diff options
| author |   Shawn Landden <shawn@git.icu> | 2019-05-10 19:02:36 -0300 |
|---|---|---|
| committer | Shawn Landden <shawn@git.icu> | 2019-05-13 16:24:32 -0500 |
| commit | 2b2e9c08b326409a28ef3d0237164e5ee67c7c6d (patch) | |
| tree | d6ead58699723d4b42286696c18b42a14e832878 /src/compat | |
| parent | socket: set ignore_df=1 on xmit (diff) | |
| download | wireguard-monolithic-historical-2b2e9c08b326409a28ef3d0237164e5ee67c7c6d.tar.xz wireguard-monolithic-historical-2b2e9c08b326409a28ef3d0237164e5ee67c7c6d.zip | |
[Zinc] Add PowerPC chacha20 implementation from openssl/cryptograms
Only runs on the out-bound path, as the in-bound path is in an interrupt,
but that can be fixed in Linux.
Otherwise this is tested (with the VSX code enabled) on Power 9 (ppc64le).
Without this I get 2GiB/s over the loopback (so 4 GiB/s), and with this
I get 2.8 GiB/s (so 5.6 GiB/s), and more time is spent in poly1305 than chacha20.
This is on a 4-thread VPS.
Signed-off-by: Shawn Landden <shawn@git.icu>
v2: more complete simd.h for PPC
benchmarks
whitespace issues
v3: honor CONFIG_ALTIVEC (CONFIG_VSX has already honored)
Diffstat (limited to 'src/compat')
| -rw-r--r-- | src/compat/simd/include/linux/simd.h | 29 |
1 files changed, 27 insertions, 2 deletions
diff --git a/src/compat/simd/include/linux/simd.h b/src/compat/simd/include/linux/simd.h index c75c724..44060a9 100644 --- a/src/compat/simd/include/linux/simd.h +++ b/src/compat/simd/include/linux/simd.h @@ -13,6 +13,9 @@ #include <asm/fpu/api.h> #elif defined(CONFIG_KERNEL_MODE_NEON) #include <asm/neon.h> +#elif defined(CONFIG_ALTIVEC) || defined(CONFIG_VSX) +#include <asm/switch_to.h> +#include <asm/cputable.h> #endif typedef enum { @@ -30,13 +33,24 @@ static inline void simd_get(simd_context_t *ctx) static inline void simd_put(simd_context_t *ctx) { + if (*ctx & HAVE_SIMD_IN_USE) { #if defined(CONFIG_X86_64) - if (*ctx & HAVE_SIMD_IN_USE) kernel_fpu_end(); #elif defined(CONFIG_KERNEL_MODE_NEON) - if (*ctx & HAVE_SIMD_IN_USE) kernel_neon_end(); +#elif defined(CONFIG_PPC32) || defined(CONFIG_PPC64) + if (cpu_has_feature(CPU_FTR_VSX_COMP)) { + disable_kernel_vsx(); + preempt_enable(); + } else if (cpu_has_feature(CPU_FTR_ALTIVEC_COMP)) { + disable_kernel_altivec(); + preempt_enable(); + } else if (!cpu_has_feature(CPU_FTR_FPU_UNAVAILABLE)) { + disable_kernel_fp(); + preempt_enable(); + } #endif + } *ctx = HAVE_NO_SIMD; } @@ -62,6 +76,17 @@ static __must_check inline bool simd_use(simd_context_t *ctx) kernel_fpu_begin(); #elif defined(CONFIG_KERNEL_MODE_NEON) kernel_neon_begin(); +#elif defined(CONFIG_PPC32) || defined(CONFIG_PPC64) + if (cpu_has_feature(CPU_FTR_VSX_COMP)) { + preempt_disable(); + enable_kernel_vsx(); + } else if (cpu_has_feature(CPU_FTR_ALTIVEC_COMP)) { + preempt_disable(); + enable_kernel_altivec(); + } else if (!cpu_has_feature(CPU_FTR_FPU_UNAVAILABLE)) { + preempt_disable(); + enable_kernel_fp(); + } #endif *ctx |= HAVE_SIMD_IN_USE; return true; |