lib/x86: disambiguate 512-bit vector from AVX-512F

crc32_x86_vpclmulqdq_avx512vl and crc32_x86_vpclmulqdq_avx512f_avx512vl actually use the same CPU features, considering that vpternlog always requires at least avx512f, and compilers consider avx512vl to imply avx512f. Rename them to *_avx512_vl256 and *_avx512_vl512 to reflect that they differ only in vector length, and fix the CPU feature checking to use a separate flag for whether 512-bit vectors are enabled.
ebiggers · Feb 24, 2024 · ea028f1 · ea028f1
1 parent 2929aea
commit ea028f1
Show file tree

Hide file tree

Showing 6 changed files with 47 additions and 36 deletions.
diff --git a/lib/x86/cpu_features.c b/lib/x86/cpu_features.c
@@ -90,6 +90,7 @@ static const struct cpu_feature x86_cpu_feature_table[] = {
 	{X86_CPU_FEATURE_AVX,		"avx"},
 	{X86_CPU_FEATURE_AVX2,		"avx2"},
 	{X86_CPU_FEATURE_BMI2,		"bmi2"},
+	{X86_CPU_FEATURE_ZMM,		"zmm"},
 	{X86_CPU_FEATURE_AVX512F,	"avx512f"},
 	{X86_CPU_FEATURE_AVX512VL,	"avx512vl"},
 	{X86_CPU_FEATURE_VPCLMULQDQ,	"vpclmulqdq"},
@@ -165,8 +166,10 @@ void libdeflate_init_x86_cpu_features(void)
 		features |= X86_CPU_FEATURE_AVX2;
 	if (b & (1 << 8))
 		features |= X86_CPU_FEATURE_BMI2;
-	if ((b & (1 << 16)) && ((xcr0 & 0xe6) == 0xe6) &&
+	if (((xcr0 & 0xe6) == 0xe6) &&
 	    allow_512bit_vectors(manufacturer, family, model))
+		features |= X86_CPU_FEATURE_ZMM;
+	if ((b & (1 << 16)) && ((xcr0 & 0xe6) == 0xe6))
 		features |= X86_CPU_FEATURE_AVX512F;
 	if ((b & (1U << 31)) && ((xcr0 & 0xa6) == 0xa6))
 		features |= X86_CPU_FEATURE_AVX512VL;

diff --git a/lib/x86/cpu_features.h b/lib/x86/cpu_features.h
@@ -44,9 +44,16 @@
 #define X86_CPU_FEATURE_AVX		0x00000004
 #define X86_CPU_FEATURE_AVX2		0x00000008
 #define X86_CPU_FEATURE_BMI2		0x00000010
-#define X86_CPU_FEATURE_AVX512F		0x00000020
-#define X86_CPU_FEATURE_AVX512VL	0x00000040
-#define X86_CPU_FEATURE_VPCLMULQDQ	0x00000080
+/*
+ * ZMM indicates whether 512-bit vectors (zmm registers) should be used.  On
+ * some CPUs, to avoid downclocking issues we don't set ZMM even if the CPU
+ * supports it, i.e. even if AVX512F is set.  On these CPUs, we may still use
+ * AVX-512 instructions, but only with ymm and xmm registers.
+ */
+#define X86_CPU_FEATURE_ZMM		0x00000020
+#define X86_CPU_FEATURE_AVX512F		0x00000040
+#define X86_CPU_FEATURE_AVX512VL	0x00000080
+#define X86_CPU_FEATURE_VPCLMULQDQ	0x00000100
 
 #define HAVE_SSE2(features)	(HAVE_SSE2_NATIVE     || ((features) & X86_CPU_FEATURE_SSE2))
 #define HAVE_PCLMULQDQ(features) (HAVE_PCLMULQDQ_NATIVE || ((features) & X86_CPU_FEATURE_PCLMULQDQ))

diff --git a/lib/x86/crc32_impl.h b/lib/x86/crc32_impl.h
@@ -95,14 +95,16 @@
 #endif
 
