Add CPU flags
cover Add CPU flags
This patch set adds CPU flags which will enable the detection of ISA features available on more recent x86 based CPUs.
The CPUID leaf information can be found in Section 1.7 of this document:
https://software.intel.com/sites/default/files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdf
Kevin Laatz (17):
eal/cpuflags: add avx512 doubleword and quadword eal/cpuflags: add avx512 integer fused multiply-add eal/cpuflags: add avx512 conflict detection eal/cpuflags: add avx512 byte and word eal/cpuflags: add avx512 vector length eal/cpuflags: add avx512 vector bit manipulation eal/cpuflags: add avx512 vector bit manipulation 2 eal/cpuflags: add galois field new instructions eal/cpuflags: add vector AES
eal/cpuflags: add vector carry-less multiply eal/cpuflags: add avx512 vector neural network instructions eal/cpuflags: add avx512 bit algorithms eal/cpuflags: add avx512 vector popcount eal/cpuflags: add cache line demote
eal/cpuflags: add direct store instructions eal/cpuflags: add direct store instructions 64B eal/cpuflags: add avx512 two register intersection
lib/librte_eal/common/arch/x86/rte_cpuflags.c | 18 ++++++++++++++++++ .../common/include/arch/x86/rte_cpuflags.h | 18 ++++++++++++++++++ 2 files changed, 36 insertions(+)
Signature · 1 line
2.17.1
PATCH 1/17 eal/cpuflags: add avx512 doubleword and quadword
Add the CPU flag for AVX-512 doubleword and quadword instructions.
--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 2 ++ lib/librte_eal/common/include/arch/x86/rte_cpuflags.h | 2 ++ 2 files changed, 4 insertions(+)
lib/librte_eal/common/arch/x86/rte_cpuflags.c +2 -0
@@ -120,6 +120,8 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(EM64T, 0x80000001, 0, RTE_REG_EDX, 29) FEAT_DEF(INVTSC, 0x80000007, 0, RTE_REG_EDX, 8) + + FEAT_DEF(AVX512DQ, 0x00000007, 0, RTE_REG_EBX, 17) };int
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +2 -0
@@ -113,6 +113,8 @@ enum rte_cpu_flag_t {/* (EAX 80000007h) EDX features */ RTE_CPUFLAG_INVTSC ,/**< INVTSC */ +RTE_CPUFLAG_AVX512DQ ,/**< AVX512 Doubleword and Quadword */ +/* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */ };
Signature · 1 line
2.17.1
PATCH 2/17 eal/cpuflags: add avx512 integer fused multiply-add
In reply to Kevin LaatzAdd the CPU flag for AVX-512 integer fused multiply-add instructions.
--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 1 + lib/librte_eal/common/include/arch/x86/rte_cpuflags.h | 1 + 2 files changed, 2 insertions(+)
lib/librte_eal/common/arch/x86/rte_cpuflags.c +1 -0
@@ -122,6 +122,7 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(INVTSC, 0x80000007, 0, RTE_REG_EDX, 8) FEAT_DEF(AVX512DQ, 0x00000007, 0, RTE_REG_EBX, 17) + FEAT_DEF(AVX512IFMA, 0x00000007, 0, RTE_REG_EBX, 21) };int
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1 -0
@@ -114,6 +114,7 @@ enum rte_cpu_flag_t {RTE_CPUFLAG_INVTSC ,/**< INVTSC */ RTE_CPUFLAG_AVX512DQ ,/**< AVX512 Doubleword and Quadword */ +RTE_CPUFLAG_AVX512IFMA ,/**< AVX512 Integer Fused Multiply-Add */ /* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */
Signature · 1 line
2.17.1
PATCH 3/17 eal/cpuflags: add avx512 conflict detection
In reply to Kevin LaatzAdd the CPU flag for AVX-512 conflict detection instructions.
--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 1 + lib/librte_eal/common/include/arch/x86/rte_cpuflags.h | 1 + 2 files changed, 2 insertions(+)
lib/librte_eal/common/arch/x86/rte_cpuflags.c +1 -0
@@ -123,6 +123,7 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(AVX512DQ, 0x00000007, 0, RTE_REG_EBX, 17) FEAT_DEF(AVX512IFMA, 0x00000007, 0, RTE_REG_EBX, 21) + FEAT_DEF(AVX512CD, 0x00000007, 0, RTE_REG_EBX, 28) };int
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1 -0
@@ -115,6 +115,7 @@ enum rte_cpu_flag_t {RTE_CPUFLAG_AVX512DQ ,/**< AVX512 Doubleword and Quadword */ RTE_CPUFLAG_AVX512IFMA ,/**< AVX512 Integer Fused Multiply-Add */ +RTE_CPUFLAG_AVX512CD ,/**< AVX512 Conflict Detection*/ /* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */
Signature · 1 line
2.17.1
PATCH 4/17 eal/cpuflags: add avx512 byte and word
In reply to Kevin LaatzAdd the CPU flag for AVX-512 byte and word instructions.
--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 1 + lib/librte_eal/common/include/arch/x86/rte_cpuflags.h | 1 + 2 files changed, 2 insertions(+)
lib/librte_eal/common/arch/x86/rte_cpuflags.c +1 -0
@@ -124,6 +124,7 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(AVX512DQ, 0x00000007, 0, RTE_REG_EBX, 17) FEAT_DEF(AVX512IFMA, 0x00000007, 0, RTE_REG_EBX, 21) FEAT_DEF(AVX512CD, 0x00000007, 0, RTE_REG_EBX, 28) + FEAT_DEF(AVX512BW, 0x00000007, 0, RTE_REG_EBX, 30) };int
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1 -0
@@ -116,6 +116,7 @@ enum rte_cpu_flag_t {RTE_CPUFLAG_AVX512DQ ,/**< AVX512 Doubleword and Quadword */ RTE_CPUFLAG_AVX512IFMA ,/**< AVX512 Integer Fused Multiply-Add */ RTE_CPUFLAG_AVX512CD ,/**< AVX512 Conflict Detection*/ +RTE_CPUFLAG_AVX512BW ,/**< AVX512 Byte and Word */ /* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */
Signature · 1 line
2.17.1
PATCH 5/17 eal/cpuflags: add avx512 vector length
In reply to Kevin LaatzAdd the CPU flag for AVX-512 vector length instructions.
--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 1 + lib/librte_eal/common/include/arch/x86/rte_cpuflags.h | 1 + 2 files changed, 2 insertions(+)
lib/librte_eal/common/arch/x86/rte_cpuflags.c +1 -0
@@ -125,6 +125,7 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(AVX512IFMA, 0x00000007, 0, RTE_REG_EBX, 21) FEAT_DEF(AVX512CD, 0x00000007, 0, RTE_REG_EBX, 28) FEAT_DEF(AVX512BW, 0x00000007, 0, RTE_REG_EBX, 30) + FEAT_DEF(AVX512VL, 0x00000007, 0, RTE_REG_EBX, 31) };int
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1 -0
@@ -117,6 +117,7 @@ enum rte_cpu_flag_t {RTE_CPUFLAG_AVX512IFMA ,/**< AVX512 Integer Fused Multiply-Add */ RTE_CPUFLAG_AVX512CD ,/**< AVX512 Conflict Detection*/ RTE_CPUFLAG_AVX512BW ,/**< AVX512 Byte and Word */ +RTE_CPUFLAG_AVX512VL ,/**< AVX512 Vector Length */ /* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */
Signature · 1 line
2.17.1
PATCH 6/17 eal/cpuflags: add avx512 vector bit manipulation
In reply to Kevin LaatzAdd the CPU flag for AVX-512 vector bit manipulation instructions.
--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 1 + lib/librte_eal/common/include/arch/x86/rte_cpuflags.h | 1 + 2 files changed, 2 insertions(+)
lib/librte_eal/common/arch/x86/rte_cpuflags.c +1 -0
@@ -126,6 +126,7 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(AVX512CD, 0x00000007, 0, RTE_REG_EBX, 28) FEAT_DEF(AVX512BW, 0x00000007, 0, RTE_REG_EBX, 30) FEAT_DEF(AVX512VL, 0x00000007, 0, RTE_REG_EBX, 31) + FEAT_DEF(AVX512VBMI, 0x00000007, 0, RTE_REG_ECX, 1) };int
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1 -0
@@ -118,6 +118,7 @@ enum rte_cpu_flag_t {RTE_CPUFLAG_AVX512CD ,/**< AVX512 Conflict Detection*/ RTE_CPUFLAG_AVX512BW ,/**< AVX512 Byte and Word */ RTE_CPUFLAG_AVX512VL ,/**< AVX512 Vector Length */ +RTE_CPUFLAG_AVX512VBMI ,/**< AVX512 Vector Bit Manipulation */ /* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */
Signature · 1 line
2.17.1
PATCH 7/17 eal/cpuflags: add avx512 vector bit manipulation 2
In reply to Kevin LaatzAdd the CPU flag for AVX-512 vector bit manipulation 2 instructions.
--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 1 + lib/librte_eal/common/include/arch/x86/rte_cpuflags.h | 1 + 2 files changed, 2 insertions(+)
lib/librte_eal/common/arch/x86/rte_cpuflags.c +1 -0
@@ -127,6 +127,7 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(AVX512BW, 0x00000007, 0, RTE_REG_EBX, 30) FEAT_DEF(AVX512VL, 0x00000007, 0, RTE_REG_EBX, 31) FEAT_DEF(AVX512VBMI, 0x00000007, 0, RTE_REG_ECX, 1) + FEAT_DEF(AVX512VBMI2, 0x00000007, 0, RTE_REG_ECX, 6) };int
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1 -0
@@ -119,6 +119,7 @@ enum rte_cpu_flag_t {RTE_CPUFLAG_AVX512BW ,/**< AVX512 Byte and Word */ RTE_CPUFLAG_AVX512VL ,/**< AVX512 Vector Length */ RTE_CPUFLAG_AVX512VBMI ,/**< AVX512 Vector Bit Manipulation */ +RTE_CPUFLAG_AVX512VBMI2 ,/**< AVX512 Vector Bit Manipulation 2 */ /* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */
Signature · 1 line
2.17.1
PATCH 8/17 eal/cpuflags: add galois field new instructions
In reply to Kevin LaatzAdd the CPU flag for Galois field new instructions.
--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 1 + lib/librte_eal/common/include/arch/x86/rte_cpuflags.h | 1 + 2 files changed, 2 insertions(+)
lib/librte_eal/common/arch/x86/rte_cpuflags.c +1 -0
@@ -128,6 +128,7 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(AVX512VL, 0x00000007, 0, RTE_REG_EBX, 31) FEAT_DEF(AVX512VBMI, 0x00000007, 0, RTE_REG_ECX, 1) FEAT_DEF(AVX512VBMI2, 0x00000007, 0, RTE_REG_ECX, 6) + FEAT_DEF(GFNI, 0x00000007, 0, RTE_REG_ECX, 8) };int
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1 -0
@@ -120,6 +120,7 @@ enum rte_cpu_flag_t {RTE_CPUFLAG_AVX512VL ,/**< AVX512 Vector Length */ RTE_CPUFLAG_AVX512VBMI ,/**< AVX512 Vector Bit Manipulation */ RTE_CPUFLAG_AVX512VBMI2 ,/**< AVX512 Vector Bit Manipulation 2 */ +RTE_CPUFLAG_GFNI ,/**< Galois Field New Instructions */ /* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */
Signature · 1 line
2.17.1
PATCH 9/17 eal/cpuflags: add vector AES
In reply to Kevin LaatzAdd the CPU flag for vector AES instructions.
--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 1 + lib/librte_eal/common/include/arch/x86/rte_cpuflags.h | 1 + 2 files changed, 2 insertions(+)
lib/librte_eal/common/arch/x86/rte_cpuflags.c +1 -0
@@ -129,6 +129,7 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(AVX512VBMI, 0x00000007, 0, RTE_REG_ECX, 1) FEAT_DEF(AVX512VBMI2, 0x00000007, 0, RTE_REG_ECX, 6) FEAT_DEF(GFNI, 0x00000007, 0, RTE_REG_ECX, 8) + FEAT_DEF(VAES, 0x00000007, 0, RTE_REG_ECX, 9) };int
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1 -0
@@ -121,6 +121,7 @@ enum rte_cpu_flag_t {RTE_CPUFLAG_AVX512VBMI ,/**< AVX512 Vector Bit Manipulation */ RTE_CPUFLAG_AVX512VBMI2 ,/**< AVX512 Vector Bit Manipulation 2 */ RTE_CPUFLAG_GFNI ,/**< Galois Field New Instructions */ +RTE_CPUFLAG_VAES ,/**< Vector AES */ /* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */
Signature · 1 line
2.17.1
PATCH 10/17 eal/cpuflags: add vector carry-less multiply
In reply to Kevin LaatzAdd the CPU flag for vector carry-less multiply instructions.
--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 1 + lib/librte_eal/common/include/arch/x86/rte_cpuflags.h | 1 + 2 files changed, 2 insertions(+)
lib/librte_eal/common/arch/x86/rte_cpuflags.c +1 -0
@@ -130,6 +130,7 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(AVX512VBMI2, 0x00000007, 0, RTE_REG_ECX, 6) FEAT_DEF(GFNI, 0x00000007, 0, RTE_REG_ECX, 8) FEAT_DEF(VAES, 0x00000007, 0, RTE_REG_ECX, 9) + FEAT_DEF(VPCLMULQDQ, 0x00000007, 0, RTE_REG_ECX, 10) };int
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1 -0
@@ -122,6 +122,7 @@ enum rte_cpu_flag_t {RTE_CPUFLAG_AVX512VBMI2 ,/**< AVX512 Vector Bit Manipulation 2 */ RTE_CPUFLAG_GFNI ,/**< Galois Field New Instructions */ RTE_CPUFLAG_VAES ,/**< Vector AES */ +RTE_CPUFLAG_VPCLMULQDQ ,/**< Vector Carry-less Multiply */ /* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */
Signature · 1 line
2.17.1
PATCH 11/17 eal/cpuflags: add avx512 vector neural network instructions
In reply to Kevin LaatzAdd the CPU flag for AVX-512 vector neural network instructions.
--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 1 + lib/librte_eal/common/include/arch/x86/rte_cpuflags.h | 1 + 2 files changed, 2 insertions(+)
lib/librte_eal/common/arch/x86/rte_cpuflags.c +1 -0
@@ -131,6 +131,7 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(GFNI, 0x00000007, 0, RTE_REG_ECX, 8) FEAT_DEF(VAES, 0x00000007, 0, RTE_REG_ECX, 9) FEAT_DEF(VPCLMULQDQ, 0x00000007, 0, RTE_REG_ECX, 10) + FEAT_DEF(AVX512VNNI, 0x00000007, 0, RTE_REG_ECX, 11) };int
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1 -0
@@ -123,6 +123,7 @@ enum rte_cpu_flag_t {RTE_CPUFLAG_GFNI ,/**< Galois Field New Instructions */ RTE_CPUFLAG_VAES ,/**< Vector AES */ RTE_CPUFLAG_VPCLMULQDQ ,/**< Vector Carry-less Multiply */ +RTE_CPUFLAG_AVX512VNNI ,/**< AVX512 Vector Neural Network Instructions */ /* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */
Signature · 1 line
2.17.1
PATCH 12/17 eal/cpuflags: add avx512 bit algorithms
In reply to Kevin LaatzAdd the CPU flag AVX-512 for bit algorithm instructions.
