2
0

ShaderProgramBinaryDumpMain.cpp 6.0 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203
  1. // Copyright (C) 2009-2022, Panagiotis Christopoulos Charitos and contributors.
  2. // All rights reserved.
  3. // Code licensed under the BSD License.
  4. // http://www.anki3d.org/LICENSE
  5. #include <AnKi/ShaderCompiler/ShaderProgramCompiler.h>
  6. #include <AnKi/ShaderCompiler/MaliOfflineCompiler.h>
  7. using namespace anki;
  8. static const char* USAGE = R"(Dump the shader binary to stdout
  9. Usage: %s [options] input_shader_program_binary
  10. Options:
  11. -stats : Print performance statistics for all shaders. By default it doesn't
  12. )";
  13. static Error parseCommandLineArgs(int argc, char** argv, Bool& dumpStats, StringAuto& filename)
  14. {
  15. // Parse config
  16. if(argc < 2)
  17. {
  18. return Error::USER_DATA;
  19. }
  20. dumpStats = false;
  21. filename = argv[argc - 1];
  22. for(I i = 1; i < argc - 1; i++)
  23. {
  24. if(strcmp(argv[i], "-stats") == 0)
  25. {
  26. dumpStats = true;
  27. }
  28. }
  29. return Error::NONE;
  30. }
  31. ANKI_USE_RESULT Error dumpStats(const ShaderProgramBinary& bin)
  32. {
  33. HeapAllocator<U8> alloc(allocAligned, nullptr);
  34. printf("\nMali offline compiler stats:\n");
  35. fflush(stdout);
  36. class Stats
  37. {
  38. public:
  39. F64 m_fma;
  40. F64 m_cvt;
  41. F64 m_sfu;
  42. F64 m_loadStore;
  43. F64 m_varying;
  44. F64 m_texture;
  45. F64 m_workRegisters;
  46. F64 m_fp16ArithmeticPercentage;
  47. Stats(F64 v)
  48. {
  49. m_fma = m_cvt = m_sfu = m_loadStore = m_varying = m_texture = m_workRegisters = m_fp16ArithmeticPercentage =
  50. v;
  51. }
  52. };
  53. class StageStats
  54. {
  55. public:
  56. Stats m_avgStats{0.0};
  57. Stats m_maxStats{-1.0};
  58. Stats m_minStats{MAX_F64};
  59. U32 m_spillingCount = 0;
  60. U32 m_count = 0;
  61. };
  62. Array<StageStats, U32(ShaderType::COUNT)> allStats;
  63. for(const ShaderProgramBinaryVariant& variant : bin.m_variants)
  64. {
  65. for(ShaderType shaderType : EnumIterable<ShaderType>())
  66. {
  67. if(variant.m_codeBlockIndices[shaderType] == MAX_U32)
  68. {
  69. continue;
  70. }
  71. const ShaderProgramBinaryCodeBlock& codeBlock = bin.m_codeBlocks[variant.m_codeBlockIndices[shaderType]];
  72. MaliOfflineCompilerOut maliocOut;
  73. const Error err = runMaliOfflineCompiler(ANKI_SOURCE_DIRECTORY "/ThirdParty/Bin/MaliOfflineCompiler/malioc",
  74. codeBlock.m_binary, shaderType, alloc, maliocOut);
  75. if(err)
  76. {
  77. ANKI_LOGE("Mali offline compiler failed");
  78. return Error::FUNCTION_FAILED;
  79. }
  80. // Appends stats
  81. StageStats& stage = allStats[shaderType];
  82. if(maliocOut.m_spilling)
  83. {
  84. ++stage.m_spillingCount;
  85. }
  86. ++stage.m_count;
  87. stage.m_avgStats.m_fma += maliocOut.m_fma;
  88. stage.m_avgStats.m_cvt += maliocOut.m_cvt;
  89. stage.m_avgStats.m_sfu += maliocOut.m_sfu;
  90. stage.m_avgStats.m_loadStore += maliocOut.m_loadStore;
  91. stage.m_avgStats.m_varying += maliocOut.m_varying;
  92. stage.m_avgStats.m_texture += maliocOut.m_texture;
  93. stage.m_avgStats.m_workRegisters += maliocOut.m_workRegisters;
  94. stage.m_avgStats.m_fp16ArithmeticPercentage += maliocOut.m_fp16ArithmeticPercentage;
  95. stage.m_maxStats.m_fma = max<F64>(stage.m_maxStats.m_fma, maliocOut.m_fma);
  96. stage.m_maxStats.m_cvt = max<F64>(stage.m_maxStats.m_cvt, maliocOut.m_cvt);
