src/gallium/drivers/swr/rasterizer/jitter/blend_jit.cpp

   1 /****************************************************************************
   2 * Copyright (C) 2014-2015 Intel Corporation.   All Rights Reserved.
   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 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8 * and/or sell copies of the Software, and to permit persons to whom the
   9 * Software is furnished to do so, subject to the following conditions:
  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 NONINFRINGEMENT.  IN NO EVENT SHALL
  18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21 * IN THE SOFTWARE.
  22 *
  23 * @file blend_jit.cpp
  24 *
  25 * @brief Implementation of the blend jitter
  26 *
  27 * Notes:
  28 *
  29 ******************************************************************************/
  30 #include "jit_api.h"
  31 #include "blend_jit.h"
  32 #include "builder.h"
  33 #include "state_llvm.h"
  34
  35 #include <sstream>
  36
  37 // components with bit-widths <= the QUANTIZE_THRESHOLD will be quantized
  38 #define QUANTIZE_THRESHOLD 2
  39
  40 using namespace llvm;
  41 using namespace SwrJit;
  42
  43 //////////////////////////////////////////////////////////////////////////
  44 /// Interface to Jitting a blend shader
  45 //////////////////////////////////////////////////////////////////////////
  46 struct BlendJit : public Builder
  47 {
  48     BlendJit(JitManager* pJitMgr) : Builder(pJitMgr){};
  49
  50     template<bool Color, bool Alpha>
  51     void GenerateBlendFactor(SWR_BLEND_FACTOR factor, Value* constColor[4], Value* src[4], Value* src1[4], Value* dst[4], Value* result[4])
  52     {
  53         Value* out[4];
  54
  55         switch (factor)
  56         {
  57         case BLENDFACTOR_ONE:
  58             out[0] = out[1] = out[2] = out[3] = VIMMED1(1.0f);
  59             break;
  60         case BLENDFACTOR_SRC_COLOR:
  61             out[0] = src[0];
  62             out[1] = src[1];
  63             out[2] = src[2];
  64             out[3] = src[3];
  65             break;
  66         case BLENDFACTOR_SRC_ALPHA:
  67             out[0] = out[1] = out[2] = out[3] = src[3];
  68             break;
  69         case BLENDFACTOR_DST_ALPHA:
  70             out[0] = out[1] = out[2] = out[3] = dst[3];
  71             break;
  72         case BLENDFACTOR_DST_COLOR:
  73             out[0] = dst[0];
  74             out[1] = dst[1];
  75             out[2] = dst[2];
  76             out[3] = dst[3];
  77             break;
  78         case BLENDFACTOR_SRC_ALPHA_SATURATE:
  79             out[0] = out[1] = out[2] = VMINPS(src[3], FSUB(VIMMED1(1.0f), dst[3]));
  80             out[3] = VIMMED1(1.0f);
  81             break;
  82         case BLENDFACTOR_CONST_COLOR:
  83             out[0] = constColor[0];
  84             out[1] = constColor[1];
  85             out[2] = constColor[2];
  86             out[3] = constColor[3];
  87             break;
  88         case BLENDFACTOR_CONST_ALPHA:
  89             out[0] = out[1] = out[2] = out[3] = constColor[3];
  90             break;
  91         case BLENDFACTOR_SRC1_COLOR:
  92             out[0] = src1[0];
  93             out[1] = src1[1];
  94             out[2] = src1[2];
  95             out[3] = src1[3];
  96             break;
  97         case BLENDFACTOR_SRC1_ALPHA:
  98             out[0] = out[1] = out[2] = out[3] = src1[3];
  99             break;
 100         case BLENDFACTOR_ZERO:
 101             out[0] = out[1] = out[2] = out[3] = VIMMED1(0.0f);
 102             break;
 103         case BLENDFACTOR_INV_SRC_COLOR:
 104             out[0] = FSUB(VIMMED1(1.0f), src[0]);
 105             out[1] = FSUB(VIMMED1(1.0f), src[1]);
 106             out[2] = FSUB(VIMMED1(1.0f), src[2]);
 107             out[3] = FSUB(VIMMED1(1.0f), src[3]);
 108             break;
 109         case BLENDFACTOR_INV_SRC_ALPHA:
 110             out[0] = out[1] = out[2] = out[3] = FSUB(VIMMED1(1.0f), src[3]);
 111             break;
 112         case BLENDFACTOR_INV_DST_ALPHA:
