ShaderProgramBinaryDumpMain.cpp 6.2 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211
  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* kUsage = 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(WeakArray<char*> argv, Bool& dumpStats, StringRaii& filename)
  14. {
  15. // Parse config
  16. if(argv.getSize() < 2)
  17. {
  18. return Error::kUserData;
  19. }
  20. dumpStats = false;
  21. filename = argv[argv.getSize() - 1];
  22. for(U32 i = 1; i < argv.getSize() - 1; i++)
  23. {
  24. if(strcmp(argv[i], "-stats") == 0)
  25. {
  26. dumpStats = true;
  27. }
  28. }
  29. return Error::kNone;
  30. }
  31. Error dumpStats(const ShaderProgramBinary& bin)
  32. {
  33. HeapMemoryPool pool(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{kMaxF64};
  59. U32 m_spillingCount = 0;
  60. U32 m_count = 0;
  61. };
  62. Array<StageStats, U32(ShaderType::kCount)> allStats;
  63. for(const ShaderProgramBinaryVariant& variant : bin.m_variants)
  64. {
  65. for(ShaderType shaderType : EnumIterable<ShaderType>())
  66. {
  67. if(variant.m_codeBlockIndices[shaderType] == kMaxU32)
  68. {
  69. continue;
  70. }
  71. const ShaderProgramBinaryCodeBlock& codeBlock = bin.m_codeBlocks[variant.m_codeBlockIndices[shaderType]];
  72. MaliOfflineCompilerOut maliocOut;
  73. const Error err = runMaliOfflineCompiler(
  74. #if ANKI_OS_LINUX
  75. ANKI_SOURCE_DIRECTORY "/ThirdParty/Bin/Linux64/MaliOfflineCompiler/malioc",
  76. #elif ANKI_OS_WINDOWS
  77. ANKI_SOURCE_DIRECTORY "/ThirdParty/Bin/Linux64/MaliOfflineCompiler/malioc.exe",
  78. #else
  79. # error "Not supported"
  80. #endif
  81. codeBlock.m_binary, shaderType, pool, maliocOut);
  82. if(err)
  83. {
  84. ANKI_LOGE("Mali offline compiler failed");
  85. return Error::kFunctionFailed;
  86. }
  87. // Appends stats
  88. StageStats& stage = allStats[shaderType];
  89. if(maliocOut.m_spilling)
  90. {
  91. ++stage.m_spillingCount;
  92. }
  93. ++stage.m_count;
  94. stage.m_avgStats.m_fma += maliocOut.m_fma;
  95. stage.m_avgStats.m_cvt += maliocOut.m_cvt;
  96. stage.m_avgStats.m_sfu += maliocOut.m_sfu;
  97. stage.m_avgStats.m_loadStore += maliocOut.m_loadStore;
  98. stage.m_avgStats.m_varying += maliocOut.m_varying;
  99. stage.m_avgStats.m_texture += maliocOut.m_texture;
  100. stage.m_avgStats.m_workRegisters += maliocOut.m_workRegisters;
  101. stage.m_avgStats.m_fp16ArithmeticPercentage += maliocOut.m_fp16ArithmeticPercentage;
  102. stage.m_maxStats.m_fma = max<F64>(stage.m_maxStats.m_fma, maliocOut.m_fma);
  103. stage.m_maxStats.m_cvt = max<F64>(stage.m_maxStats.m_cvt, maliocOut.m_cvt);
  104. stage.m_maxStats.m_sfu = max<F64>(stage.m_maxStats.m_sfu, maliocOut.m_sfu);
  105. stage.m_maxStats.m_loadStore = max<F64>(stage.m_maxStats.m_loadStore, maliocOut.m_loadStore);
  106. stage.m_maxStats.m_varying = max<F64>(stage.m_maxStats.m_varying, maliocOut.m_varying);
  107. stage.m_maxStats.m_texture = max<F64>(stage.m_maxStats.m_texture, maliocOut.m_texture);