--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 1 + lib/librte_eal/common/include/arch/x86/rte_cpuflags.h | 1 + 2 files changed, 2 insertions(+)
lib/librte_eal/common/arch/x86/rte_cpuflags.c +1 -0
@@ -132,6 +132,7 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(VAES, 0x00000007, 0, RTE_REG_ECX, 9) FEAT_DEF(VPCLMULQDQ, 0x00000007, 0, RTE_REG_ECX, 10) FEAT_DEF(AVX512VNNI, 0x00000007, 0, RTE_REG_ECX, 11) + FEAT_DEF(AVX512BITALG, 0x00000007, 0, RTE_REG_ECX, 12) };int
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1 -0
@@ -124,6 +124,7 @@ enum rte_cpu_flag_t {RTE_CPUFLAG_VAES ,/**< Vector AES */ RTE_CPUFLAG_VPCLMULQDQ ,/**< Vector Carry-less Multiply */ RTE_CPUFLAG_AVX512VNNI ,/**< AVX512 Vector Neural Network Instructions */ +RTE_CPUFLAG_AVX512BITALG ,/**< AVX512 Bit Algorithms */ /* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */
Signature · 1 line
2.17.1
PATCH 13/17 eal/cpuflags: add avx512 vector popcount
In reply to Kevin LaatzAdd the CPU flag for AVX-512 vector popcount instructions.
--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 1 + lib/librte_eal/common/include/arch/x86/rte_cpuflags.h | 1 + 2 files changed, 2 insertions(+)
lib/librte_eal/common/arch/x86/rte_cpuflags.c +1 -0
@@ -133,6 +133,7 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(VPCLMULQDQ, 0x00000007, 0, RTE_REG_ECX, 10) FEAT_DEF(AVX512VNNI, 0x00000007, 0, RTE_REG_ECX, 11) FEAT_DEF(AVX512BITALG, 0x00000007, 0, RTE_REG_ECX, 12) + FEAT_DEF(AVX512VPOPCNTDQ, 0x00000007, 0, RTE_REG_ECX, 14) };int
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1 -0
@@ -125,6 +125,7 @@ enum rte_cpu_flag_t {RTE_CPUFLAG_VPCLMULQDQ ,/**< Vector Carry-less Multiply */ RTE_CPUFLAG_AVX512VNNI ,/**< AVX512 Vector Neural Network Instructions */ RTE_CPUFLAG_AVX512BITALG ,/**< AVX512 Bit Algorithms */ +RTE_CPUFLAG_AVX512VPOPCNTDQ ,/**< AVX512 Vector Popcount */ /* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */
Signature · 1 line
2.17.1
PATCH 14/17 eal/cpuflags: add cache line demote
In reply to Kevin LaatzAdd the CPU flag for cache line demote instructions.
--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 1 + lib/librte_eal/common/include/arch/x86/rte_cpuflags.h | 1 + 2 files changed, 2 insertions(+)
lib/librte_eal/common/arch/x86/rte_cpuflags.c +1 -0
@@ -134,6 +134,7 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(AVX512VNNI, 0x00000007, 0, RTE_REG_ECX, 11) FEAT_DEF(AVX512BITALG, 0x00000007, 0, RTE_REG_ECX, 12) FEAT_DEF(AVX512VPOPCNTDQ, 0x00000007, 0, RTE_REG_ECX, 14) + FEAT_DEF(CLDEMOTE, 0x00000007, 0, RTE_REG_ECX, 25) };int
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1 -0
@@ -126,6 +126,7 @@ enum rte_cpu_flag_t {RTE_CPUFLAG_AVX512VNNI ,/**< AVX512 Vector Neural Network Instructions */ RTE_CPUFLAG_AVX512BITALG ,/**< AVX512 Bit Algorithms */ RTE_CPUFLAG_AVX512VPOPCNTDQ ,/**< AVX512 Vector Popcount */ +RTE_CPUFLAG_CLDEMOTE ,/**< Cache Line Demote */ /* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */
Signature · 1 line
2.17.1
PATCH 15/17 eal/cpuflags: add direct store instructions
In reply to Kevin LaatzAdd the CPU flag for direct store instructions.
--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 1 + lib/librte_eal/common/include/arch/x86/rte_cpuflags.h | 1 + 2 files changed, 2 insertions(+)
lib/librte_eal/common/arch/x86/rte_cpuflags.c +1 -0
@@ -135,6 +135,7 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(AVX512BITALG, 0x00000007, 0, RTE_REG_ECX, 12) FEAT_DEF(AVX512VPOPCNTDQ, 0x00000007, 0, RTE_REG_ECX, 14) FEAT_DEF(CLDEMOTE, 0x00000007, 0, RTE_REG_ECX, 25) + FEAT_DEF(MOVDIRI, 0x00000007, 0, RTE_REG_ECX, 27) };int
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1 -0
@@ -127,6 +127,7 @@ enum rte_cpu_flag_t {RTE_CPUFLAG_AVX512BITALG ,/**< AVX512 Bit Algorithms */ RTE_CPUFLAG_AVX512VPOPCNTDQ ,/**< AVX512 Vector Popcount */ RTE_CPUFLAG_CLDEMOTE ,/**< Cache Line Demote */ +RTE_CPUFLAG_MOVDIRI ,/**< Direct Store Instructions */ /* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */
Signature · 1 line
2.17.1
PATCH 16/17 eal/cpuflags: add direct store instructions 64B
In reply to Kevin LaatzAdd the CPU flag for direct store 64B instructions.
--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 1 + lib/librte_eal/common/include/arch/x86/rte_cpuflags.h | 1 + 2 files changed, 2 insertions(+)
lib/librte_eal/common/arch/x86/rte_cpuflags.c +1 -0
@@ -136,6 +136,7 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(AVX512VPOPCNTDQ, 0x00000007, 0, RTE_REG_ECX, 14) FEAT_DEF(CLDEMOTE, 0x00000007, 0, RTE_REG_ECX, 25) FEAT_DEF(MOVDIRI, 0x00000007, 0, RTE_REG_ECX, 27) + FEAT_DEF(MOVDIR64B, 0x00000007, 0, RTE_REG_ECX, 28) };int
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1 -0
@@ -128,6 +128,7 @@ enum rte_cpu_flag_t {RTE_CPUFLAG_AVX512VPOPCNTDQ ,/**< AVX512 Vector Popcount */ RTE_CPUFLAG_CLDEMOTE ,/**< Cache Line Demote */ RTE_CPUFLAG_MOVDIRI ,/**< Direct Store Instructions */ +RTE_CPUFLAG_MOVDIR64B ,/**< Direct Store Instructions 64B */ /* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */
Signature · 1 line
2.17.1
PATCH 17/17 eal/cpuflags: add avx512 two register intersection
In reply to Kevin LaatzAdd the CPU flag for AVX-512 two register intersection instructions.
--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 1 + lib/librte_eal/common/include/arch/x86/rte_cpuflags.h | 1 + 2 files changed, 2 insertions(+)
lib/librte_eal/common/arch/x86/rte_cpuflags.c +1 -0
@@ -137,6 +137,7 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(CLDEMOTE, 0x00000007, 0, RTE_REG_ECX, 25) FEAT_DEF(MOVDIRI, 0x00000007, 0, RTE_REG_ECX, 27) FEAT_DEF(MOVDIR64B, 0x00000007, 0, RTE_REG_ECX, 28) + FEAT_DEF(AVX512VP2INTERSECT, 0x00000007, 0, RTE_REG_EDX, 8) };int
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1 -0
@@ -129,6 +129,7 @@ enum rte_cpu_flag_t {RTE_CPUFLAG_CLDEMOTE ,/**< Cache Line Demote */ RTE_CPUFLAG_MOVDIRI ,/**< Direct Store Instructions */ RTE_CPUFLAG_MOVDIR64B ,/**< Direct Store Instructions 64B */ +RTE_CPUFLAG_AVX512VP2INTERSECT ,/**< AVX512 Two Register Intersection */ /* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */
Signature · 1 line
2.17.1
Quoted 26 lines · Kevin Laatz
This patch set adds CPU flags which will enable the detection of ISA features available on more recent x86 based CPUs.
The CPUID leaf information can be found in Section 1.7 of this document:
https://software.intel.com/sites/default/files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdfKevin Laatz (17):
eal/cpuflags: add avx512 doubleword and quadword eal/cpuflags: add avx512 integer fused multiply-add eal/cpuflags: add avx512 conflict detection eal/cpuflags: add avx512 byte and word eal/cpuflags: add avx512 vector length eal/cpuflags: add avx512 vector bit manipulation eal/cpuflags: add avx512 vector bit manipulation 2 eal/cpuflags: add galois field new instructions eal/cpuflags: add vector AES
eal/cpuflags: add vector carry-less multiply eal/cpuflags: add avx512 vector neural network instructions eal/cpuflags: add avx512 bit algorithms eal/cpuflags: add avx512 vector popcount eal/cpuflags: add cache line demote
eal/cpuflags: add direct store instructions eal/cpuflags: add direct store instructions 64B eal/cpuflags: add avx512 two register intersection
Quoted 3 lines
lib/librte_eal/common/arch/x86/rte_cpuflags.c | 18 ++++++++++++++++++ .../common/include/arch/x86/rte_cpuflags.h | 18 ++++++++++++++++++ 2 files changed, 36 insertions(+)
You are just adding definitions, the review should be straightforward for guys in the know.
I can see little value in splitting in that many patches.
Signature · 1 line
David Marchand
PATCH v2 1/1 eal/cpuflags: add x86 based cpu flags
In reply to Kevin LaatzThis patch adds CPU flags which will enable the detection of ISA features available on more recent x86 based CPUs.
The CPUID leaf information can be found in Section 1.7 of this document:
https://software.intel.com/sites/default/files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdf
The following CPU flags are added in this patch:
- AVX-512 doubleword and quadword instructions.
- AVX-512 integer fused multiply-add instructions.
- AVX-512 conflict detection instructions.
- AVX-512 byte and word instructions.
- AVX-512 vector length instructions.
- AVX-512 vector bit manipulation instructions.
- AVX-512 vector bit manipulation 2 instructions.
- Galois field new instructions.
- Vector AES instructions.
- Vector carry-less multiply instructions.
- AVX-512 vector neural network instructions.
- AVX-512 for bit algorithm instructions.
- AVX-512 vector popcount instructions.
- Cache line demote instructions.
- Direct store instructions.
- Direct store 64B instructions.
- AVX-512 two register intersection instructions.
--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 18 ++++++++++++++++++ .../common/include/arch/x86/rte_cpuflags.h | 18 ++++++++++++++++++ 2 files changed, 36 insertions(+)
lib/librte_eal/common/arch/x86/rte_cpuflags.c +18 -0
@@ -120,6 +120,24 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(EM64T, 0x80000001, 0, RTE_REG_EDX, 29) FEAT_DEF(INVTSC, 0x80000007, 0, RTE_REG_EDX, 8) + + FEAT_DEF(AVX512DQ, 0x00000007, 0, RTE_REG_EBX, 17) + FEAT_DEF(AVX512IFMA, 0x00000007, 0, RTE_REG_EBX, 21) + FEAT_DEF(AVX512CD, 0x00000007, 0, RTE_REG_EBX, 28) + FEAT_DEF(AVX512BW, 0x00000007, 0, RTE_REG_EBX, 30) + FEAT_DEF(AVX512VL, 0x00000007, 0, RTE_REG_EBX, 31) + FEAT_DEF(AVX512VBMI, 0x00000007, 0, RTE_REG_ECX, 1) + FEAT_DEF(AVX512VBMI2, 0x00000007, 0, RTE_REG_ECX, 6) + FEAT_DEF(GFNI, 0x00000007, 0, RTE_REG_ECX, 8) + FEAT_DEF(VAES, 0x00000007, 0, RTE_REG_ECX, 9) + FEAT_DEF(VPCLMULQDQ, 0x00000007, 0, RTE_REG_ECX, 10) + FEAT_DEF(AVX512VNNI, 0x00000007, 0, RTE_REG_ECX, 11) + FEAT_DEF(AVX512BITALG, 0x00000007, 0, RTE_REG_ECX, 12) + FEAT_DEF(AVX512VPOPCNTDQ, 0x00000007, 0, RTE_REG_ECX, 14) + FEAT_DEF(CLDEMOTE, 0x00000007, 0, RTE_REG_ECX, 25) + FEAT_DEF(MOVDIRI, 0x00000007, 0, RTE_REG_ECX, 27) + FEAT_DEF(MOVDIR64B, 0x00000007, 0, RTE_REG_ECX, 28) + FEAT_DEF(AVX512VP2INTERSECT, 0x00000007, 0, RTE_REG_EDX, 8) };int
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +18 -0
@@ -113,6 +113,24 @@ enum rte_cpu_flag_t {/* (EAX 80000007h) EDX features */ RTE_CPUFLAG_INVTSC ,/**< INVTSC */ +RTE_CPUFLAG_AVX512DQ ,/**< AVX512 Doubleword and Quadword */ +RTE_CPUFLAG_AVX512IFMA ,/**< AVX512 Integer Fused Multiply-Add */ +RTE_CPUFLAG_AVX512CD ,/**< AVX512 Conflict Detection*/ +RTE_CPUFLAG_AVX512BW ,/**< AVX512 Byte and Word */ +RTE_CPUFLAG_AVX512VL ,/**< AVX512 Vector Length */ +RTE_CPUFLAG_AVX512VBMI ,/**< AVX512 Vector Bit Manipulation */ +RTE_CPUFLAG_AVX512VBMI2 ,/**< AVX512 Vector Bit Manipulation 2 */ +RTE_CPUFLAG_GFNI ,/**< Galois Field New Instructions */ +RTE_CPUFLAG_VAES ,/**< Vector AES */ +RTE_CPUFLAG_VPCLMULQDQ ,/**< Vector Carry-less Multiply */ +RTE_CPUFLAG_AVX512VNNI ,/**< AVX512 Vector Neural Network Instructions */ +RTE_CPUFLAG_AVX512BITALG ,/**< AVX512 Bit Algorithms */ +RTE_CPUFLAG_AVX512VPOPCNTDQ ,/**< AVX512 Vector Popcount */ +RTE_CPUFLAG_CLDEMOTE ,/**< Cache Line Demote */ +RTE_CPUFLAG_MOVDIRI ,/**< Direct Store Instructions */ +RTE_CPUFLAG_MOVDIR64B ,/**< Direct Store Instructions 64B */ +RTE_CPUFLAG_AVX512VP2INTERSECT ,/**< AVX512 Two Register Intersection */ +/* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */ };
Signature · 1 line
2.17.1
Re: [dpdk-dev] [PATCH v2] eal/cpuflags: add x86 based cpu flags
Quoted 80 lines · Kevin Laatz
This patch adds CPU flags which will enable the detection of ISA features available on more recent x86 based CPUs.
The CPUID leaf information can be found in Section 1.7 of this document:
https://software.intel.com/sites/default/files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdfThe following CPU flags are added in this patch:
- AVX-512 doubleword and quadword instructions.
- AVX-512 integer fused multiply-add instructions.
- AVX-512 conflict detection instructions.
- AVX-512 byte and word instructions.
- AVX-512 vector length instructions.
- AVX-512 vector bit manipulation instructions.
- AVX-512 vector bit manipulation 2 instructions.
- Galois field new instructions.
- Vector AES instructions.
- Vector carry-less multiply instructions.
- AVX-512 vector neural network instructions.
- AVX-512 for bit algorithm instructions.
- AVX-512 vector popcount instructions.
- Cache line demote instructions.
