**ISA Comparison Table** - discussion and research at <https://bugs.libre-soc.org/show_bug.cgi?id=893>
-|ISA <br>name |Num <br>opcodes |Num <br>intrinsics |Taxonomy / <br> Class |setvl <br> scalable|Predicate <br> Masks|Twin <br> Predication|Explicit <br> Vector regs|128-bit <br> ops|Bigint |LDST <br> Fault-First|Data-dep<br> Fail-first|Pred-<br> Result |Matrix HW<br> support|DCT/FFT HW<br> support |
-|--------------------------|------------------------|------------------------|-------------------------------|-------------------|--------------------|---------------------|-------------------------|----------------|-------------|---------------------|-----------------------|------------------|---------------------|-----------------------|
-|Draft SVP64 |5 [^svp64_prefix] |see [^svp64_intrin_cnt] |Scalable [^svp64_scalable] |yes |yes |yes [^twin_pred] |no [^svp64_no_vec_regs] |see [^svp64_128]|yes [^bigint]|yes [^fail_first] |yes [^data_fail_first] |yes [^pred_result]|yes [^svp64_mat] | yes[^svp64_fft] |
-|VSX |700+ |700+? [^vsx_intrin] |Packed SIMD |no |no |no |yes [^vsx_vec_regs] |yes |no |no |no |no |yes [^ppc_mma] | no |
-|NEON |~250 [^neon_opcodes] |7088 [^neon_intrin] |Packed SIMD |no |no |no |yes |yes |no |no |no |no |no | no |
-|SVE2 |~1000 [^sve2_opcodes] |6040 [^sve2_intrin] |Predicated SIMD[^sve2_no_setvl]|no [^sve2_no_setvl]|yes |no |yes |yes |no |yes [^fail_first] |no |no |yes [^sve2_mat] | no |
-|AVX512 [^avx512_wikipedia]|~1000s [^avx512_opcodes]|7256 [^avx512_intrin] |Predicated SIMD |no |yes |no |yes |yes |no |no |no |no |yes [^x86_amx] | no |
-|RVV [^rvv_spec] |~190 [^rvv_opcodes] |~25000 [^rvv_intrin] |Scalable [^rvv_scalable] |yes |yes |no |yes |yes [^rvv_128] |no |yes |no |no |no | no |
-|Aurora SX[^sx_aurora] |~200 [^aurora_isa] |unknown [^aurora_intrin]|Scalable [^aurora_scalable] |yes |yes |no |yes |no |no |no |no |no |? | no |
+|ISA <br>name |No <br>opcodes|No <br>intrinsics|Taxonomy / <br>Class|setvl <br> scalable|Predicate <br> Masks|Twin <br>Pred|Vector <br>regs |128-bit <br> ops |Bigint |LDST <br>F/First|Data-dep<br> Fail-first|Pred-<br> Result|HW<br> Matrix|DCT/FFT <br>HW|
+|---------------|--------------|-----------------|--------------------|-------------------|--------------------|-------------|----------------|-----------------|-------|----------------|-----------------------|----------------|-------------|--------------|
+|SVP64 |5 [^1] |see [^2] |Scalable [^3] |yes |yes |yes [^4] |no [^5] |see [^6] |yes[^7]|yes [^8] |yes [^9] |yes [^10] |yes [^11] | yes[^12] |
+|VSX |700+ |700+? [^27] |Packed SIMD |no |no |no |yes [^13] |yes |no |no |no |no |yes [^14] | no |
+|NEON |~250 [^15] |7088 [^28] |Packed SIMD |no |no |no |yes |yes |no |no |no |no |no | no |
+|SVE2 |~1000 [^16] |6040 [^29] |Predicated SIMD[^17]|no [^17] |yes |no |yes |yes |no |yes [^8] |no |no |yes [^33] | no |
+|AVX512 [^18] |~1000s [^19] |7256 [^30] |Predicated SIMD |no |yes |no |yes |yes |no |no |no |no |yes [^34] | no |
+|RVV [^20] |~190 [^21] |~25000 [^31] |Scalable [^22] |yes |yes |no |yes |yes [^23] |no |yes |no |no |no | no |
+|Aurora SX[^24] |~200 [^25] |unknown [^32] |Scalable [^26] |yes |yes |no |yes |no |no |no |no |no |? | no |
-[^svp64_prefix]: plus EXT001 24-bit prefixing using 25% of EXT001 space. See [[sv/svp64]]
-[^svp64_scalable]: A 2-Dimensional Scalable Vector ISA **specifically designed for the Power ISA** with both Horizontal-First and Vertical-First Modes. See [[sv/vector_isa_comparison]]
-[^twin_pred]: on specific operations. See [[opcode_regs_deduped]] for full list. Key: 2P - Twin Predication, 1P - Single-Predicate
-[^svp64_no_vec_regs]: SVP64 provides a Vector concept on top of the **Scalar** GPR, FPR and CR Fields, extended to 128 entries.
