Message-Id: <20180914162240.7925-14-Jason@zx2c4.com>
Date:   Fri, 14 Sep 2018 18:22:33 +0200
From:   "Jason A. Donenfeld" <Jason@...c4.com>
To:     linux-kernel@...r.kernel.org, netdev@...r.kernel.org,
        linux-crypto@...r.kernel.org, davem@...emloft.net,
        gregkh@...uxfoundation.org
Cc:     "Jason A. Donenfeld" <Jason@...c4.com>,
        Samuel Neves <sneves@....uc.pt>,
        Andy Lutomirski <luto@...nel.org>,
        Jean-Philippe Aumasson <jeanphilippe.aumasson@...il.com>,
        Thomas Gleixner <tglx@...utronix.de>,
        Ingo Molnar <mingo@...hat.com>, x86@...nel.org
Subject: [PATCH net-next v4 13/20] zinc: BLAKE2s x86_64 implementation

These implementations from Samuel Neves support AVX and AVX-512VL.
Originally this used AVX-512F, but Skylake thermal throttling made
AVX-512VL more attractive and possible to do with negligable difference.

Signed-off-by: Jason A. Donenfeld <Jason@...c4.com>
Signed-off-by: Samuel Neves <sneves@....uc.pt>
Cc: Andy Lutomirski <luto@...nel.org>
Cc: Greg KH <gregkh@...uxfoundation.org>
Cc: Jean-Philippe Aumasson <jeanphilippe.aumasson@...il.com>
Cc: Thomas Gleixner <tglx@...utronix.de>
Cc: Ingo Molnar <mingo@...hat.com>
Cc: x86@...nel.org
---
 lib/zinc/Makefile                      |   4 +
 lib/zinc/blake2s/blake2s-x86_64-glue.h |  62 +++
 lib/zinc/blake2s/blake2s-x86_64.S      | 685 +++++++++++++++++++++++++
 3 files changed, 751 insertions(+)
 create mode 100644 lib/zinc/blake2s/blake2s-x86_64-glue.h
 create mode 100644 lib/zinc/blake2s/blake2s-x86_64.S

diff --git a/lib/zinc/Makefile b/lib/zinc/Makefile
index 45817ec5539c..49455abfdf5e 100644
--- a/lib/zinc/Makefile
+++ b/lib/zinc/Makefile
@@ -53,6 +53,10 @@ endif
 
 ifeq ($(CONFIG_ZINC_BLAKE2S),y)
 zinc-y += blake2s/blake2s.o
+ifeq ($(CONFIG_ZINC_ARCH_X86_64),y)
+zinc-y += blake2s/blake2s-x86_64.o
+CFLAGS_blake2s.o += -include $(srctree)/$(src)/blake2s/blake2s-x86_64-glue.h
+endif
 endif
 
 zinc-y += main.o
diff --git a/lib/zinc/blake2s/blake2s-x86_64-glue.h b/lib/zinc/blake2s/blake2s-x86_64-glue.h
new file mode 100644
index 000000000000..2f5b74bd9117
--- /dev/null
+++ b/lib/zinc/blake2s/blake2s-x86_64-glue.h
@@ -0,0 +1,62 @@
+/* SPDX-License-Identifier: GPL-2.0
+ *
+ * Copyright (C) 2015-2018 Jason A. Donenfeld <Jason@...c4.com>. All Rights Reserved.
+ */
+
+#include <zinc/blake2s.h>
+#include <asm/cpufeature.h>
+#include <asm/processor.h>
+#include <asm/fpu/api.h>
+#include <asm/simd.h>
+
+#ifdef CONFIG_AS_AVX
+asmlinkage void blake2s_compress_avx(struct blake2s_state *state,
+				     const u8 *block, const size_t nblocks,
+				     const u32 inc);
+#endif
+#ifdef CONFIG_AS_AVX512
+asmlinkage void blake2s_compress_avx512(struct blake2s_state *state,
+					const u8 *block, const size_t nblocks,
+					const u32 inc);
+#endif
+
+static bool blake2s_use_avx __ro_after_init;
+static bool blake2s_use_avx512 __ro_after_init;
+
+void __init blake2s_fpu_init(void)
+{
+	blake2s_use_avx =
+		boot_cpu_has(X86_FEATURE_AVX) &&
+		cpu_has_xfeatures(XFEATURE_MASK_SSE | XFEATURE_MASK_YMM, NULL);
+	blake2s_use_avx512 =
+		boot_cpu_has(X86_FEATURE_AVX) &&
+		boot_cpu_has(X86_FEATURE_AVX2) &&
+		boot_cpu_has(X86_FEATURE_AVX512F) &&
+		boot_cpu_has(X86_FEATURE_AVX512VL) &&
+		cpu_has_xfeatures(XFEATURE_MASK_SSE | XFEATURE_MASK_YMM |
+				  XFEATURE_MASK_AVX512, NULL);
+}
+
+static inline bool blake2s_arch(struct blake2s_state *state, const u8 *block,
+				size_t nblocks, const u32 inc)
+{
+#ifdef CONFIG_AS_AVX512
+	if (blake2s_use_avx512 && irq_fpu_usable()) {
+		kernel_fpu_begin();
+		blake2s_compress_avx512(state, block, nblocks, inc);
+		kernel_fpu_end();
+		return true;
+	}
+#endif
+#ifdef CONFIG_AS_AVX
+	if (blake2s_use_avx && irq_fpu_usable()) {
+		kernel_fpu_begin();
+		blake2s_compress_avx(state, block, nblocks, inc);
+		kernel_fpu_end();
+		return true;
+	}
+#endif
+	return false;
+}
+
+#define HAVE_BLAKE2S_ARCH_IMPLEMENTATION
diff --git a/lib/zinc/blake2s/blake2s-x86_64.S b/lib/zinc/blake2s/blake2s-x86_64.S
new file mode 100644
index 000000000000..6a84b9f1f2c4
--- /dev/null
+++ b/lib/zinc/blake2s/blake2s-x86_64.S
@@ -0,0 +1,685 @@
+/* SPDX-License-Identifier: GPL-2.0
+ *
+ * Copyright (C) 2015-2018 Jason A. Donenfeld <Jason@...c4.com>. All Rights Reserved.
+ * Copyright (C) 2017 Samuel Neves <sneves@....uc.pt>. All Rights Reserved.
+ */
+
+#include <linux/linkage.h>
+
+.section .rodata.cst32.BLAKE2S_IV, "aM", @progbits, 32
+.align 32
+IV:	.octa 0xA54FF53A3C6EF372BB67AE856A09E667
+	.octa 0x5BE0CD191F83D9AB9B05688C510E527F
+.section .rodata.cst16.ROT16, "aM", @progbits, 16
+.align 16
+ROT16:	.octa 0x0D0C0F0E09080B0A0504070601000302
+.section .rodata.cst16.ROR328, "aM", @progbits, 16
+.align 16
+ROR328:	.octa 0x0C0F0E0D080B0A090407060500030201
+#ifdef CONFIG_AS_AVX512
+.section .rodata.cst64.BLAKE2S_SIGMA, "aM", @progbits, 640
+.align 64
+SIGMA:
+.long 0, 2, 4, 6, 1, 3, 5, 7, 8, 10, 12, 14, 9, 11, 13, 15
+.long 11, 2, 12, 14, 9, 8, 15, 3, 4, 0, 13, 6, 10, 1, 7, 5
+.long 10, 12, 11, 6, 5, 9, 13, 3, 4, 15, 14, 2, 0, 7, 8, 1
+.long 10, 9, 7, 0, 11, 14, 1, 12, 6, 2, 15, 3, 13, 8, 5, 4
+.long 4, 9, 8, 13, 14, 0, 10, 11, 7, 3, 12, 1, 5, 6, 15, 2
+.long 2, 10, 4, 14, 13, 3, 9, 11, 6, 5, 7, 12, 15, 1, 8, 0
+.long 4, 11, 14, 8, 13, 10, 12, 5, 2, 1, 15, 3, 9, 7, 0, 6
+.long 6, 12, 0, 13, 15, 2, 1, 10, 4, 5, 11, 14, 8, 3, 9, 7
+.long 14, 5, 4, 12, 9, 7, 3, 10, 2, 0, 6, 15, 11, 1, 13, 8
+.long 11, 7, 13, 10, 12, 14, 0, 15, 4, 5, 6, 9, 2, 1, 8, 3
+#endif /* CONFIG_AS_AVX512 */
+
+.text
+#ifdef CONFIG_AS_AVX
+ENTRY(blake2s_compress_avx)
+	movl		%ecx, %ecx
+	testq		%rdx, %rdx
+	je		.Lendofloop
+	.align 32
+.Lbeginofloop:
+	addq		%rcx, 32(%rdi)
+	vmovdqu		IV+16(%rip), %xmm1
+	vmovdqu		(%rsi), %xmm4
+	vpxor		32(%rdi), %xmm1, %xmm1
+	vmovdqu		16(%rsi), %xmm3
+	vshufps		$136, %xmm3, %xmm4, %xmm6
+	vmovdqa		ROT16(%rip), %xmm7
+	vpaddd		(%rdi), %xmm6, %xmm6
+	vpaddd		16(%rdi), %xmm6, %xmm6
+	vpxor		%xmm6, %xmm1, %xmm1
+	vmovdqu		IV(%rip), %xmm8
+	vpshufb		%xmm7, %xmm1, %xmm1
+	vmovdqu		48(%rsi), %xmm5
+	vpaddd		%xmm1, %xmm8, %xmm8
+	vpxor		16(%rdi), %xmm8, %xmm9
+	vmovdqu		32(%rsi), %xmm2
+	vpblendw	$12, %xmm3, %xmm5, %xmm13
+	vshufps		$221, %xmm5, %xmm2, %xmm12
+	vpunpckhqdq	%xmm2, %xmm4, %xmm14
+	vpslld		$20, %xmm9, %xmm0
+	vpsrld		$12, %xmm9, %xmm9
+	vpxor		%xmm0, %xmm9, %xmm0
+	vshufps		$221, %xmm3, %xmm4, %xmm9
+	vpaddd		%xmm9, %xmm6, %xmm9
+	vpaddd		%xmm0, %xmm9, %xmm9
+	vpxor		%xmm9, %xmm1, %xmm1
+	vmovdqa		ROR328(%rip), %xmm6
+	vpshufb		%xmm6, %xmm1, %xmm1
+	vpaddd		%xmm1, %xmm8, %xmm8
+	vpxor		%xmm8, %xmm0, %xmm0
+	vpshufd		$147, %xmm1, %xmm1
+	vpshufd		$78, %xmm8, %xmm8
+	vpslld		$25, %xmm0, %xmm10
+	vpsrld		$7, %xmm0, %xmm0
+	vpxor		%xmm10, %xmm0, %xmm0
+	vshufps		$136, %xmm5, %xmm2, %xmm10
+	vpshufd		$57, %xmm0, %xmm0
+	vpaddd		%xmm10, %xmm9, %xmm9
+	vpaddd		%xmm0, %xmm9, %xmm9
+	vpxor		%xmm9, %xmm1, %xmm1
+	vpaddd		%xmm12, %xmm9, %xmm9
+	vpblendw	$12, %xmm2, %xmm3, %xmm12
+	vpshufb		%xmm7, %xmm1, %xmm1
+	vpaddd		%xmm1, %xmm8, %xmm8
+	vpxor		%xmm8, %xmm0, %xmm10
+	vpslld		$20, %xmm10, %xmm0
+	vpsrld		$12, %xmm10, %xmm10
+	vpxor		%xmm0, %xmm10, %xmm0
+	vpaddd		%xmm0, %xmm9, %xmm9
+	vpxor		%xmm9, %xmm1, %xmm1
+	vpshufb		%xmm6, %xmm1, %xmm1
+	vpaddd		%xmm1, %xmm8, %xmm8
+	vpxor		%xmm8, %xmm0, %xmm0
+	vpshufd		$57, %xmm1, %xmm1
+	vpshufd		$78, %xmm8, %xmm8
+	vpslld		$25, %xmm0, %xmm10
+	vpsrld		$7, %xmm0, %xmm0
+	vpxor		%xmm10, %xmm0, %xmm0
+	vpslldq		$4, %xmm5, %xmm10
+	vpblendw	$240, %xmm10, %xmm12, %xmm12
+	vpshufd		$147, %xmm0, %xmm0
+	vpshufd		$147, %xmm12, %xmm12
+	vpaddd		%xmm9, %xmm12, %xmm12
+	vpaddd		%xmm0, %xmm12, %xmm12
+	vpxor		%xmm12, %xmm1, %xmm1
+	vpshufb		%xmm7, %xmm1, %xmm1
+	vpaddd		%xmm1, %xmm8, %xmm8
+	vpxor		%xmm8, %xmm0, %xmm11
+	vpslld		$20, %xmm11, %xmm9
+	vpsrld		$12, %xmm11, %xmm11
+	vpxor		%xmm9, %xmm11, %xmm0
+	vpshufd		$8, %xmm2, %xmm9
+	vpblendw	$192, %xmm5, %xmm3, %xmm11
+	vpblendw	$240, %xmm11, %xmm9, %xmm9
+	vpshufd		$177, %xmm9, %xmm9
+	vpaddd		%xmm12, %xmm9, %xmm9
+	vpaddd		%xmm0, %xmm9, %xmm11
+	vpxor		%xmm11, %xmm1, %xmm1
+	vpshufb		%xmm6, %xmm1, %xmm1
+	vpaddd		%xmm1, %xmm8, %xmm8
+	vpxor		%xmm8, %xmm0, %xmm9
+	vpshufd		$147, %xmm1, %xmm1
+	vpshufd		$78, %xmm8, %xmm8
+	vpslld		$25, %xmm9, %xmm0
+	vpsrld		$7, %xmm9, %xmm9
+	vpxor		%xmm0, %xmm9, %xmm0
+	vpslldq		$4, %xmm3, %xmm9
+	vpblendw	$48, %xmm9, %xmm2, %xmm9
+	vpblendw	$240, %xmm9, %xmm4, %xmm9
+	vpshufd		$57, %xmm0, %xmm0
+	vpshufd		$177, %xmm9, %xmm9
+	vpaddd		%xmm11, %xmm9, %xmm9
+	vpaddd		%xmm0, %xmm9, %xmm9
+	vpxor		%xmm9, %xmm1, %xmm1
+	vpshufb		%xmm7, %xmm1, %xmm1
+	vpaddd		%xmm1, %xmm8, %xmm11
+	vpxor		%xmm11, %xmm0, %xmm0
+	vpslld		$20, %xmm0, %xmm8
+	vpsrld		$12, %xmm0, %xmm0
+	vpxor		%xmm8, %xmm0, %xmm0
+	vpunpckhdq	%xmm3, %xmm4, %xmm8
+	vpblendw	$12, %xmm10, %xmm8, %xmm12
+	vpshufd		$177, %xmm12, %xmm12
+	vpaddd		%xmm9, %xmm12, %xmm9
+	vpaddd		%xmm0, %xmm9, %xmm9
+	vpxor		%xmm9, %xmm1, %xmm1
+	vpshufb		%xmm6, %xmm1, %xmm1
+	vpaddd		%xmm1, %xmm11, %xmm11
+	vpxor		%xmm11, %xmm0, %xmm0
+	vpshufd		$57, %xmm1, %xmm1
+	vpshufd		$78, %xmm11, %xmm11
+	vpslld		$25, %xmm0, %xmm12
+	vpsrld		$7, %xmm0, %xmm0
+	vpxor		%xmm12, %xmm0, %xmm0
+	vpunpckhdq	%xmm5, %xmm2, %xmm12
+	vpshufd		$147, %xmm0, %xmm0
+	vpblendw	$15, %xmm13, %xmm12, %xmm12
+	vpslldq		$8, %xmm5, %xmm13
+	vpshufd		$210, %xmm12, %xmm12
+	vpaddd		%xmm9, %xmm12, %xmm9
+	vpaddd		%xmm0, %xmm9, %xmm9
+	vpxor		%xmm9, %xmm1, %xmm1
+	vpshufb		%xmm7, %xmm1, %xmm1
+	vpaddd		%xmm1, %xmm11, %xmm11
+	vpxor		%xmm11, %xmm0, %xmm0
+	vpslld		$20, %xmm0, %xmm12
+	vpsrld		$12, %xmm0, %xmm0
+	vpxor		%xmm12, %xmm0, %xmm0
+	vpunpckldq	%xmm4, %xmm2, %xmm12
+	vpblendw	$240, %xmm4, %xmm12, %xmm12
+	vpblendw	$192, %xmm13, %xmm12, %xmm12
+	vpsrldq		$12, %xmm3, %xmm13
+	vpaddd		%xmm12, %xmm9, %xmm9
+	vpaddd		%xmm0, %xmm9, %xmm9
+	vpxor		%xmm9, %xmm1, %xmm1
+	vpshufb		%xmm6, %xmm1, %xmm1
+	vpaddd		%xmm1, %xmm11, %xmm11
+	vpxor		%xmm11, %xmm0, %xmm0
+	vpshufd		$147, %xmm1, %xmm1
+	vpshufd		$78, %xmm11, %xmm11
+	vpslld		$25, %xmm0, %xmm12
+	vpsrld		$7, %xmm0, %xmm0
+	vpxor		%xmm12, %xmm0, %xmm0
+	vpblendw	$60, %xmm2, %xmm4, %xmm12
+	vpblendw	$3, %xmm13, %xmm12, %xmm12
+	vpshufd		$57, %xmm0, %xmm0
+	vpshufd		$78, %xmm12, %xmm12
+	vpaddd		%xmm9, %xmm12, %xmm9
+	vpaddd		%xmm0, %xmm9, %xmm9
+	vpxor		%xmm9, %xmm1, %xmm1
+	vpshufb		%xmm7, %xmm1, %xmm1
+	vpaddd		%xmm1, %xmm11, %xmm11
+	vpxor		%xmm11, %xmm0, %xmm12
+	vpslld		$20, %xmm12, %xmm13
+	vpsrld		$12, %xmm12, %xmm0
+	vpblendw	$51, %xmm3, %xmm4, %xmm12
+	vpxor		%xmm13, %xmm0, %xmm0
+	vpblendw	$192, %xmm10, %xmm12, %xmm10
+	vpslldq		$8, %xmm2, %xmm12
+	vpshufd		$27, %xmm10, %xmm10
+	vpaddd		%xmm9, %xmm10, %xmm9
+	vpaddd		%xmm0, %xmm9, %xmm9
+	vpxor		%xmm9, %xmm1, %xmm1
+	vpshufb		%xmm6, %xmm1, %xmm1
+	vpaddd		%xmm1, %xmm11, %xmm11
+	vpxor		%xmm11, %xmm0, %xmm0
+	vpshufd		$57, %xmm1, %xmm1
+	vpshufd		$78, %xmm11, %xmm11
+	vpslld		$25, %xmm0, %xmm10
+	vpsrld		$7, %xmm0, %xmm0
+	vpxor		%xmm10, %xmm0, %xmm0
+	vpunpckhdq	%xmm2, %xmm8, %xmm10
+	vpshufd		$147, %xmm0, %xmm0
+	vpblendw	$12, %xmm5, %xmm10, %xmm10
+	vpshufd		$210, %xmm10, %xmm10
+	vpaddd		%xmm9, %xmm10, %xmm9
+	vpaddd		%xmm0, %xmm9, %xmm9
+	vpxor		%xmm9, %xmm1, %xmm1
+	vpshufb		%xmm7, %xmm1, %xmm1
+	vpaddd		%xmm1, %xmm11, %xmm11
+	vpxor		%xmm11, %xmm0, %xmm10
+	vpslld		$20, %xmm10, %xmm0
+	vpsrld		$12, %xmm10, %xmm10
+	vpxor		%xmm0, %xmm10, %xmm0
+	vpblendw	$12, %xmm4, %xmm5, %xmm10
+	vpblendw	$192, %xmm12, %xmm10, %xmm10
+	vpunpckldq	%xmm2, %xmm4, %xmm12
+	vpshufd		$135, %xmm10, %xmm10
+	vpaddd		%xmm9, %xmm10, %xmm9
+	vpaddd		%xmm0, %xmm9, %xmm9
+	vpxor		%xmm9, %xmm1, %xmm1
+	vpshufb		%xmm6, %xmm1, %xmm1
+	vpaddd		%xmm1, %xmm11, %xmm13
+	vpxor		%xmm13, %xmm0, %xmm0
+	vpshufd		$147, %xmm1, %xmm1
+	vpshufd		$78, %xmm13, %xmm13
+	vpslld		$25, %xmm0, %xmm10
+	vpsrld		$7, %xmm0, %xmm0
+	vpxor		%xmm10, %xmm0, %xmm0
+	vpblendw	$15, %xmm3, %xmm4, %xmm10
+	vpblendw	$192, %xmm5, %xmm10, %xmm10
+	vpshufd		$57, %xmm0, %xmm0
+	vpshufd		$198, %xmm10, %xmm10
+	vpaddd		%xmm9, %xmm10, %xmm10
+	vpaddd		%xmm0, %xmm10, %xmm10
+	vpxor		%xmm10, %xmm1, %xmm1
+	vpshufb		%xmm7, %xmm1, %xmm1
+	vpaddd		%xmm1, %xmm13, %xmm13
+	vpxor		%xmm13, %xmm0, %xmm9
+	vpslld		$20, %xmm9, %xmm0
+	vpsrld		$12, %xmm9, %xmm9
+	vpxor		%xmm0, %xmm9, %xmm0
+	vpunpckhdq	%xmm2, %xmm3, %xmm9
+	vpunpcklqdq	%xmm12, %xmm9, %xmm15
+	vpunpcklqdq	%xmm12, %xmm8, %xmm12
+	vpblendw	$15, %xmm5, %xmm8, %xmm8
+	vpaddd		%xmm15, %xmm10, %xmm15
+	vpaddd		%xmm0, %xmm15, %xmm15
+	vpxor		%xmm15, %xmm1, %xmm1
+	vpshufd		$141, %xmm8, %xmm8
+	vpshufb		%xmm6, %xmm1, %xmm1
+	vpaddd		%xmm1, %xmm13, %xmm13
+	vpxor		%xmm13, %xmm0, %xmm0
+	vpshufd		$57, %xmm1, %xmm1
+	vpshufd		$78, %xmm13, %xmm13
+	vpslld		$25, %xmm0, %xmm10
+	vpsrld		$7, %xmm0, %xmm0
+	vpxor		%xmm10, %xmm0, %xmm0
+	vpunpcklqdq	%xmm2, %xmm3, %xmm10
+	vpshufd		$147, %xmm0, %xmm0
+	vpblendw	$51, %xmm14, %xmm10, %xmm14
+	vpshufd		$135, %xmm14, %xmm14
+	vpaddd		%xmm15, %xmm14, %xmm14
+	vpaddd		%xmm0, %xmm14, %xmm14
+	vpxor		%xmm14, %xmm1, %xmm1
+	vpunpcklqdq	%xmm3, %xmm4, %xmm15
+	vpshufb		%xmm7, %xmm1, %xmm1
+	vpaddd		%xmm1, %xmm13, %xmm13
+	vpxor		%xmm13, %xmm0, %xmm0
+	vpslld		$20, %xmm0, %xmm11
+	vpsrld		$12, %xmm0, %xmm0
+	vpxor		%xmm11, %xmm0, %xmm0
+	vpunpckhqdq	%xmm5, %xmm3, %xmm11
+	vpblendw	$51, %xmm15, %xmm11, %xmm11
+	vpunpckhqdq	%xmm3, %xmm5, %xmm15
+	vpaddd		%xmm11, %xmm14, %xmm11