- Direct store instructions.
- Direct store 64B instructions.
- AVX-512 two register intersection instructions.Signed-off-by Kevin Laatz <[email protected]>--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 18 ++++++++++++++++++ .../common/include/arch/x86/rte_cpuflags.h | 18 ++++++++++++++++++ 2 files changed, 36 insertions(+)lib/librte_eal/common/arch/x86/rte_cpuflags.c +18
-0@@ -120,6 +120,24 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(EM64T, 0x80000001, 0, RTE_REG_EDX, 29) FEAT_DEF(INVTSC, 0x80000007, 0, RTE_REG_EDX, 8) + + FEAT_DEF(AVX512DQ, 0x00000007, 0, RTE_REG_EBX, 17) + FEAT_DEF(AVX512IFMA, 0x00000007, 0, RTE_REG_EBX, 21) + FEAT_DEF(AVX512CD, 0x00000007, 0, RTE_REG_EBX, 28) + FEAT_DEF(AVX512BW, 0x00000007, 0, RTE_REG_EBX, 30) + FEAT_DEF(AVX512VL, 0x00000007, 0, RTE_REG_EBX, 31) + FEAT_DEF(AVX512VBMI, 0x00000007, 0, RTE_REG_ECX, 1) + FEAT_DEF(AVX512VBMI2, 0x00000007, 0, RTE_REG_ECX, 6) + FEAT_DEF(GFNI, 0x00000007, 0, RTE_REG_ECX, 8) + FEAT_DEF(VAES, 0x00000007, 0, RTE_REG_ECX, 9) + FEAT_DEF(VPCLMULQDQ, 0x00000007, 0, RTE_REG_ECX, 10) + FEAT_DEF(AVX512VNNI, 0x00000007, 0, RTE_REG_ECX, 11) + FEAT_DEF(AVX512BITALG, 0x00000007, 0, RTE_REG_ECX, 12) + FEAT_DEF(AVX512VPOPCNTDQ, 0x00000007, 0, RTE_REG_ECX, 14) + FEAT_DEF(CLDEMOTE, 0x00000007, 0, RTE_REG_ECX, 25) + FEAT_DEF(MOVDIRI, 0x00000007, 0, RTE_REG_ECX, 27) + FEAT_DEF(MOVDIR64B, 0x00000007, 0, RTE_REG_ECX, 28) + FEAT_DEF(AVX512VP2INTERSECT, 0x00000007, 0, RTE_REG_EDX, 8) };int lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +10
-0@@ -113,6 +113,24 @@ enum rte_cpu_flag_t {/* (EAX 80000007h) EDX features */ RTE_CPUFLAG_INVTSC ,/**< INVTSC */ +RTE_CPUFLAG_AVX512DQ ,/**< AVX512 Doubleword and Quadword */ +RTE_CPUFLAG_AVX512IFMA ,/**< AVX512 Integer Fused Multiply-Add */ +RTE_CPUFLAG_AVX512CD ,/**< AVX512 Conflict Detection*/ +RTE_CPUFLAG_AVX512BW ,/**< AVX512 Byte and Word */ +RTE_CPUFLAG_AVX512VL ,/**< AVX512 Vector Length */ +RTE_CPUFLAG_AVX512VBMI ,/**< AVX512 Vector Bit Manipulation */ +RTE_CPUFLAG_AVX512VBMI2 ,/**< AVX512 Vector Bit Manipulation 2 */ +RTE_CPUFLAG_GFNI ,/**< Galois Field New Instructions */ +RTE_CPUFLAG_VAES ,/**< Vector AES */ + RTE_CPUFLAG_VPCLMULQDQ, /**< Vector Carry-less Multiply */
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 135
Quoted 10 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +8
-0+RTE_CPUFLAG_AVX512VNNI ,/**< AVX512 Vector Neural Network Instructions */ +RTE_CPUFLAG_AVX512BITALG ,/**< AVX512 Bit Algorithms */ +RTE_CPUFLAG_AVX512VPOPCNTDQ ,/**< AVX512 Vector Popcount */ +RTE_CPUFLAG_CLDEMOTE ,/**< Cache Line Demote */ +RTE_CPUFLAG_MOVDIRI ,/**< Direct Store Instructions */ +RTE_CPUFLAG_MOVDIR64B ,/**< Direct Store Instructions 64B */ +RTE_CPUFLAG_AVX512VP2INTERSECT ,/**< AVX512 Two Register Intersection */ +/* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */
This is seen as an ABI break because of the change on _NUMFLAGS:
https://travis-ci.com/github/ovsrobot/dpdk/jobs/302524264#L2351
Signature · 1 line
David Marchand
Re: [dpdk-dev] [PATCH v2] eal/cpuflags: add x86 based cpu flags
Quoted 110 lines
-----Original Message-----
From: David Marchand <[email protected]>
Sent: Friday, March 27, 2020 12:24 PM
To: Laatz, Kevin <[email protected]>
Cc: dev <[email protected]>; Richardson, Bruce <[email protected]>; Van Haaren, Harry <[email protected]>; Neil Horman <[email protected]>; Thomas Monjalon <[email protected]>; Honnappa Nagarahalli <[email protected]>; Dodji Seketeli <[email protected]>
Subject: Re: [dpdk-dev] [PATCH v2] eal/cpuflags: add x86 based cpu flagsQuoted 7 lines · Kevin Laatz
This patch adds CPU flags which will enable the detection of ISA features available on more recent x86 based CPUs.
The CPUID leaf information can be found in Section 1.7 of this document:
https://software.intel.com/sites/default/files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdf
Quoted 20 lines
The following CPU flags are added in this patch:
- AVX-512 doubleword and quadword instructions.
- AVX-512 integer fused multiply-add instructions.
- AVX-512 conflict detection instructions.
- AVX-512 byte and word instructions.
- AVX-512 vector length instructions.
- AVX-512 vector bit manipulation instructions.
- AVX-512 vector bit manipulation 2 instructions.
- Galois field new instructions.
- Vector AES instructions.
- Vector carry-less multiply instructions.
- AVX-512 vector neural network instructions.
- AVX-512 for bit algorithm instructions.
- AVX-512 vector popcount instructions.
- Cache line demote instructions.
- Direct store instructions.
- Direct store 64B instructions.
- AVX-512 two register intersection instructions.Review
Quoted 7 lines
Signed-off-by Kevin Laatz <[email protected]>--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 18 ++++++++++++++++++ .../common/include/arch/x86/rte_cpuflags.h | 18 ++++++++++++++++++ 2 files changed, 36 insertions(+)(quoted diff) +0
-0
b/lib/librte_eal/common/arch/x86/rte_cpuflags.cReview
Quoted 29 lines
index
6492df556..30439e795100644lib/librte_eal/common/arch/x86/rte_cpuflags.c +18
-0@@ -120,6 +120,24 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(EM64T, 0x80000001, 0, RTE_REG_EDX, 29) FEAT_DEF(INVTSC, 0x80000007, 0, RTE_REG_EDX, 8) + + FEAT_DEF(AVX512DQ, 0x00000007, 0, RTE_REG_EBX, 17) + FEAT_DEF(AVX512IFMA, 0x00000007, 0, RTE_REG_EBX, 21) + FEAT_DEF(AVX512CD, 0x00000007, 0, RTE_REG_EBX, 28) + FEAT_DEF(AVX512BW, 0x00000007, 0, RTE_REG_EBX, 30) + FEAT_DEF(AVX512VL, 0x00000007, 0, RTE_REG_EBX, 31) + FEAT_DEF(AVX512VBMI, 0x00000007, 0, RTE_REG_ECX, 1) + FEAT_DEF(AVX512VBMI2, 0x00000007, 0, RTE_REG_ECX, 6) + FEAT_DEF(GFNI, 0x00000007, 0, RTE_REG_ECX, 8) + FEAT_DEF(VAES, 0x00000007, 0, RTE_REG_ECX, 9) + FEAT_DEF(VPCLMULQDQ, 0x00000007, 0, RTE_REG_ECX, 10) + FEAT_DEF(AVX512VNNI, 0x00000007, 0, RTE_REG_ECX, 11) + FEAT_DEF(AVX512BITALG, 0x00000007, 0, RTE_REG_ECX, 12) + FEAT_DEF(AVX512VPOPCNTDQ, 0x00000007, 0, RTE_REG_ECX, 14) + FEAT_DEF(CLDEMOTE, 0x00000007, 0, RTE_REG_ECX, 25) + FEAT_DEF(MOVDIRI, 0x00000007, 0, RTE_REG_ECX, 27) + FEAT_DEF(MOVDIR64B, 0x00000007, 0, RTE_REG_ECX, 28) + FEAT_DEF(AVX512VP2INTERSECT, 0x00000007, 0, RTE_REG_EDX, 8) };int (quoted diff) +0
-0
b/lib/librte_eal/common/include/arch/x86/rte_cpuflags.hlib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 116
Quoted 8 lines
index
25ba47b96..f8f73b19f100644lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0@@ -113,6 +113,24 @@ enum rte_cpu_flag_t {/* (EAX 80000007h) EDX features */ RTE_CPUFLAG_INVTSC ,/**< INVTSC */ + RTE_CPUFLAG_AVX512DQ, /**< AVX512 Doubleword andQuadword */
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 117
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512IFMA, /**< AVX512 Integer FusedMultiply-Add */
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t {
Quoted 4 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +4
-0+RTE_CPUFLAG_AVX512CD ,/**< AVX512 Conflict Detection*/ +RTE_CPUFLAG_AVX512BW ,/**< AVX512 Byte and Word */ +RTE_CPUFLAG_AVX512VL ,/**< AVX512 Vector Length */ + RTE_CPUFLAG_AVX512VBMI, /**< AVX512 Vector BitManipulation */
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 122
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512VBMI2, /**< AVX512 Vector BitManipulation 2 */
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 123
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_GFNI, /**< Galois Field NewInstructions */
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 125
Quoted 2 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +2
-0+RTE_CPUFLAG_VAES ,/**< Vector AES */ + RTE_CPUFLAG_VPCLMULQDQ, /**< Vector Carry-less Multiply*/
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 126
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512VNNI, /**< AVX512 Vector NeuralNetwork Instructions */
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t {
Quoted 4 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +4
-0+RTE_CPUFLAG_AVX512BITALG ,/**< AVX512 Bit Algorithms */ +RTE_CPUFLAG_AVX512VPOPCNTDQ ,/**< AVX512 Vector Popcount */ +RTE_CPUFLAG_CLDEMOTE ,/**< Cache Line Demote */ + RTE_CPUFLAG_MOVDIRI, /**< Direct Store Instructions*/
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 131
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_MOVDIR64B, /**< Direct Store Instructions64B */
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 132
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512VP2INTERSECT, /**< AVX512 Two Register
Quoted 8 lines
Intersection */
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t {
Quoted 3 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+/* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be thelast! */
This is seen as an ABI break because of the change on
_NUMFLAGS:
https://travis-ci.com/github/ovsrobot/dpdk/jobs/302524264#L2351
Correct a publicly exposed enum max value has changed - I don't believe this is an ABI break, but a backward compatible ABI change was expected with this patchset.
Code compiled against eg 19.11 or 20.02 is expected to continue operating correctly.
The new flags were only added at the end of the enum, ensuring to not change the meaning of any existing flags which would be compiled-in constants to the application binary.
The actual size of the CPU flags array is a DPDK internal structure (in a .c file), and is hidden from the application, and never allocated by an application - so no possible mismatch in ABI there? Applications compiled against the older ABI will just not know about the newer flags - but suffer no breakage.
@ABI compatibility folks, please review too - but to the best of my understanding this is not an ABI break, but a backwards compatible update of CPU flag lists?
Thanks for flagging the CI results David!
Regards, -Harry
Re: [dpdk-dev] [PATCH v2] eal/cpuflags: add x86 based cpu flags
Quoted 115 lines · Harry
Quoted 115 lines
-----Original Message-----
From: David Marchand <[email protected]>
Sent: Friday, March 27, 2020 12:24 PM
To: Laatz, Kevin <[email protected]>
Cc: dev <[email protected]>; Richardson, Bruce <[email protected]>; Van Haaren, Harry <[email protected]>; Neil Horman <[email protected]>; Thomas Monjalon <[email protected]>; Honnappa Nagarahalli <[email protected]>; Dodji Seketeli <[email protected]>
Subject: Re: [dpdk-dev] [PATCH v2] eal/cpuflags: add x86 based cpu flagsQuoted 7 lines · Kevin Laatz
This patch adds CPU flags which will enable the detection of ISA features available on more recent x86 based CPUs.
The CPUID leaf information can be found in Section 1.7 of this document:
https://software.intel.com/sites/default/files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdf
Quoted 20 lines
The following CPU flags are added in this patch:
- AVX-512 doubleword and quadword instructions.
- AVX-512 integer fused multiply-add instructions.
- AVX-512 conflict detection instructions.
- AVX-512 byte and word instructions.
- AVX-512 vector length instructions.
- AVX-512 vector bit manipulation instructions.
- AVX-512 vector bit manipulation 2 instructions.
- Galois field new instructions.
- Vector AES instructions.
- Vector carry-less multiply instructions.
- AVX-512 vector neural network instructions.
- AVX-512 for bit algorithm instructions.
- AVX-512 vector popcount instructions.
- Cache line demote instructions.
- Direct store instructions.
- Direct store 64B instructions.
