from triton.backends.compiler import BaseBackend, GPUTarget, Language from triton._C.libtriton import ir, passes, llvm, amd from triton import knobs from dataclasses import dataclass from typing import Any, Dict, Tuple from types import ModuleType import hashlib import tempfile import re import subprocess import functools from pathlib import Path def get_min_dot_size(target: GPUTarget): # We fallback to use FMA and cast arguments if certain configurations is # not supported natively by matrix core units. return lambda lhs_type, rhs_type: (1, 1, 1) def is_pingpong_schedule_enabled(arch): return (arch == "gfx942") if knobs.amd.use_block_pingpong is None else knobs.amd.use_block_pingpong def is_in_thread_transpose_enabled(arch): return (arch == "gfx942") if knobs.amd.use_in_thread_transpose is None else knobs.amd.use_in_thread_transpose @dataclass(frozen=True) class HIPOptions: num_warps: int = 4 waves_per_eu: int = 1 num_stages: int = 2 num_ctas: int = 1 extern_libs: dict = None cluster_dims: tuple = (1, 1, 1) debug: bool = False sanitize_overflow: bool = True arch: str = None supported_fp8_dtypes: Tuple[str] = ("fp8e5", ) deprecated_fp8_dot_operand_dtypes: Tuple[str] = () default_dot_input_precision: str = "ieee" allowed_dot_input_precisions: Tuple[str] = ("ieee", ) enable_fp_fusion: bool = True launch_cooperative_grid: bool = False matrix_instr_nonkdim: int = 0 kpack: int = 1 allow_flush_denorm: bool = False max_num_imprecise_acc_default: int = 0 backend_name: str = 'hip' # The following option provides hints to the AMDGPU backend regarding instruction scheduling # for all `tt.dot` operations in a kernel. The "none" variant preserves the default # instruction scheduling of the AMDGPU backend which aims at maximizing occupancy. # The option is experimental and may change at any time regarding its semantics and/or may # be gone entirely anytime. # # Current experimental scheduling variants: # # local-prefetch: implements instruction scheduling similar to the one from the ROCm Composable # Kernel library. Note, this variant requires the use of buffer load/store ops # and a special software pipelining style - i.e., 1x LDS and 1x register # prefetch buffers for each GEMM tile. # attention: enables a bunch of optimizations for attention kernels, including: # - iglp 2 and sched.barrier around it # - sink-insts-to-avoid-spills flag to avoid register spills schedule_hint: str = 'none' def __post_init__(self): gfx_major = int(self.arch[3:-2]) # Drop "gfx" prefix and minor/patch number warp_size = 32 if gfx_major >= 10 else 64 object.__setattr__(self, 'warp_size', warp_size) assert self.num_warps > 0 and (self.num_warps & (self.num_warps - 1)) == 0, \ "num_warps must be a power of 2" if self.arch == 'gfx950': assert self.kpack == 1, "gfx950 only accepts kpack == 1" default_libdir = Path(__file__).parent / 'lib' extern_libs = {} if self.extern_libs is None else dict(self.extern_libs) for lib in ["ocml", "ockl"]: extern_libs[lib] = str(default_libdir / f'{lib}.bc') object.__setattr__(self, 'extern_libs', tuple(extern_libs.items())) def hash(self): key = '_'.join([f'{name}-{val}' for name, val in self.__dict__.items()]) return hashlib.sha256(key.encode("utf-8")).hexdigest() class HIPBackend(BaseBackend): @staticmethod def supports_target(target: GPUTarget): return target.backend == 'hip' def __init__(self, target: GPUTarget) -> None: super().__init__(target) assert isinstance(target.arch, str) self.binary_ext = "hsaco" def get_target_name(self, options) -> str: return f"hip:{options.arch}" def parse_options(self, opts) -> Any: args = {'arch': knobs.runtime.override_arch or self.target.arch} # Enable XF32 (TF32) for CDNA3 GPUs if self.target.arch == 'gfx942': allowed_dot_input_precisions = set(HIPOptions.allowed_dot_input_precisions) allowed_dot_input_precisions.update({'tf32'}) args["allowed_dot_input_precisions"] = tuple(sorted(allowed_dot_input_precisions)) if "supported_fp8_dtypes" not in opts: supported_fp8_dtypes = set(HIPOptions.supported_fp8_dtypes) if self.target.arch == 'gfx942': supported_fp8_dtypes.update({'fp8e4nv', 'fp8e4b8', 'fp8e5b16'}) elif self.target.arch == 'gfx950': supported_fp8_dtypes.update({'fp8e4nv', 'fp8e5'}) elif 'gfx12' in self.target.arch: supported_fp8_dtypes.update({'fp8e4nv', 'fp8e5'}) args["supported_fp8_dtypes"] = tuple(sorted(supported_fp8_dtypes)) if "enable_fp_fusion" not in opts: args["enable_fp_fusion"] = knobs.language.default_fp_fusion args.update({k: opts[k] for k in HIPOptions.__dataclass_fields__.keys() \ if k in opts and opts[k] is not None}) return HIPOptions(**args) def pack_metadata(self, metadata): return ( metadata.num_warps, metadata.num_ctas, metadata.shared, metadata.cluster_dims[0], metadata.cluster_dims[1], metadata.cluster_dims[2], ) def get_codegen_implementation(self, options): return {"min_dot_size": get_min_dot_size(self.target)} def get_module_map(self) -> Dict[str, ModuleType]: from triton.language.extra.hip import libdevice return {"triton.language.extra.libdevice": libdevice} def load_dialects(self, ctx): amd.load_dialects(ctx) @staticmethod def is_within_2gb(arg): import torch MAX_INT_32 = 2**31 - 1 if hasattr(arg, "ptr_range"): return arg.ptr_range() <= MAX_INT_32 if isinstance(arg, torch.Tensor) and hasattr(arg, "untyped_storage"): return arg.untyped_storage().size() <= MAX_INT_32 return False @staticmethod def parse_attr(desc): ret = BaseBackend.parse_attr(desc) if "S" in desc: ret += [["tt.pointer_range", 32]] return ret @staticmethod def get_arg_specialization(arg, ty, **kwargs): ret = BaseBackend.get_arg_specialization(arg, ty, **kwargs) # Only attempt to do buffer ops specialization if buffer ops are enabled. # Otherwise the is_within_2gb check is unnecessary overhead. if knobs.amd.use_buffer_ops and ty == "tensor" and HIPBackend.is_within_2gb(arg): ret += "S" return ret @staticmethod def path_to_rocm_lld(): # Check env path for ld.lld lld_env_path = knobs.amd.lld_path if lld_env_path is not None: lld = Path(lld_env_path) if lld.is_file(): return lld # Check backend for ld.lld (used for pytorch wheels) lld = Path(__file__).parent / "llvm/bin/ld.lld" if lld.is_file(): return lld lld = Path("/opt/rocm/llvm/bin/ld.lld") if lld.is_file(): return lld lld = Path("/usr/bin/ld.lld") if lld.is_file(): return lld raise Exception("ROCm linker /opt/rocm/llvm/bin/ld.lld not found. Set 'TRITON_HIP_LLD_PATH' to its path.") @staticmethod def make_ttir(mod, metadata, options): pm = ir.pass_manager(mod.context) pm.enable_debug() passes.common.add_inliner(pm) passes.ttir.add_rewrite_tensor_pointer(pm) passes.ttir.add_rewrite_tensor_descriptor_to_pointer(pm) passes.common.add_canonicalizer(pm) passes.ttir.add_combine(pm) passes.ttir.add_reorder_broadcast(pm) passes.common.add_cse(pm) passes.ttir.add_triton_licm(pm) passes.common.add_symbol_dce(pm) passes.ttir.add_loop_unroll(pm) pm.run(mod) return mod @staticmethod def make_ttgir(mod, metadata, options): pm = ir.pass_manager(mod.context) pm.enable_debug() passes.ttir.add_convert_to_ttgpuir(pm, f"hip:{options.arch}", options.num_warps, options.warp_size, options.num_ctas) pm.run(mod) pm = ir.pass_manager(mod.context) pm.enable_debug() passes.ttgpuir.add_coalesce(pm) passes.ttgpuir.add_remove_layout_conversions(pm) passes.ttgpuir.add_optimize_thread_locality(pm) amd.passes.ttgpuir.add_accelerate_matmul(pm, options.arch, options.matrix_instr_nonkdim, options.kpack) passes.ttgpuir.add_remove_layout_conversions(pm) amd.passes.ttgpuir.add_optimize_epilogue(pm) passes.ttgpuir.add_optimize_dot_operands(pm, True) amd.passes.ttgpuir.add_hoist_layout_conversions(pm) passes.ttgpuir.add_fuse_nested_loops(pm) passes.common.add_canonicalizer(pm) passes.ttir.add_triton_licm(pm) passes.common.add_canonicalizer(pm) global_prefetch = knobs.amd.global_prefetch local_prefetch = knobs.amd.local_prefetch use_async_copy = knobs.amd.use_async_copy # The `local-prefetch` scheduling variant requires turning on buffer ops. if options.schedule_hint == "local-prefetch": global_prefetch = local_prefetch = 1 amd.passes.ttgpuir.add_stream_pipeline(pm, options.num_stages, global_prefetch, local_prefetch, use_async_copy) if use_async_copy: amd.passes.ttgpuir.add_coalesce_async_copy(pm, options.arch) passes.common.add_canonicalizer(pm) if options.schedule_hint.lower() != "none": amd.passes.ttgpuir.insert_instruction_sched_hints(pm, options.schedule_hint) passes.ttgpuir.add_optimize_dot_operands(pm, True) passes.ttgpuir.add_remove_layout_conversions(pm) passes.ttgpuir.add_reduce_data_duplication(pm) if is_in_thread_transpose_enabled(options.arch): amd.passes.ttgpuir.add_in_thread_transpose(pm) passes.ttgpuir.add_remove_layout_conversions(pm) amd.passes.ttgpuir.add_reorder_instructions(pm) use_block_pingpong = is_pingpong_schedule_enabled(options.arch) if use_block_pingpong and options.num_stages == 2: amd.passes.ttgpuir.add_block_pingpong(pm, options.num_stages) if knobs.amd.use_buffer_ops: amd.passes.ttgpuir.add_canonicalize_pointers(pm) passes.common.add_canonicalizer(pm) amd.passes.ttgpuir.add_convert_to_buffer_ops(pm, options.arch) amd.passes.ttgpuir.add_fold_true_cmpi(pm) passes.common.add_canonicalizer(pm) passes.common.add_cse(pm) passes.common.add_symbol_dce(pm) if use_async_copy: amd.passes.ttgpuir.add_update_async_wait_count(pm, options.arch) pm.run(mod) return mod @staticmethod def ttgir_opt(src, metadata, options): mod = src pm = ir.pass_manager(mod.context) pm.enable_debug() passes.ttgpuir.add_inliner(pm) passes.common.add_sccp(pm) passes.ttir.add_loop_aware_cse(pm) passes.ttgpuir.add_canonicalizer(pm) passes.ttgpuir.add_combine_tensor_select_and_if(pm) pm.run(mod) return mod @staticmethod def make_llir(src, metadata, options): mod = src # TritonGPU -> LLVM-IR (MLIR) pm = ir.pass_manager(mod.context) pm.enable_debug() # custom_lds_size is an experimental parameter that defines amount of LDS available # for one thread block. Measured in bytes. # # If custom_lds_size = 0, pass will consider all LDS is available for one threads block, # LDS size is determined by provided arch name. custom_lds_size = 0 amd.passes.ttgpuir.add_optimize_lds_usage(pm, options.arch, custom_lds_size) passes.convert.add_scf_to_cf(pm) passes.convert.add_index_to_llvmir(pm) passes.ttgpuir.add_allocate_shared_memory(pm) ## __HIP_FTZ is used to control the denorm flushing behavior of exp2 op as follows: ## 1. If __HIP_FTZ = 1, exp2 flushes denorms in input and output regardless ## of the value of kernel arg `allow_flush_denorm`. ## 2. If __HIP_FTZ = 0, whether exp2 flushes denorms in input and output ## depends on the value of kernel arg `allow_flush_denorm`. ## 3. __HIP_FTZ is default to 1 and not exposed as a kernel argument. ## For now it is used as a controller for developers only. __HIP_FTZ = True amd.passes.ttgpuir.add_to_llvmir(pm, options.arch, __HIP_FTZ) passes.common.add_canonicalizer(pm) passes.common.add_cse(pm) passes.convert.add_cf_to_llvmir(pm) passes.convert.add_arith_to_llvmir(pm) passes.common.add_canonicalizer(pm) passes.common.add_cse(pm) passes.common.add_symbol_dce(pm) if options.schedule_hint.lower() != "none": amd.passes.ttgpuir.lower_instruction_sched_hints(pm, options.arch, options.num_stages) if not knobs.compilation.disable_line_info: passes.llvmir.add_di_scope(pm) amd.passes.ttgpuir.add_builtin_func_to_llvmir(pm, __HIP_FTZ) pm.run(mod) # LLVM-IR (MLIR) -> LLVM-IR (LLVM) llvm.init_targets() context = llvm.context() llvm_mod = llvm.to_module(mod, context) amd.attach_target_triple(llvm_mod) target_features = '' if knobs.compilation.enable_asan: target_features = '+xnack' llvm.attach_datalayout(llvm_mod, amd.TARGET_TRIPLE, options.arch, target_features) # Set various control constants on the LLVM module so that device # libraries can resolve references to them. amd.set_isa_version(llvm_mod, options.arch) amd.set_abi_version(llvm_mod, 500) amd.set_bool_control_constant(llvm_mod, "__oclc_finite_only_opt", False) amd.set_bool_control_constant(llvm_mod, "__oclc_correctly_rounded_sqrt32", True) amd.set_bool_control_constant(llvm_mod, "__oclc_unsafe_math_opt", False) amd.set_bool_control_constant(llvm_mod, "__oclc_wavefrontsize64", options.warp_size == 64) # Set kernel attributes first given this may affect later optimizations. fns = [fn for fn in llvm_mod.get_functions() if not fn.is_declaration()] # The public kernel should be kernel 0. fns[0].set_calling_conv(amd.CALLING_CONV_AMDGPU_KERNEL) fns[0].add_fn_attr("amdgpu-flat-work-group-size", f"1,{options.num_warps*options.warp_size}") # LLVM AMDGPU backend supports the attribute "amdgpu-waves-per-eu"="[, ]". # This attribute may be attached to a kernel function definition and is an optimization hint. # parameter specifies the requested minimum number of waves per EU, and optional parameter # specifies the requested maximum number of waves per EU (must be greater than if specified). # If is omitted, then there is no restriction on the maximum number of waves per EU other than # the one dictated by the hardware for which the kernel is compiled. Passing 0, 0 as , # implies the default behavior (no limits). fns[0].add_fn_attr("amdgpu-waves-per-eu", f"{options.waves_per_eu}") denormal_mode = "preserve-sign" if options.allow_flush_denorm else "ieee" fns[0].add_fn_attr("denormal-fp-math-f32", denormal_mode) if knobs.compilation.enable_asan: fns[0].add_fn_target_feature("+xnack") fns[0].add_fn_asan_attr() # Hint the compiler that we'd like the firmware to set the kernel arguments # to user SGPRs so that the kernel does not need to s_load its arguments # from memory. amd.set_all_fn_arg_inreg(fns[0]) if knobs.compilation.enable_asan: default_libdir = Path(__file__).parent / 'lib' paths = [ str(default_libdir / 'asanrtl.bc'), str(default_libdir / "ocml.bc"), str(default_libdir / "ockl.bc") ] llvm.link_extern_libs(llvm_mod, paths) elif options.extern_libs: paths = [path for (name, path) in options.extern_libs if amd.need_extern_lib(llvm_mod, name)] llvm.link_extern_libs(llvm_mod, paths) llvm.optimize_module(llvm_mod, llvm.OPTIMIZE_O3, options.arch, '', [], options.enable_fp_fusion) if knobs.amd.scalarize_packed_fops: amd.add_scalarize_packed_fops_llvm_pass(fns[0]) # Get some metadata metadata["shared"] = src.get_int_attr("ttg.shared") amd.cleanup_bitcode_metadata(llvm_mod) # Disable inlining of print related functions, # because inlining of these function could slow down compilation significantly amd.disable_print_inline(llvm_mod) return str(llvm_mod) @staticmethod def make_amdgcn(src, metadata, options): # Find kernel names (there should only be one) # We get the name at the last possible step to accommodate `triton.compile` # on user-provided LLVM names = re.findall(r"define amdgpu_kernel void @([a-zA-Z_][a-zA-Z0-9_]*)", src) assert len(names) == 1 metadata["name"] = names[0] # llvm -> hsaco flags = [] # The sink-insts-to-avoid-spills flag asks LLVM backend to sink instructions # into loops to avoid register spills in the MachineSinking pass, while it # can also lead to regression in some cases. But from current observation, # the regression is not significant. It would be better to have some heuristics. if options.schedule_hint == 'attention': flags.append('sink-insts-to-avoid-spills') amdgcn = llvm.translate_to_asm(src, amd.TARGET_TRIPLE, options.arch, '', flags, options.enable_fp_fusion, False) if knobs.amd.dump_amdgcn: print("// -----// AMDGCN Dump //----- //") print(amdgcn) return amdgcn @staticmethod def make_hsaco(src, metadata, options): target_features = '' if knobs.compilation.enable_asan: target_features = '+xnack' hsaco = amd.assemble_amdgcn(src, options.arch, target_features) rocm_path = HIPBackend.path_to_rocm_lld() with tempfile.NamedTemporaryFile() as tmp_out: with tempfile.NamedTemporaryFile() as tmp_in: with open(tmp_in.name, 'wb') as fd_in: fd_in.write(hsaco) subprocess.check_call([rocm_path, '-flavor', 'gnu', '-shared', tmp_in.name, '-o', tmp_out.name]) with open(tmp_out.name, 'rb') as fd_out: ret = fd_out.read() return ret def add_stages(self, stages, options, language): if language == Language.TRITON: stages["ttir"] = lambda src, metadata: self.make_ttir(src, metadata, options) stages["ttgir"] = lambda src, metadata: self.make_ttgir(src, metadata, options) elif language == Language.GLUON: stages["ttgir"] = lambda src, metadata: self.ttgir_opt(src, metadata, options) stages["llir"] = lambda src, metadata: self.make_llir(src, metadata, options) stages["amdgcn"] = lambda src, metadata: self.make_amdgcn(src, metadata, options) stages["hsaco"] = lambda src, metadata: self.make_hsaco(src, metadata, options) @functools.lru_cache() def hash(self): version = subprocess.check_output([HIPBackend.path_to_rocm_lld(), "--version"], encoding='utf-8') return f'{version}-{self.target}'