+	vpaddd		%xmm0, %xmm11, %xmm11
+	vpxor		%xmm11, %xmm1, %xmm1
+	vpshufb		%xmm6, %xmm1, %xmm1
+	vpaddd		%xmm1, %xmm13, %xmm13
+	vpxor		%xmm13, %xmm0, %xmm0
+	vpshufd		$147, %xmm1, %xmm1
+	vpshufd		$78, %xmm13, %xmm13
+	vpslld		$25, %xmm0, %xmm14
+	vpsrld		$7, %xmm0, %xmm0
+	vpxor		%xmm14, %xmm0, %xmm14
+	vpunpckhqdq	%xmm4, %xmm2, %xmm0
+	vpshufd		$57, %xmm14, %xmm14
+	vpblendw	$51, %xmm15, %xmm0, %xmm15
+	vpaddd		%xmm15, %xmm11, %xmm15
+	vpaddd		%xmm14, %xmm15, %xmm15
+	vpxor		%xmm15, %xmm1, %xmm1
+	vpshufb		%xmm7, %xmm1, %xmm1
+	vpaddd		%xmm1, %xmm13, %xmm13
+	vpxor		%xmm13, %xmm14, %xmm14
+	vpslld		$20, %xmm14, %xmm11
+	vpsrld		$12, %xmm14, %xmm14
+	vpxor		%xmm11, %xmm14, %xmm14
+	vpblendw	$3, %xmm2, %xmm4, %xmm11
+	vpslldq		$8, %xmm11, %xmm0
+	vpblendw	$15, %xmm5, %xmm0, %xmm0
+	vpshufd		$99, %xmm0, %xmm0
+	vpaddd		%xmm15, %xmm0, %xmm15
+	vpaddd		%xmm14, %xmm15, %xmm15
+	vpxor		%xmm15, %xmm1, %xmm0
+	vpaddd		%xmm12, %xmm15, %xmm15
+	vpshufb		%xmm6, %xmm0, %xmm0
+	vpaddd		%xmm0, %xmm13, %xmm13
+	vpxor		%xmm13, %xmm14, %xmm14
+	vpshufd		$57, %xmm0, %xmm0
+	vpshufd		$78, %xmm13, %xmm13
+	vpslld		$25, %xmm14, %xmm1
+	vpsrld		$7, %xmm14, %xmm14
+	vpxor		%xmm1, %xmm14, %xmm14
+	vpblendw	$3, %xmm5, %xmm4, %xmm1
+	vpshufd		$147, %xmm14, %xmm14
+	vpaddd		%xmm14, %xmm15, %xmm15
+	vpxor		%xmm15, %xmm0, %xmm0
+	vpshufb		%xmm7, %xmm0, %xmm0
+	vpaddd		%xmm0, %xmm13, %xmm13
+	vpxor		%xmm13, %xmm14, %xmm14
+	vpslld		$20, %xmm14, %xmm12
+	vpsrld		$12, %xmm14, %xmm14
+	vpxor		%xmm12, %xmm14, %xmm14
+	vpsrldq		$4, %xmm2, %xmm12
+	vpblendw	$60, %xmm12, %xmm1, %xmm1
+	vpaddd		%xmm1, %xmm15, %xmm15
+	vpaddd		%xmm14, %xmm15, %xmm15
+	vpxor		%xmm15, %xmm0, %xmm0
+	vpblendw	$12, %xmm4, %xmm3, %xmm1
+	vpshufb		%xmm6, %xmm0, %xmm0
+	vpaddd		%xmm0, %xmm13, %xmm13
+	vpxor		%xmm13, %xmm14, %xmm14
+	vpshufd		$147, %xmm0, %xmm0
+	vpshufd		$78, %xmm13, %xmm13
+	vpslld		$25, %xmm14, %xmm12
+	vpsrld		$7, %xmm14, %xmm14
+	vpxor		%xmm12, %xmm14, %xmm14
+	vpsrldq		$4, %xmm5, %xmm12
+	vpblendw	$48, %xmm12, %xmm1, %xmm1
+	vpshufd		$33, %xmm5, %xmm12
+	vpshufd		$57, %xmm14, %xmm14
+	vpshufd		$108, %xmm1, %xmm1
+	vpblendw	$51, %xmm12, %xmm10, %xmm12
+	vpaddd		%xmm15, %xmm1, %xmm15
+	vpaddd		%xmm14, %xmm15, %xmm15
+	vpxor		%xmm15, %xmm0, %xmm0
+	vpaddd		%xmm12, %xmm15, %xmm15
+	vpshufb		%xmm7, %xmm0, %xmm0
+	vpaddd		%xmm0, %xmm13, %xmm1
+	vpxor		%xmm1, %xmm14, %xmm14
+	vpslld		$20, %xmm14, %xmm13
+	vpsrld		$12, %xmm14, %xmm14
+	vpxor		%xmm13, %xmm14, %xmm14
+	vpslldq		$12, %xmm3, %xmm13
+	vpaddd		%xmm14, %xmm15, %xmm15
+	vpxor		%xmm15, %xmm0, %xmm0
+	vpshufb		%xmm6, %xmm0, %xmm0
+	vpaddd		%xmm0, %xmm1, %xmm1
+	vpxor		%xmm1, %xmm14, %xmm14
+	vpshufd		$57, %xmm0, %xmm0
+	vpshufd		$78, %xmm1, %xmm1
+	vpslld		$25, %xmm14, %xmm12
+	vpsrld		$7, %xmm14, %xmm14
+	vpxor		%xmm12, %xmm14, %xmm14
+	vpblendw	$51, %xmm5, %xmm4, %xmm12
+	vpshufd		$147, %xmm14, %xmm14
+	vpblendw	$192, %xmm13, %xmm12, %xmm12
+	vpaddd		%xmm12, %xmm15, %xmm15
+	vpaddd		%xmm14, %xmm15, %xmm15
+	vpxor		%xmm15, %xmm0, %xmm0
