pq_crypto 0.3.2 → 0.5.0

data/ext/pqcrypto/vendor/mlkem-native/mlkem/src/native/x86_64/src/tomont.S

@@ -0,0 +1,155 @@
+/*
+ * Copyright (c) The mlkem-native project authors
+ * SPDX-License-Identifier: Apache-2.0 OR ISC OR MIT
+ */
+
+/* References
+ * ==========
+ *
+ * - [REF_AVX2]
+ *   CRYSTALS-Kyber optimized AVX2 implementation
+ *   Bos, Ducas, Kiltz, Lepoint, Lyubashevsky, Schanck, Schwabe, Seiler, Stehlé
+ *   https://github.com/pq-crystals/kyber/tree/main/avx2
+ */
+
+/*
+ * Implementation from Kyber reference repository @[REF_AVX2]
+ *
+ * Changes:
+ * - Add call to csub in reduce128_avx to produce outputs
+ *   in [0,1,...,q-1] rather than [0,1,...,q], matching the
+ *   semantics of mlk_poly_reduce(),
+ * - Use a macro instead of a local function call.
+ */
+
+#include "../../../common.h"
+#if defined(MLK_ARITH_BACKEND_X86_64_DEFAULT) && \
+    !defined(MLK_CONFIG_MULTILEVEL_NO_SHARED)
+
+/*
+ * WARNING: This file is auto-derived from the mlkem-native source file
+ *   dev/x86_64/src/tomont.S using scripts/simpasm. Do not modify it directly.
+ */
+
+#if defined(__ELF__)
+.section .note.GNU-stack,"",@progbits
+#endif
+
+.text
+.balign 4
+.global MLK_ASM_NAMESPACE(tomont_avx2)
+MLK_ASM_FN_SYMBOL(tomont_avx2)
+
+        .cfi_startproc
+        movl $0xd010d01, %eax        # imm = 0xD010D01
+        vmovd %eax, %xmm0
+        vpbroadcastd %xmm0, %ymm0
+        movl $0x50495049, %eax       # imm = 0x50495049
+        vmovd %eax, %xmm1
+        vpbroadcastd %xmm1, %ymm1
+        movl $0x5490549, %eax        # imm = 0x5490549
+        vmovd %eax, %xmm2
+        vpbroadcastd %xmm2, %ymm2
+        vmovdqa (%rdi), %ymm3
+        vmovdqa 0x20(%rdi), %ymm4
+        vmovdqa 0x40(%rdi), %ymm5
+        vmovdqa 0x60(%rdi), %ymm6
+        vmovdqa 0x80(%rdi), %ymm7
+        vmovdqa 0xa0(%rdi), %ymm8
+        vmovdqa 0xc0(%rdi), %ymm9
+        vmovdqa 0xe0(%rdi), %ymm10
+        vpmullw %ymm1, %ymm3, %ymm11
+        vpmulhw %ymm2, %ymm3, %ymm3
+        vpmulhw %ymm0, %ymm11, %ymm11
+        vpsubw %ymm11, %ymm3, %ymm3
+        vpmullw %ymm1, %ymm4, %ymm12
+        vpmulhw %ymm2, %ymm4, %ymm4
+        vpmulhw %ymm0, %ymm12, %ymm12
+        vpsubw %ymm12, %ymm4, %ymm4
+        vpmullw %ymm1, %ymm5, %ymm13
+        vpmulhw %ymm2, %ymm5, %ymm5
+        vpmulhw %ymm0, %ymm13, %ymm13
+        vpsubw %ymm13, %ymm5, %ymm5
+        vpmullw %ymm1, %ymm6, %ymm14
+        vpmulhw %ymm2, %ymm6, %ymm6
+        vpmulhw %ymm0, %ymm14, %ymm14
+        vpsubw %ymm14, %ymm6, %ymm6
+        vpmullw %ymm1, %ymm7, %ymm15
+        vpmulhw %ymm2, %ymm7, %ymm7
+        vpmulhw %ymm0, %ymm15, %ymm15
+        vpsubw %ymm15, %ymm7, %ymm7
+        vpmullw %ymm1, %ymm8, %ymm11
+        vpmulhw %ymm2, %ymm8, %ymm8
+        vpmulhw %ymm0, %ymm11, %ymm11
+        vpsubw %ymm11, %ymm8, %ymm8
+        vpmullw %ymm1, %ymm9, %ymm12
+        vpmulhw %ymm2, %ymm9, %ymm9
+        vpmulhw %ymm0, %ymm12, %ymm12
+        vpsubw %ymm12, %ymm9, %ymm9
