Skip to contentSkip to conversation

x86/cpu: Add AVX512_FP16 cpu feature

PATCH Applied Cathy Zhang 3 participants 3 messages Dec 16, 2020

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86/cpu: Add AVX512_FP16 cpu feature

AVX512 Half-precision floating point (FP16) has better performance compared to FP32 if the presicion or magnitude requirements are met. It's defined as CPUID.(EAX=7,ECX=0):EDX[bit 23].

Refer to
https://software.intel.com/content/www/us/en/develop/download/\
intel-architecture-instruction-set-extensions-programming-reference.html

---
 target/i386/cpu.c | 2 +-
 target/i386/cpu.h | 2 ++
 2 files changed, 3 insertions(+), 1 deletion(-)
target/i386/cpu.c +1 -1
index ca997a68cd..c4d623b8b9 100644
@@ -979,7 +979,7 @@ static FeatureWordInfo feature_word_info[FEATURE_WORDS] = {
             "avx512-vp2intersect", NULL, "md-clear", NULL,
             NULL, NULL, "serialize", NULL,
             "tsx-ldtrk", NULL, NULL /* pconfig */, NULL,
-            NULL, NULL, NULL, NULL,
+            NULL, NULL, NULL, "avx512-fp16",
             NULL, NULL, "spec-ctrl", "stibp",
             NULL, "arch-capabilities", "core-capability", "ssbd",
         },
target/i386/cpu.h +2 -0
index c4a49c06a8..6fd675c654 100644
@@ -784,6 +784,8 @@ typedef uint64_t FeatureWordArray[FEATURE_WORDS];
 #define CPUID_7_0_EDX_SERIALIZE         (1U << 14)
 /* TSX Suspend Load Address Tracking instruction */
 #define CPUID_7_0_EDX_TSX_LDTRK         (1U << 16)
+/* AVX512_FP16 instruction */
+#define CPUID_7_0_EDX_AVX512_FP16       (1U << 23)
 /* Speculation Control */
 #define CPUID_7_0_EDX_SPEC_CTRL         (1U << 26)
 /* Single Thread Indirect Branch Predictors */
Signature · 1 line
2.17.1
#1 lore raw

On 2020-12-16 at 23:40 CET, Cathy Zhang wrote...

Quoted 2 lines

AVX512 Half-precision floating point (FP16) has better performance compared to FP32 if the presicion or magnitude requirements are met.

spelling: precision

Quoted 30 lines

It's defined as CPUID.(EAX=7,ECX=0):EDX[bit 23].

Refer to
https://software.intel.com/content/www/us/en/develop/download/\
intel-architecture-instruction-set-extensions-programming-reference.html

---
 target/i386/cpu.c | 2 +-
 target/i386/cpu.h | 2 ++
 2 files changed, 3 insertions(+), 1 deletion(-)
target/i386/cpu.c +1 -1
index ca997a68cd..c4d623b8b9 100644
@@ -979,7 +979,7 @@ static FeatureWordInfo feature_word_info[FEATURE_WORDS] = {
             "avx512-vp2intersect", NULL, "md-clear", NULL,
             NULL, NULL, "serialize", NULL,
             "tsx-ldtrk", NULL, NULL /* pconfig */, NULL,
-            NULL, NULL, NULL, NULL,
+            NULL, NULL, NULL, "avx512-fp16",
             NULL, NULL, "spec-ctrl", "stibp",
             NULL, "arch-capabilities", "core-capability", "ssbd",
         },
target/i386/cpu.h +0 -0
index c4a49c06a8..6fd675c654 100644

target/i386/cpu.h typedef uint64_t FeatureWordArray[FEATURE_WORDS]; · line 791

Quoted 8 lines
target/i386/cpu.h +2 -0
@@ -784,6 +784,8 @@ typedef uint64_t FeatureWordArray[FEATURE_WORDS];
 #define CPUID_7_0_EDX_SERIALIZE         (1U << 14)
 /* TSX Suspend Load Address Tracking instruction */
 #define CPUID_7_0_EDX_TSX_LDTRK         (1U << 16)
+/* AVX512_FP16 instruction */
+#define CPUID_7_0_EDX_AVX512_FP16       (1U << 23)
 /* Speculation Control */
 #define CPUID_7_0_EDX_SPEC_CTRL         (1U << 26)
 /* Single Thread Indirect Branch Predictors */

--
Cheers,
Christophe de Dinechin (IRC c3d)

#2 lore raw
In reply to Cathy Zhang
Quoted 9 lines · Cathy Zhang

AVX512 Half-precision floating point (FP16) has better performance compared to FP32 if the presicion or magnitude requirements are met. It's defined as CPUID.(EAX=7,ECX=0):EDX[bit 23].

Refer to
https://software.intel.com/content/www/us/en/develop/download/\
intel-architecture-instruction-set-extensions-programming-reference.html

Queued, thanks!

Signature · 1 line
Eduardo