+	vpsrldq		$4, %xmm3, %xmm12
+	vpshufb		%xmm7, %xmm0, %xmm0
+	vpaddd		%xmm0, %xmm1, %xmm1
+	vpxor		%xmm1, %xmm14, %xmm14
+	vpslld		$20, %xmm14, %xmm13
+	vpsrld		$12, %xmm14, %xmm14
+	vpxor		%xmm13, %xmm14, %xmm14
+	vpblendw	$48, %xmm2, %xmm5, %xmm13
+	vpblendw	$3, %xmm12, %xmm13, %xmm13
+	vpshufd		$156, %xmm13, %xmm13
+	vpaddd		%xmm15, %xmm13, %xmm15
+	vpaddd		%xmm14, %xmm15, %xmm15
+	vpxor		%xmm15, %xmm0, %xmm0
+	vpshufb		%xmm6, %xmm0, %xmm0
+	vpaddd		%xmm0, %xmm1, %xmm1
+	vpxor		%xmm1, %xmm14, %xmm14
+	vpshufd		$147, %xmm0, %xmm0
+	vpshufd		$78, %xmm1, %xmm1
+	vpslld		$25, %xmm14, %xmm13
+	vpsrld		$7, %xmm14, %xmm14
+	vpxor		%xmm13, %xmm14, %xmm14
+	vpunpcklqdq	%xmm2, %xmm4, %xmm13
+	vpshufd		$57, %xmm14, %xmm14
+	vpblendw	$12, %xmm12, %xmm13, %xmm12
+	vpshufd		$180, %xmm12, %xmm12
+	vpaddd		%xmm15, %xmm12, %xmm15
+	vpaddd		%xmm14, %xmm15, %xmm15
+	vpxor		%xmm15, %xmm0, %xmm0
+	vpshufb		%xmm7, %xmm0, %xmm0
+	vpaddd		%xmm0, %xmm1, %xmm1
+	vpxor		%xmm1, %xmm14, %xmm14
+	vpslld		$20, %xmm14, %xmm12
+	vpsrld		$12, %xmm14, %xmm14
+	vpxor		%xmm12, %xmm14, %xmm14
+	vpunpckhqdq	%xmm9, %xmm4, %xmm12
+	vpshufd		$198, %xmm12, %xmm12
+	vpaddd		%xmm15, %xmm12, %xmm15
+	vpaddd		%xmm14, %xmm15, %xmm15
+	vpxor		%xmm15, %xmm0, %xmm0
+	vpaddd		%xmm15, %xmm8, %xmm15
+	vpshufb		%xmm6, %xmm0, %xmm0
+	vpaddd		%xmm0, %xmm1, %xmm1
+	vpxor		%xmm1, %xmm14, %xmm14
+	vpshufd		$57, %xmm0, %xmm0
+	vpshufd		$78, %xmm1, %xmm1
+	vpslld		$25, %xmm14, %xmm12
+	vpsrld		$7, %xmm14, %xmm14
+	vpxor		%xmm12, %xmm14, %xmm14
+	vpsrldq		$4, %xmm4, %xmm12
+	vpshufd		$147, %xmm14, %xmm14
+	vpaddd		%xmm14, %xmm15, %xmm15
+	vpxor		%xmm15, %xmm0, %xmm0
+	vpshufb		%xmm7, %xmm0, %xmm0
+	vpaddd		%xmm0, %xmm1, %xmm1
+	vpxor		%xmm1, %xmm14, %xmm14
+	vpslld		$20, %xmm14, %xmm8
+	vpsrld		$12, %xmm14, %xmm14
+	vpxor		%xmm14, %xmm8, %xmm14
+	vpblendw	$48, %xmm5, %xmm2, %xmm8
+	vpblendw	$3, %xmm12, %xmm8, %xmm8
+	vpunpckhqdq	%xmm5, %xmm4, %xmm12
+	vpshufd		$75, %xmm8, %xmm8
+	vpblendw	$60, %xmm10, %xmm12, %xmm10
+	vpaddd		%xmm15, %xmm8, %xmm15
+	vpaddd		%xmm14, %xmm15, %xmm15
+	vpxor		%xmm0, %xmm15, %xmm0
+	vpshufd		$45, %xmm10, %xmm10
+	vpshufb		%xmm6, %xmm0, %xmm0
+	vpaddd		%xmm15, %xmm10, %xmm15
+	vpaddd		%xmm0, %xmm1, %xmm1
+	vpxor		%xmm1, %xmm14, %xmm14
+	vpshufd		$147, %xmm0, %xmm0
+	vpshufd		$78, %xmm1, %xmm1
+	vpslld		$25, %xmm14, %xmm8
+	vpsrld		$7, %xmm14, %xmm14
+	vpxor		%xmm14, %xmm8, %xmm8
+	vpshufd		$57, %xmm8, %xmm8
+	vpaddd		%xmm8, %xmm15, %xmm15
+	vpxor		%xmm0, %xmm15, %xmm0
+	vpshufb		%xmm7, %xmm0, %xmm0
+	vpaddd		%xmm0, %xmm1, %xmm1
+	vpxor		%xmm8, %xmm1, %xmm8
+	vpslld		$20, %xmm8, %xmm10
+	vpsrld		$12, %xmm8, %xmm8
+	vpxor		%xmm8, %xmm10, %xmm10
+	vpunpckldq	%xmm3, %xmm4, %xmm8
+	vpunpcklqdq	%xmm9, %xmm8, %xmm9
+	vpaddd		%xmm9, %xmm15, %xmm9
+	vpaddd		%xmm10, %xmm9, %xmm9
+	vpxor		%xmm0, %xmm9, %xmm8
+	vpshufb		%xmm6, %xmm8, %xmm8
+	vpaddd		%xmm8, %xmm1, %xmm1
+	vpxor		%xmm1, %xmm10, %xmm10
+	vpshufd		$57, %xmm8, %xmm8
+	vpshufd		$78, %xmm1, %xmm1
+	vpslld		$25, %xmm10, %xmm12
