Mj(dZddlmZddlmZddlmZddlmZddl m Z ddl Z dd l m Zd d l mZgd Zedd eee ffdZd efdZd efdZd efdZd efdZd efdZd efdZd efdZd efdZd&de jj8d dfdZGddZGddZeZ e a!d&de jj8d efdZ"GddeZ#d ed efd!Z$d e%fd"Z&de jj8d dfd#Z'd efd$Z(d efd%Z)y)'zj This package implements abstractions found in ``torch.cuda`` to facilitate writing device-agnostic code. )Mapping)AbstractContextManager) lru_cache)MappingProxyType)AnyNdevice)amp) is_availableis_initialized synchronizecurrent_devicecurrent_streamstream set_device device_countStream StreamContextEventget_capabilitiesreturncdttjjj S)a Returns an immutable mapping of CPU capabilities detected at runtime. This function queries the CPU for supported instruction sets and features using cpuinfo. The result is cached after the first call for efficiency. The returned mapping contains architecture-specific capabilities: For x86/x86_64: - SSE family: sse, sse2, sse3, ssse3, sse4_1, sse4_2, sse4a - AVX family: avx, avx2, avx_vnni - AVX-512 family: avx512_f, avx512_cd, avx512_dq, avx512_bw, avx512_vl, avx512_ifma, avx512_vbmi, avx512_vbmi2, avx512_bitalg, avx512_vpopcntdq, avx512_vnni, avx512_bf16, avx512_fp16, avx512_vp2intersect, avx512_4vnniw, avx512_4fmaps - AVX10 family: avx10_1, avx10_2 - AVX-VNNI-INT: avx_vnni_int8, avx_vnni_int16, avx_ne_convert - AMX: amx_bf16, amx_tile, amx_int8, amx_fp16 - FMA: fma3, fma4 - Other: f16c, bmi, bmi2, popcnt, lzcnt, aes, sha, clflush, clflushopt, clwb For ARM64: - SIMD: neon, fp16_arith, bf16, i8mm, dot - SVE: sve, sve2, sve_bf16, sve_max_length (when supported) - SME: sme, sme2, sme_max_length (when supported) - Other: atomics, fhm, rdm, crc32, aes, sha1, sha2, pmull Common to all architectures: - architecture: string identifying the CPU architecture Returns: MappingProxyType: An immutable mapping where keys are capability names (e.g., 'avx2', 'sve') and values are booleans indicating support, or integers for properties like vector lengths. Example: >>> caps = torch.cpu.get_capabilities() >>> if caps.get("avx2", False): ... print("AVX2 is supported") >>> print(f"Architecture: {caps['architecture']}") )rtorch_C_cpu_get_cpu_capabilityQ/var/www/html/ia-enhanced/venv/lib/python3.12/site-packages/torch/cpu/__init__.pyrr#s!V EHHMM==? @@r c6tjddS)z/Returns a bool indicating if CPU supports AVX2.avx2Frgetrr r!_is_avx2_supportedr&Qs   ! !&% 00r c6tjddS)z1Returns a bool indicating if CPU supports AVX512.avx512_fFr$rr r!_is_avx512_supportedr)V   ! !*e 44r c6tjddS)z6Returns a bool indicating if CPU supports AVX512_BF16. avx512_bf16Fr$rr r!_is_avx512_bf16_supportedr-[s   ! !- 77r c6tjddS)z/Returns a bool indicating if CPU supports VNNI. avx512_vnniFr$rr r!_is_vnni_supportedr0`s   ! !- 77r c6tjddS)z3Returns a bool indicating if CPU supports AMX_TILE.amx_tileFr$rr r!_is_amx_tile_supportedr3fr*r c6tjddS)z3Returns a bool indicating if CPU supports AMX FP16.amx_fp16Fr$rr r!_is_amx_fp16_supportedr6kr*r cRtjjjS)zInitializes AMX instructions.)rrr _init_amxrr r!r8r8ps 88== " " $$r cy)zReturns a bool indicating if CPU is currently available. N.B. This function only exists to facilitate device-agnostic code Trrr r!r r us r r cy)zWaits for all kernels in all streams on the CPU device to complete. Args: device (torch.device or int, optional): ignored, there's only one CPU device. N.B. This function only exists to facilitate device-agnostic code. Nrr s r!rr~r c:eZdZdZd deddfdZd dZd dZd dZy) rzH N.B. This class only exists to facilitate device-agnostic code priorityrNcyNr)selfr=s r!__init__zStream.__init__ r cyr?rr@rs r! wait_streamzStream.wait_streamrBr cyr?rr@s r! record_eventzStream.record_eventrBr cyr?r)r@events r! wait_eventzStream.wait_eventrBr )rN) __name__ __module__ __qualname____doc__intrArErHrKrr r!rrs)  d    r rc4eZdZdefdZdddZddZdddZy) rrcy)NTrrGs r!queryz Event.querysr Ncyr?rrDs r!recordz Event.recordrBr cyr?rrGs r!rzEvent.synchronizerBr cyr?rrDs r!waitz Event.waitrBr r?rM)rNrOrPboolrUrWrrZrr r!rrst   r rctS)zReturns the currently selected :class:`Stream` for a given device. Args: device (torch.device or int, optional): Ignored. N.B. This function only exists to facilitate device-agnostic code )_current_streamr s r!rrs  r cFeZdZUdZedzed<dZdZdededed dfd Z y) rzvContext-manager that selects a given stream. N.B. This class only exists to facilitate device-agnostic code N cur_streamc(||_t|_yr?)r_default_cpu_stream prev_streamrDs r!rAzStreamContext.__init__s .r c<|j}|yt|_|ayr?)rr]rb)r@r_s r! __enter__zStreamContext.__enter__s#[[   +$r typevalue tracebackrc:|j}|y|jayr?)rrbr])r@rerfrgr_s r!__exit__zStreamContext.__exit__s![[   **r ) rNrOrPrQr__annotations__rArdrrirr r!rrs>  /%+S++++r rrct|S)zWrapper around the Context-manager StreamContext that selects a given stream. N.B. This function only exists to facilitate device-agnostic code )r)rs r!rrs   r cy)zReturns number of CPU devices (not cores). Always 1. N.B. This function only exists to facilitate device-agnostic code r rrr r!rrs r cy)zzSets the current device, in CPU we do nothing. N.B. This function only exists to facilitate device-agnostic code Nrr s r!rrr;r cy)zyReturns current device for cpu. Always 'cpu'. N.B. This function only exists to facilitate device-agnostic code cpurrr r!rrs r cy)zReturns True if the CPU is initialized. Always True. N.B. This function only exists to facilitate device-agnostic code Trrr r!rrs r r?)*rQcollections.abcr contextlibr functoolsrtypesrtypingrrr _devicer __all__strrr[r&r)r-r0r3r6r8r Devicerrrrar]rrrrRrrrrrr r!r{sv $-"    4*A'#s(+*A*AZ1D1 5d5 848 8D8 55 55 %4% d **d  $   h% 5;;--  +*+>!6!4! c u{{))dr