+        vpmullw %ymm1, %ymm10, %ymm13
+        vpmulhw %ymm2, %ymm10, %ymm10
+        vpmulhw %ymm0, %ymm13, %ymm13
+        vpsubw %ymm13, %ymm10, %ymm10
+        vmovdqa %ymm3, (%rdi)
+        vmovdqa %ymm4, 0x20(%rdi)
+        vmovdqa %ymm5, 0x40(%rdi)
+        vmovdqa %ymm6, 0x60(%rdi)
+        vmovdqa %ymm7, 0x80(%rdi)
+        vmovdqa %ymm8, 0xa0(%rdi)
+        vmovdqa %ymm9, 0xc0(%rdi)
+        vmovdqa %ymm10, 0xe0(%rdi)
+        vmovdqa 0x100(%rdi), %ymm3
+        vmovdqa 0x120(%rdi), %ymm4
+        vmovdqa 0x140(%rdi), %ymm5
+        vmovdqa 0x160(%rdi), %ymm6
+        vmovdqa 0x180(%rdi), %ymm7
+        vmovdqa 0x1a0(%rdi), %ymm8
+        vmovdqa 0x1c0(%rdi), %ymm9
+        vmovdqa 0x1e0(%rdi), %ymm10
+        vpmullw %ymm1, %ymm3, %ymm11
+        vpmulhw %ymm2, %ymm3, %ymm3
+        vpmulhw %ymm0, %ymm11, %ymm11
+        vpsubw %ymm11, %ymm3, %ymm3
+        vpmullw %ymm1, %ymm4, %ymm12
+        vpmulhw %ymm2, %ymm4, %ymm4
+        vpmulhw %ymm0, %ymm12, %ymm12
+        vpsubw %ymm12, %ymm4, %ymm4
+        vpmullw %ymm1, %ymm5, %ymm13
+        vpmulhw %ymm2, %ymm5, %ymm5
+        vpmulhw %ymm0, %ymm13, %ymm13
+        vpsubw %ymm13, %ymm5, %ymm5
+        vpmullw %ymm1, %ymm6, %ymm14
+        vpmulhw %ymm2, %ymm6, %ymm6
+        vpmulhw %ymm0, %ymm14, %ymm14
+        vpsubw %ymm14, %ymm6, %ymm6
+        vpmullw %ymm1, %ymm7, %ymm15
+        vpmulhw %ymm2, %ymm7, %ymm7
+        vpmulhw %ymm0, %ymm15, %ymm15
+        vpsubw %ymm15, %ymm7, %ymm7
+        vpmullw %ymm1, %ymm8, %ymm11
+        vpmulhw %ymm2, %ymm8, %ymm8
+        vpmulhw %ymm0, %ymm11, %ymm11
+        vpsubw %ymm11, %ymm8, %ymm8
+        vpmullw %ymm1, %ymm9, %ymm12
+        vpmulhw %ymm2, %ymm9, %ymm9
+        vpmulhw %ymm0, %ymm12, %ymm12
+        vpsubw %ymm12, %ymm9, %ymm9
+        vpmullw %ymm1, %ymm10, %ymm13
+        vpmulhw %ymm2, %ymm10, %ymm10
+        vpmulhw %ymm0, %ymm13, %ymm13
+        vpsubw %ymm13, %ymm10, %ymm10
+        vmovdqa %ymm3, 0x100(%rdi)
+        vmovdqa %ymm4, 0x120(%rdi)
+        vmovdqa %ymm5, 0x140(%rdi)
+        vmovdqa %ymm6, 0x160(%rdi)
+        vmovdqa %ymm7, 0x180(%rdi)
+        vmovdqa %ymm8, 0x1a0(%rdi)
+        vmovdqa %ymm9, 0x1c0(%rdi)
+        vmovdqa %ymm10, 0x1e0(%rdi)
+        retq
+        .cfi_endproc
+
+MLK_ASM_FN_SIZE(tomont_avx2)
+
+#endif /* MLK_ARITH_BACKEND_X86_64_DEFAULT && !MLK_CONFIG_MULTILEVEL_NO_SHARED \
+        */

data/ext/pqcrypto/vendor/mlkem-native/mlkem/src/params.h

@@ -0,0 +1,76 @@
+/*
+ * Copyright (c) The mlkem-native project authors
+ * SPDX-License-Identifier: Apache-2.0 OR ISC OR MIT
+ */
+#ifndef MLK_PARAMS_H
+#define MLK_PARAMS_H
+
+#if !defined(MLK_CONFIG_PARAMETER_SET)
+#error MLK_CONFIG_PARAMETER_SET is not defined
+#endif
+
+#if MLK_CONFIG_PARAMETER_SET == 512
+#define MLKEM_K 2
+#elif MLK_CONFIG_PARAMETER_SET == 768
+#define MLKEM_K 3
+#elif MLK_CONFIG_PARAMETER_SET == 1024
+#define MLKEM_K 4
+#else
+#error Invalid value for MLK_CONFIG_PARAMETER_SET. Must be 512, 768, or 1024.
