| Message ID | 20230905-optimize_checksum-v2-3-ccd658db743b@rivosinc.com (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | riscv: Add fine-tuned checksum functions | expand |
On Tue, Sep 05, 2023 at 09:46:52PM -0700, Charlie Jenkins wrote: > Vector code is written in assembly rather than using the GCC vector > instrinsics because they did not provide optimal code. Vector > instrinsic types are still used so the inline assembly can > appropriately select vector registers. However, this code cannot be > merged yet because it is currently not possible to use vector > instrinsics in the kernel because vector support needs to be directly > enabled by assembly. > > Signed-off-by: Charlie Jenkins <charlie@rivosinc.com> > --- > arch/riscv/include/asm/checksum.h | 87 +++++++++++++++++++++++++++++++++++++++ > 1 file changed, 87 insertions(+) > > diff --git a/arch/riscv/include/asm/checksum.h b/arch/riscv/include/asm/checksum.h > index 3f9d5a202e95..1d6c23cd1221 100644 > --- a/arch/riscv/include/asm/checksum.h > +++ b/arch/riscv/include/asm/checksum.h > @@ -10,6 +10,10 @@ > #include <linux/in6.h> > #include <linux/uaccess.h> > > +#ifdef CONFIG_RISCV_ISA_V > +#include <riscv_vector.h> > +#endif > + > #ifdef CONFIG_32BIT > typedef unsigned int csum_t; > #else > @@ -43,6 +47,89 @@ static inline __sum16 csum_fold(__wsum sum) > */ > static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl) > { > +#ifdef CONFIG_RISCV_ISA_V > + if (IS_ENABLED(CONFIG_RISCV_ALTERNATIVE)) { > + /* > + * Vector is likely available when the kernel is compiled with > + * vector support, so nop when vector is available and jump when > + * vector is not available. > + */ > + asm_volatile_goto(ALTERNATIVE("j %l[no_vector]", "nop", 0, > + RISCV_ISA_EXT_v, 1) > + : > + : > + : > + : no_vector); > + } else { > + if (!__riscv_isa_extension_available(NULL, RISCV_ISA_EXT_v)) > + goto no_vector; > + } Silly question maybe, but is this complexity required? If you were to go and do if (!has_vector()) goto no_vector is there any meaningful difference difference in performance? > + > + vuint64m1_t prev_buffer; > + vuint32m1_t curr_buffer; > + unsigned int vl; > +#ifdef CONFIG_32_BIT > + csum_t high_result, low_result; > + > + riscv_v_enable(); > + asm(".option push \n\ > + .option arch, +v \n\ > + vsetivli x0, 1, e64, ta, ma \n\ > + vmv.v.i %[prev_buffer], 0 \n\ > + 1: \n\ > + vsetvli %[vl], %[ihl], e32, m1, ta, ma \n\ > + vle32.v %[curr_buffer], (%[iph]) \n\ > + vwredsumu.vs %[prev_buffer], %[curr_buffer], %[prev_buffer] \n\ > + sub %[ihl], %[ihl], %[vl] \n\ > + slli %[vl], %[vl], 2 \n\ Also, could you please try to align the operands for asm stuff? It makes quite a difference to readability. Thanks, Conor. > + add %[iph], %[vl], %[iph] \n\ > + # If not all of iph could fit into vector reg, do another sum \n\ > + bne %[ihl], zero, 1b \n\ > + vsetivli x0, 1, e64, m1, ta, ma \n\ > + vmv.x.s %[low_result], %[prev_buffer] \n\ > + addi %[vl], x0, 32 \n\ > + vsrl.vx %[prev_buffer], %[prev_buffer], %[vl] \n\ > + vmv.x.s %[high_result], %[prev_buffer] \n\ > + .option pop" > + : [vl] "=&r" (vl), [prev_buffer] "=&vd" (prev_buffer), > + [curr_buffer] "=&vd" (curr_buffer), > + [high_result] "=&r" (high_result), > + [low_result] "=&r" (low_result) > + : [iph] "r" (iph), [ihl] "r" (ihl)); > + riscv_v_disable(); > + > + high_result += low_result; > + high_result += high_result < low_result; > +#else // !CONFIG_32_BIT > + csum_t result; > + > + riscv_v_enable(); > + asm(".option push \n\ > + .option arch, +v \n\ > + vsetivli x0, 1, e64, ta, ma \n\ > + vmv.v.i %[prev_buffer], 0 \n\ > + 1: \n\ > + # Setup 32-bit sum of iph \n\ > + vsetvli %[vl], %[ihl], e32, m1, ta, ma \n\ > + vle32.v %[curr_buffer], (%[iph]) \n\ > + # Sum each 32-bit segment of iph that can fit into a vector reg \n\ > + vwredsumu.vs %[prev_buffer], %[curr_buffer], %[prev_buffer] \n\ > + subw %[ihl], %[ihl], %[vl] \n\ > + slli %[vl], %[vl], 2 \n\ > + addw %[iph], %[vl], %[iph] \n\ > + # If not all of iph could fit into vector reg, do another sum \n\ > + bne %[ihl], zero, 1b \n\ > + vsetvli x0, x0, e64, m1, ta, ma \n\ > + vmv.x.s %[result], %[prev_buffer] \n\ > + .option pop" > + : [vl] "=&r" (vl), [prev_buffer] "=&vd" (prev_buffer), > + [curr_buffer] "=&vd" (curr_buffer), [result] "=&r" (result) > + : [iph] "r" (iph), [ihl] "r" (ihl)); > + riscv_v_disable(); > +#endif // !CONFIG_32_BIT > +no_vector: > +#endif // !CONFIG_RISCV_ISA_V > + > csum_t csum = 0; > int pos = 0; > > > -- > 2.42.0 >
On Tue, Sep 05, 2023 at 09:46:52PM -0700, Charlie Jenkins wrote: > Vector code is written in assembly rather than using the GCC vector > instrinsics because they did not provide optimal code. Vector > instrinsic types are still used so the inline assembly can > appropriately select vector registers. However, this code cannot be > merged yet because it is currently not possible to use vector > instrinsics in the kernel because vector support needs to be directly > enabled by assembly. > > Signed-off-by: Charlie Jenkins <charlie@rivosinc.com> > --- > arch/riscv/include/asm/checksum.h | 87 +++++++++++++++++++++++++++++++++++++++ > 1 file changed, 87 insertions(+) > > diff --git a/arch/riscv/include/asm/checksum.h b/arch/riscv/include/asm/checksum.h > index 3f9d5a202e95..1d6c23cd1221 100644 > --- a/arch/riscv/include/asm/checksum.h > +++ b/arch/riscv/include/asm/checksum.h > @@ -10,6 +10,10 @@ > #include <linux/in6.h> > #include <linux/uaccess.h> > > +#ifdef CONFIG_RISCV_ISA_V > +#include <riscv_vector.h> > +#endif > + > #ifdef CONFIG_32BIT > typedef unsigned int csum_t; > #else > @@ -43,6 +47,89 @@ static inline __sum16 csum_fold(__wsum sum) > */ > static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl) > { > +#ifdef CONFIG_RISCV_ISA_V > + if (IS_ENABLED(CONFIG_RISCV_ALTERNATIVE)) { > + /* > + * Vector is likely available when the kernel is compiled with > + * vector support, so nop when vector is available and jump when > + * vector is not available. > + */ > + asm_volatile_goto(ALTERNATIVE("j %l[no_vector]", "nop", 0, > + RISCV_ISA_EXT_v, 1) > + : > + : > + : > + : no_vector); > + } else { > + if (!__riscv_isa_extension_available(NULL, RISCV_ISA_EXT_v)) > + goto no_vector; > + } > + > + vuint64m1_t prev_buffer; > + vuint32m1_t curr_buffer; > + unsigned int vl; > +#ifdef CONFIG_32_BIT > + csum_t high_result, low_result; > + > + riscv_v_enable(); > + asm(".option push \n\ > + .option arch, +v \n\ > + vsetivli x0, 1, e64, ta, ma \n\ Also, I don't see that you have addressed previous review comments from Samuel: https://lore.kernel.org/linux-riscv/0a8c98bf-46da-e77a-0431-a6c1e224af2e@sifive.com/