  108. stage.m_maxStats.m_workRegisters = max<F64>(stage.m_maxStats.m_workRegisters, maliocOut.m_workRegisters);
  109. stage.m_maxStats.m_fp16ArithmeticPercentage =
  110. max<F64>(stage.m_maxStats.m_fp16ArithmeticPercentage, maliocOut.m_fp16ArithmeticPercentage);
  111. stage.m_minStats.m_fma = min<F64>(stage.m_minStats.m_fma, maliocOut.m_fma);
  112. stage.m_minStats.m_cvt = min<F64>(stage.m_minStats.m_cvt, maliocOut.m_cvt);
  113. stage.m_minStats.m_sfu = min<F64>(stage.m_minStats.m_sfu, maliocOut.m_sfu);
  114. stage.m_minStats.m_loadStore = min<F64>(stage.m_minStats.m_loadStore, maliocOut.m_loadStore);
  115. stage.m_minStats.m_varying = min<F64>(stage.m_minStats.m_varying, maliocOut.m_varying);
  116. stage.m_minStats.m_texture = min<F64>(stage.m_minStats.m_texture, maliocOut.m_texture);
  117. stage.m_minStats.m_workRegisters = min<F64>(stage.m_minStats.m_workRegisters, maliocOut.m_workRegisters);
  118. stage.m_minStats.m_fp16ArithmeticPercentage =
  119. min<F64>(stage.m_minStats.m_fp16ArithmeticPercentage, maliocOut.m_fp16ArithmeticPercentage);
  120. }
  121. }
  122. for(ShaderType shaderType : EnumIterable<ShaderType>())
  123. {
  124. const StageStats& stage = allStats[shaderType];
  125. if(stage.m_count == 0)
  126. {
  127. continue;
  128. }
  129. printf("Stage %u\n", U32(shaderType));
  130. printf("\tSpilling count %u\n", stage.m_spillingCount);
  131. const Stats& avg = stage.m_avgStats;
  132. printf("\tAvarage: Regs %f FMA %f CVT %f SFU %f LS %f VAR %f TEX %f FP16 %f%%\n",
  133. avg.m_workRegisters / F64(stage.m_count), avg.m_fma / F64(stage.m_count), avg.m_cvt / F64(stage.m_count),
  134. avg.m_sfu / F64(stage.m_count), avg.m_loadStore / F64(stage.m_count), avg.m_varying / F64(stage.m_count),
  135. avg.m_texture / F64(stage.m_count), avg.m_fp16ArithmeticPercentage / F64(stage.m_count));
  136. const Stats& maxs = stage.m_maxStats;
  137. 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,
  138. maxs.m_cvt, maxs.m_sfu, maxs.m_loadStore, maxs.m_varying, maxs.m_texture,
  139. maxs.m_fp16ArithmeticPercentage);
  140. }
  141. return Error::kNone;
  142. }
  143. Error dump(CString fname, Bool bDumpStats)
  144. {
  145. HeapMemoryPool pool(allocAligned, nullptr);
  146. ShaderProgramBinaryWrapper binw(&pool);
  147. ANKI_CHECK(binw.deserializeFromFile(fname));
  148. StringRaii txt(&pool);
  149. dumpShaderProgramBinary(binw.getBinary(), txt);
  150. printf("%s\n", txt.cstr());
  151. if(bDumpStats)
  152. {
  153. ANKI_CHECK(dumpStats(binw.getBinary()));
  154. }
  155. return Error::kNone;
  156. }
  157. int main(int argc, char** argv)
  158. {
  159. HeapMemoryPool pool(allocAligned, nullptr);
  160. StringRaii filename(&pool);
  161. Bool dumpStats;
  162. if(parseCommandLineArgs(WeakArray<char*>(argv, argc), dumpStats, filename))
  163. {
  164. ANKI_LOGE(kUsage, argv[0]);
  165. return 1;
  166. }
  167. const Error err = dump(filename, dumpStats);
  168. if(err)
  169. {
  170. ANKI_LOGE("Can't dump due to an error. Bye");
  171. return 1;
  172. }
  173. return 0;
  174. }