- Dropped all vector code since new GPU's are all scalar, makes the code much easier

- Some performance on low-end GPU may drop because of that, but only for a few hash-modes
- Dropped scalar code (aka warp) since we do not have any vector datatypes anymore
- Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4
- Still need to fix kernels to new function names, needs to be done manually
- Temperature Management needs to be rewritten partially because of conflicting datatypes names
- Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id)
- Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD)
- First tests working for -m 0, for example
- Great performance increases in general for NV so far
- Tested amp_* and markov_* kernel
- Migrated special NV optimizations for rule processor

include/constants.h

@@ -35,10 +35,10 @@
  * SipHash Constants
  */
 
-#define SIPHASHM_0 0x736f6d6570736575ull
-#define SIPHASHM_1 0x646f72616e646f6dull
-#define SIPHASHM_2 0x6c7967656e657261ull
-#define SIPHASHM_3 0x7465646279746573ull
+#define SIPHASHM_0 0x736f6d6570736575
+#define SIPHASHM_1 0x646f72616e646f6d
+#define SIPHASHM_2 0x6c7967656e657261
+#define SIPHASHM_3 0x7465646279746573
 #endif
 
 #if defined _BCRYPT_ || defined _PSAFE2_
@@ -277,95 +277,95 @@
  * SHA384 Constants (64 bits)
  */
 
-#define SHA384M_A 0xcbbb9d5dc1059ed8ull
-#define SHA384M_B 0x629a292a367cd507ull
-#define SHA384M_C 0x9159015a3070dd17ull
-#define SHA384M_D 0x152fecd8f70e5939ull
-#define SHA384M_E 0x67332667ffc00b31ull
-#define SHA384M_F 0x8eb44a8768581511ull
-#define SHA384M_G 0xdb0c2e0d64f98fa7ull
-#define SHA384M_H 0x47b5481dbefa4fa4ull
+#define SHA384M_A 0xcbbb9d5dc1059ed8
+#define SHA384M_B 0x629a292a367cd507
+#define SHA384M_C 0x9159015a3070dd17
+#define SHA384M_D 0x152fecd8f70e5939
+#define SHA384M_E 0x67332667ffc00b31
+#define SHA384M_F 0x8eb44a8768581511
+#define SHA384M_G 0xdb0c2e0d64f98fa7
+#define SHA384M_H 0x47b5481dbefa4fa4
 