 113             out[0] = out[1] = out[2] = out[3] = FSUB(VIMMED1(1.0f), dst[3]);
 114             break;
 115         case BLENDFACTOR_INV_DST_COLOR:
 116             out[0] = FSUB(VIMMED1(1.0f), dst[0]);
 117             out[1] = FSUB(VIMMED1(1.0f), dst[1]);
 118             out[2] = FSUB(VIMMED1(1.0f), dst[2]);
 119             out[3] = FSUB(VIMMED1(1.0f), dst[3]);
 120             break;
 121         case BLENDFACTOR_INV_CONST_COLOR:
 122             out[0] = FSUB(VIMMED1(1.0f), constColor[0]);
 123             out[1] = FSUB(VIMMED1(1.0f), constColor[1]);
 124             out[2] = FSUB(VIMMED1(1.0f), constColor[2]);
 125             out[3] = FSUB(VIMMED1(1.0f), constColor[3]);
 126             break;
 127         case BLENDFACTOR_INV_CONST_ALPHA:
 128             out[0] = out[1] = out[2] = out[3] = FSUB(VIMMED1(1.0f), constColor[3]);
 129             break;
 130         case BLENDFACTOR_INV_SRC1_COLOR:
 131             out[0] = FSUB(VIMMED1(1.0f), src1[0]);
 132             out[1] = FSUB(VIMMED1(1.0f), src1[1]);
 133             out[2] = FSUB(VIMMED1(1.0f), src1[2]);
 134             out[3] = FSUB(VIMMED1(1.0f), src1[3]);
 135             break;
 136         case BLENDFACTOR_INV_SRC1_ALPHA:
 137             out[0] = out[1] = out[2] = out[3] = FSUB(VIMMED1(1.0f), src1[3]);
 138             break;
 139         default:
 140             SWR_ASSERT(false, "Unsupported blend factor: %d", factor);
 141             out[0] = out[1] = out[2] = out[3] = VIMMED1(0.0f);
 142             break;
 143         }
 144
 145         if (Color)
 146         {
 147             result[0] = out[0];
 148             result[1] = out[1];
 149             result[2] = out[2];
 150         }
 151
 152         if (Alpha)
 153         {
 154             result[3] = out[3];
 155         }
 156     }
 157
 158     void Clamp(SWR_FORMAT format, Value* src[4])
 159     {
 160         const SWR_FORMAT_INFO& info = GetFormatInfo(format);
 161         SWR_TYPE type = info.type[0];
 162
 163         switch (type)
 164         {
 165         case SWR_TYPE_FLOAT:
 166             break;
 167
 168         case SWR_TYPE_UNORM:
 169             src[0] = VMINPS(VMAXPS(src[0], VIMMED1(0.0f)), VIMMED1(1.0f));
 170             src[1] = VMINPS(VMAXPS(src[1], VIMMED1(0.0f)), VIMMED1(1.0f));
 171             src[2] = VMINPS(VMAXPS(src[2], VIMMED1(0.0f)), VIMMED1(1.0f));
 172             src[3] = VMINPS(VMAXPS(src[3], VIMMED1(0.0f)), VIMMED1(1.0f));
 173             break;
 174
 175         case SWR_TYPE_SNORM:
 176             src[0] = VMINPS(VMAXPS(src[0], VIMMED1(-1.0f)), VIMMED1(1.0f));
 177             src[1] = VMINPS(VMAXPS(src[1], VIMMED1(-1.0f)), VIMMED1(1.0f));
 178             src[2] = VMINPS(VMAXPS(src[2], VIMMED1(-1.0f)), VIMMED1(1.0f));
 179             src[3] = VMINPS(VMAXPS(src[3], VIMMED1(-1.0f)), VIMMED1(1.0f));
 180             break;
 181
 182         default: SWR_ASSERT(false, "Unsupport format type: %d", type);
 183         }
 184     }
 185
 186     void ApplyDefaults(SWR_FORMAT format, Value* src[4])
 187     {
 188         const SWR_FORMAT_INFO& info = GetFormatInfo(format);
 189
 190         bool valid[] = { false, false, false, false };
 191         for (uint32_t c = 0; c < info.numComps; ++c)
 192         {
 193             valid[info.swizzle[c]] = true;
 194         }
 195
 196         for (uint32_t c = 0; c < 4; ++c)
 197         {
 198             if (!valid[c])
 199             {
 200                 src[c] = BITCAST(VIMMED1((int)info.defaults[c]), mSimdFP32Ty);
 201             }
 202         }
 203     }
 204
 205     void ApplyUnusedDefaults(SWR_FORMAT format, Value* src[4])
 206     {
 207         const SWR_FORMAT_INFO& info = GetFormatInfo(format);
 208
 209         for (uint32_t c = 0; c < info.numComps; ++c)
 210         {
 211             if (info.type[c] == SWR_TYPE_UNUSED)
 212             {
 213                 src[info.swizzle[c]] = BITCAST(VIMMED1((int)info.defaults[info.swizzle[c]]), mSimdFP32Ty);
 214             }
 215         }
 216     }
 217