On Thu, Sep 07, 2023 at 10:58:33AM +0100, Conor Dooley wrote: > On Tue, Sep 05, 2023 at 09:46:52PM -0700, Charlie Jenkins wrote: > > Vector code is written in assembly rather than using the GCC vector > > instrinsics because they did not provide optimal code. Vector > > instrinsic types are still used so the inline assembly can > > appropriately select vector registers. However, this code cannot be > > merged yet because it is currently not possible to use vector > > instrinsics in the kernel because vector support needs to be directly > > enabled by assembly. > > > > Signed-off-by: Charlie Jenkins <charlie@rivosinc.com> > > --- > > arch/riscv/include/asm/checksum.h | 87 +++++++++++++++++++++++++++++++++++++++ > > 1 file changed, 87 insertions(+) > > > > diff --git a/arch/riscv/include/asm/checksum.h b/arch/riscv/include/asm/checksum.h > > index 3f9d5a202e95..1d6c23cd1221 100644 > > --- a/arch/riscv/include/asm/checksum.h > > +++ b/arch/riscv/include/asm/checksum.h > > @@ -10,6 +10,10 @@ > > #include <linux/in6.h> > > #include <linux/uaccess.h> > > > > +#ifdef CONFIG_RISCV_ISA_V > > +#include <riscv_vector.h> > > +#endif > > + > > #ifdef CONFIG_32BIT > > typedef unsigned int csum_t; > > #else > > @@ -43,6 +47,89 @@ static inline __sum16 csum_fold(__wsum sum) > > */ > > static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl) > > { > > +#ifdef CONFIG_RISCV_ISA_V > > + if (IS_ENABLED(CONFIG_RISCV_ALTERNATIVE)) { > > + /* > > + * Vector is likely available when the kernel is compiled with > > + * vector support, so nop when vector is available and jump when > > + * vector is not available. > > + */ > > + asm_volatile_goto(ALTERNATIVE("j %l[no_vector]", "nop", 0, > > + RISCV_ISA_EXT_v, 1) > > + : > > + : > > + : > > + : no_vector); > > + } else { > > + if (!__riscv_isa_extension_available(NULL, RISCV_ISA_EXT_v)) > > + goto no_vector; > > + } > > + > > + vuint64m1_t prev_buffer; > > + vuint32m1_t curr_buffer; > > + unsigned int vl; > > +#ifdef CONFIG_32_BIT > > + csum_t high_result, low_result; > > + > > + riscv_v_enable(); > > + asm(".option push \n\ > > + .option arch, +v \n\ > > + vsetivli x0, 1, e64, ta, ma \n\ > > Also, I don't see that you have addressed previous review comments from > Samuel: > https://lore.kernel.org/linux-riscv/0a8c98bf-46da-e77a-0431-a6c1e224af2e@sifive.com/ I added the check for vector as Samuel suggested, but then I accidentally used riscv_v_enable() instead of kernel_vector_begin(), I will make that change. - Charlie
On Thu, Sep 07, 2023 at 10:47:55AM +0100, Conor Dooley wrote: > On Tue, Sep 05, 2023 at 09:46:52PM -0700, Charlie Jenkins wrote: > > Vector code is written in assembly rather than using the GCC vector > > instrinsics because they did not provide optimal code. Vector > > instrinsic types are still used so the inline assembly can > > appropriately select vector registers. However, this code cannot be > > merged yet because it is currently not possible to use vector > > instrinsics in the kernel because vector support needs to be directly > > enabled by assembly. > > > > Signed-off-by: Charlie Jenkins <charlie@rivosinc.com> > > --- > > arch/riscv/include/asm/checksum.h | 87 +++++++++++++++++++++++++++++++++++++++ > > 1 file changed, 87 insertions(+) > > > > diff --git a/arch/riscv/include/asm/checksum.h b/arch/riscv/include/asm/checksum.h > > index 3f9d5a202e95..1d6c23cd1221 100644 > > --- a/arch/riscv/include/asm/checksum.h > > +++ b/arch/riscv/include/asm/checksum.h > > @@ -10,6 +10,10 @@ > > #include <linux/in6.h> > > #include <linux/uaccess.h> > > > > +#ifdef CONFIG_RISCV_ISA_V > > +#include <riscv_vector.h> > > +#endif > > + > > #ifdef CONFIG_32BIT > > typedef unsigned int csum_t; > > #else > > @@ -43,6 +47,89 @@ static inline __sum16 csum_fold(__wsum sum) > > */ > > static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl) > > { > > +#ifdef CONFIG_RISCV_ISA_V > > + if (IS_ENABLED(CONFIG_RISCV_ALTERNATIVE)) { > > + /* > > + * Vector is likely available when the kernel is compiled with > > + * vector support, so nop when vector is available and jump when > > + * vector is not available. > > + */ > > + asm_volatile_goto(ALTERNATIVE("j %l[no_vector]", "nop", 0, > > + RISCV_ISA_EXT_v, 1) > > + : > > + : > > + : > > + : no_vector); > > + } else { > > + if (!