|
| 1 | +unit HlpXXHash3Dispatch; |
| 2 | + |
| 3 | +{$I ..\Include\HashLib.inc} |
| 4 | + |
| 5 | +interface |
| 6 | + |
| 7 | +type |
| 8 | + TXXH3Accumulate512Proc = procedure(AAcc: Pointer; AInput: Pointer; |
| 9 | + ASecret: Pointer); |
| 10 | + TXXH3AccumulateProc = procedure(AAcc: Pointer; AInput: Pointer; |
| 11 | + ASecret: Pointer; ANbStripes: Int32); |
| 12 | + TXXH3ScrambleAccProc = procedure(AAcc: Pointer; ASecret: Pointer); |
| 13 | + TXXH3InitSecretProc = procedure(ACustomSecret: Pointer; |
| 14 | + ADefaultSecret: Pointer; ASeed: UInt64); |
| 15 | + |
| 16 | +var |
| 17 | + XXH3_Accumulate512: TXXH3Accumulate512Proc; |
| 18 | + XXH3_Accumulate: TXXH3AccumulateProc; |
| 19 | + XXH3_ScrambleAcc: TXXH3ScrambleAccProc; |
| 20 | + XXH3_InitSecret: TXXH3InitSecretProc; |
| 21 | + |
| 22 | +implementation |
| 23 | + |
| 24 | +uses |
| 25 | + HlpSimd; |
| 26 | + |
| 27 | +const |
| 28 | + XXH_STRIPE_LEN = 64; |
| 29 | + XXH_ACC_NB = 8; |
| 30 | + XXH_SECRET_CONSUME_RATE = 8; |
| 31 | + XXH_PRIME32_1 = UInt32($9E3779B1); |
| 32 | + |
| 33 | +// ============================================================================= |
| 34 | +// Scalar fallback implementations |
| 35 | +// ============================================================================= |
| 36 | + |
| 37 | +procedure XXH3_accumulate_512_scalar(AAcc: Pointer; AInput: Pointer; |
| 38 | + ASecret: Pointer); |
| 39 | +var |
| 40 | + LPAcc: PUInt64; |
| 41 | + LPInput, LPSecret: PByte; |
| 42 | + I: Int32; |
| 43 | + LDataVal, LDataKey: UInt64; |
| 44 | +begin |
| 45 | + LPAcc := PUInt64(AAcc); |
| 46 | + LPInput := PByte(AInput); |
| 47 | + LPSecret := PByte(ASecret); |
| 48 | + for I := 0 to XXH_ACC_NB - 1 do |
| 49 | + begin |
| 50 | + LDataVal := PUInt64(LPInput + I * 8)^; |
| 51 | + LDataKey := LDataVal xor PUInt64(LPSecret + I * 8)^; |
| 52 | + PUInt64(PByte(LPAcc) + (I xor 1) * 8)^ := |
| 53 | + PUInt64(PByte(LPAcc) + (I xor 1) * 8)^ + LDataVal; |
| 54 | + PUInt64(PByte(LPAcc) + I * 8)^ := |
| 55 | + PUInt64(PByte(LPAcc) + I * 8)^ + |
| 56 | + UInt64(UInt32(LDataKey)) * UInt64(UInt32(LDataKey shr 32)); |
| 57 | + end; |
| 58 | +end; |
| 59 | + |
| 60 | +procedure XXH3_scrambleAcc_scalar(AAcc: Pointer; ASecret: Pointer); |
| 61 | +var |
| 62 | + LPAcc: PUInt64; |
| 63 | + LPSecret: PByte; |
| 64 | + I: Int32; |
| 65 | + LKey64, LAcc64: UInt64; |
| 66 | +begin |
| 67 | + LPAcc := PUInt64(AAcc); |
| 68 | + LPSecret := PByte(ASecret); |
| 69 | + for I := 0 to XXH_ACC_NB - 1 do |
| 70 | + begin |
| 71 | + LKey64 := PUInt64(LPSecret + I * 8)^; |
| 72 | + LAcc64 := PUInt64(PByte(LPAcc) + I * 8)^; |
| 73 | + LAcc64 := LAcc64 xor (LAcc64 shr 47); |
| 74 | + LAcc64 := LAcc64 xor LKey64; |