 218     void Quantize(SWR_FORMAT format, Value* src[4])
 219     {
 220         const SWR_FORMAT_INFO& info = GetFormatInfo(format);
 221         for (uint32_t c = 0; c < info.numComps; ++c)
 222         {
 223             if (info.bpc[c] <= QUANTIZE_THRESHOLD)
 224             {
 225                 uint32_t swizComp = info.swizzle[c];
 226                 float factor = (float)((1 << info.bpc[c]) - 1);
 227                 switch (info.type[c])
 228                 {
 229                 case SWR_TYPE_UNORM:
 230                     src[swizComp] = FADD(FMUL(src[swizComp], VIMMED1(factor)), VIMMED1(0.5f));
 231                     src[swizComp] = VROUND(src[swizComp], C(_MM_FROUND_TO_ZERO));
 232                     src[swizComp] = FMUL(src[swizComp], VIMMED1(1.0f /factor));
 233                     break;
 234                 default: SWR_ASSERT(false, "Unsupported format type: %d", info.type[c]);
 235                 }
 236             }
 237         }
 238     }
 239
 240     template<bool Color, bool Alpha>
 241     void BlendFunc(SWR_BLEND_OP blendOp, Value* src[4], Value* srcFactor[4], Value* dst[4], Value* dstFactor[4], Value* result[4])
 242     {
 243         Value* out[4];
 244         Value* srcBlend[4];
 245         Value* dstBlend[4];
 246         for (uint32_t i = 0; i < 4; ++i)
 247         {
 248             srcBlend[i] = FMUL(src[i], srcFactor[i]);
 249             dstBlend[i] = FMUL(dst[i], dstFactor[i]);
 250         }
 251
 252         switch (blendOp)
 253         {
 254         case BLENDOP_ADD:
 255             out[0] = FADD(srcBlend[0], dstBlend[0]);
 256             out[1] = FADD(srcBlend[1], dstBlend[1]);
 257             out[2] = FADD(srcBlend[2], dstBlend[2]);
 258             out[3] = FADD(srcBlend[3], dstBlend[3]);
 259             break;
 260
 261         case BLENDOP_SUBTRACT:
 262             out[0] = FSUB(srcBlend[0], dstBlend[0]);
 263             out[1] = FSUB(srcBlend[1], dstBlend[1]);
 264             out[2] = FSUB(srcBlend[2], dstBlend[2]);
 265             out[3] = FSUB(srcBlend[3], dstBlend[3]);
 266             break;
 267
 268         case BLENDOP_REVSUBTRACT:
 269             out[0] = FSUB(dstBlend[0], srcBlend[0]);
 270             out[1] = FSUB(dstBlend[1], srcBlend[1]);
 271             out[2] = FSUB(dstBlend[2], srcBlend[2]);
 272             out[3] = FSUB(dstBlend[3], srcBlend[3]);
 273             break;
 274
 275         case BLENDOP_MIN:
 276             out[0] = VMINPS(src[0], dst[0]);
 277             out[1] = VMINPS(src[1], dst[1]);
 278             out[2] = VMINPS(src[2], dst[2]);
 279             out[3] = VMINPS(src[3], dst[3]);
 280             break;
 281
 282         case BLENDOP_MAX:
 283             out[0] = VMAXPS(src[0], dst[0]);
 284             out[1] = VMAXPS(src[1], dst[1]);
 285             out[2] = VMAXPS(src[2], dst[2]);
 286             out[3] = VMAXPS(src[3], dst[3]);
 287             break;
 288
 289         default:
 290             SWR_ASSERT(false, "Unsupported blend operation: %d", blendOp);
 291             out[0] = out[1] = out[2] = out[3] = VIMMED1(0.0f);
 292             break;
 293         }
 294
 295         if (Color)
 296         {
 297             result[0] = out[0];
 298             result[1] = out[1];
 299             result[2] = out[2];
 300         }
 301
 302         if (Alpha)
 303         {
 304             result[3] = out[3];
 305         }
 306     }
 307
 308     void LogicOpFunc(SWR_LOGIC_OP logicOp, Value* src[4], Value* dst[4], Value* result[4])
 309     {
 310         // Op: (s == PS output, d = RT contents)
 311         switch(logicOp)
 312         {
 313         case LOGICOP_CLEAR:
 314             result[0] = VIMMED1(0);
 315             result[1] = VIMMED1(0);
 316             result[2] = VIMMED1(0);
 317             result[3] = VIMMED1(0);
 318             break;
 319
 320         case LOGICOP_NOR:
 321             // ~(s | d)
 322             result[0] = XOR(OR(src[0], dst[0]), VIMMED1(0xFFFFFFFF));
 323             result[1] = XOR(OR(src[1], dst[1]), VIMMED1(0xFFFFFFFF));