- AVX-512 two register intersection instructions.Review
Quoted 7 lines
Signed-off-by Kevin Laatz <[email protected]>--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 18 ++++++++++++++++++ .../common/include/arch/x86/rte_cpuflags.h | 18 ++++++++++++++++++ 2 files changed, 36 insertions(+)(quoted diff) +0
-0
b/lib/librte_eal/common/arch/x86/rte_cpuflags.cReview
Quoted 29 lines
index
6492df556..30439e795100644lib/librte_eal/common/arch/x86/rte_cpuflags.c +18
-0@@ -120,6 +120,24 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(EM64T, 0x80000001, 0, RTE_REG_EDX, 29) FEAT_DEF(INVTSC, 0x80000007, 0, RTE_REG_EDX, 8) + + FEAT_DEF(AVX512DQ, 0x00000007, 0, RTE_REG_EBX, 17) + FEAT_DEF(AVX512IFMA, 0x00000007, 0, RTE_REG_EBX, 21) + FEAT_DEF(AVX512CD, 0x00000007, 0, RTE_REG_EBX, 28) + FEAT_DEF(AVX512BW, 0x00000007, 0, RTE_REG_EBX, 30) + FEAT_DEF(AVX512VL, 0x00000007, 0, RTE_REG_EBX, 31) + FEAT_DEF(AVX512VBMI, 0x00000007, 0, RTE_REG_ECX, 1) + FEAT_DEF(AVX512VBMI2, 0x00000007, 0, RTE_REG_ECX, 6) + FEAT_DEF(GFNI, 0x00000007, 0, RTE_REG_ECX, 8) + FEAT_DEF(VAES, 0x00000007, 0, RTE_REG_ECX, 9) + FEAT_DEF(VPCLMULQDQ, 0x00000007, 0, RTE_REG_ECX, 10) + FEAT_DEF(AVX512VNNI, 0x00000007, 0, RTE_REG_ECX, 11) + FEAT_DEF(AVX512BITALG, 0x00000007, 0, RTE_REG_ECX, 12) + FEAT_DEF(AVX512VPOPCNTDQ, 0x00000007, 0, RTE_REG_ECX, 14) + FEAT_DEF(CLDEMOTE, 0x00000007, 0, RTE_REG_ECX, 25) + FEAT_DEF(MOVDIRI, 0x00000007, 0, RTE_REG_ECX, 27) + FEAT_DEF(MOVDIR64B, 0x00000007, 0, RTE_REG_ECX, 28) + FEAT_DEF(AVX512VP2INTERSECT, 0x00000007, 0, RTE_REG_EDX, 8) };int (quoted diff) +0
-0
b/lib/librte_eal/common/include/arch/x86/rte_cpuflags.hlib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 116
Quoted 8 lines
index
25ba47b96..f8f73b19f100644lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0@@ -113,6 +113,24 @@ enum rte_cpu_flag_t {/* (EAX 80000007h) EDX features */ RTE_CPUFLAG_INVTSC ,/**< INVTSC */ + RTE_CPUFLAG_AVX512DQ, /**< AVX512 Doubleword andQuadword */
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 117
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512IFMA, /**< AVX512 Integer FusedMultiply-Add */
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t {
Quoted 4 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +4
-0+RTE_CPUFLAG_AVX512CD ,/**< AVX512 Conflict Detection*/ +RTE_CPUFLAG_AVX512BW ,/**< AVX512 Byte and Word */ +RTE_CPUFLAG_AVX512VL ,/**< AVX512 Vector Length */ + RTE_CPUFLAG_AVX512VBMI, /**< AVX512 Vector BitManipulation */
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 122
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512VBMI2, /**< AVX512 Vector BitManipulation 2 */
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 123
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_GFNI, /**< Galois Field NewInstructions */
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 125
Quoted 2 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +2
-0+RTE_CPUFLAG_VAES ,/**< Vector AES */ + RTE_CPUFLAG_VPCLMULQDQ, /**< Vector Carry-less Multiply*/
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 126
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512VNNI, /**< AVX512 Vector NeuralNetwork Instructions */
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t {
Quoted 4 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +4
-0+RTE_CPUFLAG_AVX512BITALG ,/**< AVX512 Bit Algorithms */ +RTE_CPUFLAG_AVX512VPOPCNTDQ ,/**< AVX512 Vector Popcount */ +RTE_CPUFLAG_CLDEMOTE ,/**< Cache Line Demote */ + RTE_CPUFLAG_MOVDIRI, /**< Direct Store Instructions*/
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 131
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_MOVDIR64B, /**< Direct Store Instructions64B */
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 132
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512VP2INTERSECT, /**< AVX512 Two RegisterIntersection */
Quoted 3 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+/* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the
Quoted 6 lines
Quoted 4 lines
last! */
This is seen as an ABI break because of the change on
_NUMFLAGS:
https://travis-ci.com/github/ovsrobot/dpdk/jobs/302524264#L2351Correct a publicly exposed enum max value has changed - I don't believe this is an ABI break, but a backward compatible ABI change was expected with this patchset.
A change is that if an application was passing incorrect values to the functions taking this enum as input, then now it would succeed. I don't really see the point in doing this :-).
Quoted 9 lines
Code compiled against eg 19.11 or 20.02 is expected to continue operating correctly.
The new flags were only added at the end of the enum, ensuring to not change the meaning of any existing flags which would be compiled-in constants to the application binary.The actual size of the CPU flags array is a DPDK internal structure (in a .c file), and is hidden from the application, and never allocated by an application - so no possible mismatch in ABI there? Applications compiled against the older ABI will just not know about the newer flags - but suffer no breakage.
@ABI compatibility folks, please review too - but to the best of my understanding this is not an ABI break, but a backwards compatible update of CPU flag lists?
Thanks for flagging the CI results David!
I'd like people to look at this by themselves, not wait for Thomas, Aaron or me to check.
When a failure is caught by the robot, a mail is sent to the submitter afaiu (I suppose with Travis instability wrt ARM jobs, the mail might not have been sent this time).
It is then the responsibility of the submitter to either discuss the report on the mailing or/and waive it in devtools/libabigail.abignore.
Thanks.
Signature · 1 line
David Marchand
Re: [dpdk-dev] [PATCH v2] eal/cpuflags: add x86 based cpu flags
In reply to David MarchandQuoted 86 lines · David Marchand
Quoted 85 lines · Kevin Laatz
This patch adds CPU flags which will enable the detection of ISA features available on more recent x86 based CPUs.
The CPUID leaf information can be found in Section 1.7 of this document:
https://software.intel.com/sites/default/files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdfThe following CPU flags are added in this patch:
- AVX-512 doubleword and quadword instructions.
- AVX-512 integer fused multiply-add instructions.
- AVX-512 conflict detection instructions.
- AVX-512 byte and word instructions.
- AVX-512 vector length instructions.
- AVX-512 vector bit manipulation instructions.
- AVX-512 vector bit manipulation 2 instructions.
- Galois field new instructions.
- Vector AES instructions.
- Vector carry-less multiply instructions.
- AVX-512 vector neural network instructions.
- AVX-512 for bit algorithm instructions.
- AVX-512 vector popcount instructions.
- Cache line demote instructions.
- Direct store instructions.
- Direct store 64B instructions.
- AVX-512 two register intersection instructions.Signed-off-by Kevin Laatz <[email protected]>--- lib/librte_eal/common/arch/x86/rte_cpuflags.c | 18 ++++++++++++++++++ .../common/include/arch/x86/rte_cpuflags.h | 18 ++++++++++++++++++ 2 files changed, 36 insertions(+)lib/librte_eal/common/arch/x86/rte_cpuflags.c +18
-0@@ -120,6 +120,24 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(EM64T, 0x80000001, 0, RTE_REG_EDX, 29) FEAT_DEF(INVTSC, 0x80000007, 0, RTE_REG_EDX, 8) + + FEAT_DEF(AVX512DQ, 0x00000007, 0, RTE_REG_EBX, 17) + FEAT_DEF(AVX512IFMA, 0x00000007, 0, RTE_REG_EBX, 21) + FEAT_DEF(AVX512CD, 0x00000007, 0, RTE_REG_EBX, 28) + FEAT_DEF(AVX512BW, 0x00000007, 0, RTE_REG_EBX, 30) + FEAT_DEF(AVX512VL, 0x00000007, 0, RTE_REG_EBX, 31) + FEAT_DEF(AVX512VBMI, 0x00000007, 0, RTE_REG_ECX, 1) + FEAT_DEF(AVX512VBMI2, 0x00000007, 0, RTE_REG_ECX, 6) + FEAT_DEF(GFNI, 0x00000007, 0, RTE_REG_ECX, 8) + FEAT_DEF(VAES, 0x00000007, 0, RTE_REG_ECX, 9) + FEAT_DEF(VPCLMULQDQ, 0x00000007, 0, RTE_REG_ECX, 10) + FEAT_DEF(AVX512VNNI, 0x00000007, 0, RTE_REG_ECX, 11) + FEAT_DEF(AVX512BITALG, 0x00000007, 0, RTE_REG_ECX, 12) + FEAT_DEF(AVX512VPOPCNTDQ, 0x00000007, 0, RTE_REG_ECX, 14) + FEAT_DEF(CLDEMOTE, 0x00000007, 0, RTE_REG_ECX, 25) + FEAT_DEF(MOVDIRI, 0x00000007, 0, RTE_REG_ECX, 27) + FEAT_DEF(MOVDIR64B, 0x00000007, 0, RTE_REG_ECX, 28) + FEAT_DEF(AVX512VP2INTERSECT, 0x00000007, 0, RTE_REG_EDX, 8) };int lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +15
-0@@ -113,6 +113,24 @@ enum rte_cpu_flag_t {/* (EAX 80000007h) EDX features */ RTE_CPUFLAG_INVTSC ,/**< INVTSC */ +RTE_CPUFLAG_AVX512DQ ,/**< AVX512 Doubleword and Quadword */ +RTE_CPUFLAG_AVX512IFMA ,/**< AVX512 Integer Fused Multiply-Add */ +RTE_CPUFLAG_AVX512CD ,/**< AVX512 Conflict Detection*/ +RTE_CPUFLAG_AVX512BW ,/**< AVX512 Byte and Word */ +RTE_CPUFLAG_AVX512VL ,/**< AVX512 Vector Length */ +RTE_CPUFLAG_AVX512VBMI ,/**< AVX512 Vector Bit Manipulation */ +RTE_CPUFLAG_AVX512VBMI2 ,/**< AVX512 Vector Bit Manipulation 2 */ +RTE_CPUFLAG_GFNI ,/**< Galois Field New Instructions */ +RTE_CPUFLAG_VAES ,/**< Vector AES */ +RTE_CPUFLAG_VPCLMULQDQ ,/**< Vector Carry-less Multiply */ +RTE_CPUFLAG_AVX512VNNI ,/**< AVX512 Vector Neural Network Instructions */ +RTE_CPUFLAG_AVX512BITALG ,/**< AVX512 Bit Algorithms */ +RTE_CPUFLAG_AVX512VPOPCNTDQ ,/**< AVX512 Vector Popcount */ +RTE_CPUFLAG_CLDEMOTE ,/**< Cache Line Demote */ + RTE_CPUFLAG_MOVDIRI, /**< Direct Store Instructions */
Quoted 9 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 135
Quoted 5 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +3
-0+RTE_CPUFLAG_MOVDIR64B ,/**< Direct Store Instructions 64B */ +RTE_CPUFLAG_AVX512VP2INTERSECT ,/**< AVX512 Two Register Intersection */ +/* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */This is seen as an ABI break because of the change on
_NUMFLAGS:
https://travis-ci.com/github/ovsrobot/dpdk/jobs/302524264#L2351
It shouldn't be, as the only API calls we expose that use rte_cpu_flag_t accept it as an integer parameter to see if the flag is enabled. Theres no use of the enum in a public array or any struct that is sized based on the number of flags, so you should be good to go
Neil
Quoted 5 lines
Signature · 2 lines
-- David Marchand
Re: [dpdk-dev] [PATCH v2] eal/cpuflags: add x86 based cpu flags
27/03/2020 14:44, Neil Horman:
Quoted 32 lines · David Marchand
Quoted 22 lines · Kevin Laatz
Quoted 21 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +15
-0@@ -113,6 +113,24 @@ enum rte_cpu_flag_t {/* (EAX 80000007h) EDX features */ RTE_CPUFLAG_INVTSC ,/**< INVTSC */ +RTE_CPUFLAG_AVX512DQ ,/**< AVX512 Doubleword and Quadword */ +RTE_CPUFLAG_AVX512IFMA ,/**< AVX512 Integer Fused Multiply-Add */ +RTE_CPUFLAG_AVX512CD ,/**< AVX512 Conflict Detection*/ +RTE_CPUFLAG_AVX512BW ,/**< AVX512 Byte and Word */ +RTE_CPUFLAG_AVX512VL ,/**< AVX512 Vector Length */ +RTE_CPUFLAG_AVX512VBMI ,/**< AVX512 Vector Bit Manipulation */ +RTE_CPUFLAG_AVX512VBMI2 ,/**< AVX512 Vector Bit Manipulation 2 */ +RTE_CPUFLAG_GFNI ,/**< Galois Field New Instructions */ +RTE_CPUFLAG_VAES ,/**< Vector AES */ +RTE_CPUFLAG_VPCLMULQDQ ,/**< Vector Carry-less Multiply */ +RTE_CPUFLAG_AVX512VNNI ,/**< AVX512 Vector Neural Network Instructions */ +RTE_CPUFLAG_AVX512BITALG ,/**< AVX512 Bit Algorithms */ +RTE_CPUFLAG_AVX512VPOPCNTDQ ,/**< AVX512 Vector Popcount */ +RTE_CPUFLAG_CLDEMOTE ,/**< Cache Line Demote */ + RTE_CPUFLAG_MOVDIRI, /**< Direct Store Instructions */Quoted 9 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 135
Quoted 5 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +3
-0+RTE_CPUFLAG_MOVDIR64B ,/**< Direct Store Instructions 64B */ +RTE_CPUFLAG_AVX512VP2INTERSECT ,/**< AVX512 Two Register Intersection */ +/* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */This is seen as an ABI break because of the change on
_NUMFLAGS:
https://travis-ci.com/github/ovsrobot/dpdk/jobs/302524264#L2351
Quoted 4 lines
It shouldn't be, as the only API calls we expose that use
rte_cpu_flag_taccept it as an integer parameter to see if the flag is enabled. Theres no use of the enum in a public array or any struct that is sized based on the number of flags, so you should be good to go
Indeed I cannot imagine an ABI incompatibility in this case. The only behaviour change is to accept new (higher) RTE_CPUFLAG values in functions rte_cpu_get_flag_enabled() and rte_cpu_get_flag_name(). Is changing the range of valid values an ABI break?
Why is it flagged by libabigail?
Re: [dpdk-dev] [PATCH v2] eal/cpuflags: add x86 based cpu flags
Quoted 64 lines
-----Original Message-----
From: Thomas Monjalon <[email protected]>
Sent: Friday, March 27, 2020 2:16 PM
To: Neil Horman <[email protected]>; Dodji Seketeli <[email protected]>; [email protected]
Cc: David Marchand <[email protected]>; Laatz, Kevin <[email protected]>; dev <[email protected]>; Richardson, Bruce <[email protected]>; Van Haaren, Harry <[email protected]>; Honnappa Nagarahalli <[email protected]>
Subject: Re: [dpdk-dev] [PATCH v2] eal/cpuflags: add x86 based cpu flags27/03/2020 14:44, Neil Horman:
Quoted 2 lines · David Marchand
Quoted 1 line
On Wed, Mar 25, 2020 at 12:11 PM Kevin Laatz <[email protected]>
Quoted 7 lines · wrote:
Quoted 7 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 116
Quoted 7 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0@@ -113,6 +113,24 @@ enum rte_cpu_flag_t {/* (EAX 80000007h) EDX features */ RTE_CPUFLAG_INVTSC ,/**< INVTSC */ + RTE_CPUFLAG_AVX512DQ, /**< AVX512 Doubleword andQuadword */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 117
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512IFMA, /**< AVX512 Integer FusedMultiply-Add */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 118
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512CD, /**< AVX512 ConflictDetection*/
Quoted 3 lines
Quoted 3 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t {
Quoted 3 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +3
-0+RTE_CPUFLAG_AVX512BW ,/**< AVX512 Byte and Word */ +RTE_CPUFLAG_AVX512VL ,/**< AVX512 Vector Length */ + RTE_CPUFLAG_AVX512VBMI, /**< AVX512 Vector BitManipulation */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 122
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512VBMI2, /**< AVX512 Vector BitManipulation 2 */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 123
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_GFNI, /**< Galois Field NewInstructions */
Quoted 2 lines
Quoted 2 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 125
Quoted 2 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +2
-0+RTE_CPUFLAG_VAES ,/**< Vector AES */ + RTE_CPUFLAG_VPCLMULQDQ, /**< Vector Carry-lessMultiply */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 126
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512VNNI, /**< AVX512 Vector NeuralNetwork Instructions */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 127
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512BITALG, /**< AVX512 Bit Algorithms*/
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 128
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512VPOPCNTDQ, /**< AVX512 Vector Popcount*/
Quoted 2 lines
Quoted 2 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 130
Quoted 2 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +2
-0+RTE_CPUFLAG_CLDEMOTE ,/**< Cache Line Demote */ + RTE_CPUFLAG_MOVDIRI, /**< Direct StoreInstructions */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 131
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_MOVDIR64B, /**< Direct StoreInstructions 64B */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 132
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512VP2INTERSECT, /**< AVX512 Two RegisterIntersection */
Quoted 3 lines
Quoted 3 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t {
Quoted 3 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+/* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always bethe last! */
Quoted 5 lines
Quoted 4 lines
This is seen as an ABI break because of the change on
_NUMFLAGS:
https://travis-ci.com/github/ovsrobot/dpdk/jobs/302524264#L2351It shouldn't be, as the only API calls we expose that use
rte_cpu_flag_taccept
Quoted 1 line
it as an integer parameter to see if the flag is enabled. Theres no use of
the
Quoted 1 line
enum in a public array or any struct that is sized based on the number of
Quoted 8 lines
flags,
Quoted 1 line
so you should be good to go
Indeed I cannot imagine an ABI incompatibility in this case. The only behaviour change is to accept new (higher)
RTE_CPUFLAGvalues in functionsrte_cpu_get_flag_enabled()andrte_cpu_get_flag_name(). Is changing the range of valid values an ABI break?