| 75 | + LAcc64 := LAcc64 * XXH_PRIME32_1; |
| 76 | + PUInt64(PByte(LPAcc) + I * 8)^ := LAcc64; |
| 77 | + end; |
| 78 | +end; |
| 79 | + |
| 80 | +procedure XXH3_initSecret_scalar(ACustomSecret: Pointer; |
| 81 | + ADefaultSecret: Pointer; ASeed: UInt64); |
| 82 | +var |
| 83 | + I: Int32; |
| 84 | + LPSrc, LPDst: PByte; |
| 85 | +begin |
| 86 | + LPSrc := PByte(ADefaultSecret); |
| 87 | + LPDst := PByte(ACustomSecret); |
| 88 | + for I := 0 to (192 div 16) - 1 do |
| 89 | + begin |
| 90 | + PUInt64(LPDst + 16 * I)^ := PUInt64(LPSrc + 16 * I)^ + ASeed; |
| 91 | + PUInt64(LPDst + 16 * I + 8)^ := PUInt64(LPSrc + 16 * I + 8)^ - ASeed; |
| 92 | + end; |
| 93 | +end; |
| 94 | + |
| 95 | +procedure XXH3_accumulate_scalar(AAcc: Pointer; AInput: Pointer; |
| 96 | + ASecret: Pointer; ANbStripes: Int32); |
| 97 | +var |
| 98 | + N: Int32; |
| 99 | +begin |
| 100 | + for N := 0 to ANbStripes - 1 do |
| 101 | + XXH3_accumulate_512_scalar(AAcc, PByte(AInput) + N * XXH_STRIPE_LEN, |
| 102 | + PByte(ASecret) + N * XXH_SECRET_CONSUME_RATE); |
| 103 | +end; |
| 104 | + |
| 105 | +// ============================================================================= |
| 106 | +// SSE2 and AVX2 implementations (x86-64 only) |
| 107 | +// ============================================================================= |
| 108 | + |
| 109 | +{$IFDEF HASHLIB_X86_64} |
| 110 | + |
| 111 | +// ----- SSE2 ----- |
| 112 | + |
| 113 | +procedure XXH3_accumulate_512_sse2(AAcc: Pointer; AInput: Pointer; |
| 114 | + ASecret: Pointer); |
| 115 | + {$I ..\Include\Simd\Common\SimdProc3Begin.inc} |
| 116 | + {$I ..\Include\Simd\XXH3\XXH3Acc512Sse2.inc} |
| 117 | +end; |
| 118 | + |
| 119 | +procedure XXH3_scrambleAcc_sse2(AAcc: Pointer; ASecret: Pointer); |
| 120 | + {$I ..\Include\Simd\Common\SimdProc2Begin.inc} |
| 121 | + {$I ..\Include\Simd\XXH3\XXH3ScrambleSse2.inc} |
| 122 | +end; |
| 123 | + |
| 124 | +procedure XXH3_initSecret_sse2(ACustomSecret: Pointer; |
| 125 | + ADefaultSecret: Pointer; ASeed: UInt64); |
| 126 | + {$I ..\Include\Simd\Common\SimdProc3Begin.inc} |
| 127 | + {$I ..\Include\Simd\XXH3\XXH3InitSecretSse2.inc} |
| 128 | +end; |
| 129 | + |
| 130 | +procedure XXH3_accumulate_sse2(AAcc: Pointer; AInput: Pointer; |
| 131 | + ASecret: Pointer; ANbStripes: Int32); |
| 132 | +var |
| 133 | + N: Int32; |
| 134 | +begin |
| 135 | + for N := 0 to ANbStripes - 1 do |
| 136 | + XXH3_accumulate_512_sse2(AAcc, PByte(AInput) + N * XXH_STRIPE_LEN, |
| 137 | + PByte(ASecret) + N * XXH_SECRET_CONSUME_RATE); |
| 138 | +end; |
| 139 | + |
| 140 | +{$IFDEF HASHLIB_AVX2_ASM_SUPPORTED} |
| 141 | + |
| 142 | +// ----- AVX2 ----- |
| 143 | + |
| 144 | +procedure XXH3_accumulate_512_avx2(AAcc: Pointer; AInput: Pointer; |
