如何获取编译后的kernel的sgpr，vgpr使用量以及使用的shared memory大小 #196

Activity

1. xiaoyu1004

   changed the title [-]kernel使用的sgpr, vgpr，shared memory大小没有写进elf文件吗[/-] [+]如何获取编译后的kernel的sgpr，vgpr使用量以及使用的shared memory大小[/+] on Aug 26, 2025
2. Humber-186 commented on Dec 21, 2025

   

   Humber-186

   on Dec 21, 2025

   Last edited by Humber-186

   Collaborator

   More actions

   以下信息并非来自第一手开发者，只是文档中的少量描述 + AI代码分析

   ---

   有的，readelf可以看到：

   $ ./install/bin/llvm-readelf -S rodinia/opencl/gaussian/object0.riscv
   There are 15 section headers, starting at offset 0x6420:
   
   Section Headers:
     [Nr] Name              Type            Address  Off    Size   ES Flg Lk Inf Al
     [ 0]                   NULL            00000000 000000 000000 00      0   0  0
     [ 1] .text             PROGBITS        80000000 001000 000368 00  AX  0   0  4
     [ 2] .tohost           PROGBITS        80001000 002000 000048 00  WA  0   0 64
     [ 3] .eh_frame_hdr     PROGBITS        80002000 003000 000000 00  WA  0   0  1
     [ 4] .eh_frame         PROGBITS        80002000 003000 000018 00   A  0   0  4
     [ 5] .data             PROGBITS        80003000 004000 000008 00  WA  0   0  1
     [ 6] .sdata            PROGBITS        80003008 004008 000000 00  WA  0   0  1
     [ 7] .bss              NOBITS          80003008 004008 000000 00  WA  0   0  1
     [ 8] .comment          PROGBITS        00000000 004008 0000f4 01  MS  0   0  1
     [ 9] .ventus.resource.Fan1 PROGBITS    00000000 0040fc 000008 00   W  0   0  1        <= Note this
     [10] .ventus.resource.Fan2 PROGBITS    00000000 004104 000008 00   W  0   0  1        <= Note this
     [11] .riscv.attributes RISCV_ATTRIBUTES 00000000 00410c 00005f 00      0   0  1
     [12] .symtab           SYMTAB          00000000 00416c 001380 10     14 289  4
     [13] .shstrtab         STRTAB          00000000 0054ec 00009a 00      0   0  1
     [14] .strtab           STRTAB          00000000 005586 000e9a 00      0   0  1
   Key to Flags:
     W (write), A (alloc), X (execute), M (merge), S (strings), I (info),
     L (link order), O (extra OS processing required), G (group), T (TLS),
     C (compressed), x (unknown), o (OS specific), E (exclude),
     R (retain), p (processor specific)

   一个ELF中可能有多个kernel（上边的有Fan1, Fan2两个），每个kernel都有一个section名为.ventus.resource.<kernelname>记录自己的资源用量信息

   .ventus.resource.<kernelname> section的尺寸固定为8字节，实际上是4个int16（unsinged?）。生成这个section的代码在llvm/lib/Target/RISCV/RISCVAsmPrinter.cpp::RISCVAsmPrinter::runOnMachineFunction

     if (MF.getInfo<RISCVMachineFunctionInfo>()->isEntryFunction()) {
       MCSectionELF *ResourceSection = OutContext.getELFSection(
           ".ventus.resource." + MF.getName(), ELF::SHT_PROGBITS, ELF::SHF_WRITE);
       OutStreamer->switchSection(ResourceSection);
       OutStreamer->emitInt16(
                       CurrentProgramInfo->SubProgramInfoVec[FuncCount].VGPRUsage);
       OutStreamer->emitInt16(
                       CurrentProgramInfo->SubProgramInfoVec[FuncCount].SGPRUsage);
       OutStreamer->emitInt16(
                       CurrentProgramInfo->SubProgramInfoVec[FuncCount].LDSMemory);
       OutStreamer->emitInt16(
                       CurrentProgramInfo->SubProgramInfoVec[FuncCount].PDSMemory);
     }

   例如下边的ELF：

   $ ./install/bin/llvm-readelf -x .ventus.resource.matadd pocl/build/examples/matadd/object0.riscv
   Hex dump of section '.ventus.resource.matadd':
   0x00000000 05000700 10000800

   0x00000000是段偏移，其后的8字节应该是小端序的int16[]即：

   • 05 00 → 0x0005 → VGPRUsage = 5
   • 07 00 → 0x0007 → SGPRUsage = 7
   • 10 00 → 0x0010 → LDSMemory = 16
   • 08 00 → 0x0008 → PDSMemory = 8