Why is it flagged by libabigail?
If this enum _MAX value was used by the application to allocate an array, that later our DPDK code would write to it could cause out-of-bounds array accesses of the application supplied array. Abigail doesn't know what applications could use the value for, so it flags it.
IMO Abigail is right to flag it to us - a manual review to understand what that _MAX enum value is used for, and then decide on a case by case basis seems the best way forward to me.
Thanks Neil/Thomas for reviewing, as reply in this thread, I also believe this is not going to break ABI.
Re: [dpdk-dev] [PATCH v2] eal/cpuflags: add x86 based cpu flags
Quoted 73 lines · Harry
Quoted 64 lines
-----Original Message-----
From: Thomas Monjalon <[email protected]>
Sent: Friday, March 27, 2020 2:16 PM
To: Neil Horman <[email protected]>; Dodji Seketeli <[email protected]>; [email protected]
Cc: David Marchand <[email protected]>; Laatz, Kevin <[email protected]>; dev <[email protected]>; Richardson, Bruce <[email protected]>; Van Haaren, Harry <[email protected]>; Honnappa Nagarahalli <[email protected]>
Subject: Re: [dpdk-dev] [PATCH v2] eal/cpuflags: add x86 based cpu flags27/03/2020 14:44, Neil Horman:
Quoted 2 lines · David Marchand
Quoted 1 line
On Wed, Mar 25, 2020 at 12:11 PM Kevin Laatz <[email protected]>
Quoted 7 lines · wrote:
Quoted 7 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 116
Quoted 7 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0@@ -113,6 +113,24 @@ enum rte_cpu_flag_t {/* (EAX 80000007h) EDX features */ RTE_CPUFLAG_INVTSC ,/**< INVTSC */ + RTE_CPUFLAG_AVX512DQ, /**< AVX512 Doubleword andQuadword */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 117
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512IFMA, /**< AVX512 Integer FusedMultiply-Add */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 118
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512CD, /**< AVX512 ConflictDetection*/
Quoted 3 lines
Quoted 3 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t {
Quoted 3 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +3
-0+RTE_CPUFLAG_AVX512BW ,/**< AVX512 Byte and Word */ +RTE_CPUFLAG_AVX512VL ,/**< AVX512 Vector Length */ + RTE_CPUFLAG_AVX512VBMI, /**< AVX512 Vector BitManipulation */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 122
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512VBMI2, /**< AVX512 Vector BitManipulation 2 */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 123
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_GFNI, /**< Galois Field NewInstructions */
Quoted 2 lines
Quoted 2 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 125
Quoted 2 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +2
-0+RTE_CPUFLAG_VAES ,/**< Vector AES */ + RTE_CPUFLAG_VPCLMULQDQ, /**< Vector Carry-lessMultiply */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 126
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512VNNI, /**< AVX512 Vector NeuralNetwork Instructions */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 127
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512BITALG, /**< AVX512 Bit Algorithms*/
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 128
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512VPOPCNTDQ, /**< AVX512 Vector Popcount*/
Quoted 2 lines
Quoted 2 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 130
Quoted 2 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +2
-0+RTE_CPUFLAG_CLDEMOTE ,/**< Cache Line Demote */ + RTE_CPUFLAG_MOVDIRI, /**< Direct StoreInstructions */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 131
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_MOVDIR64B, /**< Direct StoreInstructions 64B */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 132
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512VP2INTERSECT, /**< AVX512 Two RegisterIntersection */
Quoted 3 lines
Quoted 3 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t {
Quoted 3 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+/* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always bethe last! */
Quoted 5 lines
Quoted 4 lines
This is seen as an ABI break because of the change on
_NUMFLAGS:
https://travis-ci.com/github/ovsrobot/dpdk/jobs/302524264#L2351It shouldn't be, as the only API calls we expose that use
rte_cpu_flag_taccept
Quoted 1 line
it as an integer parameter to see if the flag is enabled. Theres no use of
the
Quoted 1 line
enum in a public array or any struct that is sized based on the number of
Quoted 8 lines
flags,
Quoted 1 line
so you should be good to go
Indeed I cannot imagine an ABI incompatibility in this case. The only behaviour change is to accept new (higher)
RTE_CPUFLAGvalues in functionsrte_cpu_get_flag_enabled()andrte_cpu_get_flag_name(). Is changing the range of valid values an ABI break?
Why is it flagged by libabigail?
Quoted 6 lines
If this enum
_MAXvalue was used by the application to allocate an array, that later our DPDK code would write to it could cause out-of-bounds array accesses of the application supplied array. Abigail doesn't know what applications could use the value for, so it flags it.IMO Abigail is right to flag it to us - a manual review to understand what that
_MAXenum value is used for, and then decide on a case by case basis seems the best way forward to me.Thanks Neil/Thomas for reviewing, as reply in this thread, I also believe this is not going to break ABI.
+1 to Harrry's comments.
I can't immediately see how this might break the ABI either.
Re: [dpdk-dev] [PATCH v2] eal/cpuflags: add x86 based cpu flags
27/03/2020 15:36, Ray Kinsella:
Quoted 66 lines · Harry
Quoted 65 lines
From: Thomas Monjalon <[email protected]>
Quoted 53 lines
27/03/2020 14:44, Neil Horman:
Quoted 2 lines · David Marchand
Quoted 1 line
On Wed, Mar 25, 2020 at 12:11 PM Kevin Laatz <[email protected]>
Quoted 7 lines · wrote:
Quoted 7 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 116
Quoted 7 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0@@ -113,6 +113,24 @@ enum rte_cpu_flag_t {/* (EAX 80000007h) EDX features */ RTE_CPUFLAG_INVTSC ,/**< INVTSC */ + RTE_CPUFLAG_AVX512DQ, /**< AVX512 Doubleword andQuadword */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 117
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512IFMA, /**< AVX512 Integer FusedMultiply-Add */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 118
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512CD, /**< AVX512 ConflictDetection*/
Quoted 3 lines
Quoted 3 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t {
Quoted 3 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +3
-0+RTE_CPUFLAG_AVX512BW ,/**< AVX512 Byte and Word */ +RTE_CPUFLAG_AVX512VL ,/**< AVX512 Vector Length */ + RTE_CPUFLAG_AVX512VBMI, /**< AVX512 Vector BitManipulation */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 122
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512VBMI2, /**< AVX512 Vector BitManipulation 2 */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 123
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_GFNI, /**< Galois Field NewInstructions */
Quoted 2 lines
Quoted 2 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 125
Quoted 2 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +2
-0+RTE_CPUFLAG_VAES ,/**< Vector AES */ + RTE_CPUFLAG_VPCLMULQDQ, /**< Vector Carry-lessMultiply */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 126
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512VNNI, /**< AVX512 Vector NeuralNetwork Instructions */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 127
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512BITALG, /**< AVX512 Bit Algorithms*/
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 128
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512VPOPCNTDQ, /**< AVX512 Vector Popcount*/
Quoted 2 lines
Quoted 2 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 130
Quoted 2 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +2
-0+RTE_CPUFLAG_CLDEMOTE ,/**< Cache Line Demote */ + RTE_CPUFLAG_MOVDIRI, /**< Direct StoreInstructions */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 131
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_MOVDIR64B, /**< Direct StoreInstructions 64B */
Quoted 1 line
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t { · line 132
Quoted 1 line
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+ RTE_CPUFLAG_AVX512VP2INTERSECT, /**< AVX512 Two RegisterIntersection */
Quoted 3 lines
Quoted 3 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h enum rte_cpu_flag_t {
Quoted 3 lines
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +1
-0+/* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always bethe last! */
Quoted 5 lines
Quoted 4 lines
This is seen as an ABI break because of the change on
_NUMFLAGS:
https://travis-ci.com/github/ovsrobot/dpdk/jobs/302524264#L2351It shouldn't be, as the only API calls we expose that use
rte_cpu_flag_taccept
Quoted 1 line
it as an integer parameter to see if the flag is enabled. Theres no use of
the
Quoted 1 line
enum in a public array or any struct that is sized based on the number of
Quoted 8 lines
flags,
Quoted 1 line
so you should be good to go
Indeed I cannot imagine an ABI incompatibility in this case. The only behaviour change is to accept new (higher)
RTE_CPUFLAGvalues in functionsrte_cpu_get_flag_enabled()andrte_cpu_get_flag_name(). Is changing the range of valid values an ABI break?
Why is it flagged by libabigail?If this enum
_MAXvalue was used by the application to allocate an array, that later our DPDK code would write to it could cause out-of-bounds array accesses of the application supplied array. Abigail doesn't know what applications could use the value for, so it flags it.
Quoted 7 lines
Quoted 4 lines
IMO Abigail is right to flag it to us - a manual review to understand what that
_MAXenum value is used for, and then decide on a case by case basis seems the best way forward to me.Thanks Neil/Thomas for reviewing, as reply in this thread, I also believe this is not going to break ABI.
+1 to Harrry's comments.
I can't immediately see how this might break the ABI either.
So we all agree that increasing the range of valid rte_cpu_flag_t values is OK for ABI compatibility.
This conclusion must be written as a libabigail exception in this patch.
PATCH v3 1/1 eal/cpuflags: add x86 based cpu flags
In reply to Kevin LaatzThis patch adds CPU flags which will enable the detection of ISA features available on more recent x86 based CPUs.
The CPUID leaf information can be found in Section 1.7 of this document:
https://software.intel.com/sites/default/files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdf
The following CPU flags are added in this patch:
- AVX-512 doubleword and quadword instructions.
- AVX-512 integer fused multiply-add instructions.
- AVX-512 conflict detection instructions.
- AVX-512 byte and word instructions.
- AVX-512 vector length instructions.
- AVX-512 vector bit manipulation instructions.
- AVX-512 vector bit manipulation 2 instructions.
- Galois field new instructions.
- Vector AES instructions.
- Vector carry-less multiply instructions.
- AVX-512 vector neural network instructions.
- AVX-512 for bit algorithm instructions.
- AVX-512 vector popcount instructions.
- Cache line demote instructions.
- Direct store instructions.
- Direct store 64B instructions.
- AVX-512 two register intersection instructions.
---
v2:
- Squashed patch set into single patch.
v3:
- Add abignore entry for 'rte_cpu_flag_t'.
--- devtools/libabigail.abignore | 5 +++++ lib/librte_eal/common/arch/x86/rte_cpuflags.c | 18 ++++++++++++++++++ .../common/include/arch/x86/rte_cpuflags.h | 18 ++++++++++++++++++ 3 files changed, 41 insertions(+)
devtools/libabigail.abignore +5 -0
@@ -11,3 +11,8 @@ type_kind = enum name = rte_crypto_asym_xform_type changed_enumerators = RTE_CRYPTO_ASYM_XFORM_TYPE_LIST_END +; Ignore this enum update as it should not be allocated by the application +[suppress_type] + type_kind = enum + name = rte_cpu_flag_t + changed_enumerators = RTE_CPUFLAG_NUMFLAGS
lib/librte_eal/common/arch/x86/rte_cpuflags.c +18 -0
@@ -120,6 +120,24 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(EM64T, 0x80000001, 0, RTE_REG_EDX, 29) FEAT_DEF(INVTSC, 0x80000007, 0, RTE_REG_EDX, 8) + + FEAT_DEF(AVX512DQ, 0x00000007, 0, RTE_REG_EBX, 17) + FEAT_DEF(AVX512IFMA, 0x00000007, 0, RTE_REG_EBX, 21) + FEAT_DEF(AVX512CD, 0x00000007, 0, RTE_REG_EBX, 28) + FEAT_DEF(AVX512BW, 0x00000007, 0, RTE_REG_EBX, 30) + FEAT_DEF(AVX512VL, 0x00000007, 0, RTE_REG_EBX, 31) + FEAT_DEF(AVX512VBMI, 0x00000007, 0, RTE_REG_ECX, 1) + FEAT_DEF(AVX512VBMI2, 0x00000007, 0, RTE_REG_ECX, 6) + FEAT_DEF(GFNI, 0x00000007, 0, RTE_REG_ECX, 8) + FEAT_DEF(VAES, 0x00000007, 0, RTE_REG_ECX, 9) + FEAT_DEF(VPCLMULQDQ, 0x00000007, 0, RTE_REG_ECX, 10) + FEAT_DEF(AVX512VNNI, 0x00000007, 0, RTE_REG_ECX, 11) + FEAT_DEF(AVX512BITALG, 0x00000007, 0, RTE_REG_ECX, 12) + FEAT_DEF(AVX512VPOPCNTDQ, 0x00000007, 0, RTE_REG_ECX, 14) + FEAT_DEF(CLDEMOTE, 0x00000007, 0, RTE_REG_ECX, 25) + FEAT_DEF(MOVDIRI, 0x00000007, 0, RTE_REG_ECX, 27) + FEAT_DEF(MOVDIR64B, 0x00000007, 0, RTE_REG_ECX, 28) + FEAT_DEF(AVX512VP2INTERSECT, 0x00000007, 0, RTE_REG_EDX, 8) };int
lib/librte_eal/common/include/arch/x86/rte_cpuflags.h +18 -0
@@ -113,6 +113,24 @@ enum rte_cpu_flag_t {/* (EAX 80000007h) EDX features */ RTE_CPUFLAG_INVTSC ,/**< INVTSC */ +RTE_CPUFLAG_AVX512DQ ,/**< AVX512 Doubleword and Quadword */ +RTE_CPUFLAG_AVX512IFMA ,/**< AVX512 Integer Fused Multiply-Add */ +RTE_CPUFLAG_AVX512CD ,/**< AVX512 Conflict Detection*/ +RTE_CPUFLAG_AVX512BW ,/**< AVX512 Byte and Word */ +RTE_CPUFLAG_AVX512VL ,/**< AVX512 Vector Length */ +RTE_CPUFLAG_AVX512VBMI ,/**< AVX512 Vector Bit Manipulation */ +RTE_CPUFLAG_AVX512VBMI2 ,/**< AVX512 Vector Bit Manipulation 2 */ +RTE_CPUFLAG_GFNI ,/**< Galois Field New Instructions */ +RTE_CPUFLAG_VAES ,/**< Vector AES */ +RTE_CPUFLAG_VPCLMULQDQ ,/**< Vector Carry-less Multiply */ +RTE_CPUFLAG_AVX512VNNI ,/**< AVX512 Vector Neural Network Instructions */ +RTE_CPUFLAG_AVX512BITALG ,/**< AVX512 Bit Algorithms */ +RTE_CPUFLAG_AVX512VPOPCNTDQ ,/**< AVX512 Vector Popcount */ +RTE_CPUFLAG_CLDEMOTE ,/**< Cache Line Demote */ +RTE_CPUFLAG_MOVDIRI ,/**< Direct Store Instructions */ +RTE_CPUFLAG_MOVDIR64B ,/**< Direct Store Instructions 64B */ +RTE_CPUFLAG_AVX512VP2INTERSECT ,/**< AVX512 Two Register Intersection */ +/* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */ };
Signature · 1 line
2.17.1
Re: [dpdk-dev] [PATCH v3] eal/cpuflags: add x86 based cpu flags
Quoted 10 lines
-----Original Message-----
From: Laatz, Kevin <[email protected]>
Sent: Monday, March 30, 2020 1:15 PM
To: [email protected]
Cc: Richardson, Bruce <[email protected]>; Van Haaren, Harry <[email protected]>; [email protected]; Kinsella, Ray <[email protected]>; [email protected]; [email protected]; Laatz, Kevin <[email protected]>
Subject: [PATCH v3] eal/cpuflags: add x86 based cpu flags
Quoted 7 lines
This patch adds CPU flags which will enable the detection of ISA features available on more recent x86 based CPUs.