-#define SHA384C00 0x428a2f98d728ae22ull
-#define SHA384C01 0x7137449123ef65cdull
-#define SHA384C02 0xb5c0fbcfec4d3b2full
-#define SHA384C03 0xe9b5dba58189dbbcull
-#define SHA384C04 0x3956c25bf348b538ull
-#define SHA384C05 0x59f111f1b605d019ull
-#define SHA384C06 0x923f82a4af194f9bull
-#define SHA384C07 0xab1c5ed5da6d8118ull
-#define SHA384C08 0xd807aa98a3030242ull
-#define SHA384C09 0x12835b0145706fbeull
-#define SHA384C0a 0x243185be4ee4b28cull
-#define SHA384C0b 0x550c7dc3d5ffb4e2ull
-#define SHA384C0c 0x72be5d74f27b896full
-#define SHA384C0d 0x80deb1fe3b1696b1ull
-#define SHA384C0e 0x9bdc06a725c71235ull
-#define SHA384C0f 0xc19bf174cf692694ull
-#define SHA384C10 0xe49b69c19ef14ad2ull
-#define SHA384C11 0xefbe4786384f25e3ull
-#define SHA384C12 0x0fc19dc68b8cd5b5ull
-#define SHA384C13 0x240ca1cc77ac9c65ull
-#define SHA384C14 0x2de92c6f592b0275ull
-#define SHA384C15 0x4a7484aa6ea6e483ull
-#define SHA384C16 0x5cb0a9dcbd41fbd4ull
-#define SHA384C17 0x76f988da831153b5ull
-#define SHA384C18 0x983e5152ee66dfabull
-#define SHA384C19 0xa831c66d2db43210ull
-#define SHA384C1a 0xb00327c898fb213full
-#define SHA384C1b 0xbf597fc7beef0ee4ull
-#define SHA384C1c 0xc6e00bf33da88fc2ull
-#define SHA384C1d 0xd5a79147930aa725ull
-#define SHA384C1e 0x06ca6351e003826full
-#define SHA384C1f 0x142929670a0e6e70ull
-#define SHA384C20 0x27b70a8546d22ffcull
-#define SHA384C21 0x2e1b21385c26c926ull
-#define SHA384C22 0x4d2c6dfc5ac42aedull
-#define SHA384C23 0x53380d139d95b3dfull
-#define SHA384C24 0x650a73548baf63deull
-#define SHA384C25 0x766a0abb3c77b2a8ull
-#define SHA384C26 0x81c2c92e47edaee6ull
-#define SHA384C27 0x92722c851482353bull
-#define SHA384C28 0xa2bfe8a14cf10364ull
-#define SHA384C29 0xa81a664bbc423001ull
-#define SHA384C2a 0xc24b8b70d0f89791ull
-#define SHA384C2b 0xc76c51a30654be30ull
-#define SHA384C2c 0xd192e819d6ef5218ull
-#define SHA384C2d 0xd69906245565a910ull
-#define SHA384C2e 0xf40e35855771202aull
-#define SHA384C2f 0x106aa07032bbd1b8ull
-#define SHA384C30 0x19a4c116b8d2d0c8ull
-#define SHA384C31 0x1e376c085141ab53ull
-#define SHA384C32 0x2748774cdf8eeb99ull
-#define SHA384C33 0x34b0bcb5e19b48a8ull
-#define SHA384C34 0x391c0cb3c5c95a63ull
-#define SHA384C35 0x4ed8aa4ae3418acbull
-#define SHA384C36 0x5b9cca4f7763e373ull
-#define SHA384C37 0x682e6ff3d6b2b8a3ull
-#define SHA384C38 0x748f82ee5defb2fcull
-#define SHA384C39 0x78a5636f43172f60ull
-#define SHA384C3a 0x84c87814a1f0ab72ull
-#define SHA384C3b 0x8cc702081a6439ecull
-#define SHA384C3c 0x90befffa23631e28ull
-#define SHA384C3d 0xa4506cebde82bde9ull
-#define SHA384C3e 0xbef9a3f7b2c67915ull
-#define SHA384C3f 0xc67178f2e372532bull
-#define SHA384C40 0xca273eceea26619cull
-#define SHA384C41 0xd186b8c721c0c207ull
-#define SHA384C42 0xeada7dd6cde0eb1eull
-#define SHA384C43 0xf57d4f7fee6ed178ull
-#define SHA384C44 0x06f067aa72176fbaull
-#define SHA384C45 0x0a637dc5a2c898a6ull
-#define SHA384C46 0x113f9804bef90daeull
-#define SHA384C47 0x1b710b35131c471bull
-#define SHA384C48 0x28db77f523047d84ull
-#define SHA384C49 0x32caab7b40c72493ull
-#define SHA384C4a 0x3c9ebe0a15c9bebcull
-#define SHA384C4b 0x431d67c49c100d4cull
-#define SHA384C4c 0x4cc5d4becb3e42b6ull