 /*
- * VPCLMULQDQ/AVX512VL implementation.  This takes advantage of some AVX-512
- * instructions but uses 256-bit vectors rather than 512-bit.  This can be
- * useful on CPUs where 512-bit vectors cause downclocking.
+ * VPCLMULQDQ/AVX512 implementation with 256-bit vectors.  This takes advantage
+ * of some AVX-512 instructions but uses 256-bit vectors rather than 512-bit.
+ * This can be useful on CPUs where 512-bit vectors cause downclocking.
  */
-#if HAVE_VPCLMULQDQ_INTRIN && HAVE_PCLMULQDQ_INTRIN && HAVE_AVX512VL_INTRIN
-#  define crc32_x86_vpclmulqdq_avx512vl	crc32_x86_vpclmulqdq_avx512vl
-#  define SUFFIX				 _vpclmulqdq_avx512vl
-#  if HAVE_VPCLMULQDQ_NATIVE && HAVE_PCLMULQDQ_NATIVE && HAVE_AVX512VL_NATIVE
+#if HAVE_VPCLMULQDQ_INTRIN && HAVE_PCLMULQDQ_INTRIN && \
+	HAVE_AVX512F_INTRIN && HAVE_AVX512VL_INTRIN
+#  define crc32_x86_vpclmulqdq_avx512_vl256  crc32_x86_vpclmulqdq_avx512_vl256
+#  define SUFFIX				      _vpclmulqdq_avx512_vl256
+#  if HAVE_VPCLMULQDQ_NATIVE && HAVE_PCLMULQDQ_NATIVE && \
+	HAVE_AVX512F_NATIVE && HAVE_AVX512VL_NATIVE
 #    define ATTRIBUTES
 #  else
 #    define ATTRIBUTES		_target_attribute("vpclmulqdq,pclmul,avx512vl")
@@ -113,16 +115,16 @@
 #  include "crc32_pclmul_template.h"
 #endif
 
-/* VPCLMULQDQ/AVX512F/AVX512VL implementation.  Uses 512-bit vectors. */
+/* VPCLMULQDQ/AVX512 implementation with 512-bit vectors */
 #if HAVE_VPCLMULQDQ_INTRIN && HAVE_PCLMULQDQ_INTRIN && \
 	HAVE_AVX512F_INTRIN && HAVE_AVX512VL_INTRIN
-#  define crc32_x86_vpclmulqdq_avx512f_avx512vl	crc32_x86_vpclmulqdq_avx512f_avx512vl
-#  define SUFFIX					 _vpclmulqdq_avx512f_avx512vl
-#if HAVE_VPCLMULQDQ_NATIVE && HAVE_PCLMULQDQ_NATIVE && \
+#  define crc32_x86_vpclmulqdq_avx512_vl512  crc32_x86_vpclmulqdq_avx512_vl512
+#  define SUFFIX				      _vpclmulqdq_avx512_vl512
+#  if HAVE_VPCLMULQDQ_NATIVE && HAVE_PCLMULQDQ_NATIVE && \
 	HAVE_AVX512F_NATIVE && HAVE_AVX512VL_NATIVE
 #    define ATTRIBUTES
 #  else
-#    define ATTRIBUTES		_target_attribute("vpclmulqdq,pclmul,avx512f,avx512vl")
+#    define ATTRIBUTES		_target_attribute("vpclmulqdq,pclmul,avx512vl")
 #  endif
 #  define VL			64
 #  define FOLD_LESSTHAN16BYTES	1
@@ -136,15 +138,16 @@ arch_select_crc32_func(void)
 {
 	const u32 features MAYBE_UNUSED = get_x86_cpu_features();
 
