From 77541659f3a6d0fe9499889d619070ef1dd66792 Mon Sep 17 00:00:00 2001 From: Nigel Tao Date: Sun, 21 Apr 2024 13:02:34 +1000 Subject: [PATCH] std/crc64: tweak x86 to be more like std/crc32 --- release/c/wuffs-unsupported-snapshot.c | 16 ++++++------ script/print-crc64-x86-sse42-magic-numbers.go | 12 ++++++--- std/crc64/common_up_x86_sse42.wuffs | 25 +++++++++++-------- 3 files changed, 31 insertions(+), 22 deletions(-) diff --git a/release/c/wuffs-unsupported-snapshot.c b/release/c/wuffs-unsupported-snapshot.c index 3d4ed28a..63f1bf82 100644 --- a/release/c/wuffs-unsupported-snapshot.c +++ b/release/c/wuffs-unsupported-snapshot.c @@ -36609,7 +36609,7 @@ WUFFS_CRC64__SHUFFLE_8F80[16] WUFFS_BASE__POTENTIALLY_UNUSED = { }; static const uint8_t -WUFFS_CRC64__ECMA_X86_SSE42_K1K2[16] WUFFS_BASE__POTENTIALLY_UNUSED = { +WUFFS_CRC64__ECMA_X86_SSE42_K3K4[16] WUFFS_BASE__POTENTIALLY_UNUSED = { 228u, 58u, 57u, 202u, 151u, 212u, 93u, 224u, 64u, 95u, 135u, 199u, 175u, 149u, 190u, 218u, }; @@ -36904,7 +36904,7 @@ wuffs_crc64__ecma_hasher__up_x86_sse42( __m128i v_s0_8F80 = {0}; __m128i v_x0 = {0}; __m128i v_aa = {0}; - __m128i v_k1k2 = {0}; + __m128i v_k3k4 = {0}; __m128i v_t0 = {0}; __m128i v_t1 = {0}; __m128i v_t2 = {0}; @@ -36943,17 +36943,17 @@ wuffs_crc64__ecma_hasher__up_x86_sse42( v_s0_8F80 = _mm_shuffle_epi8(v_s0, _mm_lddqu_si128((const __m128i*)(const void*)(WUFFS_CRC64__SHUFFLE_8F80))); v_x0 = _mm_lddqu_si128((const __m128i*)(const void*)(a_x.ptr + 0u)); a_x = wuffs_base__slice_u8__subslice_i(a_x, 16u); - v_k1k2 = _mm_lddqu_si128((const __m128i*)(const void*)(WUFFS_CRC64__ECMA_X86_SSE42_K1K2)); + v_k3k4 = _mm_lddqu_si128((const __m128i*)(const void*)(WUFFS_CRC64__ECMA_X86_SSE42_K3K4)); v_t0 = _mm_xor_si128(v_s0_707F, v_x0); - v_t1 = _mm_clmulepi64_si128(v_t0, v_k1k2, (int32_t)(0u)); - v_t2 = _mm_clmulepi64_si128(v_t0, v_k1k2, (int32_t)(17u)); + v_t1 = _mm_clmulepi64_si128(v_t0, v_k3k4, (int32_t)(0u)); + v_t2 = _mm_clmulepi64_si128(v_t0, v_k3k4, (int32_t)(17u)); v_aa = _mm_xor_si128(_mm_xor_si128(v_t1, v_t2), v_s0_8F80); while (((uint64_t)(a_x.len)) >= 32u) { v_x0 = _mm_lddqu_si128((const __m128i*)(const void*)(a_x.ptr + 0u)); a_x = wuffs_base__slice_u8__subslice_i(a_x, 16u); v_u0 = _mm_xor_si128(v_aa, v_x0); - v_u1 = _mm_clmulepi64_si128(v_u0, v_k1k2, (int32_t)(0u)); - v_u2 = _mm_clmulepi64_si128(v_u0, v_k1k2, (int32_t)(17u)); + v_u1 = _mm_clmulepi64_si128(v_u0, v_k3k4, (int32_t)(0u)); + v_u2 = _mm_clmulepi64_si128(v_u0, v_k3k4, (int32_t)(17u)); v_aa = _mm_xor_si128(v_u1, v_u2); } if (((uint64_t)(a_x.len)) < 16u) { @@ -36962,7 +36962,7 @@ wuffs_crc64__ecma_hasher__up_x86_sse42( v_x0 = _mm_lddqu_si128((const __m128i*)(const void*)(a_x.ptr + 0u)); a_x = wuffs_base__slice_u8__subslice_i(a_x, 16u); v_v0 = _mm_xor_si128(v_aa, v_x0); - v_v1 = _mm_clmulepi64_si128(v_v0, v_k1k2, (int32_t)(16u)); + v_v1 = _mm_clmulepi64_si128(v_v0, v_k3k4, (int32_t)(16u)); v_aa = _mm_xor_si128(v_v1, _mm_srli_si128(v_v0, (int32_t)(8u))); v_pxmu = _mm_lddqu_si128((const __m128i*)(const void*)(WUFFS_CRC64__ECMA_X86_SSE42_PXMU)); v_w1 = _mm_clmulepi64_si128(v_aa, v_pxmu, (int32_t)(16u)); diff --git a/script/print-crc64-x86-sse42-magic-numbers.go b/script/print-crc64-x86-sse42-magic-numbers.go index adc0695a..90d5c431 100644 --- a/script/print-crc64-x86-sse42-magic-numbers.go +++ b/script/print-crc64-x86-sse42-magic-numbers.go @@ -23,8 +23,10 @@ package main // // Output: // Px' = 0x92D8_AF2B_AF0E_1E85 -// k1' = 0xE05D_D497_CA39_3AE4 -// k2' = 0xDABE_95AF_C787_5F40 +// k1' = 0xF83B_4090_32DA_5F76 +// k2' = 0x081F_6054_A784_2DF4 +// k3' = 0xE05D_D497_CA39_3AE4 +// k4' = 0xDABE_95AF_C787_5F40 // μ' = 0x9C3E_466C_1729_63D5 import ( @@ -104,7 +106,9 @@ func calcMu(name string) { func main() { show("Px'", px) - calcKn("k1'", 128+64) - calcKn("k2'", 128) + calcKn("k1'", 512+64) + calcKn("k2'", 512) + calcKn("k3'", 128+64) + calcKn("k4'", 128) calcMu("μ' ") } diff --git a/std/crc64/common_up_x86_sse42.wuffs b/std/crc64/common_up_x86_sse42.wuffs index 189a175f..117b8a9e 100644 --- a/std/crc64/common_up_x86_sse42.wuffs +++ b/std/crc64/common_up_x86_sse42.wuffs @@ -25,7 +25,7 @@ pri func ecma_hasher.up_x86_sse42!(x: roslice base.u8), var s0_8F80 : base.x86_m128i var x0 : base.x86_m128i var aa : base.x86_m128i - var k1k2 : base.x86_m128i + var k3k4 : base.x86_m128i var t0 : base.x86_m128i var t1 : base.x86_m128i var t2 : base.x86_m128i @@ -71,10 +71,10 @@ pri func ecma_hasher.up_x86_sse42!(x: roslice base.u8), x0 = util.make_m128i_slice128(a: args.x[0x00 .. 0x10]) args.x = args.x[0x10 ..] - k1k2 = util.make_m128i_slice128(a: ECMA_X86_SSE42_K1K2[.. 16]) + k3k4 = util.make_m128i_slice128(a: ECMA_X86_SSE42_K3K4[.. 16]) t0 = s0_707F._mm_xor_si128(b: x0) - t1 = t0._mm_clmulepi64_si128(b: k1k2, imm8: 0x00) - t2 = t0._mm_clmulepi64_si128(b: k1k2, imm8: 0x11) + t1 = t0._mm_clmulepi64_si128(b: k3k4, imm8: 0x00) + t2 = t0._mm_clmulepi64_si128(b: k3k4, imm8: 0x11) aa = t1._mm_xor_si128(b: t2)._mm_xor_si128(b: s0_8F80) // Main loop. @@ -84,8 +84,8 @@ pri func ecma_hasher.up_x86_sse42!(x: roslice base.u8), args.x = args.x[0x10 ..] u0 = aa._mm_xor_si128(b: x0) - u1 = u0._mm_clmulepi64_si128(b: k1k2, imm8: 0x00) - u2 = u0._mm_clmulepi64_si128(b: k1k2, imm8: 0x11) + u1 = u0._mm_clmulepi64_si128(b: k3k4, imm8: 0x00) + u2 = u0._mm_clmulepi64_si128(b: k3k4, imm8: 0x11) aa = u1._mm_xor_si128(b: u2) } endwhile @@ -99,7 +99,7 @@ pri func ecma_hasher.up_x86_sse42!(x: roslice base.u8), args.x = args.x[0x10 ..] v0 = aa._mm_xor_si128(b: x0) - v1 = v0._mm_clmulepi64_si128(b: k1k2, imm8: 0x10) + v1 = v0._mm_clmulepi64_si128(b: k3k4, imm8: 0x10) aa = v1._mm_xor_si128(b: v0._mm_srli_si128(imm8: 8)) // Barrett reduction. @@ -134,9 +134,14 @@ pri const SHUFFLE_8F80 : roarray[16] base.u8 = [ // These constants are reproduced by // script/print-crc64-x86-sse42-magic-numbers.go -pri const ECMA_X86_SSE42_K1K2 : roarray[16] base.u8 = [ - 0xE4, 0x3A, 0x39, 0xCA, 0x97, 0xD4, 0x5D, 0xE0, // k1' = 0xE05D_D497_CA39_3AE4 - 0x40, 0x5F, 0x87, 0xC7, 0xAF, 0x95, 0xBE, 0xDA, // k2' = 0xDABE_95AF_C787_5F40 +// pri const ECMA_X86_SSE42_K1K2 : roarray[16] base.u8 = [ +// 0x76, 0x5F, 0xDA, 0x32, 0x90, 0x40, 0x3B, 0xF8, // k1' = 0xF83B_4090_32DA_5F76 +// 0xF4, 0x2D, 0x84, 0xA7, 0x54, 0x60, 0x1F, 0x08, // k2' = 0x081F_6054_A784_2DF4 +// ] + +pri const ECMA_X86_SSE42_K3K4 : roarray[16] base.u8 = [ + 0xE4, 0x3A, 0x39, 0xCA, 0x97, 0xD4, 0x5D, 0xE0, // k3' = 0xE05D_D497_CA39_3AE4 + 0x40, 0x5F, 0x87, 0xC7, 0xAF, 0x95, 0xBE, 0xDA, // k4' = 0xDABE_95AF_C787_5F40 ] pri const ECMA_X86_SSE42_PXMU : roarray[16] base.u8 = [