| 145 | + ASecret: Pointer); |
| 146 | + {$I ..\Include\Simd\Common\SimdProc3Begin.inc} |
| 147 | + {$I ..\Include\Simd\XXH3\XXH3Acc512Avx2.inc} |
| 148 | +end; |
| 149 | + |
| 150 | +procedure XXH3_scrambleAcc_avx2(AAcc: Pointer; ASecret: Pointer); |
| 151 | + {$I ..\Include\Simd\Common\SimdProc2Begin.inc} |
| 152 | + {$I ..\Include\Simd\XXH3\XXH3ScrambleAvx2.inc} |
| 153 | +end; |
| 154 | + |
| 155 | +procedure XXH3_initSecret_avx2(ACustomSecret: Pointer; |
| 156 | + ADefaultSecret: Pointer; ASeed: UInt64); |
| 157 | + {$I ..\Include\Simd\Common\SimdProc3Begin.inc} |
| 158 | + {$I ..\Include\Simd\XXH3\XXH3InitSecretAvx2.inc} |
| 159 | +end; |
| 160 | + |
| 161 | +procedure XXH3_accumulate_avx2(AAcc: Pointer; AInput: Pointer; |
| 162 | + ASecret: Pointer; ANbStripes: Int32); |
| 163 | +var |
| 164 | + N: Int32; |
| 165 | +begin |
| 166 | + for N := 0 to ANbStripes - 1 do |
| 167 | + XXH3_accumulate_512_avx2(AAcc, PByte(AInput) + N * XXH_STRIPE_LEN, |
| 168 | + PByte(ASecret) + N * XXH_SECRET_CONSUME_RATE); |
| 169 | +end; |
| 170 | + |
| 171 | +{$ENDIF HASHLIB_AVX2_ASM_SUPPORTED} |
| 172 | + |
| 173 | +{$ENDIF HASHLIB_X86_64} |
| 174 | + |
| 175 | +// ============================================================================= |
| 176 | +// Dispatch initialization |
| 177 | +// ============================================================================= |
| 178 | + |
| 179 | +procedure InitDispatch(); |
| 180 | +begin |
| 181 | + case TSimd.GetActiveLevel() of |
| 182 | +{$IFDEF HASHLIB_X86_64} |
| 183 | + {$IFDEF HASHLIB_AVX2_ASM_SUPPORTED} |
| 184 | + TSimdLevel.AVX2: |
| 185 | + begin |
| 186 | + XXH3_Accumulate512 := @XXH3_accumulate_512_avx2; |
| 187 | + XXH3_Accumulate := @XXH3_accumulate_avx2; |
| 188 | + XXH3_ScrambleAcc := @XXH3_scrambleAcc_avx2; |
| 189 | + XXH3_InitSecret := @XXH3_initSecret_avx2; |
| 190 | + end; |
| 191 | + {$ENDIF HASHLIB_AVX2_ASM_SUPPORTED} |
| 192 | + TSimdLevel.SSE2: |
| 193 | + begin |
| 194 | + XXH3_Accumulate512 := @XXH3_accumulate_512_sse2; |
| 195 | + XXH3_Accumulate := @XXH3_accumulate_sse2; |
| 196 | + XXH3_ScrambleAcc := @XXH3_scrambleAcc_sse2; |
| 197 | + XXH3_InitSecret := @XXH3_initSecret_sse2; |
| 198 | + end; |
| 199 | +{$ENDIF} |
| 200 | + TSimdLevel.Scalar: |
| 201 | + begin |
| 202 | + XXH3_Accumulate512 := @XXH3_accumulate_512_scalar; |
| 203 | + XXH3_Accumulate := @XXH3_accumulate_scalar; |
| 204 | + XXH3_ScrambleAcc := @XXH3_scrambleAcc_scalar; |
| 205 | + XXH3_InitSecret := @XXH3_initSecret_scalar; |
| 206 | + end; |
| 207 | + end; |
| 208 | +end; |
| 209 | + |
| 210 | +initialization |
| 211 | + InitDispatch(); |
| 212 | + |
| 213 | +end. |
0 commit comments