+	vpsrld		$7, %xmm10, %xmm10
+	vpxor		%xmm10, %xmm12, %xmm10
+	vpblendw	$48, %xmm4, %xmm3, %xmm12
+	vpshufd		$147, %xmm10, %xmm0
+	vpunpckhdq	%xmm5, %xmm3, %xmm10
+	vpshufd		$78, %xmm12, %xmm12
+	vpunpcklqdq	%xmm4, %xmm10, %xmm10
+	vpblendw	$192, %xmm2, %xmm10, %xmm10
+	vpshufhw	$78, %xmm10, %xmm10
+	vpaddd		%xmm10, %xmm9, %xmm10
+	vpaddd		%xmm0, %xmm10, %xmm10
+	vpxor		%xmm8, %xmm10, %xmm8
+	vpshufb		%xmm7, %xmm8, %xmm8
+	vpaddd		%xmm8, %xmm1, %xmm1
+	vpxor		%xmm0, %xmm1, %xmm9
+	vpslld		$20, %xmm9, %xmm0
+	vpsrld		$12, %xmm9, %xmm9
+	vpxor		%xmm9, %xmm0, %xmm0
+	vpunpckhdq	%xmm5, %xmm4, %xmm9
+	vpblendw	$240, %xmm9, %xmm2, %xmm13
+	vpshufd		$39, %xmm13, %xmm13
+	vpaddd		%xmm10, %xmm13, %xmm10
+	vpaddd		%xmm0, %xmm10, %xmm10
+	vpxor		%xmm8, %xmm10, %xmm8
+	vpblendw	$12, %xmm4, %xmm2, %xmm13
+	vpshufb		%xmm6, %xmm8, %xmm8
+	vpslldq		$4, %xmm13, %xmm13
+	vpblendw	$15, %xmm5, %xmm13, %xmm13
+	vpaddd		%xmm8, %xmm1, %xmm1
+	vpxor		%xmm1, %xmm0, %xmm0
+	vpaddd		%xmm13, %xmm10, %xmm13
+	vpshufd		$147, %xmm8, %xmm8
+	vpshufd		$78, %xmm1, %xmm1
+	vpslld		$25, %xmm0, %xmm14
+	vpsrld		$7, %xmm0, %xmm0
+	vpxor		%xmm0, %xmm14, %xmm14
+	vpshufd		$57, %xmm14, %xmm14
+	vpaddd		%xmm14, %xmm13, %xmm13
+	vpxor		%xmm8, %xmm13, %xmm8
+	vpaddd		%xmm13, %xmm12, %xmm12
+	vpshufb		%xmm7, %xmm8, %xmm8
+	vpaddd		%xmm8, %xmm1, %xmm1
+	vpxor		%xmm14, %xmm1, %xmm14
+	vpslld		$20, %xmm14, %xmm10
+	vpsrld		$12, %xmm14, %xmm14
+	vpxor		%xmm14, %xmm10, %xmm10
+	vpaddd		%xmm10, %xmm12, %xmm12
+	vpxor		%xmm8, %xmm12, %xmm8
+	vpshufb		%xmm6, %xmm8, %xmm8
+	vpaddd		%xmm8, %xmm1, %xmm1
+	vpxor		%xmm1, %xmm10, %xmm0
+	vpshufd		$57, %xmm8, %xmm8
+	vpshufd		$78, %xmm1, %xmm1
+	vpslld		$25, %xmm0, %xmm10
+	vpsrld		$7, %xmm0, %xmm0
+	vpxor		%xmm0, %xmm10, %xmm10
+	vpblendw	$48, %xmm2, %xmm3, %xmm0
+	vpblendw	$15, %xmm11, %xmm0, %xmm0
+	vpshufd		$147, %xmm10, %xmm10
+	vpshufd		$114, %xmm0, %xmm0
+	vpaddd		%xmm12, %xmm0, %xmm0
+	vpaddd		%xmm10, %xmm0, %xmm0
+	vpxor		%xmm8, %xmm0, %xmm8
+	vpshufb		%xmm7, %xmm8, %xmm8
+	vpaddd		%xmm8, %xmm1, %xmm1
+	vpxor		%xmm10, %xmm1, %xmm10
+	vpslld		$20, %xmm10, %xmm11
+	vpsrld		$12, %xmm10, %xmm10
+	vpxor		%xmm10, %xmm11, %xmm10
+	vpslldq		$4, %xmm4, %xmm11
+	vpblendw	$192, %xmm11, %xmm3, %xmm3
+	vpunpckldq	%xmm5, %xmm4, %xmm4
+	vpshufd		$99, %xmm3, %xmm3
+	vpaddd		%xmm0, %xmm3, %xmm3
+	vpaddd		%xmm10, %xmm3, %xmm3
+	vpxor		%xmm8, %xmm3, %xmm11
+	vpunpckldq	%xmm5, %xmm2, %xmm0
+	vpblendw	$192, %xmm2, %xmm5, %xmm2
+	vpshufb		%xmm6, %xmm11, %xmm11
+	vpunpckhqdq	%xmm0, %xmm9, %xmm0
+	vpblendw	$15, %xmm4, %xmm2, %xmm4
+	vpaddd		%xmm11, %xmm1, %xmm1
+	vpxor		%xmm1, %xmm10, %xmm10
+	vpshufd		$147, %xmm11, %xmm11
+	vpshufd		$201, %xmm0, %xmm0
+	vpslld		$25, %xmm10, %xmm8
+	vpsrld		$7, %xmm10, %xmm10
+	vpxor		%xmm10, %xmm8, %xmm10
+	vpshufd		$78, %xmm1, %xmm1
+	vpaddd		%xmm3, %xmm0, %xmm0
+	vpshufd		$27, %xmm4, %xmm4
+	vpshufd		$57, %xmm10, %xmm10
+	vpaddd		%xmm10, %xmm0, %xmm0
+	vpxor		%xmm11, %xmm0, %xmm11
+	vpaddd		%xmm0, %xmm4, %xmm0