-#ifdef crc32_x86_vpclmulqdq_avx512f_avx512vl
-	if (HAVE_VPCLMULQDQ(features) && HAVE_PCLMULQDQ(features) &&
+#ifdef crc32_x86_vpclmulqdq_avx512_vl512
+	if ((features & X86_CPU_FEATURE_ZMM) &&
+	    HAVE_VPCLMULQDQ(features) && HAVE_PCLMULQDQ(features) &&
 	    HAVE_AVX512F(features) && HAVE_AVX512VL(features))
-		return crc32_x86_vpclmulqdq_avx512f_avx512vl;
+		return crc32_x86_vpclmulqdq_avx512_vl512;
 #endif
-#ifdef crc32_x86_vpclmulqdq_avx512vl
+#ifdef crc32_x86_vpclmulqdq_avx512_vl256
 	if (HAVE_VPCLMULQDQ(features) && HAVE_PCLMULQDQ(features) &&
-	    HAVE_AVX512VL(features))
-		return crc32_x86_vpclmulqdq_avx512vl;
+	    HAVE_AVX512F(features) && HAVE_AVX512VL(features))
+		return crc32_x86_vpclmulqdq_avx512_vl256;
 #endif
 #ifdef crc32_x86_vpclmulqdq_avx2
 	if (HAVE_VPCLMULQDQ(features) && HAVE_PCLMULQDQ(features) &&

diff --git a/lib/x86/crc32_pclmul_template.h b/lib/x86/crc32_pclmul_template.h
@@ -38,7 +38,7 @@
  *	   VL=16 && FOLD_LESSTHAN16BYTES=1: at least pclmul,sse4.1
  *	   VL=32 && USE_TERNARYLOGIC=0: at least vpclmulqdq,pclmul,avx2
  *	   VL=32 && USE_TERNARYLOGIC=1: at least vpclmulqdq,pclmul,avx512vl
- *	   VL=64: at least vpclmulqdq,pclmul,avx512f,avx512vl
+ *	   VL=64: at least vpclmulqdq,pclmul,avx512vl
  * VL:
  *	Vector length in bytes.  Supported values are 16, 32, and 64.
  * FOLD_LESSTHAN16BYTES:

diff --git a/scripts/checksum_benchmarks.sh b/scripts/checksum_benchmarks.sh
@@ -67,12 +67,11 @@ sort_by_speed() {
 }
 
 disable_cpu_feature() {
-	local name="$1"
+	LIBDEFLATE_DISABLE_CPU_FEATURES+=",$1"
 	shift
-	local extra_cflags=("$@")
-
-	LIBDEFLATE_DISABLE_CPU_FEATURES+=",$name"
-	EXTRA_CFLAGS+=("${extra_cflags[@]}")
+	if (( $# > 0 )); then
+		EXTRA_CFLAGS+=("$@")
+	fi
 }
 
 cleanup() {
@@ -114,15 +113,14 @@ export LIBDEFLATE_DISABLE_CPU_FEATURES=""
 {
 case $ARCH in
 i386|x86_64)
-	if have_cpu_features vpclmulqdq avx512f avx512vl; then
-		do_benchmark "VPCLMULQDQ/AVX512F/AVX512VL"
-		disable_cpu_feature "avx512f" "-mno-avx512f"
-	fi
-	if have_cpu_features vpclmulqdq avx512vl; then
-		do_benchmark "VPCLMULQDQ/AVX512VL"
+	if have_cpu_features vpclmulqdq pclmulqdq avx512f avx512vl; then
+		do_benchmark "VPCLMULQDQ/AVX512/VL512"
+		disable_cpu_feature "zmm"
+		do_benchmark "VPCLMULQDQ/AVX512/VL256"
 		disable_cpu_feature "avx512vl" "-mno-avx512vl"
+		disable_cpu_feature "avx512f" "-mno-avx512f"
 	fi
-	if have_cpu_features vpclmulqdq avx2; then
+	if have_cpu_features vpclmulqdq pclmulqdq avx2; then
 		do_benchmark "VPCLMULQDQ/AVX2"
 		disable_cpu_feature "vpclmulqdq" "-mno-vpclmulqdq"
 	fi

diff --git a/scripts/run_tests.sh b/scripts/run_tests.sh
@@ -142,7 +142,7 @@ build_and_run_tests()
 	if ! [[ "$CFLAGS" =~ "-march=native" ]] && ! $quick; then
 		case "$ARCH" in
 		i386|x86_64)
-			features+=(vpclmulqdq avx512vl avx512f
+			features+=(zmm vpclmulqdq avx512vl avx512f
 				   avx2 avx bmi2 pclmulqdq sse2)
 			;;
 		arm*|aarch*)