 324             result[2] = XOR(OR(src[2], dst[2]), VIMMED1(0xFFFFFFFF));
 325             result[3] = XOR(OR(src[3], dst[3]), VIMMED1(0xFFFFFFFF));
 326             break;
 327
 328         case LOGICOP_AND_INVERTED:
 329             // ~s & d
 330             // todo: use avx andnot instr when I can find the intrinsic to call
 331             result[0] = AND(XOR(src[0], VIMMED1(0xFFFFFFFF)), dst[0]);
 332             result[1] = AND(XOR(src[1], VIMMED1(0xFFFFFFFF)), dst[1]);
 333             result[2] = AND(XOR(src[2], VIMMED1(0xFFFFFFFF)), dst[2]);
 334             result[3] = AND(XOR(src[3], VIMMED1(0xFFFFFFFF)), dst[3]);
 335             break;
 336
 337         case LOGICOP_COPY_INVERTED:
 338             // ~s
 339             result[0] = XOR(src[0], VIMMED1(0xFFFFFFFF));
 340             result[1] = XOR(src[1], VIMMED1(0xFFFFFFFF));
 341             result[2] = XOR(src[2], VIMMED1(0xFFFFFFFF));
 342             result[3] = XOR(src[3], VIMMED1(0xFFFFFFFF));
 343             break;
 344
 345         case LOGICOP_AND_REVERSE:
 346             // s & ~d
 347             // todo: use avx andnot instr when I can find the intrinsic to call
 348             result[0] = AND(XOR(dst[0], VIMMED1(0xFFFFFFFF)), src[0]);
 349             result[1] = AND(XOR(dst[1], VIMMED1(0xFFFFFFFF)), src[1]);
 350             result[2] = AND(XOR(dst[2], VIMMED1(0xFFFFFFFF)), src[2]);
 351             result[3] = AND(XOR(dst[3], VIMMED1(0xFFFFFFFF)), src[3]);
 352             break;
 353
 354         case LOGICOP_INVERT:
 355             // ~d
 356             result[0] = XOR(dst[0], VIMMED1(0xFFFFFFFF));
 357             result[1] = XOR(dst[1], VIMMED1(0xFFFFFFFF));
 358             result[2] = XOR(dst[2], VIMMED1(0xFFFFFFFF));
 359             result[3] = XOR(dst[3], VIMMED1(0xFFFFFFFF));
 360             break;
 361
 362         case LOGICOP_XOR:
 363             // s ^ d
 364             result[0] = XOR(src[0], dst[0]);
 365             result[1] = XOR(src[1], dst[1]);
 366             result[2] = XOR(src[2], dst[2]);
 367             result[3] = XOR(src[3], dst[3]);
 368             break;
 369
 370         case LOGICOP_NAND:
 371             // ~(s & d)
 372             result[0] = XOR(AND(src[0], dst[0]), VIMMED1(0xFFFFFFFF));
 373             result[1] = XOR(AND(src[1], dst[1]), VIMMED1(0xFFFFFFFF));
 374             result[2] = XOR(AND(src[2], dst[2]), VIMMED1(0xFFFFFFFF));
 375             result[3] = XOR(AND(src[3], dst[3]), VIMMED1(0xFFFFFFFF));
 376             break;
 377
 378         case LOGICOP_AND:
 379             // s & d
 380             result[0] = AND(src[0], dst[0]);
 381             result[1] = AND(src[1], dst[1]);
 382             result[2] = AND(src[2], dst[2]);
 383             result[3] = AND(src[3], dst[3]);
 384             break;
 385
 386         case LOGICOP_EQUIV:
 387             // ~(s ^ d)
 388             result[0] = XOR(XOR(src[0], dst[0]), VIMMED1(0xFFFFFFFF));
 389             result[1] = XOR(XOR(src[1], dst[1]), VIMMED1(0xFFFFFFFF));
 390             result[2] = XOR(XOR(src[2], dst[2]), VIMMED1(0xFFFFFFFF));
 391             result[3] = XOR(XOR(src[3], dst[3]), VIMMED1(0xFFFFFFFF));
 392             break;
 393
 394         case LOGICOP_NOOP:
 395             result[0] = dst[0];
 396             result[1] = dst[1];
 397             result[2] = dst[2];
 398             result[3] = dst[3];
 399             break;
 400
 401         case LOGICOP_OR_INVERTED:
 402             // ~s | d
 403             result[0] = OR(XOR(src[0], VIMMED1(0xFFFFFFFF)), dst[0]);
 404             result[1] = OR(XOR(src[1], VIMMED1(0xFFFFFFFF)), dst[1]);
 405             result[2] = OR(XOR(src[2], VIMMED1(0xFFFFFFFF)), dst[2]);
 406             result[3] = OR(XOR(src[3], VIMMED1(0xFFFFFFFF)), dst[3]);
 407             break;
 408
 409         case LOGICOP_COPY:
 410             result[0] = src[0];
 411             result[1] = src[1];
 412             result[2] = src[2];
 413             result[3] = src[3];
 414             break;
 415
 416         case LOGICOP_OR_REVERSE:
 417             // s | ~d
 418             result[0] = OR(XOR(dst[0], VIMMED1(0xFFFFFFFF)), src[0]);
 419             result[1] = OR(XOR(dst[1], VIMMED1(0xFFFFFFFF)), src[1]);
 420             result[2] = OR(XOR(dst[2], VIMMED1(0xFFFFFFFF)), src[2]);
 421             result[3] = OR(XOR(dst[3], VIMMED1(0xFFFFFFFF)), src[3]);
 422             break;
 423
 424         case LOGICOP_OR:
 425             // s | d
 426             result[0] = OR(src[0], dst[0]);
 427             result[1] = OR(src[1], dst[1]);
 428             result[2] = OR(src[2], dst[2]);
 429             result[3] = OR(src[3], dst[3]);
 430             break;
 431
 432         case LOGICOP_SET:
 433             result[0] = VIMMED1(0xFFFFFFFF);
 434             result[1] = VIMMED1(0xFFFFFFFF);
 435             result[2] = VIMMED1(0xFFFFFFFF);
 436             result[3] = VIMMED1(0xFFFFFFFF);
 437             break;
 438
 439         default:
 440             SWR_ASSERT(false, "Unsupported logic operation: %d", logicOp);
 441             result[0] = result[1] = result[2] = result[3] = VIMMED1(0.0f);
 442             break;
 443         }
 444     }
 445
 446     void AlphaTest(const BLEND_COMPILE_STATE& state, Value* pBlendState, Value* pAlpha, Value* ppMask)
 447     {
 448         // load uint32_t reference
 449         Value* pRef = VBROADCAST(LOAD(pBlendState, { 0, SWR_BLEND_STATE_alphaTestReference }));
 450
 451         Value* pTest = nullptr;
 452         if (state.alphaTestFormat == ALPHA_TEST_UNORM8)
 453         {
 454             // convert float alpha to unorm8
 455             Value* pAlphaU8 = FMUL(pAlpha, VIMMED1(256.0f));
 456             pAlphaU8 = FP_TO_UI(pAlphaU8, mSimdInt32Ty);
 457
 458             // compare
 459             switch (state.alphaTestFunction)
 460             {
 461             case ZFUNC_ALWAYS:  pTest = VIMMED1(true); break;
 462             case ZFUNC_NEVER:   pTest = VIMMED1(false); break;
 463             case ZFUNC_LT:      pTest = ICMP_ULT(pAlphaU8, pRef); break;
 464             case ZFUNC_EQ:      pTest = ICMP_EQ(pAlphaU8, pRef); break;
 465             case ZFUNC_LE:      pTest = ICMP_ULE(pAlphaU8, pRef); break;
 466             case ZFUNC_GT:      pTest = ICMP_UGT(pAlphaU8, pRef); break;
 467             case ZFUNC_NE:      pTest = ICMP_NE(pAlphaU8, pRef); break;
 468             case ZFUNC_GE:      pTest = ICMP_UGE(pAlphaU8, pRef); break;
 469             default:
 470                 SWR_ASSERT(false, "Invalid alpha test function");
 471                 break;
 472             }
 473         }
 474         else
 475         {
 476             // cast ref to float
 477             pRef = BITCAST(pRef, mSimdFP32Ty);
 478
 479             // compare
 480             switch (state.alphaTestFunction)
 481             {
 482             case ZFUNC_ALWAYS:  pTest = VIMMED1(true); break;
 483             case ZFUNC_NEVER:   pTest = VIMMED1(false); break;
 484             case ZFUNC_LT:      pTest = FCMP_OLT(pAlpha, pRef); break;
 485             case ZFUNC_EQ:      pTest = FCMP_OEQ(pAlpha, pRef); break;
 486             case ZFUNC_LE:      pTest = FCMP_OLE(pAlpha, pRef); break;
 487             case ZFUNC_GT:      pTest = FCMP_OGT(pAlpha, pRef); break;
 488             case ZFUNC_NE:      pTest = FCMP_ONE(pAlpha, pRef); break;
 489             case ZFUNC_GE:      pTest = FCMP_OGE(pAlpha, pRef); break;
 490             default:
 491                 SWR_ASSERT(false, "Invalid alpha test function");
 492                 break;
 493             }
 494         }
 495
 496         // load current mask
 497         Value* pMask = LOAD(ppMask);
 498
 499         // convert to int1 mask
 500         pMask = MASK(pMask);
 501
 502         // and with alpha test result
 503         pMask = AND(pMask, pTest);
 504
 505         // convert back to vector mask
 506         pMask = VMASK(pMask);
 507
 508         // store new mask
 509         STORE(pMask, ppMask);
 510     }
 511
 512     Function* Create(const BLEND_COMPILE_STATE& state)
 513     {
 514         static std::size_t jitNum = 0;
