src/gallium/drivers/radeonsi/radeonsi_shader.c

   1
   2 /*
   3  * Copyright 2012 Advanced Micro Devices, Inc.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * on the rights to use, copy, modify, merge, publish, distribute, sub
   9  * license, and/or sell copies of the Software, and to permit persons to whom
  10  * the Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  19  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  20  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  21  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  22  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  23  *
  24  * Authors:
  25  *      Tom Stellard <thomas.stellard@amd.com>
  26  *      Michel Dänzer <michel.daenzer@amd.com>
  27  *      Christian König <christian.koenig@amd.com>
  28  */
  29
  30 #include "gallivm/lp_bld_tgsi_action.h"
  31 #include "gallivm/lp_bld_const.h"
  32 #include "gallivm/lp_bld_gather.h"
  33 #include "gallivm/lp_bld_intr.h"
  34 #include "gallivm/lp_bld_tgsi.h"
  35 #include "radeon_llvm.h"
  36 #include "radeon_llvm_emit.h"
  37 #include "tgsi/tgsi_info.h"
  38 #include "tgsi/tgsi_parse.h"
  39 #include "tgsi/tgsi_scan.h"
  40 #include "tgsi/tgsi_dump.h"
  41
  42 #include "radeonsi_pipe.h"
  43 #include "radeonsi_shader.h"
  44 #include "si_state.h"
  45 #include "sid.h"
  46
  47 #include <assert.h>
  48 #include <errno.h>
  49 #include <stdio.h>
  50
  51 /*
  52 static ps_remap_inputs(
  53         struct tgsi_llvm_context * tl_ctx,
  54         unsigned tgsi_index,
  55         unsigned tgsi_chan)
  56 {
  57         :
  58 }
  59
  60 struct si_input
  61 {
  62         struct list_head head;
  63         unsigned tgsi_index;
  64         unsigned tgsi_chan;
  65         unsigned order;
  66 };
  67 */
  68
  69
  70 struct si_shader_context
  71 {
  72         struct radeon_llvm_context radeon_bld;
  73         struct r600_context *rctx;
  74         struct tgsi_parse_context parse;
  75         struct tgsi_token * tokens;
  76         struct si_pipe_shader *shader;
  77         unsigned type; /* TGSI_PROCESSOR_* specifies the type of shader. */
  78 /*      unsigned num_inputs; */
  79 /*      struct list_head inputs; */
  80 /*      unsigned * input_mappings *//* From TGSI to SI hw */
  81 /*      struct tgsi_shader_info info;*/
  82 };
  83
  84 static struct si_shader_context * si_shader_context(
  85         struct lp_build_tgsi_context * bld_base)
  86 {
  87         return (struct si_shader_context *)bld_base;
  88 }
  89
  90
  91 #define PERSPECTIVE_BASE 0
  92 #define LINEAR_BASE 9
  93
  94 #define SAMPLE_OFFSET 0
  95 #define CENTER_OFFSET 2
  96 #define CENTROID_OFSET 4
  97
  98 #define USE_SGPR_MAX_SUFFIX_LEN 5
  99 #define CONST_ADDR_SPACE 2
 100 #define USER_SGPR_ADDR_SPACE 8
 101
 102 enum sgpr_type {
 103         SGPR_CONST_PTR_F32,
 104         SGPR_CONST_PTR_V4I32,
 105         SGPR_CONST_PTR_V8I32,
 106         SGPR_I32,
 107         SGPR_I64
 108 };
 109
 110 /**
 111  * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad
 112  *
 113  * @param offset The offset parameter specifies the number of
 114  * elements to offset, not the number of bytes or dwords.  An element is the
 115  * the type pointed to by the base_ptr parameter (e.g. int is the element of
 116  * an int* pointer)
 117  *
 118  * When LLVM lowers the load instruction, it will convert the element offset
 119  * into a dword offset automatically.