+#endif
+
+#define MLKEM_N 256
+#define MLKEM_Q 3329
+#define MLKEM_Q_HALF ((MLKEM_Q + 1) / 2) /* 1665 */
+#define MLKEM_UINT12_LIMIT 4096
+
+#define MLKEM_SYMBYTES 32 /* size in bytes of hashes, and seeds */
+#define MLKEM_SSBYTES 32  /* size in bytes of shared key */
+
+#define MLKEM_POLYBYTES 384
+#define MLKEM_POLYVECBYTES (MLKEM_K * MLKEM_POLYBYTES)
+
+#define MLKEM_POLYCOMPRESSEDBYTES_D4 128
+#define MLKEM_POLYCOMPRESSEDBYTES_D5 160
+#define MLKEM_POLYCOMPRESSEDBYTES_D10 320
+#define MLKEM_POLYCOMPRESSEDBYTES_D11 352
+
+#if MLKEM_K == 2
+#define MLKEM_ETA1 3
+#define MLKEM_DU 10
+#define MLKEM_DV 4
+#define MLKEM_POLYCOMPRESSEDBYTES_DV MLKEM_POLYCOMPRESSEDBYTES_D4
+#define MLKEM_POLYCOMPRESSEDBYTES_DU MLKEM_POLYCOMPRESSEDBYTES_D10
+#define MLKEM_POLYVECCOMPRESSEDBYTES_DU (MLKEM_K * MLKEM_POLYCOMPRESSEDBYTES_DU)
+#elif MLKEM_K == 3
+#define MLKEM_ETA1 2
+#define MLKEM_DU 10
+#define MLKEM_DV 4
+#define MLKEM_POLYCOMPRESSEDBYTES_DV MLKEM_POLYCOMPRESSEDBYTES_D4
+#define MLKEM_POLYCOMPRESSEDBYTES_DU MLKEM_POLYCOMPRESSEDBYTES_D10
+#define MLKEM_POLYVECCOMPRESSEDBYTES_DU (MLKEM_K * MLKEM_POLYCOMPRESSEDBYTES_DU)
+#elif MLKEM_K == 4
+#define MLKEM_ETA1 2
+#define MLKEM_DU 11
+#define MLKEM_DV 5
+#define MLKEM_POLYCOMPRESSEDBYTES_DV MLKEM_POLYCOMPRESSEDBYTES_D5
+#define MLKEM_POLYCOMPRESSEDBYTES_DU MLKEM_POLYCOMPRESSEDBYTES_D11
+#define MLKEM_POLYVECCOMPRESSEDBYTES_DU (MLKEM_K * MLKEM_POLYCOMPRESSEDBYTES_DU)
+#endif /* MLKEM_K == 4 */
+
+#define MLKEM_ETA2 2
+
+#define MLKEM_INDCPA_MSGBYTES (MLKEM_SYMBYTES)
+#define MLKEM_INDCPA_PUBLICKEYBYTES (MLKEM_POLYVECBYTES + MLKEM_SYMBYTES)
+#define MLKEM_INDCPA_SECRETKEYBYTES (MLKEM_POLYVECBYTES)
+#define MLKEM_INDCPA_BYTES \
+  (MLKEM_POLYVECCOMPRESSEDBYTES_DU + MLKEM_POLYCOMPRESSEDBYTES_DV)
+
+#define MLKEM_INDCCA_PUBLICKEYBYTES (MLKEM_INDCPA_PUBLICKEYBYTES)
+/* 32 bytes of additional space to save H(pk) */
+#define MLKEM_INDCCA_SECRETKEYBYTES                            \
+  (MLKEM_INDCPA_SECRETKEYBYTES + MLKEM_INDCPA_PUBLICKEYBYTES + \
+   2 * MLKEM_SYMBYTES)
+#define MLKEM_INDCCA_CIPHERTEXTBYTES (MLKEM_INDCPA_BYTES)
+
+#endif /* !MLK_PARAMS_H */