Open ashvardanian opened 7 months ago
There is now a better reference for how to implement this, also coming from SimSIMD:
// Read CPUID registers directly
unsigned long id_aa64isar0_el1 = 0, id_aa64isar1_el1 = 0, id_aa64pfr0_el1 = 0, id_aa64zfr0_el1 = 0;
// Now let's unpack the status flags from ID_AA64ISAR0_EL1
// https://developer.arm.com/documentation/ddi0601/2024-03/AArch64-Registers/ID-AA64ISAR0-EL1--AArch64-Instruction-Set-Attribute-Register-0?lang=en
__asm__ __volatile__("mrs %0, ID_AA64ISAR0_EL1" : "=r"(id_aa64isar0_el1));
// DP, bits [47:44] of ID_AA64ISAR0_EL1
unsigned supports_integer_dot_products = ((id_aa64isar0_el1 >> 44) & 0xF) >= 1;
// Now let's unpack the status flags from ID_AA64ISAR1_EL1
// https://developer.arm.com/documentation/ddi0601/2024-03/AArch64-Registers/ID-AA64ISAR1-EL1--AArch64-Instruction-Set-Attribute-Register-1?lang=en
__asm__ __volatile__("mrs %0, ID_AA64ISAR1_EL1" : "=r"(id_aa64isar1_el1));
// I8MM, bits [55:52] of ID_AA64ISAR1_EL1
unsigned supports_i8mm = ((id_aa64isar1_el1 >> 52) & 0xF) >= 1;
// BF16, bits [47:44] of ID_AA64ISAR1_EL1
unsigned supports_bf16 = ((id_aa64isar1_el1 >> 44) & 0xF) >= 1;
// Now let's unpack the status flags from ID_AA64PFR0_EL1
// https://developer.arm.com/documentation/ddi0601/2024-03/AArch64-Registers/ID-AA64PFR0-EL1--AArch64-Processor-Feature-Register-0?lang=en
__asm__ __volatile__("mrs %0, ID_AA64PFR0_EL1" : "=r"(id_aa64pfr0_el1));
// SVE, bits [35:32] of ID_AA64PFR0_EL1
unsigned supports_sve = ((id_aa64pfr0_el1 >> 32) & 0xF) >= 1;
// AdvSIMD, bits [23:20] of ID_AA64PFR0_EL1 can be used to check for `fp16` support
// - 0b0000: integers, single, double precision arithmetic
// - 0b0001: includes support for half-precision floating-point arithmetic
unsigned supports_fp16 = ((id_aa64pfr0_el1 >> 20) & 0xF) == 1;
// Now let's unpack the status flags from ID_AA64ZFR0_EL1
// https://developer.arm.com/documentation/ddi0601/2024-03/AArch64-Registers/ID-AA64ZFR0-EL1--SVE-Feature-ID-Register-0?lang=en
if (supports_sve)
__asm__ __volatile__("mrs %0, ID_AA64ZFR0_EL1" : "=r"(id_aa64zfr0_el1));
// I8MM, bits [47:44] of ID_AA64ZFR0_EL1
unsigned supports_sve_i8mm = ((id_aa64zfr0_el1 >> 44) & 0xF) >= 1;
// BF16, bits [23:20] of ID_AA64ZFR0_EL1
unsigned supports_sve_bf16 = ((id_aa64zfr0_el1 >> 20) & 0xF) >= 1;
// SVEver, bits [3:0] can be used to check for capability levels:
// - 0b0000: SVE is implemented
// - 0b0001: SVE2 is implemented
// - 0b0010: SVE2.1 is implemented
// This value must match the existing indicator obtained from ID_AA64PFR0_EL1:
// unsigned supports_sve = ((id_aa64zfr0_el1) & 0xF) >= 1;
// unsigned supports_sve2 = ((id_aa64zfr0_el1) & 0xF) >= 2;
unsigned supports_neon = 1; // NEON is always supported
return (simsimd_capability_t)( //
(simsimd_cap_neon_k * (supports_neon)) | //
(simsimd_cap_neon_f16_k * (supports_neon && supports_fp16)) | //
(simsimd_cap_neon_bf16_k * (supports_neon && supports_bf16)) | //
(simsimd_cap_neon_i8_k * (supports_neon && supports_i8mm && supports_integer_dot_products)) | //
(simsimd_cap_sve_k * (supports_sve)) | //
(simsimd_cap_sve_f16_k * (supports_sve && supports_fp16)) | //
(simsimd_cap_sve_bf16_k * (supports_sve && supports_sve_bf16)) | //
(simsimd_cap_sve_i8_k * (supports_sve && supports_sve_i8mm)) | //
(simsimd_cap_serial_k));
Hi @ashvardanian , I can take a shot at this, now that I know the SimSIMD codebase a bit better :)
Following the discussion in #137, it would be great to reach some uniformity in feature detection on x86 and Arm. On the latter, we can't yet use SVE, and only differentiate "extended NEON" and serial code. Assuming broader adoption among Arm devices, we need to isolate the features we use in different sub-generations of Arm v8, and consider bit-level operations from SVE.
https://github.com/ashvardanian/SimSIMD/blob/18d17686124ddebd9fe55eee56b2e0273a613d4b/include/simsimd/simsimd.h#L208-L228
In other libraries, like SimSIMD, I currently use the Linux API to check for those capabilities. But that is less portable than using inline Assembly, and we may need to detect those features on the upcoming Apple M-series chips.