 120  *
 121  */
 122 static LLVMValueRef build_indexed_load(
 123         struct gallivm_state * gallivm,
 124         LLVMValueRef base_ptr,
 125         LLVMValueRef offset)
 126 {
 127         LLVMValueRef computed_ptr = LLVMBuildGEP(
 128                 gallivm->builder, base_ptr, &offset, 1, "");
 129
 130         return LLVMBuildLoad(gallivm->builder, computed_ptr, "");
 131 }
 132
 133 /**
 134  * Load a value stored in one of the user SGPRs
 135  *
 136  * @param sgpr This is the sgpr to load the value from.  If you need to load a
 137  * value that is stored in consecutive SGPR registers (e.g. a 64-bit pointer),
 138  * then you should pass the index of the first SGPR that holds the value.  For
 139  * example, if you want to load a pointer that is stored in SGPRs 2 and 3, then
 140  * use pass 2 for the sgpr parameter.
 141  *
 142  * The value of the sgpr parameter must also be aligned to the width of the type
 143  * being loaded, so that the sgpr parameter is divisible by the dword width of the
 144  * type.  For example, if the value being loaded is two dwords wide, then the sgpr
 145  * parameter must be divisible by two.
 146  */
 147 static LLVMValueRef use_sgpr(
 148         struct gallivm_state * gallivm,
 149         enum sgpr_type type,
 150         unsigned sgpr)
 151 {
 152         LLVMValueRef sgpr_index;
 153         LLVMTypeRef ret_type;
 154         LLVMValueRef ptr;
 155
 156         sgpr_index = lp_build_const_int32(gallivm, sgpr);
 157
 158         switch (type) {
 159         case SGPR_CONST_PTR_F32:
 160                 assert(sgpr % 2 == 0);
 161                 ret_type = LLVMFloatTypeInContext(gallivm->context);
 162                 ret_type = LLVMPointerType(ret_type, CONST_ADDR_SPACE);
 163                 break;
 164
 165         case SGPR_I32:
 166                 ret_type = LLVMInt32TypeInContext(gallivm->context);
 167                 break;
 168
 169         case SGPR_I64:
 170                 assert(sgpr % 2 == 0);
 171                 ret_type= LLVMInt64TypeInContext(gallivm->context);
 172                 break;
 173
 174         case SGPR_CONST_PTR_V4I32:
 175                 assert(sgpr % 2 == 0);
 176                 ret_type = LLVMInt32TypeInContext(gallivm->context);
 177                 ret_type = LLVMVectorType(ret_type, 4);
 178                 ret_type = LLVMPointerType(ret_type, CONST_ADDR_SPACE);
 179                 break;
 180
 181         case SGPR_CONST_PTR_V8I32:
 182                 assert(sgpr % 2 == 0);
 183                 ret_type = LLVMInt32TypeInContext(gallivm->context);
 184                 ret_type = LLVMVectorType(ret_type, 8);
 185                 ret_type = LLVMPointerType(ret_type, CONST_ADDR_SPACE);
 186                 break;
 187
 188         default:
 189                 assert(!"Unsupported SGPR type in use_sgpr()");
 190                 return NULL;
 191         }
 192
 193         ret_type = LLVMPointerType(ret_type, USER_SGPR_ADDR_SPACE);
 194         ptr = LLVMBuildIntToPtr(gallivm->builder, sgpr_index, ret_type, "");
 195         return LLVMBuildLoad(gallivm->builder, ptr, "");
 196 }
 197
 198 static void declare_input_vs(
 199         struct si_shader_context * si_shader_ctx,
 200         unsigned input_index,
 201         const struct tgsi_full_declaration *decl)
 202 {
 203         LLVMValueRef t_list_ptr;
 204         LLVMValueRef t_offset;
 205         LLVMValueRef t_list;
 206         LLVMValueRef attribute_offset;
 207         LLVMValueRef buffer_index_reg;
 208         LLVMValueRef args[3];
 209         LLVMTypeRef vec4_type;
 210         LLVMValueRef input;
 211         struct lp_build_context * uint = &si_shader_ctx->radeon_bld.soa.bld_base.uint_bld;
 212         struct lp_build_context * base = &si_shader_ctx->radeon_bld.soa.bld_base.base;
 213         struct r600_context *rctx = si_shader_ctx->rctx;
 214         //struct pipe_vertex_element *velem = &rctx->vertex_elements->elements[input_index];
 215         unsigned chan;
 216
 217         /* Load the T list */
 218         /* XXX: Communicate with the rest of the driver about which SGPR the T#
 219          * list pointer is going to be stored in.  Hard code to SGPR[6:7] for
 220          * now */
 221         t_list_ptr = use_sgpr(base->gallivm, SGPR_CONST_PTR_V4I32, 6);
 222
 223         t_offset = lp_build_const_int32(base->gallivm, input_index);
