From 950d8de00e1e7dd62aee03efbf52be81ce13286f Mon Sep 17 00:00:00 2001 From: Christopher Milan Date: Thu, 11 Dec 2025 21:02:44 -0800 Subject: [PATCH] automatically inline anonymous (#13652) --- tinygrad/runtime/autogen/hsa.py | 89 +++++++++++++++-------------- tinygrad/runtime/support/autogen.py | 10 +++- 2 files changed, 55 insertions(+), 44 deletions(-) diff --git a/tinygrad/runtime/autogen/hsa.py b/tinygrad/runtime/autogen/hsa.py index 82e418d37f..fe68dd9b0d 100644 --- a/tinygrad/runtime/autogen/hsa.py +++ b/tinygrad/runtime/autogen/hsa.py @@ -1294,18 +1294,13 @@ struct_hsa_amd_aie_ert_start_kernel_data_s._fields_ = [ ] hsa_amd_aie_ert_start_kernel_data_t = struct_hsa_amd_aie_ert_start_kernel_data_s class struct_hsa_amd_aie_ert_packet_s(Struct): pass -class struct_hsa_amd_aie_ert_packet_s_0(Struct): pass -struct_hsa_amd_aie_ert_packet_s_0._fields_ = [ +struct_hsa_amd_aie_ert_packet_s._fields_ = [ + ('header', hsa_amd_vendor_packet_header_t), ('state', uint32_t,4), ('custom', uint32_t,8), ('count', uint32_t,11), ('opcode', uint32_t,5), ('type', uint32_t,4), -] -struct_hsa_amd_aie_ert_packet_s._anonymous_ = ['_0'] -struct_hsa_amd_aie_ert_packet_s._fields_ = [ - ('header', hsa_amd_vendor_packet_header_t), - ('_0', struct_hsa_amd_aie_ert_packet_s_0), ('reserved0', uint64_t), ('reserved1', uint64_t), ('reserved2', uint64_t), @@ -1314,16 +1309,24 @@ struct_hsa_amd_aie_ert_packet_s._fields_ = [ ('reserved5', uint64_t), ('payload_data', uint64_t), ] +class _anonstruct0(Struct): pass +_anonstruct0._fields_ = [ + ('state', uint32_t,4), + ('custom', uint32_t,8), + ('count', uint32_t,11), + ('opcode', uint32_t,5), + ('type', uint32_t,4), +] hsa_amd_aie_ert_packet_t = struct_hsa_amd_aie_ert_packet_s -_anonenum0 = CEnum(ctypes.c_uint32) -HSA_STATUS_ERROR_INVALID_MEMORY_POOL = _anonenum0.define('HSA_STATUS_ERROR_INVALID_MEMORY_POOL', 40) -HSA_STATUS_ERROR_MEMORY_APERTURE_VIOLATION = _anonenum0.define('HSA_STATUS_ERROR_MEMORY_APERTURE_VIOLATION', 41) -HSA_STATUS_ERROR_ILLEGAL_INSTRUCTION = _anonenum0.define('HSA_STATUS_ERROR_ILLEGAL_INSTRUCTION', 42) -HSA_STATUS_ERROR_MEMORY_FAULT = _anonenum0.define('HSA_STATUS_ERROR_MEMORY_FAULT', 43) -HSA_STATUS_CU_MASK_REDUCED = _anonenum0.define('HSA_STATUS_CU_MASK_REDUCED', 44) -HSA_STATUS_ERROR_OUT_OF_REGISTERS = _anonenum0.define('HSA_STATUS_ERROR_OUT_OF_REGISTERS', 45) -HSA_STATUS_ERROR_RESOURCE_BUSY = _anonenum0.define('HSA_STATUS_ERROR_RESOURCE_BUSY', 46) -HSA_STATUS_ERROR_NOT_SUPPORTED = _anonenum0.define('HSA_STATUS_ERROR_NOT_SUPPORTED', 47) +_anonenum1 = CEnum(ctypes.c_uint32) +HSA_STATUS_ERROR_INVALID_MEMORY_POOL = _anonenum1.define('HSA_STATUS_ERROR_INVALID_MEMORY_POOL', 40) +HSA_STATUS_ERROR_MEMORY_APERTURE_VIOLATION = _anonenum1.define('HSA_STATUS_ERROR_MEMORY_APERTURE_VIOLATION', 41) +HSA_STATUS_ERROR_ILLEGAL_INSTRUCTION = _anonenum1.define('HSA_STATUS_ERROR_ILLEGAL_INSTRUCTION', 42) +HSA_STATUS_ERROR_MEMORY_FAULT = _anonenum1.define('HSA_STATUS_ERROR_MEMORY_FAULT', 43) +HSA_STATUS_CU_MASK_REDUCED = _anonenum1.define('HSA_STATUS_CU_MASK_REDUCED', 44) +HSA_STATUS_ERROR_OUT_OF_REGISTERS = _anonenum1.define('HSA_STATUS_ERROR_OUT_OF_REGISTERS', 45) +HSA_STATUS_ERROR_RESOURCE_BUSY = _anonenum1.define('HSA_STATUS_ERROR_RESOURCE_BUSY', 46) +HSA_STATUS_ERROR_NOT_SUPPORTED = _anonenum1.define('HSA_STATUS_ERROR_NOT_SUPPORTED', 47) hsa_amd_iommu_version_t = CEnum(ctypes.c_uint32) HSA_IOMMU_SUPPORT_NONE = hsa_amd_iommu_version_t.define('HSA_IOMMU_SUPPORT_NONE', 0) @@ -2382,14 +2385,14 @@ struct_amd_runtime_loader_debug_info_s._fields_ = [ amd_runtime_loader_debug_info_t = struct_amd_runtime_loader_debug_info_s class struct_BrigModuleHeader(Struct): pass BrigModule_t = ctypes.POINTER(struct_BrigModuleHeader) -_anonenum1 = CEnum(ctypes.c_uint32) -HSA_EXT_STATUS_ERROR_INVALID_PROGRAM = _anonenum1.define('HSA_EXT_STATUS_ERROR_INVALID_PROGRAM', 8192) -HSA_EXT_STATUS_ERROR_INVALID_MODULE = _anonenum1.define('HSA_EXT_STATUS_ERROR_INVALID_MODULE', 8193) -HSA_EXT_STATUS_ERROR_INCOMPATIBLE_MODULE = _anonenum1.define('HSA_EXT_STATUS_ERROR_INCOMPATIBLE_MODULE', 8194) -HSA_EXT_STATUS_ERROR_MODULE_ALREADY_INCLUDED = _anonenum1.define('HSA_EXT_STATUS_ERROR_MODULE_ALREADY_INCLUDED', 8195) -HSA_EXT_STATUS_ERROR_SYMBOL_MISMATCH = _anonenum1.define('HSA_EXT_STATUS_ERROR_SYMBOL_MISMATCH', 8196) -HSA_EXT_STATUS_ERROR_FINALIZATION_FAILED = _anonenum1.define('HSA_EXT_STATUS_ERROR_FINALIZATION_FAILED', 8197) -HSA_EXT_STATUS_ERROR_DIRECTIVE_MISMATCH = _anonenum1.define('HSA_EXT_STATUS_ERROR_DIRECTIVE_MISMATCH', 8198) +_anonenum2 = CEnum(ctypes.c_uint32) +HSA_EXT_STATUS_ERROR_INVALID_PROGRAM = _anonenum2.define('HSA_EXT_STATUS_ERROR_INVALID_PROGRAM', 8192) +HSA_EXT_STATUS_ERROR_INVALID_MODULE = _anonenum2.define('HSA_EXT_STATUS_ERROR_INVALID_MODULE', 8193) +HSA_EXT_STATUS_ERROR_INCOMPATIBLE_MODULE = _anonenum2.define('HSA_EXT_STATUS_ERROR_INCOMPATIBLE_MODULE', 8194) +HSA_EXT_STATUS_ERROR_MODULE_ALREADY_INCLUDED = _anonenum2.define('HSA_EXT_STATUS_ERROR_MODULE_ALREADY_INCLUDED', 8195) +HSA_EXT_STATUS_ERROR_SYMBOL_MISMATCH = _anonenum2.define('HSA_EXT_STATUS_ERROR_SYMBOL_MISMATCH', 8196) +HSA_EXT_STATUS_ERROR_FINALIZATION_FAILED = _anonenum2.define('HSA_EXT_STATUS_ERROR_FINALIZATION_FAILED', 8197) +HSA_EXT_STATUS_ERROR_DIRECTIVE_MISMATCH = _anonenum2.define('HSA_EXT_STATUS_ERROR_DIRECTIVE_MISMATCH', 8198) hsa_ext_module_t = ctypes.POINTER(struct_BrigModuleHeader) class struct_hsa_ext_program_s(Struct): pass @@ -2448,26 +2451,26 @@ struct_hsa_ext_finalizer_1_00_pfn_s._fields_ = [ ('hsa_ext_program_finalize', ctypes.CFUNCTYPE(hsa_status_t, hsa_ext_program_t, hsa_isa_t, int32_t, hsa_ext_control_directives_t, ctypes.POINTER(ctypes.c_char), hsa_code_object_type_t, ctypes.POINTER(hsa_code_object_t))), ] hsa_ext_finalizer_1_00_pfn_t = struct_hsa_ext_finalizer_1_00_pfn_s -_anonenum2 = CEnum(ctypes.c_uint32) -HSA_EXT_STATUS_ERROR_IMAGE_FORMAT_UNSUPPORTED = _anonenum2.define('HSA_EXT_STATUS_ERROR_IMAGE_FORMAT_UNSUPPORTED', 12288) -HSA_EXT_STATUS_ERROR_IMAGE_SIZE_UNSUPPORTED = _anonenum2.define('HSA_EXT_STATUS_ERROR_IMAGE_SIZE_UNSUPPORTED', 12289) -HSA_EXT_STATUS_ERROR_IMAGE_PITCH_UNSUPPORTED = _anonenum2.define('HSA_EXT_STATUS_ERROR_IMAGE_PITCH_UNSUPPORTED', 12290) -HSA_EXT_STATUS_ERROR_SAMPLER_DESCRIPTOR_UNSUPPORTED = _anonenum2.define('HSA_EXT_STATUS_ERROR_SAMPLER_DESCRIPTOR_UNSUPPORTED', 12291) - _anonenum3 = CEnum(ctypes.c_uint32) -HSA_EXT_AGENT_INFO_IMAGE_1D_MAX_ELEMENTS = _anonenum3.define('HSA_EXT_AGENT_INFO_IMAGE_1D_MAX_ELEMENTS', 12288) -HSA_EXT_AGENT_INFO_IMAGE_1DA_MAX_ELEMENTS = _anonenum3.define('HSA_EXT_AGENT_INFO_IMAGE_1DA_MAX_ELEMENTS', 12289) -HSA_EXT_AGENT_INFO_IMAGE_1DB_MAX_ELEMENTS = _anonenum3.define('HSA_EXT_AGENT_INFO_IMAGE_1DB_MAX_ELEMENTS', 12290) -HSA_EXT_AGENT_INFO_IMAGE_2D_MAX_ELEMENTS = _anonenum3.define('HSA_EXT_AGENT_INFO_IMAGE_2D_MAX_ELEMENTS', 12291) -HSA_EXT_AGENT_INFO_IMAGE_2DA_MAX_ELEMENTS = _anonenum3.define('HSA_EXT_AGENT_INFO_IMAGE_2DA_MAX_ELEMENTS', 12292) -HSA_EXT_AGENT_INFO_IMAGE_2DDEPTH_MAX_ELEMENTS = _anonenum3.define('HSA_EXT_AGENT_INFO_IMAGE_2DDEPTH_MAX_ELEMENTS', 12293) -HSA_EXT_AGENT_INFO_IMAGE_2DADEPTH_MAX_ELEMENTS = _anonenum3.define('HSA_EXT_AGENT_INFO_IMAGE_2DADEPTH_MAX_ELEMENTS', 12294) -HSA_EXT_AGENT_INFO_IMAGE_3D_MAX_ELEMENTS = _anonenum3.define('HSA_EXT_AGENT_INFO_IMAGE_3D_MAX_ELEMENTS', 12295) -HSA_EXT_AGENT_INFO_IMAGE_ARRAY_MAX_LAYERS = _anonenum3.define('HSA_EXT_AGENT_INFO_IMAGE_ARRAY_MAX_LAYERS', 12296) -HSA_EXT_AGENT_INFO_MAX_IMAGE_RD_HANDLES = _anonenum3.define('HSA_EXT_AGENT_INFO_MAX_IMAGE_RD_HANDLES', 12297) -HSA_EXT_AGENT_INFO_MAX_IMAGE_RORW_HANDLES = _anonenum3.define('HSA_EXT_AGENT_INFO_MAX_IMAGE_RORW_HANDLES', 12298) -HSA_EXT_AGENT_INFO_MAX_SAMPLER_HANDLERS = _anonenum3.define('HSA_EXT_AGENT_INFO_MAX_SAMPLER_HANDLERS', 12299) -HSA_EXT_AGENT_INFO_IMAGE_LINEAR_ROW_PITCH_ALIGNMENT = _anonenum3.define('HSA_EXT_AGENT_INFO_IMAGE_LINEAR_ROW_PITCH_ALIGNMENT', 12300) +HSA_EXT_STATUS_ERROR_IMAGE_FORMAT_UNSUPPORTED = _anonenum3.define('HSA_EXT_STATUS_ERROR_IMAGE_FORMAT_UNSUPPORTED', 12288) +HSA_EXT_STATUS_ERROR_IMAGE_SIZE_UNSUPPORTED = _anonenum3.define('HSA_EXT_STATUS_ERROR_IMAGE_SIZE_UNSUPPORTED', 12289) +HSA_EXT_STATUS_ERROR_IMAGE_PITCH_UNSUPPORTED = _anonenum3.define('HSA_EXT_STATUS_ERROR_IMAGE_PITCH_UNSUPPORTED', 12290) +HSA_EXT_STATUS_ERROR_SAMPLER_DESCRIPTOR_UNSUPPORTED = _anonenum3.define('HSA_EXT_STATUS_ERROR_SAMPLER_DESCRIPTOR_UNSUPPORTED', 12291) + +_anonenum4 = CEnum(ctypes.c_uint32) +HSA_EXT_AGENT_INFO_IMAGE_1D_MAX_ELEMENTS = _anonenum4.define('HSA_EXT_AGENT_INFO_IMAGE_1D_MAX_ELEMENTS', 12288) +HSA_EXT_AGENT_INFO_IMAGE_1DA_MAX_ELEMENTS = _anonenum4.define('HSA_EXT_AGENT_INFO_IMAGE_1DA_MAX_ELEMENTS', 12289) +HSA_EXT_AGENT_INFO_IMAGE_1DB_MAX_ELEMENTS = _anonenum4.define('HSA_EXT_AGENT_INFO_IMAGE_1DB_MAX_ELEMENTS', 12290) +HSA_EXT_AGENT_INFO_IMAGE_2D_MAX_ELEMENTS = _anonenum4.define('HSA_EXT_AGENT_INFO_IMAGE_2D_MAX_ELEMENTS', 12291) +HSA_EXT_AGENT_INFO_IMAGE_2DA_MAX_ELEMENTS = _anonenum4.define('HSA_EXT_AGENT_INFO_IMAGE_2DA_MAX_ELEMENTS', 12292) +HSA_EXT_AGENT_INFO_IMAGE_2DDEPTH_MAX_ELEMENTS = _anonenum4.define('HSA_EXT_AGENT_INFO_IMAGE_2DDEPTH_MAX_ELEMENTS', 12293) +HSA_EXT_AGENT_INFO_IMAGE_2DADEPTH_MAX_ELEMENTS = _anonenum4.define('HSA_EXT_AGENT_INFO_IMAGE_2DADEPTH_MAX_ELEMENTS', 12294) +HSA_EXT_AGENT_INFO_IMAGE_3D_MAX_ELEMENTS = _anonenum4.define('HSA_EXT_AGENT_INFO_IMAGE_3D_MAX_ELEMENTS', 