diff mbox series

kselftest/arm64: Size sycall-abi buffers for the actual maximum VL

Message ID 20230810-arm64-syscall-abi-perf-v1-1-6a0d7656359c@kernel.org (mailing list archive)
State Accepted
Commit 358b763ee64b11f45287c286e1f6145729ed49ab
Headers show
Series kselftest/arm64: Size sycall-abi buffers for the actual maximum VL | expand

Commit Message

Mark Brown Aug. 10, 2023, 11:39 a.m. UTC 
Our ABI opts to provide future proofing by defining a much larger
SVE_VQ_MAX than the architecture actually supports.  Since we use
this define to control the size of our vector data buffers this results
in a lot of overhead when we initialise which can be a very noticable
problem in emulation, we fill buffers that are orders of magnitude
larger than we will ever actually use even with virtual platforms that
provide the full range of architecturally supported vector lengths.

Define and use the actual architecture maximum to mitigate this.

Signed-off-by: Mark Brown <broonie@kernel.org>
---
 tools/testing/selftests/arm64/abi/syscall-abi.c | 38 +++++++++++++++----------
 1 file changed, 23 insertions(+), 15 deletions(-)


---
base-commit: 52a93d39b17dc7eb98b6aa3edb93943248e03b2f
change-id: 20230809-arm64-syscall-abi-perf-1e5876d161b2

Best regards,

Comments

Will Deacon Aug. 11, 2023, 11:44 a.m. UTC | #1 
On Thu, 10 Aug 2023 12:39:19 +0100, Mark Brown wrote:
> Our ABI opts to provide future proofing by defining a much larger
> SVE_VQ_MAX than the architecture actually supports.  Since we use
> this define to control the size of our vector data buffers this results
> in a lot of overhead when we initialise which can be a very noticable
> problem in emulation, we fill buffers that are orders of magnitude
> larger than we will ever actually use even with virtual platforms that
> provide the full range of architecturally supported vector lengths.
> 
> [...]

Applied to arm64 (for-next/selftests), thanks!

[1/1] kselftest/arm64: Size sycall-abi buffers for the actual maximum VL
      https://git.kernel.org/arm64/c/358b763ee64b

Cheers,
diff mbox series

Patch

diff --git a/tools/testing/selftests/arm64/abi/syscall-abi.c b/tools/testing/selftests/arm64/abi/syscall-abi.c
index 18cc123e2347..d704511a0955 100644
--- a/tools/testing/selftests/arm64/abi/syscall-abi.c
+++ b/tools/testing/selftests/arm64/abi/syscall-abi.c
@@ -20,12 +20,20 @@ 
 
 #include "syscall-abi.h"
 
+/*
+ * The kernel defines a much larger SVE_VQ_MAX than is expressable in
+ * the architecture, this creates a *lot* of overhead filling the
+ * buffers (especially ZA) on emulated platforms so use the actual
+ * architectural maximum instead.
+ */
+#define ARCH_SVE_VQ_MAX 16
+
 static int default_sme_vl;
 
 static int sve_vl_count;
-static unsigned int sve_vls[SVE_VQ_MAX];
+static unsigned int sve_vls[ARCH_SVE_VQ_MAX];
 static int sme_vl_count;
-static unsigned int sme_vls[SVE_VQ_MAX];
+static unsigned int sme_vls[ARCH_SVE_VQ_MAX];
 
 extern void do_syscall(int sve_vl, int sme_vl);
 
@@ -130,9 +138,9 @@  static int check_fpr(struct syscall_cfg *cfg, int sve_vl, int sme_vl,
 
 #define SVE_Z_SHARED_BYTES (128 / 8)
 