 224
 225         t_list = build_indexed_load(base->gallivm, t_list_ptr, t_offset);
 226
 227         /* Build the attribute offset */
 228         attribute_offset = lp_build_const_int32(base->gallivm, 0);
 229
 230         /* Load the buffer index is always, which is always stored in VGPR0
 231          * for Vertex Shaders */
 232         buffer_index_reg = lp_build_intrinsic(base->gallivm->builder,
 233                 "llvm.SI.vs.load.buffer.index", uint->elem_type, NULL, 0);
 234
 235         vec4_type = LLVMVectorType(base->elem_type, 4);
 236         args[0] = t_list;
 237         args[1] = attribute_offset;
 238         args[2] = buffer_index_reg;
 239         input = lp_build_intrinsic(base->gallivm->builder,
 240                 "llvm.SI.vs.load.input", vec4_type, args, 3);
 241
 242         /* Break up the vec4 into individual components */
 243         for (chan = 0; chan < 4; chan++) {
 244                 LLVMValueRef llvm_chan = lp_build_const_int32(base->gallivm, chan);
 245                 /* XXX: Use a helper function for this.  There is one in
 246                  * tgsi_llvm.c. */
 247                 si_shader_ctx->radeon_bld.inputs[radeon_llvm_reg_index_soa(input_index, chan)] =
 248                                 LLVMBuildExtractElement(base->gallivm->builder,
 249                                 input, llvm_chan, "");
 250         }
 251 }
 252
 253 static void declare_input_fs(
 254         struct si_shader_context * si_shader_ctx,
 255         unsigned input_index,
 256         const struct tgsi_full_declaration *decl)
 257 {
 258         const char * intr_name;
 259         unsigned chan;
 260         struct lp_build_context * base =
 261                                 &si_shader_ctx->radeon_bld.soa.bld_base.base;
 262         struct gallivm_state * gallivm = base->gallivm;
 263
 264         /* This value is:
 265          * [15:0] NewPrimMask (Bit mask for each quad.  It is set it the
 266          *                     quad begins a new primitive.  Bit 0 always needs
 267          *                     to be unset)
 268          * [32:16] ParamOffset
 269          *
 270          */
 271         /* XXX: This register number must be identical to the S_00B02C_USER_SGPR
 272          * register field value
 273          */
 274         LLVMValueRef params = use_sgpr(base->gallivm, SGPR_I32, 6);
 275
 276
 277         /* XXX: Is this the input_index? */
 278         LLVMValueRef attr_number = lp_build_const_int32(gallivm, input_index);
 279
 280         /* XXX: Handle all possible interpolation modes */
 281         switch (decl->Interp.Interpolate) {
 282         case TGSI_INTERPOLATE_COLOR:
 283                 /* XXX: Flat shading hangs the GPU */
 284                 if (si_shader_ctx->rctx->queued.named.rasterizer->flatshade) {
 285 #if 0
 286                         intr_name = "llvm.SI.fs.interp.constant";
 287 #else
 288                         intr_name = "llvm.SI.fs.interp.linear.center";
 289 #endif
 290                 } else {
 291                         if (decl->Interp.Centroid)
 292                                 intr_name = "llvm.SI.fs.interp.persp.centroid";
 293                         else
 294                                 intr_name = "llvm.SI.fs.interp.persp.center";
 295                 }
 296                 break;
 297         case TGSI_INTERPOLATE_CONSTANT:
 298                 /* XXX: Flat shading hangs the GPU */
 299 #if 0
 300                 intr_name = "llvm.SI.fs.interp.constant";
 301                 break;
 302 #endif
 303         case TGSI_INTERPOLATE_LINEAR:
 304                 if (decl->Interp.Centroid)
 305                         intr_name = "llvm.SI.fs.interp.linear.centroid";
 306                 else
 307                         intr_name = "llvm.SI.fs.interp.linear.center";
 308                 break;
 309         case TGSI_INTERPOLATE_PERSPECTIVE:
 310                 if (decl->Interp.Centroid)
 311                         intr_name = "llvm.SI.fs.interp.persp.centroid";
 312                 else
 313                         intr_name = "llvm.SI.fs.interp.persp.center";
 314                 break;
 315         default:
 316                 fprintf(stderr, "Warning: Unhandled interpolation mode.\n");
 317                 return;
 318         }
 319