 515
 516         std::stringstream fnName("BlendShader", std::ios_base::in | std::ios_base::out | std::ios_base::ate);
 517         fnName << jitNum++;
 518
 519         // blend function signature
 520         //typedef void(*PFN_BLEND_JIT_FUNC)(const SWR_BLEND_STATE*, simdvector&, simdvector&, uint32_t, BYTE*, simdvector&, simdscalari*, simdscalari*);
 521
 522         std::vector<Type*> args{
 523             PointerType::get(Gen_SWR_BLEND_STATE(JM()), 0), // SWR_BLEND_STATE*
 524             PointerType::get(mSimdFP32Ty, 0),               // simdvector& src
 525             PointerType::get(mSimdFP32Ty, 0),               // simdvector& src1
 526             Type::getInt32Ty(JM()->mContext),               // sampleNum
 527             PointerType::get(mSimdFP32Ty, 0),               // uint8_t* pDst
 528             PointerType::get(mSimdFP32Ty, 0),               // simdvector& result
 529             PointerType::get(mSimdInt32Ty, 0),              // simdscalari* oMask
 530             PointerType::get(mSimdInt32Ty, 0),              // simdscalari* pMask
 531         };
 532
 533         FunctionType* fTy = FunctionType::get(IRB()->getVoidTy(), args, false);
 534         Function* blendFunc = Function::Create(fTy, GlobalValue::ExternalLinkage, fnName.str(), JM()->mpCurrentModule);
 535
 536         BasicBlock* entry = BasicBlock::Create(JM()->mContext, "entry", blendFunc);
 537
 538         IRB()->SetInsertPoint(entry);
 539
 540         // arguments
 541         auto argitr = blendFunc->getArgumentList().begin();
 542         Value* pBlendState = &*argitr++;
 543         pBlendState->setName("pBlendState");
 544         Value* pSrc = &*argitr++;
 545         pSrc->setName("src");
 546         Value* pSrc1 = &*argitr++;
 547         pSrc1->setName("src1");
 548         Value* sampleNum = &*argitr++;
 549         sampleNum->setName("sampleNum");
 550         Value* pDst = &*argitr++;
 551         pDst->setName("pDst");
 552         Value* pResult = &*argitr++;
 553         pResult->setName("result");
 554         Value* ppoMask = &*argitr++;
 555         ppoMask->setName("ppoMask");
 556         Value* ppMask = &*argitr++;
 557         ppMask->setName("pMask");
 558
 559         static_assert(KNOB_COLOR_HOT_TILE_FORMAT == R32G32B32A32_FLOAT, "Unsupported hot tile format");
 560         Value* dst[4];
 561         Value* constantColor[4];
 562         Value* src[4];
 563         Value* src1[4];
 564         Value* result[4];
 565         for (uint32_t i = 0; i < 4; ++i)
 566         {
 567             // load hot tile
 568             dst[i] = LOAD(pDst, { i });
 569
 570             // load constant color
 571             constantColor[i] = VBROADCAST(LOAD(pBlendState, { 0, SWR_BLEND_STATE_constantColor, i }));
 572
 573             // load src
 574             src[i] = LOAD(pSrc, { i });
 575
 576             // load src1
 577             src1[i] = LOAD(pSrc1, { i });
 578         }
 579         Value* currentMask = VIMMED1(-1);
 580         if (state.desc.alphaToCoverageEnable)
 581         {
 582             Value* pClampedSrc = FCLAMP(src[3], 0.0f, 1.0f);
 583             uint32_t bits = (1 << state.desc.numSamples) - 1;
 584             currentMask = FMUL(pClampedSrc, VBROADCAST(C((float)bits)));
 585             currentMask = FP_TO_SI(FADD(currentMask, VIMMED1(0.5f)), mSimdInt32Ty);
 586         }
 587
 588         // alpha test
 589         if (state.desc.alphaTestEnable)
 590         {
 591             AlphaTest(state, pBlendState, src[3], ppMask);
 592         }
 593
 594         // color blend
 595         if (state.blendState.blendEnable)
 596         {
 597             // clamp sources
 598             Clamp(state.format, src);
 599             Clamp(state.format, src1);
 600             Clamp(state.format, dst);
 601             Clamp(state.format, constantColor);
 602
 603             // apply defaults to hottile contents to take into account missing components
 604             ApplyDefaults(state.format, dst);
 605