-#define SHA384C4d 0x597f299cfc657e2aull
-#define SHA384C4e 0x5fcb6fab3ad6faecull
-#define SHA384C4f 0x6c44198c4a475817ull
+#define SHA384C00 0x428a2f98d728ae22
+#define SHA384C01 0x7137449123ef65cd
+#define SHA384C02 0xb5c0fbcfec4d3b2f
+#define SHA384C03 0xe9b5dba58189dbbc
+#define SHA384C04 0x3956c25bf348b538
+#define SHA384C05 0x59f111f1b605d019
+#define SHA384C06 0x923f82a4af194f9b
+#define SHA384C07 0xab1c5ed5da6d8118
+#define SHA384C08 0xd807aa98a3030242
+#define SHA384C09 0x12835b0145706fbe
+#define SHA384C0a 0x243185be4ee4b28c
+#define SHA384C0b 0x550c7dc3d5ffb4e2
+#define SHA384C0c 0x72be5d74f27b896f
+#define SHA384C0d 0x80deb1fe3b1696b1
+#define SHA384C0e 0x9bdc06a725c71235
+#define SHA384C0f 0xc19bf174cf692694
+#define SHA384C10 0xe49b69c19ef14ad2
+#define SHA384C11 0xefbe4786384f25e3
+#define SHA384C12 0x0fc19dc68b8cd5b5
+#define SHA384C13 0x240ca1cc77ac9c65
+#define SHA384C14 0x2de92c6f592b0275
+#define SHA384C15 0x4a7484aa6ea6e483
+#define SHA384C16 0x5cb0a9dcbd41fbd4
+#define SHA384C17 0x76f988da831153b5
+#define SHA384C18 0x983e5152ee66dfab
+#define SHA384C19 0xa831c66d2db43210
+#define SHA384C1a 0xb00327c898fb213f
+#define SHA384C1b 0xbf597fc7beef0ee4
+#define SHA384C1c 0xc6e00bf33da88fc2
+#define SHA384C1d 0xd5a79147930aa725
+#define SHA384C1e 0x06ca6351e003826f
+#define SHA384C1f 0x142929670a0e6e70
+#define SHA384C20 0x27b70a8546d22ffc
+#define SHA384C21 0x2e1b21385c26c926
+#define SHA384C22 0x4d2c6dfc5ac42aed
+#define SHA384C23 0x53380d139d95b3df
+#define SHA384C24 0x650a73548baf63de
+#define SHA384C25 0x766a0abb3c77b2a8
+#define SHA384C26 0x81c2c92e47edaee6
+#define SHA384C27 0x92722c851482353b
+#define SHA384C28 0xa2bfe8a14cf10364
+#define SHA384C29 0xa81a664bbc423001
+#define SHA384C2a 0xc24b8b70d0f89791
+#define SHA384C2b 0xc76c51a30654be30
+#define SHA384C2c 0xd192e819d6ef5218
+#define SHA384C2d 0xd69906245565a910
+#define SHA384C2e 0xf40e35855771202a
+#define SHA384C2f 0x106aa07032bbd1b8
+#define SHA384C30 0x19a4c116b8d2d0c8
+#define SHA384C31 0x1e376c085141ab53
+#define SHA384C32 0x2748774cdf8eeb99
+#define SHA384C33 0x34b0bcb5e19b48a8
+#define SHA384C34 0x391c0cb3c5c95a63
+#define SHA384C35 0x4ed8aa4ae3418acb
+#define SHA384C36 0x5b9cca4f7763e373
+#define SHA384C37 0x682e6ff3d6b2b8a3
+#define SHA384C38 0x748f82ee5defb2fc
+#define SHA384C39 0x78a5636f43172f60
+#define SHA384C3a 0x84c87814a1f0ab72
+#define SHA384C3b 0x8cc702081a6439ec
+#define SHA384C3c 0x90befffa23631e28
+#define SHA384C3d 0xa4506cebde82bde9
+#define SHA384C3e 0xbef9a3f7b2c67915
+#define SHA384C3f 0xc67178f2e372532b
+#define SHA384C40 0xca273eceea26619c
+#define SHA384C41 0xd186b8c721c0c207
+#define SHA384C42 0xeada7dd6cde0eb1e
+#define SHA384C43 0xf57d4f7fee6ed178
+#define SHA384C44 0x06f067aa72176fba
+#define SHA384C45 0x0a637dc5a2c898a6
+#define SHA384C46 0x113f9804bef90dae
+#define SHA384C47 0x1b710b35131c471b
+#define SHA384C48 0x28db77f523047d84
+#define SHA384C49 0x32caab7b40c72493
+#define SHA384C4a 0x3c9ebe0a15c9bebc
+#define SHA384C4b 0x431d67c49c100d4c
+#define SHA384C4c 0x4cc5d4becb3e42b6
+#define SHA384C4d 0x597f299cfc657e2a
+#define SHA384C4e 0x5fcb6fab3ad6faec
+#define SHA384C4f 0x6c44198c4a475817
 