 320         /* XXX: Could there be more than TGSI_NUM_CHANNELS (4) ? */
 321         for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
 322                 LLVMValueRef args[3];
 323                 LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan);
 324                 unsigned soa_index = radeon_llvm_reg_index_soa(input_index, chan);
 325                 LLVMTypeRef input_type = LLVMFloatTypeInContext(gallivm->context);
 326                 args[0] = llvm_chan;
 327                 args[1] = attr_number;
 328                 args[2] = params;
 329                 si_shader_ctx->radeon_bld.inputs[soa_index] =
 330                         lp_build_intrinsic(gallivm->builder, intr_name,
 331                                                 input_type, args, 3);
 332         }
 333 }
 334
 335 static void declare_input(
 336         struct radeon_llvm_context * radeon_bld,
 337         unsigned input_index,
 338         const struct tgsi_full_declaration *decl)
 339 {
 340         struct si_shader_context * si_shader_ctx =
 341                                 si_shader_context(&radeon_bld->soa.bld_base);
 342         if (si_shader_ctx->type == TGSI_PROCESSOR_VERTEX) {
 343                 declare_input_vs(si_shader_ctx, input_index, decl);
 344         } else if (si_shader_ctx->type == TGSI_PROCESSOR_FRAGMENT) {
 345                 declare_input_fs(si_shader_ctx, input_index, decl);
 346         } else {
 347                 fprintf(stderr, "Warning: Unsupported shader type,\n");
 348         }
 349 }
 350
 351 static LLVMValueRef fetch_constant(
 352         struct lp_build_tgsi_context * bld_base,
 353         const struct tgsi_full_src_register *reg,
 354         enum tgsi_opcode_type type,
 355         unsigned swizzle)
 356 {
 357         struct lp_build_context * base = &bld_base->base;
 358
 359         LLVMValueRef const_ptr;
 360         LLVMValueRef offset;
 361         LLVMValueRef load;
 362
 363         /* XXX: Assume the pointer to the constant buffer is being stored in
 364          * SGPR[0:1] */
 365         const_ptr = use_sgpr(base->gallivm, SGPR_CONST_PTR_F32, 0);
 366
 367         /* XXX: This assumes that the constant buffer is not packed, so
 368          * CONST[0].x will have an offset of 0 and CONST[1].x will have an
 369          * offset of 4. */
 370         offset = lp_build_const_int32(base->gallivm,
 371                                         (reg->Register.Index * 4) + swizzle);
 372
 373         load = build_indexed_load(base->gallivm, const_ptr, offset);
 374         return bitcast(bld_base, type, load);
 375 }
 376
 377 /* XXX: This is partially implemented for VS only at this point.  It is not complete */
 378 static void si_llvm_emit_epilogue(struct lp_build_tgsi_context * bld_base)
 379 {
 380         struct si_shader_context * si_shader_ctx = si_shader_context(bld_base);
 381         struct si_shader * shader = &si_shader_ctx->shader->shader;
 382         struct lp_build_context * base = &bld_base->base;
 383         struct lp_build_context * uint =
 384                                 &si_shader_ctx->radeon_bld.soa.bld_base.uint_bld;
 385         struct tgsi_parse_context *parse = &si_shader_ctx->parse;
 386         LLVMValueRef last_args[9] = { 0 };
 387         unsigned color_count = 0;
 388         unsigned param_count = 0;
 389
 390         while (!tgsi_parse_end_of_tokens(parse)) {
 391                 /* XXX: component_bits controls which components of the output
 392                  * registers actually get exported. (e.g bit 0 means export
 393                  * X component, bit 1 means export Y component, etc.)  I'm
 394                  * hard coding this to 0xf for now.  In the future, we might
 395                  * want to do something else. */
 396                 unsigned component_bits = 0xf;
 397                 unsigned chan;
 398                 struct tgsi_full_declaration *d =
 399                                         &parse->FullToken.FullDeclaration;
 400                 LLVMValueRef args[9];
 401                 unsigned target;
 402                 unsigned index;
 403                 int i;
 404
 405                 tgsi_parse_token(parse);
 406                 if (parse->FullToken.Token.Type != TGSI_TOKEN_TYPE_DECLARATION)
 407                         continue;
 408
 409                 switch (d->Declaration.File) {
