| Message ID | 20220118104338.2081259-3-giulio.benetti@benettiengineering.com |
|---|---|
| State | Rejected |
| Headers | show |
| Series | Use the best FPU strategies on 32-bits Arm Cortex | expand |
Hi! [Replying to list only, hoping it is the right etiquette.] Giulio Benetti wrote: > --- a/configs/acmesystems_acqua_a5_512mb_defconfig > +++ b/configs/acmesystems_acqua_a5_512mb_defconfig > @@ -1,6 +1,7 @@ > BR2_arm=y > BR2_cortex_a5=y > BR2_ARM_ENABLE_VFP=y > +BR2_ARM_FPU_NEON_VFPV4=y > BR2_ARM_INSTRUCTIONS_THUMB2=y > BR2_PACKAGE_HOST_LINUX_HEADERS_CUSTOM_5_4=y > BR2_SYSTEM_DHCP="eth0" I think this may be missing some dependencies, as this change does not survive `save defconfig'. I tried the following from a clean tree: git pull # from master git am ~/tmp/saved-patch.email # the patch above make acmesystems_acqua_a5_512mb_defconfig make nconfig # no changes to configuration <F6> Save <F9> Exit make savedefconfig git diff And git showed me that the patch had just been reverted on the worktree. According to arch/Config.in.arm, BR2_ARM_FPU_NEON_VFPV4 depends on BR2_ARM_CPU_HAS_NEON, which is not mentioned in my .config. It also appears that BR2_ARM_ENABLE_NEON selects BR2_ARM_CPU_HAS_NEON. Would BR2_ARM_ENABLE_NEON be the canonical way of making BR2_ARM_FPU_NEON_VFPV4 acceptable? I am a bit confused by the VFP selection logic in arch/Config.in.arm, so I am not sure that would be the right approach. As a follow up question, if this issue is fixed, what would be the simplest way of testing the floating point implementation on the board? I was thinking about adding gnuplot, which doesn't have any dependencies, and having it do some computations. I do not know, however, whether this would exercise the NEON SIMD extensions. Best regards, Edgar.
Hi again! Giulio Benetti wrote: > As pointed by SAMA5D2 Datasheet [...] So let's enable VFPV4/NEON FPU > strategy instead of the default VFPV4-D16. BTW, the Acqua is a SAMA5D3[1] (SAMA5D31 or SAMA5D36, depending on the version), not a SAMA5D2. The datasheet[2] of the SAMA5D3 states: The Floating-Point Unit (FPU) supports the ARMv7 VFPv4-D16 architecture without Advanced SIMD extensions (NEON). So I guess the patch is not relevant to this board. [1] https://www.acmesystems.it/acqua [2] https://ww1.microchip.com/downloads/en/DeviceDoc/SAMA5D3-Series-Data-sheet-DS60001609b.pdf
Hi Edgar, On 18/01/22 14:28, Edgar Bonet wrote: > Hi again! > > Giulio Benetti wrote: >> As pointed by SAMA5D2 Datasheet [...] So let's enable VFPV4/NEON FPU >> strategy instead of the default VFPV4-D16. > > BTW, the Acqua is a SAMA5D3[1] (SAMA5D31 or SAMA5D36, depending on the > version), not a SAMA5D2. The datasheet[2] of the SAMA5D3 states: > > The Floating-Point Unit (FPU) supports the ARMv7 VFPv4-D16 > architecture without Advanced SIMD extensions (NEON). Ah, thank you. I had the SAMA5D3 opened and I couldn't find anymore which board used it. I thought it was this one. Thanks for pointing. > So I guess the patch is not relevant to this board. Exactly. So please drop this patch and I've set it as rejected in Patchwork. I'll go in depth later with patch 01/28. Thank you for the answers! Best regards
diff --git a/configs/acmesystems_acqua_a5_512mb_defconfig b/configs/acmesystems_acqua_a5_512mb_defconfig index e399d4a9ff..56885ea5a1 100644 --- a/configs/acmesystems_acqua_a5_512mb_defconfig +++ b/configs/acmesystems_acqua_a5_512mb_defconfig @@ -1,6 +1,7 @@ BR2_arm=y BR2_cortex_a5=y BR2_ARM_ENABLE_VFP=y +BR2_ARM_FPU_NEON_VFPV4=y BR2_ARM_INSTRUCTIONS_THUMB2=y BR2_PACKAGE_HOST_LINUX_HEADERS_CUSTOM_5_4=y BR2_SYSTEM_DHCP="eth0"
As pointed by SAMA5D2 Datasheet[1]: ``` The Cortex-A5 NEON Media Processing Engine (MPE) extends the Cortex-A5 functionality to provide support for the ARM v7 Advanced SIMD v2 and Vector Floating-Point v4 (VFPv4) instruction sets. ``` So let's enable VFPV4/NEON FPU strategy instead of the default VFPV4-D16. [1]: https://ww1.microchip.com/downloads/aemDocuments/documents/MPU32/ProductDocuments/DataSheets/SAMA5D2-Series-Data-sheet-ds60001476G.pdf Signed-off-by: Giulio Benetti <giulio.benetti@benettiengineering.com> --- configs/acmesystems_acqua_a5_512mb_defconfig | 1 + 1 file changed, 1 insertion(+)