 #endif
 
@@ -374,98 +374,98 @@
  * SHA512 Constants (64 bits)
  */
 
-#define SHA512M_A 0x6a09e667f3bcc908ull
-#define SHA512M_B 0xbb67ae8584caa73bull
-#define SHA512M_C 0x3c6ef372fe94f82bull
-#define SHA512M_D 0xa54ff53a5f1d36f1ull
-#define SHA512M_E 0x510e527fade682d1ull
-#define SHA512M_F 0x9b05688c2b3e6c1full
-#define SHA512M_G 0x1f83d9abfb41bd6bull
-#define SHA512M_H 0x5be0cd19137e2179ull
+#define SHA512M_A 0x6a09e667f3bcc908
+#define SHA512M_B 0xbb67ae8584caa73b
+#define SHA512M_C 0x3c6ef372fe94f82b
+#define SHA512M_D 0xa54ff53a5f1d36f1
+#define SHA512M_E 0x510e527fade682d1
+#define SHA512M_F 0x9b05688c2b3e6c1f
+#define SHA512M_G 0x1f83d9abfb41bd6b
+#define SHA512M_H 0x5be0cd19137e2179
 
-#define SHA512C00 0x428a2f98d728ae22ull
-#define SHA512C01 0x7137449123ef65cdull
-#define SHA512C02 0xb5c0fbcfec4d3b2full
-#define SHA512C03 0xe9b5dba58189dbbcull
-#define SHA512C04 0x3956c25bf348b538ull
-#define SHA512C05 0x59f111f1b605d019ull
-#define SHA512C06 0x923f82a4af194f9bull
-#define SHA512C07 0xab1c5ed5da6d8118ull
-#define SHA512C08 0xd807aa98a3030242ull
-#define SHA512C09 0x12835b0145706fbeull
-#define SHA512C0a 0x243185be4ee4b28cull
-#define SHA512C0b 0x550c7dc3d5ffb4e2ull
-#define SHA512C0c 0x72be5d74f27b896full
-#define SHA512C0d 0x80deb1fe3b1696b1ull
-#define SHA512C0e 0x9bdc06a725c71235ull
-#define SHA512C0f 0xc19bf174cf692694ull
-#define SHA512C10 0xe49b69c19ef14ad2ull
-#define SHA512C11 0xefbe4786384f25e3ull
-#define SHA512C12 0x0fc19dc68b8cd5b5ull
-#define SHA512C13 0x240ca1cc77ac9c65ull
-#define SHA512C14 0x2de92c6f592b0275ull
-#define SHA512C15 0x4a7484aa6ea6e483ull
-#define SHA512C16 0x5cb0a9dcbd41fbd4ull
-#define SHA512C17 0x76f988da831153b5ull
-#define SHA512C18 0x983e5152ee66dfabull
-#define SHA512C19 0xa831c66d2db43210ull
-#define SHA512C1a 0xb00327c898fb213full
-#define SHA512C1b 0xbf597fc7beef0ee4ull
-#define SHA512C1c 0xc6e00bf33da88fc2ull
-#define SHA512C1d 0xd5a79147930aa725ull
-#define SHA512C1e 0x06ca6351e003826full
-#define SHA512C1f 0x142929670a0e6e70ull
-#define SHA512C20 0x27b70a8546d22ffcull
-#define SHA512C21 0x2e1b21385c26c926ull
-#define SHA512C22 0x4d2c6dfc5ac42aedull
-#define SHA512C23 0x53380d139d95b3dfull
-#define SHA512C24 0x650a73548baf63deull
-#define SHA512C25 0x766a0abb3c77b2a8ull
-#define SHA512C26 0x81c2c92e47edaee6ull
-#define SHA512C27 0x92722c851482353bull
-#define SHA512C28 0xa2bfe8a14cf10364ull
-#define SHA512C29 0xa81a664bbc423001ull
-#define SHA512C2a 0xc24b8b70d0f89791ull
-#define SHA512C2b 0xc76c51a30654be30ull
-#define SHA512C2c 0xd192e819d6ef5218ull
-#define SHA512C2d 0xd69906245565a910ull
-#define SHA512C2e 0xf40e35855771202aull
-#define SHA512C2f 0x106aa07032bbd1b8ull
-#define SHA512C30 0x19a4c116b8d2d0c8ull
-#define SHA512C31 0x1e376c085141ab53ull
-#define SHA512C32 0x2748774cdf8eeb99ull
-#define SHA512C33 0x34b0bcb5e19b48a8ull
-#define SHA512C34 0x391c0cb3c5c95a63ull
-#define SHA512C35 0x4ed8aa4ae3418acbull
-#define SHA512C36 0x5b9cca4f7763e373ull
-#define SHA512C37 0x682e6ff3d6b2b8a3ull
-#define SHA512C38 0x748f82ee5defb2fcull
-#define SHA512C39 0x78a5636f43172f60ull
-#define SHA512C3a 0x84c87814a1f0ab72ull
-#define SHA512C3b 0x8cc702081a6439ecull
-#define SHA512C3c 0x90befffa23631e28ull
-#define SHA512C3d 0xa4506cebde82bde9ull
-#define SHA512C3e 0xbef9a3f7b2c67915ull
-#define SHA512C3f 0xc67178f2e372532bull
-#define SHA512C40 0xca273eceea26619cull
-#define SHA512C41 0xd186b8c721c0c207ull
-#define SHA512C42 0xeada7dd6cde0eb1eull
-#define SHA512C43 0xf57d4f7fee6ed178ull
-#define SHA512C44 0x06f067aa72176fbaull