 410                 case TGSI_FILE_INPUT:
 411                         i = shader->ninput++;
 412                         shader->input[i].name = d->Semantic.Name;
 413                         shader->input[i].sid = d->Semantic.Index;
 414                         shader->input[i].interpolate = d->Interp.Interpolate;
 415                         shader->input[i].centroid = d->Interp.Centroid;
 416                         continue;
 417
 418                 case TGSI_FILE_OUTPUT:
 419                         i = shader->noutput++;
 420                         shader->output[i].name = d->Semantic.Name;
 421                         shader->output[i].sid = d->Semantic.Index;
 422                         shader->output[i].interpolate = d->Interp.Interpolate;
 423                         break;
 424
 425                 default:
 426                         continue;
 427                 }
 428
 429                 for (index = d->Range.First; index <= d->Range.Last; index++) {
 430                         for (chan = 0; chan < 4; chan++ ) {
 431                                 LLVMValueRef out_ptr =
 432                                         si_shader_ctx->radeon_bld.soa.outputs
 433                                         [index][chan];
 434                                 /* +5 because the first output value will be
 435                                  * the 6th argument to the intrinsic. */
 436                                 args[chan + 5]= LLVMBuildLoad(
 437                                         base->gallivm->builder, out_ptr, "");
 438                         }
 439
 440                         /* XXX: We probably need to keep track of the output
 441                          * values, so we know what we are passing to the next
 442                          * stage. */
 443
 444                         /* Select the correct target */
 445                         switch(d->Semantic.Name) {
 446                         case TGSI_SEMANTIC_POSITION:
 447                                 target = V_008DFC_SQ_EXP_POS;
 448                                 break;
 449                         case TGSI_SEMANTIC_COLOR:
 450                                 if (si_shader_ctx->type == TGSI_PROCESSOR_VERTEX) {
 451                                         target = V_008DFC_SQ_EXP_PARAM + param_count;
 452                                         shader->output[i].param_offset = param_count;
 453                                         param_count++;
 454                                 } else {
 455                                         target = V_008DFC_SQ_EXP_MRT + color_count;
 456                                         color_count++;
 457                                 }
 458                                 break;
 459                         case TGSI_SEMANTIC_GENERIC:
 460                                 target = V_008DFC_SQ_EXP_PARAM + param_count;
 461                                 shader->output[i].param_offset = param_count;
 462                                 param_count++;
 463                                 break;
 464                         default:
 465                                 target = 0;
 466                                 fprintf(stderr,
 467                                         "Warning: SI unhandled output type:%d\n",
 468                                         d->Semantic.Name);
 469                         }
 470
 471                         /* Specify which components to enable */
 472                         args[0] = lp_build_const_int32(base->gallivm,
 473                                                                 component_bits);
 474
 475                         /* Specify whether the EXEC mask represents the valid mask */
 476                         args[1] = lp_build_const_int32(base->gallivm, 0);
 477
 478                         /* Specify whether this is the last export */
 479                         args[2] = lp_build_const_int32(base->gallivm, 0);
 480
 481                         /* Specify the target we are exporting */
 482                         args[3] = lp_build_const_int32(base->gallivm, target);
 483
 484                         /* Set COMPR flag to zero to export data as 32-bit */
 485                         args[4] = uint->zero;
 486
 487                         if (si_shader_ctx->type == TGSI_PROCESSOR_VERTEX ?