 606             // Force defaults for unused 'X' components
 607             ApplyUnusedDefaults(state.format, dst);
 608
 609             // Quantize low precision components
 610             Quantize(state.format, dst);
 611
 612             // special case clamping for R11G11B10_float which has no sign bit
 613             if (state.format == R11G11B10_FLOAT)
 614             {
 615                 dst[0] = VMAXPS(dst[0], VIMMED1(0.0f));
 616                 dst[1] = VMAXPS(dst[1], VIMMED1(0.0f));
 617                 dst[2] = VMAXPS(dst[2], VIMMED1(0.0f));
 618                 dst[3] = VMAXPS(dst[3], VIMMED1(0.0f));
 619             }
 620
 621             Value* srcFactor[4];
 622             Value* dstFactor[4];
 623             if (state.desc.independentAlphaBlendEnable)
 624             {
 625                 GenerateBlendFactor<true, false>(state.blendState.sourceBlendFactor, constantColor, src, src1, dst, srcFactor);
 626                 GenerateBlendFactor<false, true>(state.blendState.sourceAlphaBlendFactor, constantColor, src, src1, dst, srcFactor);
 627
 628                 GenerateBlendFactor<true, false>(state.blendState.destBlendFactor, constantColor, src, src1, dst, dstFactor);
 629                 GenerateBlendFactor<false, true>(state.blendState.destAlphaBlendFactor, constantColor, src, src1, dst, dstFactor);
 630
 631                 BlendFunc<true, false>(state.blendState.colorBlendFunc, src, srcFactor, dst, dstFactor, result);
 632                 BlendFunc<false, true>(state.blendState.alphaBlendFunc, src, srcFactor, dst, dstFactor, result);
 633             }
 634             else
 635             {
 636                 GenerateBlendFactor<true, true>(state.blendState.sourceBlendFactor, constantColor, src, src1, dst, srcFactor);
 637                 GenerateBlendFactor<true, true>(state.blendState.destBlendFactor, constantColor, src, src1, dst, dstFactor);
 638
 639                 BlendFunc<true, true>(state.blendState.colorBlendFunc, src, srcFactor, dst, dstFactor, result);
 640             }
 641
 642             // store results out
 643             for (uint32_t i = 0; i < 4; ++i)
 644             {
 645                 STORE(result[i], pResult, { i });
 646             }
 647         }
 648
 649         if(state.blendState.logicOpEnable)
 650         {
 651             const SWR_FORMAT_INFO& info = GetFormatInfo(state.format);
 652             SWR_ASSERT(info.type[0] == SWR_TYPE_UINT);
 653             Value* vMask[4];
 654             for(uint32_t i = 0; i < 4; i++)
 655             {
 656                 switch(info.bpc[i])
 657                 {
 658                 case 0: vMask[i] = VIMMED1(0x00000000); break;
 659                 case 2: vMask[i] = VIMMED1(0x00000003); break;
 660                 case 5: vMask[i] = VIMMED1(0x0000001F); break;
 661                 case 6: vMask[i] = VIMMED1(0x0000003F); break;
 662                 case 8: vMask[i] = VIMMED1(0x000000FF); break;
 663                 case 10: vMask[i] = VIMMED1(0x000003FF); break;
 664                 case 11: vMask[i] = VIMMED1(0x000007FF); break;
 665                 case 16: vMask[i] = VIMMED1(0x0000FFFF); break;
 666                 case 24: vMask[i] = VIMMED1(0x00FFFFFF); break;
 667                 case 32: vMask[i] = VIMMED1(0xFFFFFFFF); break;
 668                 default:
 669                     vMask[i] = VIMMED1(0x0);
 670                     SWR_ASSERT(0, "Unsupported bpc for logic op\n");
 671                     break;
 672                 }
 673                 src[i] = BITCAST(src[i], mSimdInt32Ty);//, vMask[i]);
 674                 dst[i] = BITCAST(dst[i], mSimdInt32Ty);
 675             }
 676
 677             LogicOpFunc(state.blendState.logicOpFunc, src, dst, result);
 678
 679             // store results out
 680             for(uint32_t i = 0; i < 4; ++i)
 681             {
 682                 // clear upper bits from PS output not in RT format after doing logic op
 683                 result[i] = AND(result[i], vMask[i]);
 684