-static uint8_t z_zero[__SVE_ZREG_SIZE(SVE_VQ_MAX)];
-uint8_t z_in[SVE_NUM_ZREGS * __SVE_ZREG_SIZE(SVE_VQ_MAX)];
-uint8_t z_out[SVE_NUM_ZREGS * __SVE_ZREG_SIZE(SVE_VQ_MAX)];
+static uint8_t z_zero[__SVE_ZREG_SIZE(ARCH_SVE_VQ_MAX)];
+uint8_t z_in[SVE_NUM_ZREGS * __SVE_ZREG_SIZE(ARCH_SVE_VQ_MAX)];
+uint8_t z_out[SVE_NUM_ZREGS * __SVE_ZREG_SIZE(ARCH_SVE_VQ_MAX)];
 
 static void setup_z(struct syscall_cfg *cfg, int sve_vl, int sme_vl,
 		    uint64_t svcr)
@@ -190,8 +198,8 @@  static int check_z(struct syscall_cfg *cfg, int sve_vl, int sme_vl,
 	return errors;
 }
 
-uint8_t p_in[SVE_NUM_PREGS * __SVE_PREG_SIZE(SVE_VQ_MAX)];
-uint8_t p_out[SVE_NUM_PREGS * __SVE_PREG_SIZE(SVE_VQ_MAX)];
+uint8_t p_in[SVE_NUM_PREGS * __SVE_PREG_SIZE(ARCH_SVE_VQ_MAX)];
+uint8_t p_out[SVE_NUM_PREGS * __SVE_PREG_SIZE(ARCH_SVE_VQ_MAX)];
 
 static void setup_p(struct syscall_cfg *cfg, int sve_vl, int sme_vl,
 		    uint64_t svcr)
@@ -222,8 +230,8 @@  static int check_p(struct syscall_cfg *cfg, int sve_vl, int sme_vl,
 	return errors;
 }
 
-uint8_t ffr_in[__SVE_PREG_SIZE(SVE_VQ_MAX)];
-uint8_t ffr_out[__SVE_PREG_SIZE(SVE_VQ_MAX)];
+uint8_t ffr_in[__SVE_PREG_SIZE(ARCH_SVE_VQ_MAX)];
+uint8_t ffr_out[__SVE_PREG_SIZE(ARCH_SVE_VQ_MAX)];
 
 static void setup_ffr(struct syscall_cfg *cfg, int sve_vl, int sme_vl,
 		      uint64_t svcr)
@@ -300,8 +308,8 @@  static int check_svcr(struct syscall_cfg *cfg, int sve_vl, int sme_vl,
 	return errors;
 }
 
-uint8_t za_in[ZA_SIG_REGS_SIZE(SVE_VQ_MAX)];
-uint8_t za_out[ZA_SIG_REGS_SIZE(SVE_VQ_MAX)];
+uint8_t za_in[ZA_SIG_REGS_SIZE(ARCH_SVE_VQ_MAX)];
+uint8_t za_out[ZA_SIG_REGS_SIZE(ARCH_SVE_VQ_MAX)];
 
 static void setup_za(struct syscall_cfg *cfg, int sve_vl, int sme_vl,
 		     uint64_t svcr)
@@ -470,9 +478,9 @@  void sve_count_vls(void)
 		return;
 
 	/*
-	 * Enumerate up to SVE_VQ_MAX vector lengths
+	 * Enumerate up to ARCH_SVE_VQ_MAX vector lengths
 	 */
-	for (vq = SVE_VQ_MAX; vq > 0; vq /= 2) {
+	for (vq = ARCH_SVE_VQ_MAX; vq > 0; vq /= 2) {
 		vl = prctl(PR_SVE_SET_VL, vq * 16);
 		if (vl == -1)
 			ksft_exit_fail_msg("PR_SVE_SET_VL failed: %s (%d)\n",
@@ -496,9 +504,9 @@  void sme_count_vls(void)
 		return;
 
 	/*
-	 * Enumerate up to SVE_VQ_MAX vector lengths
+	 * Enumerate up to ARCH_SVE_VQ_MAX vector lengths
 	 */
-	for (vq = SVE_VQ_MAX; vq > 0; vq /= 2) {
+	for (vq = ARCH_SVE_VQ_MAX; vq > 0; vq /= 2) {
 		vl = prctl(PR_SME_SET_VL, vq * 16);
 		if (vl == -1)
 			ksft_exit_fail_msg("PR_SME_SET_VL failed: %s (%d)\n",