  97. stage.m_maxStats.m_sfu = max<F64>(stage.m_maxStats.m_sfu, maliocOut.m_sfu);
  98. stage.m_maxStats.m_loadStore = max<F64>(stage.m_maxStats.m_loadStore, maliocOut.m_loadStore);
  99. stage.m_maxStats.m_varying = max<F64>(stage.m_maxStats.m_varying, maliocOut.m_varying);
  100. stage.m_maxStats.m_texture = max<F64>(stage.m_maxStats.m_texture, maliocOut.m_texture);
  101. stage.m_maxStats.m_workRegisters = max<F64>(stage.m_maxStats.m_workRegisters, maliocOut.m_workRegisters);
  102. stage.m_maxStats.m_fp16ArithmeticPercentage =
  103. max<F64>(stage.m_maxStats.m_fp16ArithmeticPercentage, maliocOut.m_fp16ArithmeticPercentage);
  104. stage.m_minStats.m_fma = min<F64>(stage.m_minStats.m_fma, maliocOut.m_fma);
  105. stage.m_minStats.m_cvt = min<F64>(stage.m_minStats.m_cvt, maliocOut.m_cvt);
  106. stage.m_minStats.m_sfu = min<F64>(stage.m_minStats.m_sfu, maliocOut.m_sfu);
  107. stage.m_minStats.m_loadStore = min<F64>(stage.m_minStats.m_loadStore, maliocOut.m_loadStore);
  108. stage.m_minStats.m_varying = min<F64>(stage.m_minStats.m_varying, maliocOut.m_varying);
  109. stage.m_minStats.m_texture = min<F64>(stage.m_minStats.m_texture, maliocOut.m_texture);
  110. stage.m_minStats.m_workRegisters = min<F64>(stage.m_minStats.m_workRegisters, maliocOut.m_workRegisters);
  111. stage.m_minStats.m_fp16ArithmeticPercentage =
  112. min<F64>(stage.m_minStats.m_fp16ArithmeticPercentage, maliocOut.m_fp16ArithmeticPercentage);
  113. }
  114. }
  115. for(ShaderType shaderType : EnumIterable<ShaderType>())
  116. {
  117. const StageStats& stage = allStats[shaderType];
  118. if(stage.m_count == 0)
  119. {
  120. continue;
  121. }
  122. printf("Stage %u\n", U32(shaderType));
  123. printf("\tSpilling count %u\n", stage.m_spillingCount);
  124. const Stats& avg = stage.m_avgStats;
  125. printf("\tAvarage: Regs %f FMA %f CVT %f SFU %f LS %f VAR %f TEX %f FP16 %f%%\n",
  126. avg.m_workRegisters / F64(stage.m_count), avg.m_fma / F64(stage.m_count), avg.m_cvt / F64(stage.m_count),
  127. avg.m_sfu / F64(stage.m_count), avg.m_loadStore / F64(stage.m_count), avg.m_varying / F64(stage.m_count),
  128. avg.m_texture / F64(stage.m_count), avg.m_fp16ArithmeticPercentage / F64(stage.m_count));
  129. const Stats& maxs = stage.m_maxStats;
  130. printf("\tMax: Regs %f FMA %f CVT %f SFU %f LS %f VAR %f TEX %f FP16 %f%%\n", maxs.m_workRegisters, maxs.m_fma,
  131. maxs.m_cvt, maxs.m_sfu, maxs.m_loadStore, maxs.m_varying, maxs.m_texture,
  132. maxs.m_fp16ArithmeticPercentage);
  133. }
  134. return Error::NONE;
  135. }
  136. Error dump(CString fname, Bool bDumpStats)
  137. {
  138. HeapAllocator<U8> alloc(allocAligned, nullptr);
  139. ShaderProgramBinaryWrapper binw(alloc);
  140. ANKI_CHECK(binw.deserializeFromFile(fname));
  141. StringAuto txt(alloc);
  142. dumpShaderProgramBinary(binw.getBinary(), txt);
  143. printf("%s\n", txt.cstr());
  144. if(bDumpStats)
  145. {
  146. ANKI_CHECK(dumpStats(binw.getBinary()));
  147. }
  148. return Error::NONE;
  149. }
  150. int main(int argc, char** argv)
  151. {
  152. StringAuto filename(HeapAllocator<U8>(allocAligned, nullptr));
  153. Bool dumpStats;
  154. if(parseCommandLineArgs(argc, argv, dumpStats, filename))
  155. {
  156. ANKI_LOGE(USAGE, argv[0]);
  157. return 1;
  158. }
  159. const Error err = dump(filename, dumpStats);
  160. if(err)
  161. {
  162. ANKI_LOGE("Can't dump due to an error. Bye");
  163. return 1;
  164. }
  165. return 0;
  166. }