diff options
| author | fireice-uk <fireice-uk@users.noreply.github.com> | 2017-12-04 15:48:58 +0000 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2017-12-04 15:48:58 +0000 |
| commit | c6f56b24ec8c32de594afc2f99ee571b1485426e (patch) | |
| tree | 4b48cbace306b54a0455d98d744798d834a396a5 | |
| parent | a78b62d19756865879c54eac00c1c3f4d983d3a1 (diff) | |
| parent | efc8dc447adc8fcbf2c1faf0de2dacf8d38889ac (diff) | |
| download | xmr-stak-c6f56b24ec8c32de594afc2f99ee571b1485426e.zip xmr-stak-c6f56b24ec8c32de594afc2f99ee571b1485426e.tar.gz | |
Merge pull request #332 from psychocrypt/topic-tuneAMD
optimize aes round command queue
| -rw-r--r-- | xmrstak/backend/amd/amd_gpu/opencl/wolf-aes.cl | 30 |
1 files changed, 22 insertions, 8 deletions
diff --git a/xmrstak/backend/amd/amd_gpu/opencl/wolf-aes.cl b/xmrstak/backend/amd/amd_gpu/opencl/wolf-aes.cl index 996944b..81e1644 100644 --- a/xmrstak/backend/amd/amd_gpu/opencl/wolf-aes.cl +++ b/xmrstak/backend/amd/amd_gpu/opencl/wolf-aes.cl @@ -74,15 +74,29 @@ static const __constant uint AES0_C[256] = #define BYTE(x, y) (amd_bfe((x), (y) << 3U, 8U)) -uint4 AES_Round(const __local uint *AES0, const __local uint *AES1, const __local uint *AES2, const __local uint *AES3, const uint4 X, const uint4 key) +uint4 AES_Round(const __local uint *AES0, const __local uint *AES1, const __local uint *AES2, const __local uint *AES3, const uint4 X, uint4 key) { - uint4 Y; - Y.s0 = AES0[BYTE(X.s0, 0)] ^ AES1[BYTE(X.s1, 1)] ^ AES2[BYTE(X.s2, 2)] ^ AES3[BYTE(X.s3, 3)]; - Y.s1 = AES0[BYTE(X.s1, 0)] ^ AES1[BYTE(X.s2, 1)] ^ AES2[BYTE(X.s3, 2)] ^ AES3[BYTE(X.s0, 3)]; - Y.s2 = AES0[BYTE(X.s2, 0)] ^ AES1[BYTE(X.s3, 1)] ^ AES2[BYTE(X.s0, 2)] ^ AES3[BYTE(X.s1, 3)]; - Y.s3 = AES0[BYTE(X.s3, 0)] ^ AES1[BYTE(X.s0, 1)] ^ AES2[BYTE(X.s1, 2)] ^ AES3[BYTE(X.s2, 3)]; - Y ^= key; - return(Y); + key.s0 ^= AES0[BYTE(X.s0, 0)]; + key.s1 ^= AES0[BYTE(X.s1, 0)]; + key.s2 ^= AES0[BYTE(X.s2, 0)]; + key.s3 ^= AES0[BYTE(X.s3, 0)]; + + key.s0 ^= AES2[BYTE(X.s2, 2)]; + key.s1 ^= AES2[BYTE(X.s3, 2)]; + key.s2 ^= AES2[BYTE(X.s0, 2)]; + key.s3 ^= AES2[BYTE(X.s1, 2)]; + + key.s0 ^= AES1[BYTE(X.s1, 1)]; + key.s1 ^= AES1[BYTE(X.s2, 1)]; + key.s2 ^= AES1[BYTE(X.s3, 1)]; + key.s3 ^= AES1[BYTE(X.s0, 1)]; + + key.s0 ^= AES3[BYTE(X.s3, 3)]; + key.s1 ^= AES3[BYTE(X.s0, 3)]; + key.s2 ^= AES3[BYTE(X.s1, 3)]; + key.s3 ^= AES3[BYTE(X.s2, 3)]; + + return key; } #endif |
