Merge pull request #188 from RadeonOpenCompute/fix_extractkernel

Fix extractkernel
ROCm · Dec 12, 2016 · 71c6149 · 71c6149
2 parents 10979fc + 13ba8e1
commit 71c6149
Showing 1 changed file with 22 additions and 36 deletions.
diff --git a/lib/extractkernel.in b/lib/extractkernel.in
@@ -9,9 +9,9 @@ sub usage {
   print("-h          \t\t\t\tshow this help message\n");
   print("-i <input>  \t\t\t\tinput file\n");
   print("-t <target> \t\t\t\ttarget architecture (optional)\n");
-  print("            \t\t\t\t 7:0:0 for CI (Kaveri)\n");
-  print("            \t\t\t\t 8:0:1 for subset of VI (Carizzo)\n");
-  print("            \t\t\t\t 8:0:3 for subset of VI (Fiji) (default)\n");
+  print("            \t\t\t\t gfx700 for CI (Kaveri)\n");
+  print("            \t\t\t\t gfx801 for subset of VI (Carizzo)\n");
+  print("            \t\t\t\t gfx803 for subset of VI (Fiji) (default)\n");
   exit;
 }
 
@@ -29,51 +29,37 @@ $input_file = $options{i};
 (-f $input_file) || die("can't find $input_file");
 
 # Set default target architecture to Fiji
-my $target = "8:0:3";
+my $target = "gfx803";
 if (defined $options{t}) {
   $target = $options{t};
 }
 
-my $brig_start_tag = "_binary_kernel_brig_start";
-my $brig_size_tag  = "_binary_kernel_brig_size";
-my $brig_start = hex(`objdump -t $input_file | grep $brig_start_tag | awk '{print \$1}'`);
-my $brig_size  = hex(`objdump -t $input_file | grep $brig_size_tag | awk '{print \$1}'`);
+my $command;
 
-if ($brig_start == 0 || $brig_size == 0) {
+my $binary_start_tag = "_binary_kernel_brig_start";
+my $binary_size_tag  = "_binary_kernel_brig_size";
+
+$command = "objdump -t $input_file | grep $binary_start_tag | awk '{print \$1}'";
+my $binary_start = hex(`$command`);
+$command = "objdump -t $input_file | grep $binary_size_tag | awk '{print \$1}'";
+my $binary_size  = hex(`$command`);
+
+if ($binary_start == 0 || $binary_size == 0) {
   print("can't find any hsail kernel in $input_file\n");
 }  
 else {
 
-  # extract the hsail brig file from the binary
+  # extract the GPU binary file from the host binary
   my $data_lma         = hex(`objdump -h $input_file | grep ".data.rel.ro" | awk '{print \$5}'`);
   my $data_file_offset = hex(`objdump -h $input_file | grep ".data.rel.ro" | awk '{print \$6}'`);
-  my $brig_file_offset = $data_file_offset + ($brig_start - $data_lma);
-  my $brig_file_name = "$input_file.brig";
-  system("dd if=$input_file of=$brig_file_name skip=$brig_file_offset count=$brig_size bs=1 status=none");
+  my $binary_file_offset = $data_file_offset + ($binary_start - $data_lma);
+  my $binary_file_name = "$input_file.hsaco";
+  system("dd if=$input_file of=$binary_file_name skip=$binary_file_offset count=$binary_size bs=1 status=none");
 
-  my $hsa_path = "@HSA_ROOT@";
   my $rocm_path = "@ROCM_ROOT@";
+  my $llvm_objdump = "$rocm_path/hcc-lc/llvm/bin/llvm-objdump";
+  (-f $llvm_objdump) || die("can't find llvm-objdump to diassemble the GPU binary");
 
-  # set the default to rocm path
-  my $hsafin = "$rocm_path/bin/amdhsafin";
-
-  if (defined $ENV{'HSA_HOME'}) {
-    $hsafin = "$ENV{'HSA_HOME'}/bin/amdhsafin";
-  }
-  elsif (!(-f $hsafin)) {
-    # can't find amdhsafin in rocm, let's try with hsa root
-    $hsafin = "$hsa_path/bin/amdhsafin";
-  }
-
-  if (-f $hsafin) {
-
-    # use the offline finalizer to dump out the hsail and isa
-    system("$hsafin -target=$target -brig $brig_file_name -output=$input_file.gpu.o -O2 -dump-isa -dump-hsail");
-    move("./amdhsa001.hsail","./$input_file.hsail");
-    move("./amdhsa001.isa","./$input_file.isa");
-  }
-  else {
-    print("can't find HSAIL offline finalizer\n");
-  }
-
+  $command = "$llvm_objdump -disassemble -mcpu=$target $binary_file_name > $input_file.$target.isa";
+  system($command);
 }