12295) +HSA_EXT_AGENT_INFO_IMAGE_ARRAY_MAX_LAYERS = _anonenum4.define('HSA_EXT_AGENT_INFO_IMAGE_ARRAY_MAX_LAYERS', 12296) +HSA_EXT_AGENT_INFO_MAX_IMAGE_RD_HANDLES = _anonenum4.define('HSA_EXT_AGENT_INFO_MAX_IMAGE_RD_HANDLES', 12297) +HSA_EXT_AGENT_INFO_MAX_IMAGE_RORW_HANDLES = _anonenum4.define('HSA_EXT_AGENT_INFO_MAX_IMAGE_RORW_HANDLES', 12298) +HSA_EXT_AGENT_INFO_MAX_SAMPLER_HANDLERS = _anonenum4.define('HSA_EXT_AGENT_INFO_MAX_SAMPLER_HANDLERS', 12299) +HSA_EXT_AGENT_INFO_IMAGE_LINEAR_ROW_PITCH_ALIGNMENT = _anonenum4.define('HSA_EXT_AGENT_INFO_IMAGE_LINEAR_ROW_PITCH_ALIGNMENT', 12300) hsa_ext_image_channel_type_t = CEnum(ctypes.c_uint32) HSA_EXT_IMAGE_CHANNEL_TYPE_SNORM_INT8 = hsa_ext_image_channel_type_t.define('HSA_EXT_IMAGE_CHANNEL_TYPE_SNORM_INT8', 0) diff --git a/tinygrad/runtime/support/autogen.py b/tinygrad/runtime/support/autogen.py index d55e4d06c3..728ae9dea8 100644 --- a/tinygrad/runtime/support/autogen.py +++ b/tinygrad/runtime/support/autogen.py @@ -27,6 +27,14 @@ def fields(t: clang.CXType) -> list[clang.CXCursor]: clang.clang_Type_visitFields(t, visitor, None) return ret +# flattens anonymous fields +def all_fields(t, kind): + for f in fields(t): + if (clang.clang_Cursor_isAnonymousRecordDecl(clang.clang_getTypeDeclaration(clang.clang_getCursorType(f))) and + clang.clang_getTypeDeclaration(clang.clang_getCursorType(f)).kind == kind): + yield from all_fields(clang.clang_getCursorType(f), kind) + else: yield f + def walk(c: clang.CXCursor) -> Iterator[clang.CXCursor]: yield c for child in children(c): yield from walk(child) @@ -137,7 +145,7 @@ def gen(dll, files, args=[], prolog=[], rules=[], epilog=[], recsym=False, use_e def is_anon(f): return clang.clang_Cursor_isAnonymousRecordDecl(clang.clang_getTypeDeclaration(clang.clang_getCursorType(f))) ll=[" ("+((fn:=f"'_{acnt()}'")+f", {tname(clang.clang_getCursorType(f), tnm+fn[1:-1])}" if is_anon(f) else f"'{nm(f)}', "+ tname(clang.clang_getCursorType(f), f'{tnm}_{nm(f)}'))+(f',{clang.clang_getFieldDeclBitWidth(f)}' * clang.clang_Cursor_isBitField(f))+")," - for f in fields(t)] + for f in all_fields(t, decl.kind)] lines.extend(([f"{tnm}._anonymous_ = ["+", ".join(f"'_{i}'" for i in range(n))+"]"] if (n:=acnt()) else [])+ ([f"{tnm}._packed_ = True"] * is_packed)+([f"{tnm}._fields_ = [",*ll,"]"] if ll else [])) return tnm