 488                             (d->Semantic.Name == TGSI_SEMANTIC_POSITION) :
 489                             (d->Semantic.Name == TGSI_SEMANTIC_COLOR)) {
 490                                 if (last_args[0]) {
 491                                         lp_build_intrinsic(base->gallivm->builder,
 492                                                            "llvm.SI.export",
 493                                                            LLVMVoidTypeInContext(base->gallivm->context),
 494                                                            last_args, 9);
 495                                 }
 496
 497                                 memcpy(last_args, args, sizeof(args));
 498                         } else {
 499                                 lp_build_intrinsic(base->gallivm->builder,
 500                                                    "llvm.SI.export",
 501                                                    LLVMVoidTypeInContext(base->gallivm->context),
 502                                                    args, 9);
 503                         }
 504
 505                 }
 506         }
 507
 508         if (!last_args[0]) {
 509                 assert(si_shader_ctx->type == TGSI_PROCESSOR_FRAGMENT);
 510
 511                 /* Specify which components to enable */
 512                 last_args[0] = lp_build_const_int32(base->gallivm, 0x0);
 513
 514                 /* Specify the target we are exporting */
 515                 last_args[3] = lp_build_const_int32(base->gallivm, V_008DFC_SQ_EXP_MRT);
 516
 517                 /* Set COMPR flag to zero to export data as 32-bit */
 518                 last_args[4] = uint->zero;
 519
 520                 /* dummy bits */
 521                 last_args[5]= uint->zero;
 522                 last_args[6]= uint->zero;
 523                 last_args[7]= uint->zero;
 524                 last_args[8]= uint->zero;
 525         }
 526
 527         /* Specify whether the EXEC mask represents the valid mask */
 528         last_args[1] = lp_build_const_int32(base->gallivm,
 529                                             si_shader_ctx->type == TGSI_PROCESSOR_FRAGMENT);
 530
 531         /* Specify that this is the last export */
 532         last_args[2] = lp_build_const_int32(base->gallivm, 1);
 533
 534         lp_build_intrinsic(base->gallivm->builder,
 535                            "llvm.SI.export",
 536                            LLVMVoidTypeInContext(base->gallivm->context),
 537                            last_args, 9);
 538
 539 /* XXX: Look up what this function does */
 540 /*              ctx->shader->output[i].spi_sid = r600_spi_sid(&ctx->shader->output[i]);*/
 541 }
 542
 543 static void tex_fetch_args(
 544         struct lp_build_tgsi_context * bld_base,
 545         struct lp_build_emit_data * emit_data)
 546 {
 547         const struct tgsi_full_instruction * inst = emit_data->inst;
 548         LLVMValueRef ptr;
 549         LLVMValueRef offset;
 550
 551         /* WriteMask */
 552         /* XXX: should be optimized using emit_data->inst->Dst[0].Register.WriteMask*/
 553         emit_data->args[0] = lp_build_const_int32(bld_base->base.gallivm, 0xf);
 554
 555         /* Coordinates */
 556         /* XXX: Not all sample instructions need 4 address arguments. */
 557         if (inst->Instruction.Opcode == TGSI_OPCODE_TXP) {
 558                 LLVMValueRef src_w;
 559                 unsigned chan;
 560                 LLVMValueRef coords[4];