-#define SHA512C45 0x0a637dc5a2c898a6ull
-#define SHA512C46 0x113f9804bef90daeull
-#define SHA512C47 0x1b710b35131c471bull
-#define SHA512C48 0x28db77f523047d84ull
-#define SHA512C49 0x32caab7b40c72493ull
-#define SHA512C4a 0x3c9ebe0a15c9bebcull
-#define SHA512C4b 0x431d67c49c100d4cull
-#define SHA512C4c 0x4cc5d4becb3e42b6ull
-#define SHA512C4d 0x597f299cfc657e2aull
-#define SHA512C4e 0x5fcb6fab3ad6faecull
-#define SHA512C4f 0x6c44198c4a475817ull
+#define SHA512C00 0x428a2f98d728ae22
+#define SHA512C01 0x7137449123ef65cd
+#define SHA512C02 0xb5c0fbcfec4d3b2f
+#define SHA512C03 0xe9b5dba58189dbbc
+#define SHA512C04 0x3956c25bf348b538
+#define SHA512C05 0x59f111f1b605d019
+#define SHA512C06 0x923f82a4af194f9b
+#define SHA512C07 0xab1c5ed5da6d8118
+#define SHA512C08 0xd807aa98a3030242
+#define SHA512C09 0x12835b0145706fbe
+#define SHA512C0a 0x243185be4ee4b28c
+#define SHA512C0b 0x550c7dc3d5ffb4e2
+#define SHA512C0c 0x72be5d74f27b896f
+#define SHA512C0d 0x80deb1fe3b1696b1
+#define SHA512C0e 0x9bdc06a725c71235
+#define SHA512C0f 0xc19bf174cf692694
+#define SHA512C10 0xe49b69c19ef14ad2
+#define SHA512C11 0xefbe4786384f25e3
+#define SHA512C12 0x0fc19dc68b8cd5b5
+#define SHA512C13 0x240ca1cc77ac9c65
+#define SHA512C14 0x2de92c6f592b0275
+#define SHA512C15 0x4a7484aa6ea6e483
+#define SHA512C16 0x5cb0a9dcbd41fbd4
+#define SHA512C17 0x76f988da831153b5
+#define SHA512C18 0x983e5152ee66dfab
+#define SHA512C19 0xa831c66d2db43210
+#define SHA512C1a 0xb00327c898fb213f
+#define SHA512C1b 0xbf597fc7beef0ee4
+#define SHA512C1c 0xc6e00bf33da88fc2
+#define SHA512C1d 0xd5a79147930aa725
+#define SHA512C1e 0x06ca6351e003826f
+#define SHA512C1f 0x142929670a0e6e70
+#define SHA512C20 0x27b70a8546d22ffc
+#define SHA512C21 0x2e1b21385c26c926
+#define SHA512C22 0x4d2c6dfc5ac42aed
+#define SHA512C23 0x53380d139d95b3df
+#define SHA512C24 0x650a73548baf63de
+#define SHA512C25 0x766a0abb3c77b2a8
+#define SHA512C26 0x81c2c92e47edaee6
+#define SHA512C27 0x92722c851482353b
+#define SHA512C28 0xa2bfe8a14cf10364
+#define SHA512C29 0xa81a664bbc423001
+#define SHA512C2a 0xc24b8b70d0f89791
+#define SHA512C2b 0xc76c51a30654be30
+#define SHA512C2c 0xd192e819d6ef5218
+#define SHA512C2d 0xd69906245565a910
+#define SHA512C2e 0xf40e35855771202a
+#define SHA512C2f 0x106aa07032bbd1b8
+#define SHA512C30 0x19a4c116b8d2d0c8
+#define SHA512C31 0x1e376c085141ab53
+#define SHA512C32 0x2748774cdf8eeb99
+#define SHA512C33 0x34b0bcb5e19b48a8
+#define SHA512C34 0x391c0cb3c5c95a63
+#define SHA512C35 0x4ed8aa4ae3418acb
+#define SHA512C36 0x5b9cca4f7763e373
+#define SHA512C37 0x682e6ff3d6b2b8a3
+#define SHA512C38 0x748f82ee5defb2fc
+#define SHA512C39 0x78a5636f43172f60
+#define SHA512C3a 0x84c87814a1f0ab72
+#define SHA512C3b 0x8cc702081a6439ec
+#define SHA512C3c 0x90befffa23631e28
+#define SHA512C3d 0xa4506cebde82bde9
+#define SHA512C3e 0xbef9a3f7b2c67915
+#define SHA512C3f 0xc67178f2e372532b
+#define SHA512C40 0xca273eceea26619c
+#define SHA512C41 0xd186b8c721c0c207
+#define SHA512C42 0xeada7dd6cde0eb1e
+#define SHA512C43 0xf57d4f7fee6ed178
+#define SHA512C44 0x06f067aa72176fba
+#define SHA512C45 0x0a637dc5a2c898a6
+#define SHA512C46 0x113f9804bef90dae
+#define SHA512C47 0x1b710b35131c471b
+#define SHA512C48 0x28db77f523047d84
+#define SHA512C49 0x32caab7b40c72493
+#define SHA512C4a 0x3c9ebe0a15c9bebc
+#define SHA512C4b 0x431d67c49c100d4c
+#define SHA512C4c 0x4cc5d4becb3e42b6
+#define SHA512C4d 0x597f299cfc657e2a
+#define SHA512C4e 0x5fcb6fab3ad6faec
+#define SHA512C4f 0x6c44198c4a475817
 
