UINT, tgsi_op2},
5222 {TGSI_OPCODE_UMIN, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MIN_UINT, tgsi_op2},
5223 {TGSI_OPCODE_UMOD, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_umod},
5224 {TGSI_OPCODE_UMUL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MULLO_INT, cayman_mul_int_instr},
5225 {TGSI_OPCODE_USEQ, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETE_INT, tgsi_op2},
5226 {TGSI_OPCODE_USGE, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGE_UINT, tgsi_op2},
5227 {TGSI_OPCODE_USHR, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LSHR_INT, tgsi_op2},
5228 {TGSI_OPCODE_USLT, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGT_UINT, tgsi_op2_swap},
5229 {TGSI_OPCODE_USNE, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETNE_INT, tgsi_op2},
5230 {TGSI_OPCODE_SWITCH, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5231 {TGSI_OPCODE_CASE, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5232 {TGSI_OPCODE_DEFAULT, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5233 {TGSI_OPCODE_ENDSWITCH, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5234 {TGSI_OPCODE_LOAD, 0, 0, tgsi_unsupported},
5235 {TGSI_OPCODE_LOAD_MS, 0, 0, tgsi_unsupported},
5236 {TGSI_OPCODE_SAMPLE, 0, 0, tgsi_unsupported},
5237 {TGSI_OPCODE_SAMPLE_B, 0, 0, tgsi_unsupported},
5238 {TGSI_OPCODE_SAMPLE_C, 0, 0, tgsi_unsupported},
5239 {TGSI_OPCODE_SAMPLE_C_LZ, 0, 0, tgsi_unsupported},
5240 {TGSI_OPCODE_SAMPLE_D, 0, 0, tgsi_unsupported},
5241 {TGSI_OPCODE_SAMPLE_L, 0, 0, tgsi_unsupported},
5242 {TGSI_OPCODE_GATHER4, 0, 0, tgsi_unsupported},
5243 {TGSI_OPCODE_RESINFO, 0, 0, tgsi_unsupported},
5244 {TGSI_OPCODE_SAMPLE_POS, 0, 0, tgsi_unsupported},
5245 {TGSI_OPCODE_SAMPLE_INFO, 0, 0, tgsi_unsupported},
5246 {TGSI_OPCODE_UARL, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOVA_INT, tgsi_eg_arl},
5247 {TGSI_OPCODE_UCMP, 0, 0, tgsi_unsupported},
5248 {TGSI_OPCODE_IABS, 0, 0, tgsi_iabs},
5249 {TGSI_OPCODE_ISSG, 0, 0, tgsi_issg},
5250 {TGSI_OPCODE_LAST, 0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
5251 };