-[^svp64_128]: SVP64 Vectorises Scalar ops. It is up to the **implementor** to choose (**optionally**) whether to apply SVP64 to e.g. VSX Quad-Precision (128-bit) instructions, to create 128-bit Vector ops.
-[^bigint]: big-integer add is just `sv.adde`. For optimal performance Bigint Mul and divide first require addition of two scalar operations (in turn, naturally Vectorised by SVP64). See [[sv/biginteger/analysis]]
-[^fail_first]: See [[sv/svp64/appendix]] and [ARM SVE Fault-First](https://alastairreid.github.io/papers/sve-ieee-micro-2017.pdf)
-[^data_fail_first]: Based on LD/ST Fail-first, extended to data. See [[sv/svp64/appendix]]
-[^pred_result]: Predicate-result effectively turns any standard op into a type of "cmp". See [[sv/svp64/appendix]]
-[^svp64_mat]: Any non-power-of-two Matrices up to 127 FMACs (or other FMA-style op), full triple-loop Schedule. See [[sv/remap]]
-[^svp64_fft]: DCT (Lee) and FFT Full Triple-loops supported, RADIX2-only. Normally only found in VLIW DSPs (TI MSP320, Qualcom Hexagon). See [[sv/remap]]
-[^vsx_vec_regs]: VSX's Vector Registers are mis-named: they are 100% PackedSIMD. AVX-512 is not a Vector ISA either. See [Flynn's Taxonomy](https://en.wikipedia.org/wiki/Flynn%27s_taxonomy)
-[^ppc_mma]: Power ISA v3.1 contains "Matrix Multiply Assist" (MMA) which due to PackedSIMD is restricted to RADIX2 and requires inline assembler loop-unrolling for non-power-of-two Matrix dimensions
-[^neon_opcodes]: difficult to ascertain, see [NEON/VFP](https://developer.arm.com/documentation/den0018/a/NEON-and-VFP-Instruction-Summary/List-of-all-NEON-and-VFP-instructions).
+[^1]: plus EXT001 24-bit prefixing using 25% of EXT001 space. See [[sv/svp64]]
+[^3]: A 2-Dimensional Scalable Vector ISA **specifically designed for the Power ISA** with both Horizontal-First and Vertical-First Modes. See [[sv/vector_isa_comparison]]
+[^4]: on specific operations. See [[opcode_regs_deduped]] for full list. Key: 2P - Twin Predication, 1P - Single-Predicate
+[^5]: SVP64 provides a Vector concept on top of the **Scalar** GPR, FPR and CR Fields, extended to 128 entries.
+[^6]: SVP64 Vectorises Scalar ops. It is up to the **implementor** to choose (**optionally**) whether to apply SVP64 to e.g. VSX Quad-Precision (128-bit) instructions, to create 128-bit Vector ops.
+[^7]: big-integer add is just `sv.adde`. For optimal performance Bigint Mul and divide first require addition of two scalar operations (in turn, naturally Vectorised by SVP64). See [[sv/biginteger/analysis]]
+[^8]: LD/ST Fault-First: see [[sv/svp64/appendix]] and [ARM SVE Fault-First](https://alastairreid.github.io/papers/sve-ieee-micro-2017.pdf)
+[^9]: Based on LD/ST Fail-first, extended to data. See [[sv/svp64/appendix]]
+[^10]: Predicate-result effectively turns any standard op into a type of "cmp". See [[sv/svp64/appendix]]
+[^11]: Any non-power-of-two Matrices up to 127 FMACs (or other FMA-style op), full triple-loop Schedule. See [[sv/remap]]
+[^12]: DCT (Lee) and FFT Full Triple-loops supported, RADIX2-only. Normally only found in VLIW DSPs (TI MSP320, Qualcom Hexagon). See [[sv/remap]]
+[^13]: VSX's Vector Registers are mis-named: they are 100% PackedSIMD. AVX-512 is not a Vector ISA either. See [Flynn's Taxonomy](https://en.wikipedia.org/wiki/Flynn%27s_taxonomy)
+[^14]: Power ISA v3.1 contains "Matrix Multiply Assist" (MMA) which due to PackedSIMD is restricted to RADIX2 and requires inline assembler loop-unrolling for non-power-of-two Matrix dimensions
+[^15]: difficult to ascertain, see [NEON/VFP](https://developer.arm.com/documentation/den0018/a/NEON-and-VFP-Instruction-Summary/List-of-all-NEON-and-VFP-instructions).