-#define SHA512REV0 0x5218a97a1b97e8a0ull
-#define SHA512REV1 0x4334c1bea164f555ull
+#define SHA512REV0 0x5218a97a1b97e8a0
+#define SHA512REV1 0x4334c1bea164f555
 
 #endif
 

include/ext_ADL.h

@@ -14,7 +14,7 @@ typedef int bool;
 
 #include <adl_sdk.h>
 
-typedef int HM_ADAPTER;
+//typedef int HM_ADAPTER;
 
 #ifdef _POSIX
 void *GetProcAddress (void *pLibrary, const char *name);

include/ext_OpenCL.h

@@ -24,8 +24,8 @@
 
 void hc_clBuildProgram (cl_program program, cl_uint num_devices, const cl_device_id *device_list, const char *options, void (CL_CALLBACK *pfn_notify) (cl_program program, void *user_data), void *user_data);
 cl_mem hc_clCreateBuffer (cl_context context, cl_mem_flags flags, size_t size, void *host_ptr);
-//cl_command_queue hc_clCreateCommandQueue (cl_context context, cl_device_id device, cl_command_queue_properties properties);
-cl_command_queue hc_clCreateCommandQueueWithProperties (cl_context context, cl_device_id device, const cl_queue_properties *properties);
+cl_command_queue hc_clCreateCommandQueue (cl_context context, cl_device_id device, cl_command_queue_properties properties);
+//cl_command_queue hc_clCreateCommandQueueWithProperties (cl_context context, cl_device_id device, const cl_queue_properties *properties);
 cl_context hc_clCreateContext (cl_context_properties *properties, cl_uint num_devices, const cl_device_id *devices, void (CL_CALLBACK *pfn_notify) (const char *, const void *, size_t, void *), void *user_data);
 cl_kernel hc_clCreateKernel (cl_program program, const char *kernel_name);
 cl_program hc_clCreateProgramWithSource (cl_context context, cl_uint count, const char **strings, const size_t *lengths);

include/ext_nvapi.h

@@ -54,7 +54,7 @@
 
 #include <nvapi.h>
 
-typedef NvPhysicalGpuHandle HM_ADAPTER;
+//typedef NvPhysicalGpuHandle HM_ADAPTER;
 