 561
 562                 emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4);
 563                 src_w = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_W);
 564
 565                 for (chan = 0; chan < 3; chan++ ) {
 566                         LLVMValueRef arg = lp_build_emit_fetch(bld_base,
 567                                                                emit_data->inst, 0, chan);
 568                         coords[chan] = lp_build_emit_llvm_binary(bld_base,
 569                                                                  TGSI_OPCODE_DIV,
 570                                                                  arg, src_w);
 571                 }
 572                 coords[3] = bld_base->base.one;
 573                 emit_data->args[1] = lp_build_gather_values(bld_base->base.gallivm,
 574                                                             coords, 4);
 575         } else
 576                 emit_data->args[1] = lp_build_emit_fetch(bld_base, emit_data->inst,
 577                                                          0, LP_CHAN_ALL);
 578
 579         /* Resource */
 580         ptr = use_sgpr(bld_base->base.gallivm, SGPR_CONST_PTR_V8I32, 4);
 581         offset = lp_build_const_int32(bld_base->base.gallivm,
 582                                   emit_data->inst->Src[1].Register.Index);
 583         emit_data->args[2] = build_indexed_load(bld_base->base.gallivm,
 584                                                 ptr, offset);
 585
 586         /* Sampler */
 587         ptr = use_sgpr(bld_base->base.gallivm, SGPR_CONST_PTR_V4I32, 2);
 588         offset = lp_build_const_int32(bld_base->base.gallivm,
 589                                   emit_data->inst->Src[1].Register.Index);
 590         emit_data->args[3] = build_indexed_load(bld_base->base.gallivm,
 591                                                 ptr, offset);
 592
 593         /* Dimensions */
 594         /* XXX: We might want to pass this information to the shader at some. */
 595 /*      emit_data->args[4] = lp_build_const_int32(bld_base->base.gallivm,
 596                                         emit_data->inst->Texture.Texture);
 597 */
 598
 599         emit_data->arg_count = 4;
 600         /* XXX: To optimize, we could use a float or v2f32, if the last bits of
 601          * the writemask are clear */
 602         emit_data->dst_type = LLVMVectorType(
 603                         LLVMFloatTypeInContext(bld_base->base.gallivm->context),
 604                         4);
 605 }
 606
 607 static const struct lp_build_tgsi_action tex_action = {
 608         .fetch_args = tex_fetch_args,
 609         .emit = lp_build_tgsi_intrinsic,
 610         .intr_name = "llvm.SI.sample"
 611 };
 612
 613
 614 int si_pipe_shader_create(
 615         struct pipe_context *ctx,
 616         struct si_pipe_shader *shader)
 617 {
 618         struct r600_context *rctx = (struct r600_context*)ctx;
 619         struct si_shader_context si_shader_ctx;
 620         struct tgsi_shader_info shader_info;
 621         struct lp_build_tgsi_context * bld_base;
 622         LLVMModuleRef mod;
 623         unsigned char * inst_bytes;
 624         unsigned inst_byte_count;
 625         unsigned i;
 626         uint32_t *ptr;
 627         bool dump;
 628
 629         dump = debug_get_bool_option("RADEON_DUMP_SHADERS", FALSE);
 630
 631         memset(&si_shader_ctx.radeon_bld, 0, sizeof(si_shader_ctx.radeon_bld));
