diff --git a/tinygrad/runtime/support/compiler_amd.py b/tinygrad/runtime/support/compiler_amd.py index 8f26780d92..d0f7ec6682 100644 --- a/tinygrad/runtime/support/compiler_amd.py +++ b/tinygrad/runtime/support/compiler_amd.py @@ -13,8 +13,9 @@ from tinygrad.runtime.support.compiler_cpu import LLVMCompiler from tinygrad.helpers import OSX, to_char_p_p def amdgpu_disassemble(lib:bytes): - asm = subprocess.check_output(["llvm-objdump" if OSX else "/opt/rocm/llvm/bin/llvm-objdump", '-d', '-'], input=lib) - print('\n'.join([x for x in asm.decode('utf-8').split("\n") if 's_code_end' not in x])) + asm = subprocess.check_output(["llvm-objdump" if OSX else "/opt/rocm/llvm/bin/llvm-objdump", '-d', '-'], input=lib).decode("utf-8").splitlines() + while asm and ("s_nop 0" in asm[-1] or "s_code_end" in asm[-1]): asm.pop() + print("\n".join(asm)) def check(status): if status != 0: