@@ -20,9 +20,10 @@ SIMD := 3dnow sse sse2 sse4 avx avx2 xop
FMA := fma4 fma
SG := avx2-sg avx512f-sg avx512vl-sg
AES := ssse3-aes avx-aes avx2-vaes avx512bw-vaes
+CLMUL := ssse3-pclmul avx-pclmul avx2-vpclmulqdq avx512bw-vpclmulqdq avx512vbmi2-vpclmulqdq
SHA := sse4-sha avx-sha avx512f-sha
GF := sse2-gf avx2-gf avx512bw-gf
-TESTCASES := blowfish $(SIMD) $(FMA) $(SG) $(AES) $(SHA) $(GF)
+TESTCASES := blowfish $(SIMD) $(FMA) $(SG) $(AES) $(CLMUL) $(SHA) $(GF)
OPMASK := avx512f avx512dq avx512bw
@@ -89,6 +90,7 @@ avx512er-flts := 4 8
avx512vbmi-vecs := $(avx512bw-vecs)
avx512vbmi-ints := $(avx512bw-ints)
avx512vbmi-flts := $(avx512bw-flts)
+avx512vbmi2-vecs := $(avx512bw-vecs)
avx512f-opmask-vecs := 2
avx512dq-opmask-vecs := 1 2
@@ -149,6 +151,10 @@ define simd-aes-defs
$(1)-cflags := $(foreach vec,$($(patsubst %-aes,sse,$(1))-vecs) $($(patsubst %-vaes,%,$(1))-vecs), \
"-D_$(vec) -maes $(addprefix -m,$(subst -,$(space),$(1))) $(call non-sse,$(1)) -Os -DVEC_SIZE=$(vec)")
endef
+define simd-clmul-defs
+$(1)-cflags := $(foreach vec,$($(patsubst %-pclmul,sse,$(1))-vecs) $($(patsubst %-vpclmulqdq,%,$(1))-vecs), \
+ "-D_$(vec) -mpclmul $(addprefix -m,$(subst -,$(space),$(1))) $(call non-sse,$(1)) -Os -DVEC_SIZE=$(vec)")
+endef
define simd-sha-defs
$(1)-cflags := $(foreach vec,$(sse-vecs), \
"-D_$(vec) $(addprefix -m,$(subst -,$(space),$(1))) -Os -DVEC_SIZE=$(vec)")
@@ -164,6 +170,7 @@ endef
$(foreach flavor,$(SIMD) $(FMA),$(eval $(call simd-defs,$(flavor))))
$(foreach flavor,$(SG),$(eval $(call simd-sg-defs,$(flavor))))
$(foreach flavor,$(AES),$(eval $(call simd-aes-defs,$(flavor))))
+$(foreach flavor,$(CLMUL),$(eval $(call simd-clmul-defs,$(flavor))))
$(foreach flavor,$(SHA),$(eval $(call simd-sha-defs,$(flavor))))
$(foreach flavor,$(GF),$(eval $(call simd-gf-defs,$(flavor))))
$(foreach flavor,$(OPMASK),$(eval $(call opmask-defs,$(flavor))))
@@ -218,13 +225,16 @@ $(addsuffix .c,$(SG)):
$(addsuffix .c,$(AES)):
ln -sf simd-aes.c $@
+$(addsuffix .c,$(CLMUL)):
+ ln -sf simd-clmul.c $@
+
$(addsuffix .c,$(SHA)):
ln -sf simd-sha.c $@
$(addsuffix .c,$(GF)):
ln -sf simd-gf.c $@
-$(addsuffix .h,$(SIMD) $(FMA) $(SG) $(AES) $(SHA) $(GF)): simd.h
+$(addsuffix .h,$(SIMD) $(FMA) $(SG) $(AES) $(CLMUL) $(SHA) $(GF)): simd.h
xop.h avx512f.h: simd-fma.c
@@ -0,0 +1,150 @@
+#define UINT_SIZE 8
+
+#include "simd.h"
+ENTRY(clmul_test);
+
+#ifdef __AVX512F__ /* AVX512BW may get enabled only below */
+# define ALL_TRUE (~0ULL >> (64 - ELEM_COUNT))
+# define eq(x, y) (B(pcmpeqq, _mask, (vdi_t)(x), (vdi_t)(y), -1) == ALL_TRUE)
+# define lane_shr_unit(x) \
+ ((vec_t)B(palignr, _mask, (vdi_t)(x), (vdi_t)(x), 64, (vdi_t){}, \
+ 0x00ff00ff00ff00ffULL & (~0ULL >> (64 - VEC_SIZE))))
+#else
+# if defined(__AVX2__) && VEC_SIZE == 32
+# define to_bool(cmp) B(ptestc, , cmp, (vdi_t){} == 0)
+# else
+# define to_bool(cmp) (__builtin_ia32_pmovmskb128(cmp) == 0xffff)
+# endif
+# define eq(x, y) to_bool((x) == (y))
+# define lane_shr_unit(x) ((vec_t)B(palignr, , (vdi_t){}, (vdi_t)(x), 64))
+#endif
+
+#define CLMUL(op, x, y, c) (vec_t)(__builtin_ia32_ ## op((vdi_t)(x), (vdi_t)(y), c))
+
+#if VEC_SIZE == 16
+# define clmul(x, y, c) CLMUL(pclmulqdq128, x, y, c)
+# define vpshrd __builtin_ia32_vpshrd_v2di
+#elif VEC_SIZE == 32
+# define clmul(x, y, c) CLMUL(vpclmulqdq_v4di, x, y, c)
+# define vpshrd __builtin_ia32_vpshrd_v4di
+#elif VEC_SIZE == 64
+# define clmul(x, y, c) CLMUL(vpclmulqdq_v8di, x, y, c)
+# define vpshrd __builtin_ia32_vpshrd_v8di
+#endif
+
+#define clmul_ll(x, y) clmul(x, y, 0x00)
+#define clmul_hl(x, y) clmul(x, y, 0x01)
+#define clmul_lh(x, y) clmul(x, y, 0x10)
+#define clmul_hh(x, y) clmul(x, y, 0x11)
+
+#if defined(__AVX512VBMI2__)
+# pragma GCC target ( "avx512bw" )
+# define lane_shr_i(x, n) ({ \
+ vec_t h_ = lane_shr_unit(x); \
+ touch(h_); \
+ (n) < 64 ? (vec_t)vpshrd((vdi_t)(x), (vdi_t)(h_), n) : h_ >> ((n) - 64); \
+})
+# define lane_shr_v(x, n) ({ \
+ vec_t t_ = (x), h_ = lane_shr_unit(x); \
+ typeof(t_[0]) n_ = (n); \
+ if ( (n) < 64 ) \
+ /* gcc does not support embedded broadcast */ \
+ asm ( "vpshrdvq %2%{1to%c3%}, %1, %0" \
+ : "+v" (t_) : "v" (h_), "m" (n_), "i" (ELEM_COUNT) ); \
+ else \
+ t_ = h_ >> ((n) - 64); \
+ t_; \
+})
+#else
+# define lane_shr_i lane_shr_v
+# define lane_shr_v(x, n) ({ \
+ vec_t t_ = (n) > 0 ? lane_shr_unit(x) : (x); \
+ (n) < 64 ? ((x) >> (n)) | (t_ << (-(n) & 0x3f)) \
+ : t_ >> ((n) - 64); \
+})
+#endif
+
+int clmul_test(void)
+{
+ unsigned int i;
+ vec_t src;
+ vqi_t raw = {};
+
+ for ( i = 1; i < VEC_SIZE; ++i )
+ raw[i] = i;
+ src = (vec_t)raw;
+
+ for ( i = 0; i < 256; i += VEC_SIZE )
+ {
+ vec_t x = {}, y, z, lo, hi;
+ unsigned int j;
+
+ touch(x);
+ y = clmul_ll(src, x);
+ touch(x);
+
+ if ( !eq(y, x) ) return __LINE__;
+
+ for ( j = 0; j < ELEM_COUNT; j += 2 )
+ x[j] = 1;
+
+ touch(src);
+ y = clmul_ll(x, src);
+ touch(src);
+ z = clmul_lh(x, src);
+ touch(src);
+
+ for ( j = 0; j < ELEM_COUNT; j += 2 )
+ y[j + 1] = z[j];
+
+ if ( !eq(y, src) ) return __LINE__;
+
+ /*
+ * Besides the obvious property of the low and high half products
+ * being the same either direction, the "square" of a number has the
+ * property of simply being the original bit pattern with a zero bit
+ * inserted between any two bits. This is what the code below checks.
+ */
+
+ x = src;
+ touch(src);
+ y = clmul_lh(x, src);
+ touch(src);
+ z = clmul_hl(x, src);
+
+ if ( !eq(y, z) ) return __LINE__;
+
+ touch(src);
+ y = lo = clmul_ll(x, src);
+ touch(src);
+ z = hi = clmul_hh(x, src);
+ touch(src);
+
+ for ( j = 0; j < 64; ++j )
+ {
+ vec_t l = lane_shr_v(lo, 2 * j);
+ vec_t h = lane_shr_v(hi, 2 * j);
+ unsigned int n;
+
+ if ( !eq(l, y) ) return __LINE__;
+ if ( !eq(h, z) ) return __LINE__;
+
+ x = src >> j;
+
+ for ( n = 0; n < ELEM_COUNT; n += 2 )
+ {
+ if ( (x[n + 0] & 1) != (l[n] & 3) ) return __LINE__;
+ if ( (x[n + 1] & 1) != (h[n] & 3) ) return __LINE__;
+ }
+
+ touch(y);
+ y = lane_shr_i(y, 2);
+ touch(z);
+ z = lane_shr_i(z, 2);
+ }
+
+ src += 0x0101010101010101ULL * VEC_SIZE;
+ }
+
+ return 0;
+}
@@ -381,6 +381,7 @@ OVR(movntdq);
OVR(movntdqa);
OVR(movshdup);
OVR(movsldup);
+OVR(pclmulqdq);
OVR(permd);
OVR(permq);
OVR(pmovsxbd);
@@ -13,16 +13,19 @@ asm ( ".pushsection .test, \"ax\", @prog
#include "sse2.h"
#include "sse2-gf.h"
#include "ssse3-aes.h"
+#include "ssse3-pclmul.h"
#include "sse4.h"
#include "sse4-sha.h"
#include "avx.h"
#include "avx-aes.h"
+#include "avx-pclmul.h"
#include "avx-sha.h"
#include "fma4.h"
#include "fma.h"
#include "avx2.h"
#include "avx2-sg.h"
#include "avx2-vaes.h"
+#include "avx2-vpclmulqdq.h"
#include "avx2-gf.h"
#include "xop.h"
#include "avx512f-opmask.h"
@@ -34,10 +37,12 @@ asm ( ".pushsection .test, \"ax\", @prog
#include "avx512vl-sg.h"
#include "avx512bw.h"
#include "avx512bw-vaes.h"
+#include "avx512bw-vpclmulqdq.h"
#include "avx512bw-gf.h"
#include "avx512dq.h"
#include "avx512er.h"
#include "avx512vbmi.h"
+#include "avx512vbmi2-vpclmulqdq.h"
#define verbose false /* Switch to true for far more logging. */
@@ -108,6 +113,16 @@ static bool simd_check_avx_aes(void)
return cpu_has_aesni && cpu_has_avx;
}
+static bool simd_check_ssse3_pclmul(void)
+{
+ return cpu_has_pclmulqdq && cpu_has_ssse3;
+}
+
+static bool simd_check_avx_pclmul(void)
+{
+ return cpu_has_pclmulqdq && cpu_has_avx;
+}
+
static bool simd_check_avx512f(void)
{
return cpu_has_avx512f;
@@ -189,6 +204,31 @@ static bool simd_check_avx512bw_vaes_vl(
cpu_has_avx512bw && cpu_has_avx512vl;
}
+static bool simd_check_avx2_vpclmulqdq(void)
+{
+ return cpu_has_vpclmulqdq && cpu_has_avx2;
+}
+
+static bool simd_check_avx512bw_vpclmulqdq(void)
+{
+ return cpu_has_vpclmulqdq && cpu_has_avx512bw;
+}
+
+static bool simd_check_avx512bw_vpclmulqdq_vl(void)
+{
+ return cpu_has_vpclmulqdq && cpu_has_avx512bw && cpu_has_avx512vl;
+}
+
+static bool simd_check_avx512vbmi2_vpclmulqdq(void)
+{
+ return cpu_has_avx512_vbmi2 && simd_check_avx512bw_vpclmulqdq();
+}
+
+static bool simd_check_avx512vbmi2_vpclmulqdq_vl(void)
+{
+ return cpu_has_avx512_vbmi2 && simd_check_avx512bw_vpclmulqdq_vl();
+}
+
static bool simd_check_sse2_gf(void)
{
return cpu_has_gfni && cpu_has_sse2;
@@ -369,6 +409,8 @@ static const struct {
SIMD(XOP i64x4, xop, 32i8),
SIMD(AES (legacy), ssse3_aes, 16),
SIMD(AES (VEX/x16), avx_aes, 16),
+ SIMD(PCLMUL (legacy), ssse3_pclmul, 16),
+ SIMD(PCLMUL (VEX/x2), avx_pclmul, 16),
SIMD(OPMASK/w, avx512f_opmask, 2),
SIMD(OPMASK+DQ/b, avx512dq_opmask, 1),
SIMD(OPMASK+DQ/w, avx512dq_opmask, 2),
@@ -475,6 +517,13 @@ static const struct {
SIMD(VAES (EVEX/x64), avx512bw_vaes, 64),
AVX512VL(VL+VAES (x16), avx512bw_vaes, 16),
AVX512VL(VL+VAES (x32), avx512bw_vaes, 32),
+ SIMD(VPCLMUL (VEX/x4), avx2_vpclmulqdq, 32),
+ SIMD(VPCLMUL (EVEX/x8), avx512bw_vpclmulqdq, 64),
+ AVX512VL(VL+VPCLMUL (x4), avx512bw_vpclmulqdq, 16),
+ AVX512VL(VL+VPCLMUL (x8), avx512bw_vpclmulqdq, 32),
+ SIMD(AVX512_VBMI2+VPCLMUL (x8), avx512vbmi2_vpclmulqdq, 64),
+ AVX512VL(_VBMI2+VL+VPCLMUL (x2), avx512vbmi2_vpclmulqdq, 16),
+ AVX512VL(_VBMI2+VL+VPCLMUL (x4), avx512vbmi2_vpclmulqdq, 32),
SIMD(GFNI (legacy), sse2_gf, 16),
SIMD(GFNI (VEX/x16), avx2_gf, 16),
SIMD(GFNI (VEX/x32), avx2_gf, 32),
@@ -125,6 +125,7 @@ static inline bool xcr0_mask(uint64_t ma
#define cpu_has_sse cp.basic.sse
#define cpu_has_sse2 cp.basic.sse2
#define cpu_has_sse3 cp.basic.sse3
+#define cpu_has_pclmulqdq cp.basic.pclmulqdq
#define cpu_has_ssse3 cp.basic.ssse3
#define cpu_has_fma (cp.basic.fma && xcr0_mask(6))
#define cpu_has_sse4_1 cp.basic.sse4_1
Also use this for AVX512_VBMI2 VPSH{L,R}D{,V}{D,Q,W} testing (only the quad word right shifts get actually used; the assumption is that their "left" counterparts as well as the double word and word forms then work as well). Signed-off-by: Jan Beulich <jbeulich@suse.com> --- v8: New.