__riscv_isa_extension_available(NULL, RISCV_ISA_EXT_v)) > > + goto no_vector; > > + } > > Silly question maybe, but is this complexity required? > If you were to go and do > if (!has_vector()) > goto no_vector > is there any meaningful difference difference in performance? Yes I should use that instead. > > > > + > > + vuint64m1_t prev_buffer; > > + vuint32m1_t curr_buffer; > > + unsigned int vl; > > +#ifdef CONFIG_32_BIT > > + csum_t high_result, low_result; > > + > > + riscv_v_enable(); > > + asm(".option push \n\ > > + .option arch, +v \n\ > > + vsetivli x0, 1, e64, ta, ma \n\ > > + vmv.v.i %[prev_buffer], 0 \n\ > > + 1: \n\ > > + vsetvli %[vl], %[ihl], e32, m1, ta, ma \n\ > > + vle32.v %[curr_buffer], (%[iph]) \n\ > > + vwredsumu.vs %[prev_buffer], %[curr_buffer], %[prev_buffer] \n\ > > + sub %[ihl], %[ihl], %[vl] \n\ > > + slli %[vl], %[vl], 2 \n\ > > Also, could you please try to align the operands for asm stuff? > It makes quite a difference to readability. > > Thanks, > Conor. > Will do. - Charlie > > + add %[iph], %[vl], %[iph] \n\ > > + # If not all of iph could fit into vector reg, do another sum \n\ > > + bne %[ihl], zero, 1b \n\ > > + vsetivli x0, 1, e64, m1, ta, ma \n\ > > + vmv.x.s %[low_result], %[prev_buffer] \n\ > > + addi %[vl], x0, 32 \n\ > > + vsrl.vx %[prev_buffer], %[prev_buffer], %[vl] \n\ > > + vmv.x.s %[high_result], %[prev_buffer] \n\ > > + .option pop" > > + : [vl] "=&r" (vl), [prev_buffer] "=&vd" (prev_buffer), > > + [curr_buffer] "=&vd" (curr_buffer), > > + [high_result] "=&r" (high_result), > > + [low_result] "=&r" (low_result) > > + : [iph] "r" (iph), [ihl] "r" (ihl)); > > + riscv_v_disable(); > > + > > + high_result += low_result; > > + high_result += high_result < low_result; > > +#else // !CONFIG_32_BIT > > + csum_t result; > > + > > + riscv_v_enable(); > > + asm(".option push \n\ > > + .option arch, +v \n\ > > + vsetivli x0, 1, e64, ta, ma \n\ > > + vmv.v.i %[prev_buffer], 0 \n\ > > + 1: \n\ > > + # Setup 32-bit sum of iph \n\ > > + vsetvli %[vl], %[ihl], e32, m1, ta, ma \n\ > > + vle32.v %[curr_buffer], (%[iph]) \n\ > > + # Sum each 32-bit segment of iph that can fit into a vector reg \n\ > > + vwredsumu.vs %[prev_buffer], %[curr_buffer], %[prev_buffer] \n\ > > + subw %[ihl], %[ihl], %[vl] \n\ > > + slli %[vl], %[vl], 2 \n\ > > + addw %[iph], %[vl], %[iph] \n\ > > + # If not all of iph could fit into vector reg, do another sum \n\ > > + bne %[ihl], zero, 1b \n\ > > + vsetvli x0, x0, e64, m1, ta, ma \n\ > > + vmv.x.s %[result], %[prev_buffer] \n\ > > + .option pop" > > + : [vl] "=&r" (vl), [prev_buffer] "=&vd" (prev_buffer), > > + [curr_buffer] "=&vd" (curr_buffer), [result] "=&r" (result) > > + : [iph] "r" (iph), [ihl] "r" (ihl)); > > + riscv_v_disable(); > > +#endif // !CONFIG_32_BIT > > +no_vector: > > +#endif // !CONFIG_RISCV_ISA_V > > + > > csum_t csum = 0; > > int pos = 0; > > > > > > -- > > 2.42.0 > >