The CPUID leaf information can be found in Section 1.7 of this document:
https://software.intel.com/sites/default/files/managed/c5/15/architecture-
instruction-set-extensions-programming-reference.pdf
Section numbering has been updated, and there doesn't seem to be a stable link to a specific version of the doc, recommend to refer to table name instead? Table 1-2. "Information Returned by CPUID Instruction"
Quoted 19 lines
The following CPU flags are added in this patch:
- AVX-512 doubleword and quadword instructions.
- AVX-512 integer fused multiply-add instructions.
- AVX-512 conflict detection instructions.
- AVX-512 byte and word instructions.
- AVX-512 vector length instructions.
- AVX-512 vector bit manipulation instructions.
- AVX-512 vector bit manipulation 2 instructions.
- Galois field new instructions.
- Vector AES instructions.
- Vector carry-less multiply instructions.
- AVX-512 vector neural network instructions.
- AVX-512 for bit algorithm instructions.
- AVX-512 vector popcount instructions.
- Cache line demote instructions.
- Direct store instructions.
- Direct store 64B instructions.
- AVX-512 two register intersection instructions.
Quoted 1 line
Signed-off-by Kevin Laatz <[email protected]>
Required a git am --3way due to some file movements in master, but applied cleanly on auto-merge.
Add a VNNI documentation line wrap on new line:RTE_CPU_FLAG_AVX512VNNI,
/**< AVX512 Vector ... */
With the above changes/updates;
PATCH v4 1/1 eal/cpuflags: add x86 based cpu flags
In reply to Kevin LaatzThis patch adds CPU flags which will enable the detection of ISA features available on more recent x86 based CPUs.
The CPUID leaf information can be found in
Table 1-2. "Information Returned by CPUID Instruction" of this document:
https://software.intel.com/sites/default/files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdf
The following CPU flags are added in this patch:
- AVX-512 doubleword and quadword instructions.
- AVX-512 integer fused multiply-add instructions.
- AVX-512 conflict detection instructions.
- AVX-512 byte and word instructions.
- AVX-512 vector length instructions.
- AVX-512 vector bit manipulation instructions.
- AVX-512 vector bit manipulation 2 instructions.
- Galois field new instructions.
- Vector AES instructions.
- Vector carry-less multiply instructions.
- AVX-512 vector neural network instructions.
- AVX-512 for bit algorithm instructions.
- AVX-512 vector popcount instructions.
- Cache line demote instructions.
- Direct store instructions.
- Direct store 64B instructions.
- AVX-512 two register intersection instructions.
---
v2:
- Squashed patch set into single patch.
v3:
- Add abignore entry for 'rte_cpu_flag_t'.
v4:
- Updated commit message to reflect updated ISA doc linked.
- Fixed line wrap for VNNI comment.
- Rebased on master.
--- devtools/libabigail.abignore | 5 +++++ lib/librte_eal/x86/include/rte_cpuflags.h | 19 +++++++++++++++++++ lib/librte_eal/x86/rte_cpuflags.c | 18 ++++++++++++++++++ 3 files changed, 42 insertions(+)
devtools/libabigail.abignore +5 -0
@@ -11,3 +11,8 @@ type_kind = enum name = rte_crypto_asym_xform_type changed_enumerators = RTE_CRYPTO_ASYM_XFORM_TYPE_LIST_END +; Ignore this enum update as it should not be allocated by the application +[suppress_type] + type_kind = enum + name = rte_cpu_flag_t + changed_enumerators = RTE_CPUFLAG_NUMFLAGS
lib/librte_eal/x86/include/rte_cpuflags.h +19 -0
@@ -113,6 +113,25 @@ enum rte_cpu_flag_t {/* (EAX 80000007h) EDX features */ RTE_CPUFLAG_INVTSC ,/**< INVTSC */ +RTE_CPUFLAG_AVX512DQ ,/**< AVX512 Doubleword and Quadword */ +RTE_CPUFLAG_AVX512IFMA ,/**< AVX512 Integer Fused Multiply-Add */ +RTE_CPUFLAG_AVX512CD ,/**< AVX512 Conflict Detection*/ +RTE_CPUFLAG_AVX512BW ,/**< AVX512 Byte and Word */ +RTE_CPUFLAG_AVX512VL ,/**< AVX512 Vector Length */ +RTE_CPUFLAG_AVX512VBMI ,/**< AVX512 Vector Bit Manipulation */ +RTE_CPUFLAG_AVX512VBMI2 ,/**< AVX512 Vector Bit Manipulation 2 */ +RTE_CPUFLAG_GFNI ,/**< Galois Field New Instructions */ +RTE_CPUFLAG_VAES ,/**< Vector AES */ +RTE_CPUFLAG_VPCLMULQDQ ,/**< Vector Carry-less Multiply */ +RTE_CPUFLAG_AVX512VNNI , +/**< AVX512 Vector Neural Network Instructions */ +RTE_CPUFLAG_AVX512BITALG ,/**< AVX512 Bit Algorithms */ +RTE_CPUFLAG_AVX512VPOPCNTDQ ,/**< AVX512 Vector Popcount */ +RTE_CPUFLAG_CLDEMOTE ,/**< Cache Line Demote */ +RTE_CPUFLAG_MOVDIRI ,/**< Direct Store Instructions */ +RTE_CPUFLAG_MOVDIR64B ,/**< Direct Store Instructions 64B */ +RTE_CPUFLAG_AVX512VP2INTERSECT ,/**< AVX512 Two Register Intersection */ +/* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */ };
lib/librte_eal/x86/rte_cpuflags.c +18 -0
@@ -120,6 +120,24 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(EM64T, 0x80000001, 0, RTE_REG_EDX, 29) FEAT_DEF(INVTSC, 0x80000007, 0, RTE_REG_EDX, 8) + + FEAT_DEF(AVX512DQ, 0x00000007, 0, RTE_REG_EBX, 17) + FEAT_DEF(AVX512IFMA, 0x00000007, 0, RTE_REG_EBX, 21) + FEAT_DEF(AVX512CD, 0x00000007, 0, RTE_REG_EBX, 28) + FEAT_DEF(AVX512BW, 0x00000007, 0, RTE_REG_EBX, 30) + FEAT_DEF(AVX512VL, 0x00000007, 0, RTE_REG_EBX, 31) + FEAT_DEF(AVX512VBMI, 0x00000007, 0, RTE_REG_ECX, 1) + FEAT_DEF(AVX512VBMI2, 0x00000007, 0, RTE_REG_ECX, 6) + FEAT_DEF(GFNI, 0x00000007, 0, RTE_REG_ECX, 8) + FEAT_DEF(VAES, 0x00000007, 0, RTE_REG_ECX, 9) + FEAT_DEF(VPCLMULQDQ, 0x00000007, 0, RTE_REG_ECX, 10) + FEAT_DEF(AVX512VNNI, 0x00000007, 0, RTE_REG_ECX, 11) + FEAT_DEF(AVX512BITALG, 0x00000007, 0, RTE_REG_ECX, 12) + FEAT_DEF(AVX512VPOPCNTDQ, 0x00000007, 0, RTE_REG_ECX, 14) + FEAT_DEF(CLDEMOTE, 0x00000007, 0, RTE_REG_ECX, 25) + FEAT_DEF(MOVDIRI, 0x00000007, 0, RTE_REG_ECX, 27) + FEAT_DEF(MOVDIR64B, 0x00000007, 0, RTE_REG_ECX, 28) + FEAT_DEF(AVX512VP2INTERSECT, 0x00000007, 0, RTE_REG_EDX, 8) };int
Signature · 1 line
2.25.1
Re: [dpdk-dev] [PATCH v4] eal/cpuflags: add x86 based cpu flags
16/04/2020 13:00, Kevin Laatz:
Quoted 2 lines
This patch adds CPU flags which will enable the detection of ISA features available on more recent x86 based CPUs.
[...]
devtools/libabigail.abignore line 18
Quoted 7 lines
devtools/libabigail.abignore +5
-0+; Ignore this enum update as it should not be allocated by the application +[suppress_type] + type_kind = enum + name = rte_cpu_flag_t + changed_enumerators = RTE_CPUFLAG_NUMFLAGS
The justification is not correct.
The application is allowed to use RTE_CPUFLAG_NUMFLAGS in array allocation. But no API is returning a CPU flag, so the new flags will remain unknown to the application.
However, there is a behaviour change:
The functions rte_cpu_get_flag_name() and rte_cpu_get_flag_enabled() will now accept new values, which were previously considered as an error. Is it an ABI breakage? I would say no.
PS: Who is REALLY maintaining the ABI?
We really miss someone who carefully check all these things, and take care of the doc and tooling.
Re: [dpdk-dev] [PATCH v4] eal/cpuflags: add x86 based cpu flags
(replying this time to the list)
Quoted 17 lines · Thomas Monjalon
16/04/2020 13:00, Kevin Laatz:
Quoted 2 lines
This patch adds CPU flags which will enable the detection of ISA features available on more recent x86 based CPUs.
[...]
devtools/libabigail.abignore line 18
Quoted 7 lines
devtools/libabigail.abignore +5
-0+; Ignore this enum update as it should not be allocated by the application +[suppress_type] + type_kind = enum + name = rte_cpu_flag_t + changed_enumerators = RTE_CPUFLAG_NUMFLAGSThe justification is not correct.
The application is allowed to useRTE_CPUFLAG_NUMFLAGSin array allocation. But no API is returning a CPU flag, so the new flags will remain unknown to the application.
Quoted 4 lines
However, there is a behaviour change:
The functionsrte_cpu_get_flag_name()andrte_cpu_get_flag_enabled()will now accept new values, which were previously considered as an error. Is it an ABI breakage? I would say no.
We saw something similar with the Cryptodev's rte_crypto_sym_xform_type also. Libabigail appears to be particularly sensitive to changes to enumerations. Leaving it to the user to decide if there is a problem.
I am seeing a bit of weirdness though between versions of libabigail. 1.7.1 seems to fine with the change, however 1.2 is reporting an issue.
Kevin - what version are you using?
Quoted 6 lines
PS: Who is REALLY maintaining the ABI?
We really miss someone who carefully check all these things, and take care of the doc and tooling.
I would say that I am missing these changes to libabigail.ignore, which would be useful. Should we consolidate the ABI Policy and ABI Versioning sections of the MAINTAINERS file?
Re: [dpdk-dev] [PATCH v4] eal/cpuflags: add x86 based cpu flags
Quoted 32 lines
(replying this time to the list)
Quoted 4 lines · Thomas Monjalon
16/04/2020 13:00, Kevin Laatz:
Quoted 2 lines
This patch adds CPU flags which will enable the detection of ISA features available on more recent x86 based CPUs.
[...]
Quoted 10 lines
devtools/libabigail.abignore line 18
Quoted 7 lines
devtools/libabigail.abignore +5
-0+; Ignore this enum update as it should not be allocated by the +application [suppress_type] + type_kind = enum + name = rte_cpu_flag_t + changed_enumerators = RTE_CPUFLAG_NUMFLAGSThe justification is not correct.
The application is allowed to useRTE_CPUFLAG_NUMFLAGSin arrayallocation.
Quoted 7 lines
But no API is returning a CPU flag, so the new flags will remain unknown to the application.
However, there is a behaviour change:
The functionsrte_cpu_get_flag_name()andrte_cpu_get_flag_enabled()will now accept new values, which were previously considered as an error. Is it an ABI breakage? I would say no.We saw something similar with the Cryptodev's
rte_crypto_sym_xform_typealso.
Libabigail appears to be particularly sensitive to changes to enumerations. Leaving it to the user to decide if there is a problem.
Quoted 4 lines
I am seeing a bit of weirdness though between versions of libabigail. 1.7.1 seems to fine with the change, however 1.2 is reporting an issue.
Kevin - what version are you using?
I'm using version 1.6.0
Quoted 12 lines
Quoted 7 lines
PS: Who is REALLY maintaining the ABI?
We really miss someone who carefully check all these things, and take care of the doc and tooling.I would say that I am missing these changes to libabigail.ignore, which would be useful.
Should we consolidate the ABI Policy and ABI Versioning sections of theMAINTAINERSfile?
Re: [dpdk-dev] [PATCH v4] eal/cpuflags: add x86 based cpu flags
Quoted 32 lines · Kevin
Quoted 32 lines
(replying this time to the list)
Quoted 4 lines · Thomas Monjalon
16/04/2020 13:00, Kevin Laatz:
Quoted 2 lines
This patch adds CPU flags which will enable the detection of ISA features available on more recent x86 based CPUs.
[...]
Quoted 10 lines
devtools/libabigail.abignore line 18
Quoted 7 lines
devtools/libabigail.abignore +5
-0+; Ignore this enum update as it should not be allocated by the +application [suppress_type] + type_kind = enum + name = rte_cpu_flag_t + changed_enumerators = RTE_CPUFLAG_NUMFLAGSThe justification is not correct.
The application is allowed to useRTE_CPUFLAG_NUMFLAGSin arrayallocation.
Quoted 7 lines
But no API is returning a CPU flag, so the new flags will remain unknown to the application.
However, there is a behaviour change:
The functionsrte_cpu_get_flag_name()andrte_cpu_get_flag_enabled()will now accept new values, which were previously considered as an error. Is it an ABI breakage? I would say no.We saw something similar with the Cryptodev's
rte_crypto_sym_xform_typealso.
Libabigail appears to be particularly sensitive to changes to enumerations. Leaving it to the user to decide if there is a problem.
Quoted 6 lines
Quoted 4 lines
I am seeing a bit of weirdness though between versions of libabigail. 1.7.1 seems to fine with the change, however 1.2 is reporting an issue.
Kevin - what version are you using?
I'm using version 1.6.0
right you are either on Fedora 31 or some Ubuntu v19.xx, right?
Quoted 14 lines
Quoted 13 lines
Quoted 8 lines
PS: Who is REALLY maintaining the ABI?
We really miss someone who carefully check all these things, and take care of the doc and tooling.I would say that I am missing these changes to libabigail.ignore, which would be useful.
Should we consolidate the ABI Policy and ABI Versioning sections of theMAINTAINERSfile?
Re: [dpdk-dev] [PATCH v4] eal/cpuflags: add x86 based cpu flags
Quoted 35 lines · Kevin
Quoted 18 lines
Quoted 18 lines
(replying this time to the list)
Quoted 14 lines · Thomas Monjalon
16/04/2020 13:00, Kevin Laatz:
Quoted 2 lines
This patch adds CPU flags which will enable the detection of ISA features available on more recent x86 based CPUs.