Critically depends on ARM Scalar instructions
-[^sve2_opcodes]: difficult to exactly ascertain, see ARM Architecture Reference Manual Supplement, DDI 0584. Critically depends on ARM Scalar instructions.
-[^sve2_no_setvl]: ARM states that the Scalability is a [Silicon-partner choice](https://developer.arm.com/-/media/Arm%20Developer%20Community/PDF/102340_0001_00_en_introduction-to-sve2.pdf?revision=aae96dd2-5334-4ad3-9a47-393086a20fea).
+[^16]: difficult to exactly ascertain, see ARM Architecture Reference Manual Supplement, DDI 0584. Critically depends on ARM Scalar instructions.
+[^17]: ARM states that the Scalability is a [Silicon-partner choice](https://developer.arm.com/-/media/Arm%20Developer%20Community/PDF/102340_0001_00_en_introduction-to-sve2.pdf?revision=aae96dd2-5334-4ad3-9a47-393086a20fea).
Scalability in the ISA is **not available to the programmer**: there is no `setvl` instruction in SVE2, which is already causing assembler programmer difficulties.
[quote](https://gist.github.com/zingaburga/805669eb891c820bd220418ee3f0d6bd#file-sve2-md) **"you may be stuck with only using the bottom 128 bits of the vector, or need to code specifically for each width"**
-[^avx512_wikipedia]: [AVX512 Wikipedia](https://en.wikipedia.org/wiki/AVX-512), [Lifecycle of an instruction set](https://media.handmade-seattle.com/tom-forsyth/) including full slides
-[^avx512_opcodes]: difficult to exactly ascertain, contains subsets. Critically depends on ISA support from earlier x86 ISA subsets (several more thousand instructions). See [SIMD ISA listing](https://www.officedaytime.com/simd512e/)
-[^rvv_spec]: [RVV Spec](https://github.com/riscv/riscv-v-spec/blob/master/v-spec.adoc)
-[^rvv_opcodes]: RISC-V Vectors are not stand-alone, i.e. like SVE2 and AVX-512 are critically dependent on the Scalar ISA (an additional ~96 instructions for the Scalar RV64GC set, needed for Linux).
-[^rvv_scalable]: Like the original Cray RVV is a truly scalable Vector ISA (Cray setvl instruction). However, like SVE2, the Maximum Vector length is a Silicon-partner choice, which creates similar limitations that SVP64 does not have.
+[^18]: [AVX512 Wikipedia](https://en.wikipedia.org/wiki/AVX-512), [Lifecycle of an instruction set](https://media.handmade-seattle.com/tom-forsyth/) including full slides
+[^19]: difficult to exactly ascertain, contains subsets. Critically depends on ISA support from earlier x86 ISA subsets (several more thousand instructions). See [SIMD ISA listing](https://www.officedaytime.com/simd512e/)
+[^20]: [RVV Spec](https://github.com/riscv/riscv-v-spec/blob/master/v-spec.adoc)
+[^21]: RISC-V Vectors are not stand-alone, i.e. like SVE2 and AVX-512 are critically dependent on the Scalar ISA (an additional ~96 instructions for the Scalar RV64GC set, needed for Linux).
+[^22]: Like the original Cray RVV is a truly scalable Vector ISA (Cray setvl instruction). However, like SVE2, the Maximum Vector length is a Silicon-partner choice, which creates similar limitations that SVP64 does not have.
The RISC-V Founders strongly discourage efforts by programmers to find out the Silicon's Maximum Vector Length, as an effort to steer programmers towards Silicon-independent assembler. This requires **all** algorithms to contain a loop construct.
MAXVL in SVP64 is a Spec-hard-fixed quantity therefore loop constructs are not necessary 100% of the time.
-[^rvv_128]: like SVP64 it is up to the hardware implementor (Silicon partner) to choose whether to support 128-bit elements.
-[^sx_aurora]: [NEC SX Aurora](https://ftp.libre-soc.org/NEC_SX_Aurora_TSUBASA_VectorEngine-as-manual-v1.2.pdf) is based on the original Cray Vectors
-[^aurora_isa]: [Aurora ISA guide](https://sxauroratsubasa.sakura.ne.jp/documents/guide/pdfs/Aurora_ISA_guide.pdf) Appendix-3 11.1 p508
-[^aurora_scalable]: Like the original Cray Vectors, the ISA Vector Length is independent of the underlying hardware, however Generation 1 has 256 elements per Vector register (3.2.4 p24, Aurora ISA guide)
-[^svp64_intrin_cnt]: If treated as a 1-Dimensional ISA, and designed badly, the 24-bit Prefix expands 200+ scalar instructions to well over a million intrinsics (N~=10^4 **times** M~=10^2).
