[1/3] net/hinic: restore vectorised code
Checks
Commit Message
Following make support removal, the vectorised code is not built
anymore, fix the build flag check.
Fixes: 3cc6ecfdfe85 ("build: remove makefiles")
Cc: stable@dpdk.org
Signed-off-by: David Marchand <david.marchand@redhat.com>
---
drivers/net/hinic/hinic_pmd_rx.c | 6 +++---
drivers/net/hinic/hinic_pmd_tx.c | 10 +++++-----
2 files changed, 8 insertions(+), 8 deletions(-)
Comments
> -----Original Message-----
> From: David Marchand <david.marchand@redhat.com>
> Sent: Friday, January 15, 2021 9:40 PM
> To: dev@dpdk.org
> Cc: ferruh.yigit@intel.com; stable@dpdk.org; Ziyang Xuan
> <xuanziyang2@huawei.com>; Xiaoyun Wang
> <cloud.wangxiaoyun@huawei.com>; Guoyang Zhou
> <zhouguoyang@huawei.com>; Ciara Power <ciara.power@intel.com>;
> Ruifeng Wang <Ruifeng.Wang@arm.com>; thomas@monjalon.net
> Subject: [PATCH 1/3] net/hinic: restore vectorised code
>
> Following make support removal, the vectorised code is not built anymore,
> fix the build flag check.
>
> Fixes: 3cc6ecfdfe85 ("build: remove makefiles")
> Cc: stable@dpdk.org
>
> Signed-off-by: David Marchand <david.marchand@redhat.com>
> ---
> drivers/net/hinic/hinic_pmd_rx.c | 6 +++---
> drivers/net/hinic/hinic_pmd_tx.c | 10 +++++-----
> 2 files changed, 8 insertions(+), 8 deletions(-)
>
> diff --git a/drivers/net/hinic/hinic_pmd_rx.c
> b/drivers/net/hinic/hinic_pmd_rx.c
> index a49769a863..842399cc4c 100644
> --- a/drivers/net/hinic/hinic_pmd_rx.c
> +++ b/drivers/net/hinic/hinic_pmd_rx.c
> @@ -4,7 +4,7 @@
>
> #include <rte_ether.h>
> #include <rte_mbuf.h>
> -#ifdef __ARM64_NEON__
> +#ifdef RTE_ARCH_ARM64
We can test '__ARM_NEON' which will be defined by compilers.
https://developer.arm.com/documentation/ihi0053/latest/
> #include <arm_neon.h>
> #endif
>
> @@ -762,7 +762,7 @@ void hinic_free_all_rx_mbufs(struct hinic_rxq *rxq)
> static inline void hinic_rq_cqe_be_to_cpu32(void *dst_le32,
> volatile void *src_be32)
> {
> -#if defined(__X86_64_SSE__)
> +#if defined(RTE_ARCH_X86_64)
> volatile __m128i *wqe_be = (volatile __m128i *)src_be32;
> __m128i *wqe_le = (__m128i *)dst_le32;
> __m128i shuf_mask = _mm_set_epi8(12, 13, 14, 15, 8, 9, 10, @@ -
> 770,7 +770,7 @@ static inline void hinic_rq_cqe_be_to_cpu32(void *dst_le32,
>
> /* l2nic just use first 128 bits */
> wqe_le[0] = _mm_shuffle_epi8(wqe_be[0], shuf_mask); -#elif
> defined(__ARM64_NEON__)
> +#elif defined(RTE_ARCH_ARM64)
> volatile uint8x16_t *wqe_be = (volatile uint8x16_t *)src_be32;
> uint8x16_t *wqe_le = (uint8x16_t *)dst_le32;
> const uint8x16_t shuf_mask = {3, 2, 1, 0, 7, 6, 5, 4, 11, 10, diff --git
> a/drivers/net/hinic/hinic_pmd_tx.c b/drivers/net/hinic/hinic_pmd_tx.c
> index 9d0264e67a..669f82389c 100644
> --- a/drivers/net/hinic/hinic_pmd_tx.c
> +++ b/drivers/net/hinic/hinic_pmd_tx.c
> @@ -7,7 +7,7 @@
> #include <rte_sctp.h>
> #include <rte_udp.h>
> #include <rte_ip.h>
> -#ifdef __ARM64_NEON__
> +#ifdef RTE_ARCH_ARM64
> #include <arm_neon.h>
> #endif
>
> @@ -203,7 +203,7 @@
>
> static inline void hinic_sq_wqe_cpu_to_be32(void *data, int nr_wqebb) { -
> #if defined(__X86_64_SSE__)
> +#if defined(RTE_ARCH_X86_64)
> int i;
> __m128i *wqe_line = (__m128i *)data;
> __m128i shuf_mask = _mm_set_epi8(12, 13, 14, 15, 8, 9, 10, @@ -