[...]
devtools/libabigail.abignore line 18
Quoted 7 lines
devtools/libabigail.abignore +5
-0+; Ignore this enum update as it should not be allocated by the +application [suppress_type] + type_kind = enum + name = rte_cpu_flag_t + changed_enumerators = RTE_CPUFLAG_NUMFLAGSThe justification is not correct.
The application is allowed to useRTE_CPUFLAG_NUMFLAGSin arrayQuoted 6 lines
Quoted 6 lines
allocation.
Quoted 5 lines
But no API is returning a CPU flag, so the new flags will remain unknown to the application.
However, there is a behaviour change:
The functionsrte_cpu_get_flag_name()and
rte_cpu_get_flag_enabled()Quoted 1 line
Quoted 1 line
Quoted 1 line
will now accept new values, which were previously considered as an
error.
Quoted 7 lines
Quoted 7 lines
Quoted 1 line
Is it an ABI breakage? I would say no.
We saw something similar with the Cryptodev's
rte_crypto_sym_xform_typealso.
Libabigail appears to be particularly sensitive to changes to enumerations. Leaving it to the user to decide if there is a problem.
Quoted 8 lines
Quoted 6 lines
Quoted 4 lines
I am seeing a bit of weirdness though between versions of libabigail. 1.7.1 seems to fine with the change, however 1.2 is reporting an issue.
Kevin - what version are you using?
I'm using version 1.6.0
right you are either on Fedora 31 or some Ubuntu v19.xx, right?
At the time of making the patch: Ubuntu 18.04 with a manually upgraded libabigail Currently Ubuntu 20.04 (beta).
Quoted 15 lines
Quoted 15 lines
Quoted 14 lines
Quoted 9 lines
PS: Who is REALLY maintaining the ABI?
We really miss someone who carefully check all these things, and take care of the doc and tooling.I would say that I am missing these changes to libabigail.ignore, which would be useful.
Should we consolidate the ABI Policy and ABI Versioning sections of theMAINTAINERSfile?
Re: [dpdk-dev] [PATCH v4] eal/cpuflags: add x86 based cpu flags
In reply to Ray Kinsella27/04/2020 11:27, Ray Kinsella:
Quoted 7 lines · Thomas Monjalon
Quoted 3 lines
PS: Who is REALLY maintaining the ABI?
We really miss someone who carefully check all these things, and take care of the doc and tooling.I would say that I am missing these changes to libabigail.ignore, which would be useful. Should we consolidate the ABI Policy and ABI Versioning sections of the
MAINTAINERSfile?
Yes, I think it does not make sense spliting ABI topic in 2 sections in MAINTAINERS file.
We need to have a clear ownership covering policy, libs, tooling and doc. Let's agree to merge all in one section please.
Re: [dpdk-dev] [PATCH v4] eal/cpuflags: add x86 based cpu flags
Quoted 6 lines · Thomas Monjalon
27/04/2020 11:27, Ray Kinsella:
Quoted 5 lines · Thomas Monjalon
Quoted 3 lines
PS: Who is REALLY maintaining the ABI?
We really miss someone who carefully check all these things, and take care of the doc and tooling.
Quoted 8 lines
Quoted 2 lines
I would say that I am missing these changes to libabigail.ignore, which would be useful. Should we consolidate the ABI Policy and ABI Versioning sections of the
MAINTAINERSfile?Yes, I think it does not make sense spliting ABI topic in 2 sections in
MAINTAINERSfile.
We need to have a clear ownership covering policy, libs, tooling and doc. Let's agree to merge all in one section please.
I would suggest merging and listing myself and Neil as maintainers? Unless you are aware of another potential owner?
Ray K
Re: [dpdk-dev] [PATCH v4] eal/cpuflags: add x86 based cpu flags
Quoted 12 lines · Ray Kinsella
Quoted 9 lines · Thomas Monjalon
27/04/2020 11:27, Ray Kinsella:
Quoted 7 lines · Thomas Monjalon
Quoted 3 lines
PS: Who is REALLY maintaining the ABI?
We really miss someone who carefully check all these things, and take care of the doc and tooling.I would say that I am missing these changes to libabigail.ignore, which would be useful. Should we consolidate the ABI Policy and ABI Versioning sections of the
MAINTAINERSfile?
Quoted 9 lines
Quoted 5 lines
Yes, I think it does not make sense spliting ABI topic in 2 sections in
MAINTAINERSfile.
We need to have a clear ownership covering policy, libs, tooling and doc. Let's agree to merge all in one section please.I would suggest merging and listing myself and Neil as maintainers? Unless you are aware of another potential owner?
I'm ok with this
Neil
Quoted 3 lines
Ray K
Re: [dpdk-dev] [PATCH v4] eal/cpuflags: add x86 based cpu flags
Quoted 19 lines · Neil Horman
Quoted 18 lines · Ray Kinsella
Quoted 14 lines · Thomas Monjalon
27/04/2020 11:27, Ray Kinsella:
Quoted 7 lines · Thomas Monjalon
Quoted 3 lines
PS: Who is REALLY maintaining the ABI?
We really miss someone who carefully check all these things, and take care of the doc and tooling.I would say that I am missing these changes to libabigail.ignore, which would be useful. Should we consolidate the ABI Policy and ABI Versioning sections of the
MAINTAINERSfile?Yes, I think it does not make sense spliting ABI topic in 2 sections in
MAINTAINERSfile.
We need to have a clear ownership covering policy, libs, tooling and doc. Let's agree to merge all in one section please.
Quoted 5 lines
Quoted 3 lines
I would suggest merging and listing myself and Neil as maintainers? Unless you are aware of another potential owner?
I'm ok with this
Neil
ok I will take care of it in the next rev of the
"[PATCH] abi: change references to abi 20.0.1 to abi v21"
Quoted 3 lines
Quoted 3 lines
Ray K
Re: [dpdk-dev] [PATCH v4] eal/cpuflags: add x86 based cpu flags
In reply to Thomas MonjalonQuoted 30 lines
-----Original Message-----
From: Thomas Monjalon <[email protected]>
Sent: Saturday 25 April 2020 17:04
To: Kinsella, Ray <[email protected]>; [email protected]; Laatz, Kevin <[email protected]>
Cc: [email protected]; Richardson, Bruce <[email protected]>; Van Haaren, Harry <[email protected]>; [email protected]
Subject: Re: [dpdk-dev] [PATCH v4] eal/cpuflags: add x86 based cpu flags16/04/2020 13:00, Kevin Laatz:
Quoted 2 lines
This patch adds CPU flags which will enable the detection of ISA features available on more recent x86 based CPUs.
[...]
devtools/libabigail.abignore line 18
Quoted 7 lines
devtools/libabigail.abignore +5
-0+; Ignore this enum update as it should not be allocated by the +application [suppress_type] + type_kind = enum + name = rte_cpu_flag_t + changed_enumerators = RTE_CPUFLAG_NUMFLAGSThe justification is not correct.
The application is allowed to useRTE_CPUFLAG_NUMFLAGSin array allocation.
But no API is returning a CPU flag, so the new flags will remain unknown to the application.
Quoted 5 lines
However, there is a behaviour change:
The functionsrte_cpu_get_flag_name()andrte_cpu_get_flag_enabled()will now accept new values, which were previously considered as an error.
Is it an ABI breakage? I would say no.
We saw something similar with the Cryptodev's rte_crypto_sym_xform_type also. Libabigail appears to be particularly sensitive to changes to enumerations. Leaving it to the user to decide if there is a problem.
I am seeing a bit of weirdness though between versions of libabigail. 1.7.1 seems to fine with the change, however 1.2 is reporting an issue.
Kevin - what version are you using?
Quoted 4 lines
PS: Who is REALLY maintaining the ABI?
We really miss someone who carefully check all these things, and take care of the doc and tooling.
I would say that I am missing these changes to libabigail.ignore, which would be useful. Should we consolidate the ABI Policy and ABI Versioning sections of the MAINTAINERS file?
PATCH v5 1/1 eal/cpuflags: add x86 based cpu flags
In reply to Kevin LaatzThis patch adds CPU flags which will enable the detection of ISA features available on more recent x86 based CPUs.
The CPUID leaf information can be found in
Table 1-2. "Information Returned by CPUID Instruction" of this document:
https://software.intel.com/sites/default/files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdf
The following CPU flags are added in this patch:
- AVX-512 doubleword and quadword instructions.
- AVX-512 integer fused multiply-add instructions.
- AVX-512 conflict detection instructions.
- AVX-512 byte and word instructions.
- AVX-512 vector length instructions.
- AVX-512 vector bit manipulation instructions.
- AVX-512 vector bit manipulation 2 instructions.
- Galois field new instructions.
- Vector AES instructions.
- Vector carry-less multiply instructions.
- AVX-512 vector neural network instructions.
- AVX-512 for bit algorithm instructions.
- AVX-512 vector popcount instructions.
- Cache line demote instructions.
- Direct store instructions.
- Direct store 64B instructions.
- AVX-512 two register intersection instructions.
---
v2:
- Squashed patch set into single patch.
v3:
- Add abignore entry for 'rte_cpu_flag_t'.
v4:
- Updated commit message to reflect updated ISA doc linked.
- Fixed line wrap for VNNI comment.
- Rebased on master.
v5:
- Update abignore entry justification.
--- devtools/libabigail.abignore | 5 +++++ lib/librte_eal/x86/include/rte_cpuflags.h | 19 +++++++++++++++++++ lib/librte_eal/x86/rte_cpuflags.c | 18 ++++++++++++++++++ 3 files changed, 42 insertions(+)
devtools/libabigail.abignore +5 -0
@@ -11,3 +11,8 @@ type_kind = enum name = rte_crypto_asym_xform_type changed_enumerators = RTE_CRYPTO_ASYM_XFORM_TYPE_LIST_END +; Ignore this enum update as new flags remain unknown to applications +[suppress_type] + type_kind = enum + name = rte_cpu_flag_t + changed_enumerators = RTE_CPUFLAG_NUMFLAGS
lib/librte_eal/x86/include/rte_cpuflags.h +19 -0
@@ -113,6 +113,25 @@ enum rte_cpu_flag_t {/* (EAX 80000007h) EDX features */ RTE_CPUFLAG_INVTSC ,/**< INVTSC */ +RTE_CPUFLAG_AVX512DQ ,/**< AVX512 Doubleword and Quadword */ +RTE_CPUFLAG_AVX512IFMA ,/**< AVX512 Integer Fused Multiply-Add */ +RTE_CPUFLAG_AVX512CD ,/**< AVX512 Conflict Detection*/ +RTE_CPUFLAG_AVX512BW ,/**< AVX512 Byte and Word */ +RTE_CPUFLAG_AVX512VL ,/**< AVX512 Vector Length */ +RTE_CPUFLAG_AVX512VBMI ,/**< AVX512 Vector Bit Manipulation */ +RTE_CPUFLAG_AVX512VBMI2 ,/**< AVX512 Vector Bit Manipulation 2 */ +RTE_CPUFLAG_GFNI ,/**< Galois Field New Instructions */ +RTE_CPUFLAG_VAES ,/**< Vector AES */ +RTE_CPUFLAG_VPCLMULQDQ ,/**< Vector Carry-less Multiply */ +RTE_CPUFLAG_AVX512VNNI , +/**< AVX512 Vector Neural Network Instructions */ +RTE_CPUFLAG_AVX512BITALG ,/**< AVX512 Bit Algorithms */ +RTE_CPUFLAG_AVX512VPOPCNTDQ ,/**< AVX512 Vector Popcount */ +RTE_CPUFLAG_CLDEMOTE ,/**< Cache Line Demote */ +RTE_CPUFLAG_MOVDIRI ,/**< Direct Store Instructions */ +RTE_CPUFLAG_MOVDIR64B ,/**< Direct Store Instructions 64B */ +RTE_CPUFLAG_AVX512VP2INTERSECT ,/**< AVX512 Two Register Intersection */ +/* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */ };
lib/librte_eal/x86/rte_cpuflags.c +18 -0
@@ -120,6 +120,24 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(EM64T, 0x80000001, 0, RTE_REG_EDX, 29) FEAT_DEF(INVTSC, 0x80000007, 0, RTE_REG_EDX, 8) + + FEAT_DEF(AVX512DQ, 0x00000007, 0, RTE_REG_EBX, 17) + FEAT_DEF(AVX512IFMA, 0x00000007, 0, RTE_REG_EBX, 21) + FEAT_DEF(AVX512CD, 0x00000007, 0, RTE_REG_EBX, 28) + FEAT_DEF(AVX512BW, 0x00000007, 0, RTE_REG_EBX, 30) + FEAT_DEF(AVX512VL, 0x00000007, 0, RTE_REG_EBX, 31) + FEAT_DEF(AVX512VBMI, 0x00000007, 0, RTE_REG_ECX, 1) + FEAT_DEF(AVX512VBMI2, 0x00000007, 0, RTE_REG_ECX, 6) + FEAT_DEF(GFNI, 0x00000007, 0, RTE_REG_ECX, 8) + FEAT_DEF(VAES, 0x00000007, 0, RTE_REG_ECX, 9) + FEAT_DEF(VPCLMULQDQ, 0x00000007, 0, RTE_REG_ECX, 10) + FEAT_DEF(AVX512VNNI, 0x00000007, 0, RTE_REG_ECX, 11) + FEAT_DEF(AVX512BITALG, 0x00000007, 0, RTE_REG_ECX, 12) + FEAT_DEF(AVX512VPOPCNTDQ, 0x00000007, 0, RTE_REG_ECX, 14) + FEAT_DEF(CLDEMOTE, 0x00000007, 0, RTE_REG_ECX, 25) + FEAT_DEF(MOVDIRI, 0x00000007, 0, RTE_REG_ECX, 27) + FEAT_DEF(MOVDIR64B, 0x00000007, 0, RTE_REG_ECX, 28) + FEAT_DEF(AVX512VP2INTERSECT, 0x00000007, 0, RTE_REG_EDX, 8) };int
Signature · 1 line
2.25.1
Re: [dpdk-dev] [PATCH v5] eal/cpuflags: add x86 based cpu flags
devtools/libabigail.abignore
Quoted 53 lines · Kevin Laatz
This patch adds CPU flags which will enable the detection of ISA features available on more recent x86 based CPUs.
The CPUID leaf information can be found in
Table 1-2. "Information Returned by CPUID Instruction" of this document:
https://software.intel.com/sites/default/files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdfThe following CPU flags are added in this patch:
- AVX-512 doubleword and quadword instructions.
- AVX-512 integer fused multiply-add instructions.
- AVX-512 conflict detection instructions.
- AVX-512 byte and word instructions.
- AVX-512 vector length instructions.
- AVX-512 vector bit manipulation instructions.
- AVX-512 vector bit manipulation 2 instructions.
- Galois field new instructions.
- Vector AES instructions.
- Vector carry-less multiply instructions.
- AVX-512 vector neural network instructions.
- AVX-512 for bit algorithm instructions.
- AVX-512 vector popcount instructions.
- Cache line demote instructions.
- Direct store instructions.
- Direct store 64B instructions.
- AVX-512 two register intersection instructions.Signed-off-by Kevin Laatz <[email protected]>Acked-by Harry van Haaren <[email protected]>---
v2:
- Squashed patch set into single patch.v3:
- Add abignore entry for 'rte_cpu_flag_t'.v4:
- Updated commit message to reflect updated ISA doc linked.
- Fixed line wrap for VNNI comment.