+[^23]: like SVP64 it is up to the hardware implementor (Silicon partner) to choose whether to support 128-bit elements.
+[^24]: [NEC SX Aurora](https://ftp.libre-soc.org/NEC_SX_Aurora_TSUBASA_VectorEngine-as-manual-v1.2.pdf) is based on the original Cray Vectors
+[^25]: [Aurora ISA guide](https://sxauroratsubasa.sakura.ne.jp/documents/guide/pdfs/Aurora_ISA_guide.pdf) Appendix-3 11.1 p508
+[^26]: Like the original Cray Vectors, the ISA Vector Length is independent of the underlying hardware, however Generation 1 has 256 elements per Vector register (3.2.4 p24, Aurora ISA guide)
+[^2]: If treated as a 1-Dimensional ISA, and designed badly, the 24-bit Prefix expands 200+ scalar instructions to well over a million intrinsics (N~=10^4 **times** M~=10^2).
If treated as a 2-Dimensional ISA and designed well, there are far less. N prefix intrinsics **plus** M scalar instruction intrinsics, where N is likely to be of the order of 10^2 and M of the order of 10^2.
-[^vsx_intrin]: [Altivec gcc intrinsics](https://gcc.gnu.org/onlinedocs/gcc/PowerPC-AltiVec_002fVSX-Built-in-Functions.html), contains links to additional VSX intrinsics for ISA 2.05/6/7, 3.0 and 3.1
-[^neon_intrin]: NEON 32-bit 2754 intrinsics, NEON 64-bit 4334 intrinsics.
-[^sve2_intrin]: SVE: 4140 intrinsics, SVE2 1900 intrinsics
-[^avx512_intrin]: Count includes SSE, SSE2, AVX, AVX2 and all AVX512 variants
-[^rvv_intrin]: [RVV intrinsics listing](https://raw.githubusercontent.com/riscv-non-isa/rvv-intrinsic-doc/master/intrinsic_funcs.md) page is 25,000 lines long.
-[^aurora_intrin]: Unknown. estimated to be of the order of length of RVV due to also being a Cray-style Scalable ISA, NEC maintains an [LLVM hard fork](https://github.com/sx-aurora-dev)
-[^sve2_mat]: [Scalable Matrix Optional Extension](https://community.arm.com/arm-community-blogs/b/architectures-and-processors-blog/posts/scalable-matrix-extension-armv9-a-architecture)
+[^27]: [Altivec gcc intrinsics](https://gcc.gnu.org/onlinedocs/gcc/PowerPC-AltiVec_002fVSX-Built-in-Functions.html), contains links to additional VSX intrinsics for ISA 2.05/6/7, 3.0 and 3.1
+[^28]: NEON 32-bit 2754 intrinsics, NEON 64-bit 4334 intrinsics.
+[^29]: SVE: 4140 intrinsics, SVE2 1900 intrinsics
+[^30]: Count includes SSE, SSE2, AVX, AVX2 and all AVX512 variants
+[^31]: [RVV intrinsics listing](https://raw.githubusercontent.com/riscv-non-isa/rvv-intrinsic-doc/master/intrinsic_funcs.md) page is 25,000 lines long.
+[^32]: Unknown. estimated to be of the order of length of RVV due to also being a Cray-style Scalable ISA, NEC maintains an [LLVM hard fork](https://github.com/sx-aurora-dev)
+[^33]: [Scalable Matrix Optional Extension](https://community.arm.com/arm-community-blogs/b/architectures-and-processors-blog/posts/scalable-matrix-extension-armv9-a-architecture)
the key is an outer-product instruction [SMOPA](https://developer.arm.com/documentation/ddi0602/2022-06/SME-Instructions/SMOPA--Signed-integer-sum-of-outer-products-and-accumulate-?lang=en) which is very hard to tell at a glance if it is power-2 or non-power-2
-[^x86_amx]: [Advanced matrix Extensions](https://en.wikipedia.org/wiki/Advanced_Matrix_Extensions) supports BF16 and INT8 only. Separate regfile, power-of-two "tiles". Not general-purpose at all.
+[^34]: [Advanced matrix Extensions](https://en.wikipedia.org/wiki/Advanced_Matrix_Extensions) supports BF16 and INT8 only. Separate regfile, power-of-two "tiles". Not general-purpose at all.
projects / libreriscv.git / commitdiff