 632         radeon_llvm_context_init(&si_shader_ctx.radeon_bld);
 633         bld_base = &si_shader_ctx.radeon_bld.soa.bld_base;
 634
 635         tgsi_scan_shader(shader->tokens, &shader_info);
 636         bld_base->info = &shader_info;
 637         bld_base->emit_fetch_funcs[TGSI_FILE_CONSTANT] = fetch_constant;
 638         bld_base->emit_epilogue = si_llvm_emit_epilogue;
 639
 640         bld_base->op_actions[TGSI_OPCODE_TEX] = tex_action;
 641         bld_base->op_actions[TGSI_OPCODE_TXP] = tex_action;
 642
 643         si_shader_ctx.radeon_bld.load_input = declare_input;
 644         si_shader_ctx.tokens = shader->tokens;
 645         tgsi_parse_init(&si_shader_ctx.parse, si_shader_ctx.tokens);
 646         si_shader_ctx.shader = shader;
 647         si_shader_ctx.type = si_shader_ctx.parse.FullHeader.Processor.Processor;
 648         si_shader_ctx.rctx = rctx;
 649
 650         shader->shader.nr_cbufs = rctx->framebuffer.nr_cbufs;
 651
 652         /* Dump TGSI code before doing TGSI->LLVM conversion in case the
 653          * conversion fails. */
 654         if (dump) {
 655                 tgsi_dump(shader->tokens, 0);
 656         }
 657
 658         if (!lp_build_tgsi_llvm(bld_base, shader->tokens)) {
 659                 fprintf(stderr, "Failed to translate shader from TGSI to LLVM\n");
 660                 return -EINVAL;
 661         }
 662
 663         radeon_llvm_finalize_module(&si_shader_ctx.radeon_bld);
 664
 665         mod = bld_base->base.gallivm->module;
 666         if (dump) {
 667                 LLVMDumpModule(mod);
 668         }
 669         radeon_llvm_compile(mod, &inst_bytes, &inst_byte_count, "SI", dump);
 670         if (dump) {
 671                 fprintf(stderr, "SI CODE:\n");
 672                 for (i = 0; i < inst_byte_count; i+=4 ) {
 673                         fprintf(stderr, "%02x%02x%02x%02x\n", inst_bytes[i + 3],
 674                                 inst_bytes[i + 2], inst_bytes[i + 1],
 675                                 inst_bytes[i]);
 676                 }
 677         }
 678
 679         shader->num_sgprs = util_le32_to_cpu(*(uint32_t*)inst_bytes);
 680         shader->num_vgprs = util_le32_to_cpu(*(uint32_t*)(inst_bytes + 4));
 681         shader->spi_ps_input_ena = util_le32_to_cpu(*(uint32_t*)(inst_bytes + 8));
 682
 683         tgsi_parse_free(&si_shader_ctx.parse);
 684
 685         /* copy new shader */
 686         si_resource_reference(&shader->bo, NULL);
 687         shader->bo = si_resource_create_custom(ctx->screen, PIPE_USAGE_IMMUTABLE,
 688                                                inst_byte_count - 12);
 689         if (shader->bo == NULL) {
 690                 return -ENOMEM;
 691         }
 692
 693         ptr = (uint32_t*)rctx->ws->buffer_map(shader->bo->cs_buf, rctx->cs, PIPE_TRANSFER_WRITE);
 694         if (0 /*R600_BIG_ENDIAN*/) {
 695                 for (i = 0; i < (inst_byte_count-12)/4; ++i) {
 696                         ptr[i] = util_bswap32(*(uint32_t*)(inst_bytes+12 + i*4));
 697                 }
 698         } else {
 699                 memcpy(ptr, inst_bytes + 12, inst_byte_count - 12);
 700         }
 701         rctx->ws->buffer_unmap(shader->bo->cs_buf);
 702
 703         free(inst_bytes);
 704
 705         return 0;
 706 }
 707
 708 void si_pipe_shader_destroy(struct pipe_context *ctx, struct si_pipe_shader *shader)
 709 {
 710         si_resource_reference(&shader->bo, NULL);
 711
 712         memset(&shader->shader,0,sizeof(struct si_shader));
 713 }