 685                 STORE(BITCAST(result[i], mSimdFP32Ty), pResult, {i});
 686             }
 687         }
 688
 689         if(state.desc.oMaskEnable)
 690         {
 691             assert(!(state.desc.alphaToCoverageEnable));
 692             // load current mask
 693             Value* oMask = LOAD(ppoMask);
 694             Value* sampleMasked = VBROADCAST(SHL(C(1), sampleNum));
 695             oMask = AND(oMask, sampleMasked);
 696             currentMask = AND(oMask, currentMask);
 697         }
 698
 699         if(state.desc.sampleMaskEnable)
 700         {
 701             Value* sampleMask = LOAD(pBlendState, { 0, SWR_BLEND_STATE_sampleMask});
 702             Value* sampleMasked = SHL(C(1), sampleNum);
 703             sampleMask = AND(sampleMask, sampleMasked);
 704             sampleMask = VBROADCAST(ICMP_SGT(sampleMask, C(0)));
 705             sampleMask = S_EXT(sampleMask, mSimdInt32Ty);
 706             currentMask = AND(sampleMask, currentMask);
 707         }
 708
 709         if (state.desc.alphaToCoverageEnable)
 710         {
 711             Value* sampleMasked = SHL(C(1), sampleNum);
 712             currentMask = AND(currentMask, VBROADCAST(sampleMasked));
 713         }
 714
 715         if(state.desc.sampleMaskEnable || state.desc.alphaToCoverageEnable ||
 716            state.desc.oMaskEnable)
 717         {
 718             // load current mask
 719             Value* pMask = LOAD(ppMask);
 720             currentMask = S_EXT(ICMP_SGT(currentMask, VBROADCAST(C(0))), mSimdInt32Ty);
 721             Value* outputMask = AND(pMask, currentMask);
 722             // store new mask
 723             STORE(outputMask, GEP(ppMask, C(0)));
 724         }
 725
 726         RET_VOID();
 727
 728         JitManager::DumpToFile(blendFunc, "");
 729
 730         ::FunctionPassManager passes(JM()->mpCurrentModule);
 731
 732         passes.add(createBreakCriticalEdgesPass());
 733         passes.add(createCFGSimplificationPass());
 734         passes.add(createEarlyCSEPass());
 735         passes.add(createPromoteMemoryToRegisterPass());
 736         passes.add(createCFGSimplificationPass());
 737         passes.add(createEarlyCSEPass());
 738         passes.add(createInstructionCombiningPass());
 739         passes.add(createInstructionSimplifierPass());
 740         passes.add(createConstantPropagationPass());
 741         passes.add(createSCCPPass());
 742         passes.add(createAggressiveDCEPass());
 743
 744         passes.run(*blendFunc);
 745
 746         JitManager::DumpToFile(blendFunc, "optimized");
 747
 748         return blendFunc;
 749     }
 750 };
 751
 752 //////////////////////////////////////////////////////////////////////////
 753 /// @brief JITs from fetch shader IR
 754 /// @param hJitMgr - JitManager handle
 755 /// @param func   - LLVM function IR
 756 /// @return PFN_FETCH_FUNC - pointer to fetch code
 757 PFN_BLEND_JIT_FUNC JitBlendFunc(HANDLE hJitMgr, const HANDLE hFunc)
 758 {
 759     const llvm::Function *func = (const llvm::Function*)hFunc;
 760     JitManager* pJitMgr = reinterpret_cast<JitManager*>(hJitMgr);
 761     PFN_BLEND_JIT_FUNC pfnBlend;
 762     pfnBlend = (PFN_BLEND_JIT_FUNC)(pJitMgr->mpExec->getFunctionAddress(func->getName().str()));
 763     // MCJIT finalizes modules the first time you JIT code from them. After finalized, you cannot add new IR to the module
 764     pJitMgr->mIsModuleFinalized = true;
 765
 766     return pfnBlend;
 767 }
 768
 769 //////////////////////////////////////////////////////////////////////////
 770 /// @brief JIT compiles blend shader
 771 /// @param hJitMgr - JitManager handle
 772 /// @param state   - blend state to build function from
 773 extern "C" PFN_BLEND_JIT_FUNC JITCALL JitCompileBlend(HANDLE hJitMgr, const BLEND_COMPILE_STATE& state)
 774 {
 775     JitManager* pJitMgr = reinterpret_cast<JitManager*>(hJitMgr);
 776
 777     pJitMgr->SetupNewModule();
 778
 779     BlendJit theJit(pJitMgr);
 780     HANDLE hFunc = theJit.Create(state);
 781
 782     return JitBlendFunc(hJitMgr, hFunc);
 783 }