4 ms·

> It lacks many features required for high performance without hardware of excessive complexity, the most obvious being that RISC-V does not have decent address

by hajile 5y ago

> It lacks many features required for high performance without hardware of excessive complexity, the most obvious being that RISC-V does not have decent addressing modes, so that the highest performance implementation reported until now (by Alibaba at the 2020 Hot Chips Conference) had to add a non-standard ISA extension to correct this. Do you have any proof that additional addressing modes would increase performance? What I remember from college and papers I've read is that more modes are generally a detriment to pipelining. Maybe you're talking about "needing more instructions". You'd be wrong in this case. Using the compact instruction set gives an average of 15% more dense code compared to x86 and somewhere around 25-35% more dense compared to aarch64 (about equivalent density to thumb, but without the switching overhead that makes thumb slow).

acidbaseextract 5y ago
https://conferences.computer.org/isca/pdfs/ISCA2020-4QlDegUf3fKiwUXfV0KdCm/466100a052/466100a052.pdf https://conferences.computer.org/isca/pdfs/ISCA2020-4QlDegUf... "Xuantie-910: A Commercial Multi-Core 12-Stage Pipeline Out-of-Order 64-bit High Performance RISC-V Processor with Vector Extension" From "VIII. NON-STANDARD INSTRUCTION SET EXTENSION" Targeting at various industrial applications, XT-910 enables a set of custom non-standard instructions, additional to the standard RISC-V instructions. The non-standard instruction extension can be categorized into two groups based on the purposes - memory access enhancement, and basic arithmetic operation enhancement. A. Memory access enhancement Memory access instructions usually account for a high proportion in the total number of instructions, so enhancing memory access instructions can directly benefit the overall performance. By analyzing the mainstream applications running on RISC-V, we observed that for the basic RISC-V instructions there is still quite some room for improvement in memory access related instructions. First, we support register + register addressing mode, and support indexed load and store instructions. This type of instruction extension reduces the usage of the registers for calculation and reduces the number of instructions for address generation, thereby effectively accelerating the data access of a loop body. Second, unsigned extension during address generation is supported. Otherwise, the basic instruction set does not support direct unsigned extension from 32-bit data to 64-bit data, resulting in too many shift instructions. I'm not certain if they break out the results by individual optimization, but in Fig. 20 in the paper it looks like all their tweaks add up to a 20% boost over the standard RISC-V ISA. Pretty huge.
- snvzz 5y ago
  >in Fig. 20 in the paper it looks like all their tweaks add up to a 20% boost over the standard RISC-V ISA. Pretty huge. >I'm not certain if they break out the results by individual optimization They don't, so it's not clear where the 20% comes from. They're using a compiler that optimizes for their microarchitecture, which might account for a lot of that. As for the mention of "extensions", it's not even clear to me whether they mean official RISC-V extensions such as IMAFD or custom extensions.

by hajile 5y ago

acidbaseextract 5y ago

https://conferences.computer.org/isca/pdfs/ISCA2020-4QlDegUf3fKiwUXfV0KdCm/466100a052/466100a052.pdf https://conferences.computer.org/isca/pdfs/ISCA2020-4QlDegUf... "Xuantie-910: A Commercial Multi-Core 12-Stage Pipeline Out-of-Order 64-bit High Performance RISC-V Processor with Vector Extension" From "VIII. NON-STANDARD INSTRUCTION SET EXTENSION" Targeting at various industrial applications, XT-910 enables a set of custom non-standard instructions, additional to the standard RISC-V instructions. The non-standard instruction extension can be categorized into two groups based on the purposes - memory access enhancement, and basic arithmetic operation enhancement. A. Memory access enhancement Memory access instructions usually account for a high proportion in the total number of instructions, so enhancing memory access instructions can directly benefit the overall performance. By analyzing the mainstream applications running on RISC-V, we observed that for the basic RISC-V instructions there is still quite some room for improvement in memory access related instructions. First, we support register + register addressing mode, and support indexed load and store instructions. This type of instruction extension reduces the usage of the registers for calculation and reduces the number of instructions for address generation, thereby effectively accelerating the data access of a loop body. Second, unsigned extension during address generation is supported. Otherwise, the basic instruction set does not support direct unsigned extension from 32-bit data to 64-bit data, resulting in too many shift instructions. I'm not certain if they break out the results by individual optimization, but in Fig. 20 in the paper it looks like all their tweaks add up to a 20% boost over the standard RISC-V ISA. Pretty huge.

snvzz 5y ago
>in Fig. 20 in the paper it looks like all their tweaks add up to a 20% boost over the standard RISC-V ISA. Pretty huge. >I'm not certain if they break out the results by individual optimization They don't, so it's not clear where the 20% comes from. They're using a compiler that optimizes for their microarchitecture, which might account for a lot of that. As for the mention of "extensions", it's not even clear to me whether they mean official RISC-V extensions such as IMAFD or custom extensions.