- Rebased on master.v5:
- Update abignore entry justification.--- devtools/libabigail.abignore | 5 +++++ lib/librte_eal/x86/include/rte_cpuflags.h | 19 +++++++++++++++++++ lib/librte_eal/x86/rte_cpuflags.c | 18 ++++++++++++++++++ 3 files changed, 42 insertions(+)devtools/libabigail.abignore +0
-0
Kevin - you still have the surpession.
I am testing locally with 1.7.1, and it doesn't complain when I disable the supression. Are you seeing something different?
Quoted 69 lines
devtools/libabigail.abignore +5
-0@@ -11,3 +11,8 @@ type_kind = enum name = rte_crypto_asym_xform_type changed_enumerators = RTE_CRYPTO_ASYM_XFORM_TYPE_LIST_END +; Ignore this enum update as new flags remain unknown to applications +[suppress_type] + type_kind = enum + name = rte_cpu_flag_t + changed_enumerators = RTE_CPUFLAG_NUMFLAGSlib/librte_eal/x86/include/rte_cpuflags.h +19
-0@@ -113,6 +113,25 @@ enum rte_cpu_flag_t {/* (EAX 80000007h) EDX features */ RTE_CPUFLAG_INVTSC ,/**< INVTSC */ +RTE_CPUFLAG_AVX512DQ ,/**< AVX512 Doubleword and Quadword */ +RTE_CPUFLAG_AVX512IFMA ,/**< AVX512 Integer Fused Multiply-Add */ +RTE_CPUFLAG_AVX512CD ,/**< AVX512 Conflict Detection*/ +RTE_CPUFLAG_AVX512BW ,/**< AVX512 Byte and Word */ +RTE_CPUFLAG_AVX512VL ,/**< AVX512 Vector Length */ +RTE_CPUFLAG_AVX512VBMI ,/**< AVX512 Vector Bit Manipulation */ +RTE_CPUFLAG_AVX512VBMI2 ,/**< AVX512 Vector Bit Manipulation 2 */ +RTE_CPUFLAG_GFNI ,/**< Galois Field New Instructions */ +RTE_CPUFLAG_VAES ,/**< Vector AES */ +RTE_CPUFLAG_VPCLMULQDQ ,/**< Vector Carry-less Multiply */ +RTE_CPUFLAG_AVX512VNNI , +/**< AVX512 Vector Neural Network Instructions */ +RTE_CPUFLAG_AVX512BITALG ,/**< AVX512 Bit Algorithms */ +RTE_CPUFLAG_AVX512VPOPCNTDQ ,/**< AVX512 Vector Popcount */ +RTE_CPUFLAG_CLDEMOTE ,/**< Cache Line Demote */ +RTE_CPUFLAG_MOVDIRI ,/**< Direct Store Instructions */ +RTE_CPUFLAG_MOVDIR64B ,/**< Direct Store Instructions 64B */ +RTE_CPUFLAG_AVX512VP2INTERSECT ,/**< AVX512 Two Register Intersection */ +/* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */ };lib/librte_eal/x86/rte_cpuflags.c +18
-0@@ -120,6 +120,24 @@ const struct feature_entry rte_cpu_feature_table[] = { FEAT_DEF(EM64T, 0x80000001, 0, RTE_REG_EDX, 29) FEAT_DEF(INVTSC, 0x80000007, 0, RTE_REG_EDX, 8) + + FEAT_DEF(AVX512DQ, 0x00000007, 0, RTE_REG_EBX, 17) + FEAT_DEF(AVX512IFMA, 0x00000007, 0, RTE_REG_EBX, 21) + FEAT_DEF(AVX512CD, 0x00000007, 0, RTE_REG_EBX, 28) + FEAT_DEF(AVX512BW, 0x00000007, 0, RTE_REG_EBX, 30) + FEAT_DEF(AVX512VL, 0x00000007, 0, RTE_REG_EBX, 31) + FEAT_DEF(AVX512VBMI, 0x00000007, 0, RTE_REG_ECX, 1) + FEAT_DEF(AVX512VBMI2, 0x00000007, 0, RTE_REG_ECX, 6) + FEAT_DEF(GFNI, 0x00000007, 0, RTE_REG_ECX, 8) + FEAT_DEF(VAES, 0x00000007, 0, RTE_REG_ECX, 9) + FEAT_DEF(VPCLMULQDQ, 0x00000007, 0, RTE_REG_ECX, 10) + FEAT_DEF(AVX512VNNI, 0x00000007, 0, RTE_REG_ECX, 11) + FEAT_DEF(AVX512BITALG, 0x00000007, 0, RTE_REG_ECX, 12) + FEAT_DEF(AVX512VPOPCNTDQ, 0x00000007, 0, RTE_REG_ECX, 14) + FEAT_DEF(CLDEMOTE, 0x00000007, 0, RTE_REG_ECX, 25) + FEAT_DEF(MOVDIRI, 0x00000007, 0, RTE_REG_ECX, 27) + FEAT_DEF(MOVDIR64B, 0x00000007, 0, RTE_REG_ECX, 28) + FEAT_DEF(AVX512VP2INTERSECT, 0x00000007, 0, RTE_REG_EDX, 8) };int
Re: [dpdk-dev] [PATCH v5] eal/cpuflags: add x86 based cpu flags
<snip>
Quoted 6 lines
Quoted 6 lines
--- devtools/libabigail.abignore | 5 +++++ lib/librte_eal/x86/include/rte_cpuflags.h | 19 +++++++++++++++++++ lib/librte_eal/x86/rte_cpuflags.c | 18 ++++++++++++++++++ 3 files changed, 42 insertions(+)
Quoted 10 lines
devtools/libabigail.abignore
Quoted 4 lines
(quoted diff) +0
-0b/devtools/libabigail.abignore index
a59df8f13..045f436fb100644devtools/libabigail.abignore +0
-0Kevin - you still have the surpession.
I am testing locally with 1.7.1, and it doesn't complain when I disable the supression.
Are you seeing something different?
Ray,
I have re-tested and with libabigail 1.6 and it reports the addition of the flags as an ABI break without the abignore suppression. With the suppression, it will still report changes to existing flags (e.g. inserting a new flag somewhere in the middle) in the enum as an ABI break, as expected.
The Travis CI is also based on Ubuntu 18.04 LTS, which uses libabigail 1.2-1. Without the suppression the community Travis CI builds fail on this false positive.
-Kevin
Re: [dpdk-dev] [PATCH v5] eal/cpuflags: add x86 based cpu flags
28/04/2020 20:11, Laatz, Kevin:
Quoted 9 lines
Quoted 8 lines
devtools/libabigail.abignore
Quoted 2 lines
devtools/libabigail.abignore +0
-0Kevin - you still have the surpession.
I am testing locally with 1.7.1, and it doesn't complain when I disable the supression.
Are you seeing something different?
Quoted 5 lines
Ray,
I have re-tested and with libabigail 1.6 and it reports the addition of the flags as an ABI break without the abignore suppression. With the suppression, it will still report changes to existing flags (e.g. inserting a new flag somewhere in the middle) in the enum as an ABI break, as expected.The Travis CI is also based on Ubuntu 18.04 LTS, which uses libabigail 1.2-1. Without the suppression the community Travis CI builds fail on this false positive.
I think Travis uses libabigail 1.6:
http://git.dpdk.org/dpdk/tree/.ci/linux-build.sh#n61
Re: [dpdk-dev] [PATCH v5] eal/cpuflags: add x86 based cpu flags
In reply to Ray KinsellaQuoted 4 lines · On Tue, 28 Apr 2020 17:39:24 +0100 Ray Kinsella <[email protected]> wrote:
lib/librte_eal/x86/include/rte_cpuflags.h enum rte_cpu_flag_t { · line 137
Quoted 4 lines
lib/librte_eal/x86/include/rte_cpuflags.h +1
-0+/* The last item */ RTE_CPUFLAG_NUMFLAGS, /**< This should always be the last! */ };
These kind of enums break API's. We should remove them all in 20.11.
Re: [dpdk-dev] [PATCH v5] eal/cpuflags: add x86 based cpu flags
In reply to Ray KinsellaQuoted 8 lines · Ray Kinsella
devtools/libabigail.abignore
Quoted 4 lines
devtools/libabigail.abignore +0
-0Kevin - you still have the surpession.
I am testing locally with 1.7.1, and it doesn't complain when I disable the supression. Are you seeing something different?
Using current master libabigail, without the rule Kevin included, I get the warning:
1 function with some indirect sub-type change:
[C] 'function int rte_cpu_get_flag_enabled(rte_cpu_flag_t)' at
rte_cpuflags.c:144:1 has some indirect sub-type changes:
parameter 1 of type 'enum rte_cpu_flag_t' has sub-type changes:
type size hasn't changed
17 enumerator insertions:
'rte_cpu_flag_t::RTE_CPUFLAG_AVX512DQ' value '87'
'rte_cpu_flag_t::RTE_CPUFLAG_AVX512IFMA' value '88'
'rte_cpu_flag_t::RTE_CPUFLAG_AVX512CD' value '89'
'rte_cpu_flag_t::RTE_CPUFLAG_AVX512BW' value '90'
'rte_cpu_flag_t::RTE_CPUFLAG_AVX512VL' value '91'
'rte_cpu_flag_t::RTE_CPUFLAG_AVX512VBMI' value '92'
'rte_cpu_flag_t::RTE_CPUFLAG_AVX512VBMI2' value '93'
'rte_cpu_flag_t::RTE_CPUFLAG_GFNI' value '94'
'rte_cpu_flag_t::RTE_CPUFLAG_VAES' value '95'
'rte_cpu_flag_t::RTE_CPUFLAG_VPCLMULQDQ' value '96'
'rte_cpu_flag_t::RTE_CPUFLAG_AVX512VNNI' value '97'
'rte_cpu_flag_t::RTE_CPUFLAG_AVX512BITALG' value '98'
'rte_cpu_flag_t::RTE_CPUFLAG_AVX512VPOPCNTDQ' value '99'
'rte_cpu_flag_t::RTE_CPUFLAG_CLDEMOTE' value '100'
'rte_cpu_flag_t::RTE_CPUFLAG_MOVDIRI' value '101'
'rte_cpu_flag_t::RTE_CPUFLAG_MOVDIR64B' value '102'
'rte_cpu_flag_t::RTE_CPUFLAG_AVX512VP2INTERSECT' value '103'
1 enumerator change:
'rte_cpu_flag_t::RTE_CPUFLAG_NUMFLAGS' from value '87' to'104' at rte_cpuflags.h:12:1
Ray, could you check that the reference and new dumps in your env contain this enum?
$ grep RTE_CPUFLAG_NUMFLAGS
$HOME/abi/v20.02/x86_64-native-linux-gcc+shared+debug+ASSERT+RTE_IBVERBS_LINK_DLOPEN/dump/librte_eal.dump
<enumerator name='RTE_CPUFLAG_NUMFLAGS' value='87'/>
$ grep RTE_CPUFLAG_NUMFLAGS
$HOME/builds/x86_64-native-linux-gcc+shared+debug+ASSERT+RTE_IBVERBS_LINK_DLOPEN/install/dump/librte_eal.dump
<enumerator name='RTE_CPUFLAG_NUMFLAGS' value='104'/>If you are missing those, you might have built dpdk without debuginfo.
Signature · 1 line
David Marchand
Ray Kinsella
Acked-byRe: [dpdk-dev] [PATCH v5] eal/cpuflags: add x86 based cpu flags
So that isn't the issue either.
$ grep RTE_CPUFLAG_NUMFLAGS build-gcc-shared/install/dump/librte_eal.dump
4646: <enumerator name='RTE_CPUFLAG_NUMFLAGS' value='104'/>
$ grep RTE_CPUFLAG_NUMFLAGS /build/dpdk/reference/v20.02/build-gcc-shared/dump/librte_eal.dump
3296: <enumerator name='RTE_CPUFLAG_NUMFLAGS' value='87'/>
1.7-1.fc31 @updatesI finally got libabigail complaining about rte_cpu_flag_t, after doing a complete clear down. So the suppression _is_ required.
I then spent the following hour trying to identify the gremlin in the system with no luck. In anycase, added my ack below.
Quoted 45 lines · David Marchand
Quoted 8 lines · Ray Kinsella
devtools/libabigail.abignore
Quoted 4 lines
devtools/libabigail.abignore +0
-0Kevin - you still have the surpession.
I am testing locally with 1.7.1, and it doesn't complain when I disable the supression. Are you seeing something different?Using current master libabigail, without the rule Kevin included, I get the warning:
1 function with some indirect sub-type change:
[C] 'function int
rte_cpu_get_flag_enabled(rte_cpu_flag_t)' at
rte_cpuflags.c:144:1 has some indirect sub-type changes:parameter 1 of type 'enum rte_cpu_flag_t' has sub-type changes: type size hasn't changed 17 enumerator insertions: 'rte_cpu_flag_t::RTE_CPUFLAG_AVX512DQ' value '87' 'rte_cpu_flag_t::RTE_CPUFLAG_AVX512IFMA' value '88' 'rte_cpu_flag_t::RTE_CPUFLAG_AVX512CD' value '89' 'rte_cpu_flag_t::RTE_CPUFLAG_AVX512BW' value '90' 'rte_cpu_flag_t::RTE_CPUFLAG_AVX512VL' value '91' 'rte_cpu_flag_t::RTE_CPUFLAG_AVX512VBMI' value '92' 'rte_cpu_flag_t::RTE_CPUFLAG_AVX512VBMI2' value '93' 'rte_cpu_flag_t::RTE_CPUFLAG_GFNI' value '94' 'rte_cpu_flag_t::RTE_CPUFLAG_VAES' value '95' 'rte_cpu_flag_t::RTE_CPUFLAG_VPCLMULQDQ' value '96' 'rte_cpu_flag_t::RTE_CPUFLAG_AVX512VNNI' value '97' 'rte_cpu_flag_t::RTE_CPUFLAG_AVX512BITALG' value '98' 'rte_cpu_flag_t::RTE_CPUFLAG_AVX512VPOPCNTDQ' value '99' 'rte_cpu_flag_t::RTE_CPUFLAG_CLDEMOTE' value '100' 'rte_cpu_flag_t::RTE_CPUFLAG_MOVDIRI' value '101' 'rte_cpu_flag_t::RTE_CPUFLAG_MOVDIR64B' value '102' 'rte_cpu_flag_t::RTE_CPUFLAG_AVX512VP2INTERSECT' value '103' 1 enumerator change: 'rte_cpu_flag_t::RTE_CPUFLAG_NUMFLAGS' from value '87' to'104' at rte_cpuflags.h:12:1
Ray, could you check that the reference and new dumps in your env contain this enum?
Quoted 9 lines
$ grep RTE_CPUFLAG_NUMFLAGS $HOME/abi/v20.02/x86_64-native-linux-gcc+shared+debug+ASSERT+RTE_IBVERBS_LINK_DLOPEN/dump/librte_eal.dump <enumerator name='RTE_CPUFLAG_NUMFLAGS' value='87'/> $ grep RTE_CPUFLAG_NUMFLAGS $HOME/builds/x86_64-native-linux-gcc+shared+debug+ASSERT+RTE_IBVERBS_LINK_DLOPEN/install/dump/librte_eal.dump <enumerator name='RTE_CPUFLAG_NUMFLAGS' value='104'/>If you are missing those, you might have built dpdk without debuginfo.
Thomas Monjalon
AppliedRe: [dpdk-dev] [PATCH v5] eal/cpuflags: add x86 based cpu flags
Quoted 1 line
Acked-by Ray Kinsella <[email protected]>
Applied, thanks