+	vpshufb		%xmm7, %xmm11, %xmm7
+	vpaddd		%xmm7, %xmm1, %xmm1
+	vpxor		%xmm10, %xmm1, %xmm10
+	vpslld		$20, %xmm10, %xmm8
+	vpsrld		$12, %xmm10, %xmm10
+	vpxor		%xmm10, %xmm8, %xmm8
+	vpaddd		%xmm8, %xmm0, %xmm0
+	vpxor		%xmm7, %xmm0, %xmm7
+	vpshufb		%xmm6, %xmm7, %xmm6
+	vpaddd		%xmm6, %xmm1, %xmm1
+	vpxor		%xmm1, %xmm8, %xmm8
+	vpshufd		$78, %xmm1, %xmm1
+	vpshufd		$57, %xmm6, %xmm6
+	vpslld		$25, %xmm8, %xmm2
+	vpsrld		$7, %xmm8, %xmm8
+	vpxor		%xmm8, %xmm2, %xmm8
+	vpxor		(%rdi), %xmm1, %xmm1
+	vpshufd		$147, %xmm8, %xmm8
+	vpxor		%xmm0, %xmm1, %xmm0
+	vmovups		%xmm0, (%rdi)
+	vpxor		16(%rdi), %xmm8, %xmm0
+	vpxor		%xmm6, %xmm0, %xmm6
+	vmovups		%xmm6, 16(%rdi)
+	addq		$64, %rsi
+	decq		%rdx
+	jnz .Lbeginofloop
+.Lendofloop:
+	ret
+ENDPROC(blake2s_compress_avx)
+#endif /* CONFIG_AS_AVX */
+
+#ifdef CONFIG_AS_AVX512
+ENTRY(blake2s_compress_avx512)
+	vmovdqu		(%rdi),%xmm0
+	vmovdqu		0x10(%rdi),%xmm1
+	vmovdqu		0x20(%rdi),%xmm4
+	vmovq		%rcx,%xmm5
+	vmovdqa		IV(%rip),%xmm14
+	vmovdqa		IV+16(%rip),%xmm15
+	jmp		.Lblake2s_compress_avx512_mainloop
+.align 32
+.Lblake2s_compress_avx512_mainloop:
+	vmovdqa		%xmm0,%xmm10
+	vmovdqa		%xmm1,%xmm11
+	vpaddq		%xmm5,%xmm4,%xmm4
+	vmovdqa		%xmm14,%xmm2
+	vpxor		%xmm15,%xmm4,%xmm3
+	vmovdqu		(%rsi),%ymm6
+	vmovdqu		0x20(%rsi),%ymm7
+	addq		$0x40,%rsi
+	leaq		SIGMA(%rip),%rax
+	movb		$0xa,%cl
+.Lblake2s_compress_avx512_roundloop:
+	addq		$0x40,%rax
+	vmovdqa		-0x40(%rax),%ymm8
+	vmovdqa		-0x20(%rax),%ymm9
+	vpermi2d	%ymm7,%ymm6,%ymm8
+	vpermi2d	%ymm7,%ymm6,%ymm9
+	vmovdqa		%ymm8,%ymm6
+	vmovdqa		%ymm9,%ymm7
+	vpaddd		%xmm8,%xmm0,%xmm0
+	vpaddd		%xmm1,%xmm0,%xmm0
+	vpxor		%xmm0,%xmm3,%xmm3
+	vprord		$0x10,%xmm3,%xmm3
+	vpaddd		%xmm3,%xmm2,%xmm2
+	vpxor		%xmm2,%xmm1,%xmm1
+	vprord		$0xc,%xmm1,%xmm1
+	vextracti128	$0x1,%ymm8,%xmm8
+	vpaddd		%xmm8,%xmm0,%xmm0
+	vpaddd		%xmm1,%xmm0,%xmm0
+	vpxor		%xmm0,%xmm3,%xmm3
+	vprord		$0x8,%xmm3,%xmm3
+	vpaddd		%xmm3,%xmm2,%xmm2
+	vpxor		%xmm2,%xmm1,%xmm1
+	vprord		$0x7,%xmm1,%xmm1
+	vpshufd		$0x39,%xmm1,%xmm1
+	vpshufd		$0x4e,%xmm2,%xmm2
+	vpshufd		$0x93,%xmm3,%xmm3
+	vpaddd		%xmm9,%xmm0,%xmm0
+	vpaddd		%xmm1,%xmm0,%xmm0
+	vpxor		%xmm0,%xmm3,%xmm3
+	vprord		$0x10,%xmm3,%xmm3
+	vpaddd		%xmm3,%xmm2,%xmm2
+	vpxor		%xmm2,%xmm1,%xmm1
+	vprord		$0xc,%xmm1,%xmm1
+	vextracti128	$0x1,%ymm9,%xmm9
+	vpaddd		%xmm9,%xmm0,%xmm0
+	vpaddd		%xmm1,%xmm0,%xmm0
+	vpxor		%xmm0,%xmm3,%xmm3
+	vprord		$0x8,%xmm3,%xmm3
+	vpaddd		%xmm3,%xmm2,%xmm2
+	vpxor		%xmm2,%xmm1,%xmm1
+	vprord		$0x7,%xmm1,%xmm1
+	vpshufd		$0x93,%xmm1,%xmm1
+	vpshufd		$0x4e,%xmm2,%xmm2
+	vpshufd		$0x39,%xmm3,%xmm3
+	decb		%cl
+	jne		.Lblake2s_compress_avx512_roundloop
+	vpxor		%xmm10,%xmm0,%xmm0
+	vpxor		%xmm11,%xmm1,%xmm1
+	vpxor		%xmm2,%xmm0,%xmm0
+	vpxor		%xmm3,%xmm1,%xmm1
+	decq		%rdx
+	jne		.Lblake2s_compress_avx512_mainloop
+	vmovdqu		%xmm0,(%rdi)
+	vmovdqu		%xmm1,0x10(%rdi)
+	vmovdqu		%xmm4,0x20(%rdi)
+	vzeroupper
+	retq
+ENDPROC(blake2s_compress_avx512)
+#endif /* CONFIG_AS_AVX512 */
-- 
2.19.0

Powered by blists - more mailing lists