> 217,7 +217,7 @@ static inline void hinic_sq_wqe_cpu_to_be32(void *data,
> int nr_wqebb)
> wqe_line[3] = _mm_shuffle_epi8(wqe_line[3], shuf_mask);
> wqe_line += 4;
> }
> -#elif defined(__ARM64_NEON__)
> +#elif defined(RTE_ARCH_ARM64)
> int i;
> uint8x16_t *wqe_line = (uint8x16_t *)data;
> const uint8x16_t shuf_mask = {3, 2, 1, 0, 7, 6, 5, 4, 11, 10, @@ -237,7
> +237,7 @@ static inline void hinic_sq_wqe_cpu_to_be32(void *data, int
> nr_wqebb)
>
> static inline void hinic_sge_cpu_to_be32(void *data, int nr_sge) { -#if
> defined(__X86_64_SSE__)
> +#if defined(RTE_ARCH_X86_64)
> int i;
> __m128i *sge_line = (__m128i *)data;
> __m128i shuf_mask = _mm_set_epi8(12, 13, 14, 15, 8, 9, 10, @@ -
> 248,7 +248,7 @@ static inline void hinic_sge_cpu_to_be32(void *data, int
> nr_sge)
> *sge_line = _mm_shuffle_epi8(*sge_line, shuf_mask);
> sge_line++;
> }
> -#elif defined(__ARM64_NEON__)
> +#elif defined(RTE_ARCH_ARM64)
> int i;
> uint8x16_t *sge_line = (uint8x16_t *)data;
> const uint8x16_t shuf_mask = {3, 2, 1, 0, 7, 6, 5, 4, 11, 10,
> --
> 2.23.0
On Sat, Jan 16, 2021 at 3:39 PM Ruifeng Wang <Ruifeng.Wang@arm.com> wrote:
> > Subject: [PATCH 1/3] net/hinic: restore vectorised code
> >
> > Following make support removal, the vectorised code is not built anymore,
> > fix the build flag check.
> >
> > Fixes: 3cc6ecfdfe85 ("build: remove makefiles")
> > Cc: stable@dpdk.org
> >
> > Signed-off-by: David Marchand <david.marchand@redhat.com>
> > ---
> > drivers/net/hinic/hinic_pmd_rx.c | 6 +++---
> > drivers/net/hinic/hinic_pmd_tx.c | 10 +++++-----
> > 2 files changed, 8 insertions(+), 8 deletions(-)
> >
> > diff --git a/drivers/net/hinic/hinic_pmd_rx.c
> > b/drivers/net/hinic/hinic_pmd_rx.c
> > index a49769a863..842399cc4c 100644
> > --- a/drivers/net/hinic/hinic_pmd_rx.c
> > +++ b/drivers/net/hinic/hinic_pmd_rx.c
> > @@ -4,7 +4,7 @@
> >
> > #include <rte_ether.h>
> > #include <rte_mbuf.h>
> > -#ifdef __ARM64_NEON__
> > +#ifdef RTE_ARCH_ARM64
>
> We can test '__ARM_NEON' which will be defined by compilers.
> https://developer.arm.com/documentation/ihi0053/latest/
On the principle, I agree, but this was not what was tested before.
The activation was only gated by checking for the arm architecture.
See: https://git.dpdk.org/dpdk/tree/drivers/net/hinic/Makefile?id=05b6eee7bdabf7c17ed69c44515e0cd7d6e1da23#n14
So I restored the test as it was before make removal.
What you propose can be done in a followup from the hinic developers,
but the minimal fix is still this current patch.
> -----Original Message-----
> From: David Marchand <david.marchand@redhat.com>
> Sent: Monday, January 18, 2021 4:18 PM
> To: Ruifeng Wang <Ruifeng.Wang@arm.com>
> Cc: dev@dpdk.org; ferruh.yigit@intel.com; stable@dpdk.org; Ziyang Xuan
> <xuanziyang2@huawei.com>; Xiaoyun Wang
> <cloud.wangxiaoyun@huawei.com>; Guoyang Zhou
> <zhouguoyang@huawei.com>; Ciara Power <ciara.power@intel.com>;
> thomas@monjalon.net; nd <nd@arm.com>
> Subject: Re: [PATCH 1/3] net/hinic: restore vectorised code
>
> On Sat, Jan 16, 2021 at 3:39 PM Ruifeng Wang <Ruifeng.Wang@arm.com>
> wrote:
> > > Subject: [PATCH 1/3] net/hinic: restore vectorised code
> > >
> > > Following make support removal, the vectorised code is not built
> > > anymore, fix the build flag check.