diff --git a/arch/riscv/include/asm/checksum.h b/arch/riscv/include/asm/checksum.h index 3f9d5a202e95..1d6c23cd1221 100644 --- a/arch/riscv/include/asm/checksum.h +++ b/arch/riscv/include/asm/checksum.h @@ -10,6 +10,10 @@ #include <linux/in6.h> #include <linux/uaccess.h> +#ifdef CONFIG_RISCV_ISA_V +#include <riscv_vector.h> +#endif + #ifdef CONFIG_32BIT typedef unsigned int csum_t; #else @@ -43,6 +47,89 @@ static inline __sum16 csum_fold(__wsum sum) */ static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl) { +#ifdef CONFIG_RISCV_ISA_V + if (IS_ENABLED(CONFIG_RISCV_ALTERNATIVE)) { + /* + * Vector is likely available when the kernel is compiled with + * vector support, so nop when vector is available and jump when + * vector is not available. + */ + asm_volatile_goto(ALTERNATIVE("j %l[no_vector]", "nop", 0, + RISCV_ISA_EXT_v, 1) + : + : + : + : no_vector); + } else { + if (!__riscv_isa_extension_available(NULL, RISCV_ISA_EXT_v)) + goto no_vector; + } + + vuint64m1_t prev_buffer; + vuint32m1_t curr_buffer; + unsigned int vl; +#ifdef CONFIG_32_BIT + csum_t high_result, low_result; + + riscv_v_enable(); + asm(".option push \n\ + .option arch, +v \n\ + vsetivli x0, 1, e64, ta, ma \n\ + vmv.v.i %[prev_buffer], 0 \n\ + 1: \n\ + vsetvli %[vl], %[ihl], e32, m1, ta, ma \n\ + vle32.v %[curr_buffer], (%[iph]) \n\ + vwredsumu.vs %[prev_buffer], %[curr_buffer], %[prev_buffer] \n\ + sub %[ihl], %[ihl], %[vl] \n\ + slli %[vl], %[vl], 2 \n\ + add %[iph], %[vl], %[iph] \n\ + # If not all of iph could fit into vector reg, do another sum \n\ + bne %[ihl], zero, 1b \n\ + vsetivli x0, 1, e64, m1, ta, ma \n\ + vmv.x.s %[low_result], %[prev_buffer] \n\ + addi %[vl], x0, 32 \n\ + vsrl.vx %[prev_buffer], %[prev_buffer], %[vl] \n\ + vmv.x.s %[high_result], %[prev_buffer] \n\ + .option pop" + : [vl] "=&r" (vl), [prev_buffer] "=&vd" (prev_buffer), + [curr_buffer] "=&vd" (curr_buffer), + [high_result] "=&r" (high_result), + [low_result] "=&r" (low_result) + : [iph] "r" (iph), [ihl] "r" (ihl)); + riscv_v_disable(); + + high_result += low_result; + high_result += high_result < low_result; +#else // !CONFIG_32_BIT + csum_t result; + + riscv_v_enable(); + asm(".option push \n\ + .option arch, +v \n\ + vsetivli x0, 1, e64, ta, ma \n\ + vmv.v.i %[prev_buffer], 0 \n\ + 1: \n\ + # Setup 32-bit sum of iph \n\ + vsetvli %[vl], %[ihl], e32, m1, ta, ma \n\ + vle32.v %[curr_buffer], (%[iph]) \n\ + # Sum each 32-bit segment of iph that can fit into a vector reg \n\ + vwredsumu.vs %[prev_buffer], %[curr_buffer], %[prev_buffer] \n\ + subw %[ihl], %[ihl], %[vl] \n\ + slli %[vl], %[vl], 2 \n\ + addw %[iph], %[vl], %[iph] \n\ + # If not all of iph could fit into vector reg, do another sum \n\ + bne %[ihl], zero, 1b \n\ + vsetvli x0, x0, e64, m1, ta, ma \n\ + vmv.x.s %[result], %[prev_buffer] \n\ + .option pop" + : [vl] "=&r" (vl), [prev_buffer] "=&vd" (prev_buffer), + [curr_buffer] "=&vd" (curr_buffer), [result] "=&r" (result) + : [iph] "r" (iph), [ihl] "r" (ihl)); + riscv_v_disable(); +#endif // !CONFIG_32_BIT +no_vector: +#endif // !CONFIG_RISCV_ISA_V + csum_t csum = 0; int pos = 0;
Vector code is written in assembly rather than using the GCC vector instrinsics because they did not provide optimal code. Vector instrinsic types are still used so the inline assembly can appropriately select vector registers. However, this code cannot be merged yet because it is currently not possible to use vector instrinsics in the kernel because vector support needs to be directly enabled by assembly. Signed-off-by: Charlie Jenkins <charlie@rivosinc.com> --- arch/riscv/include/asm/checksum.h | 87 +++++++++++++++++++++++++++++++++++++++ 1 file changed, 87 insertions(+)