> > >
> > > Fixes: 3cc6ecfdfe85 ("build: remove makefiles")
> > > Cc: stable@dpdk.org
> > >
> > > Signed-off-by: David Marchand <david.marchand@redhat.com>
> > > ---
> > > drivers/net/hinic/hinic_pmd_rx.c | 6 +++---
> > > drivers/net/hinic/hinic_pmd_tx.c | 10 +++++-----
> > > 2 files changed, 8 insertions(+), 8 deletions(-)
> > >
> > > diff --git a/drivers/net/hinic/hinic_pmd_rx.c
> > > b/drivers/net/hinic/hinic_pmd_rx.c
> > > index a49769a863..842399cc4c 100644
> > > --- a/drivers/net/hinic/hinic_pmd_rx.c
> > > +++ b/drivers/net/hinic/hinic_pmd_rx.c
> > > @@ -4,7 +4,7 @@
> > >
> > > #include <rte_ether.h>
> > > #include <rte_mbuf.h>
> > > -#ifdef __ARM64_NEON__
> > > +#ifdef RTE_ARCH_ARM64
> >
> > We can test '__ARM_NEON' which will be defined by compilers.
> > https://developer.arm.com/documentation/ihi0053/latest/
>
> On the principle, I agree, but this was not what was tested before.
>
> The activation was only gated by checking for the arm architecture.
> See:
> https://git.dpdk.org/dpdk/tree/drivers/net/hinic/Makefile?id=05b6eee7bda
> bf7c17ed69c44515e0cd7d6e1da23#n14
> So I restored the test as it was before make removal.
>
> What you propose can be done in a followup from the hinic developers, but
> the minimal fix is still this current patch.
I'm OK with this.
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
>
>
> --
> David Marchand
@@ -4,7 +4,7 @@
#include <rte_ether.h>
#include <rte_mbuf.h>
-#ifdef __ARM64_NEON__
+#ifdef RTE_ARCH_ARM64
#include <arm_neon.h>
#endif
@@ -762,7 +762,7 @@ void hinic_free_all_rx_mbufs(struct hinic_rxq *rxq)
static inline void hinic_rq_cqe_be_to_cpu32(void *dst_le32,
volatile void *src_be32)
{
-#if defined(__X86_64_SSE__)
+#if defined(RTE_ARCH_X86_64)
volatile __m128i *wqe_be = (volatile __m128i *)src_be32;
__m128i *wqe_le = (__m128i *)dst_le32;
__m128i shuf_mask = _mm_set_epi8(12, 13, 14, 15, 8, 9, 10,
@@ -770,7 +770,7 @@ static inline void hinic_rq_cqe_be_to_cpu32(void *dst_le32,
/* l2nic just use first 128 bits */
wqe_le[0] = _mm_shuffle_epi8(wqe_be[0], shuf_mask);
-#elif defined(__ARM64_NEON__)
+#elif defined(RTE_ARCH_ARM64)
volatile uint8x16_t *wqe_be = (volatile uint8x16_t *)src_be32;
uint8x16_t *wqe_le = (uint8x16_t *)dst_le32;
const uint8x16_t shuf_mask = {3, 2, 1, 0, 7, 6, 5, 4, 11, 10,
@@ -7,7 +7,7 @@
#include <rte_sctp.h>
#include <rte_udp.h>
#include <rte_ip.h>
-#ifdef __ARM64_NEON__
+#ifdef RTE_ARCH_ARM64
#include <arm_neon.h>
#endif
@@ -203,7 +203,7 @@
static inline void hinic_sq_wqe_cpu_to_be32(void *data, int nr_wqebb)
{
-#if defined(__X86_64_SSE__)
+#if defined(RTE_ARCH_X86_64)
int i;
__m128i *wqe_line = (__m128i *)data;
__m128i shuf_mask = _mm_set_epi8(12, 13, 14, 15, 8, 9, 10,
@@ -217,7 +217,7 @@ static inline void hinic_sq_wqe_cpu_to_be32(void *data, int nr_wqebb)
wqe_line[3] = _mm_shuffle_epi8(wqe_line[3], shuf_mask);
wqe_line += 4;
}
-#elif defined(__ARM64_NEON__)
+#elif defined(RTE_ARCH_ARM64)
int i;
uint8x16_t *wqe_line = (uint8x16_t *)data;
const uint8x16_t shuf_mask = {3, 2, 1, 0, 7, 6, 5, 4, 11, 10,
@@ -237,7 +237,7 @@ static inline void hinic_sq_wqe_cpu_to_be32(void *data, int nr_wqebb)
static inline void hinic_sge_cpu_to_be32(void *data, int nr_sge)
{
-#if defined(__X86_64_SSE__)
+#if defined(RTE_ARCH_X86_64)
int i;
__m128i *sge_line = (__m128i *)data;
__m128i shuf_mask = _mm_set_epi8(12, 13, 14, 15, 8, 9, 10,
@@ -248,7 +248,7 @@ static inline void hinic_sge_cpu_to_be32(void *data, int nr_sge)
*sge_line = _mm_shuffle_epi8(*sge_line, shuf_mask);
sge_line++;
}
-#elif defined(__ARM64_NEON__)
+#elif defined(RTE_ARCH_ARM64)
int i;
uint8x16_t *sge_line = (uint8x16_t *)data;
const uint8x16_t shuf_mask = {3, 2, 1, 0, 7, 6, 5, 4, 11, 10,