 int hc_NvAPI_EnumPhysicalGPUs (NvPhysicalGpuHandle nvGPUHandle[NVAPI_MAX_PHYSICAL_GPUS], NvU32 *pGpuCount);
 int hc_NvAPI_GPU_GetThermalSettings (NvPhysicalGpuHandle hPhysicalGpu, NvU32 sensorIndex, NV_GPU_THERMAL_SETTINGS *pThermalSettings);

include/ext_nvml.h

@@ -10,7 +10,7 @@
 
 #include <nvml.h>
 
-typedef nvmlDevice_t HM_ADAPTER;
+//typedef nvmlDevice_t HM_ADAPTER;
 
 nvmlReturn_t hc_NVML_nvmlInit (void);
 nvmlReturn_t hc_NVML_nvmlShutdown (void);

include/kernel_functions.c

@@ -233,7 +233,7 @@
 
 #define SHA384_STEP(F0,F1,a,b,c,d,e,f,g,h,x,K)  \
 {                                               \
-  u64x   temp0;                                 \
+  u64   temp0;                                 \
   temp0  = K;                                   \
   temp0 += x;                                   \
   temp0 += h;                                   \
@@ -273,7 +273,7 @@
 
 #define SHA512_STEP(F0,F1,a,b,c,d,e,f,g,h,x,K)  \
 {                                               \
-  u64x   temp0;                                 \
+  u64   temp0;                                 \
   temp0  = K;                                   \
   temp0 += x;                                   \
   temp0 += h;                                   \

include/kernel_vendor.h

@@ -3,15 +3,17 @@
  * License.....: MIT
  */
 
+#pragma OPENCL EXTENSION cl_khr_byte_addressable_store : enable
+
 /**
  * vendor specific
  */
 
-#ifdef __GPU__
+#if VENDOR_ID == 4096
 #define IS_AMD
 #endif
 
-#ifdef __CUDACC__
+#if VENDOR_ID == 4318
 #define IS_NV
 #endif
 
@@ -19,92 +21,14 @@
  * AMD specific
  */
 
-/*
 #ifdef IS_AMD
-#ifdef __ATI_RV710__
-#define VLIW1
-#elif __ATI_RV730__
-#define VLIW1
-#elif __ATI_RV770__
-#define VLIW4
-#elif __Barts__
-#define VLIW5
-#elif __BeaverCreek__
-#define VLIW5
-#elif __Caicos__
-#define VLIW5
-#elif __Capeverde__
-#define VLIW1
-#elif __Cayman__
-#define VLIW4
-#elif __Cedar__
-#define VLIW5
-#elif __Cypress__
-#define VLIW5
-#elif __Devastator__
-#define VLIW4
-#elif __Juniper__
-#define VLIW5
-#elif __Loveland__
-#define VLIW5
-#elif __Pitcairn__
-#define VLIW1
-#elif __Redwood__
-#define VLIW5
-#elif __Tahiti__
-#define VLIW1
-#elif __Turks__
-#define VLIW5
-#elif __Scrapper__
-#define VLIW4
-#elif __WinterPark__
-#define VLIW5
-#endif
-#endif
-*/
-
-#ifdef IS_AMD
-
-#pragma OPENCL EXTENSION cl_khr_byte_addressable_store : enable
-
-#ifdef OSX
-
-#else
-
-#ifdef cl_amd_media_ops
 #pragma OPENCL EXTENSION cl_amd_media_ops : enable
-#endif
-
-#ifdef cl_amd_media_ops2
 #pragma OPENCL EXTENSION cl_amd_media_ops2 : enable
 #endif
 
-#endif
-
-#endif
+/**
+ * NV specific
+ */
 
 #ifdef IS_NV
-#ifdef sm_10
-#define VLIW1
-#elif sm_11
-#define VLIW1
-#elif sm_12
-#define VLIW1
-#elif sm_13
-#define VLIW1
-#elif sm_20
-#define VLIW1
-#elif sm_21
-#define VLIW2
-#elif sm_30
-#define VLIW2
-#elif sm_35
-#define VLIW2
-#elif sm_37
-#define VLIW2
-#elif sm_50
-#define VLIW2
-#elif sm_52
-#define VLIW2
-#endif
 #endif

include/shared.h

@@ -54,39 +54,26 @@
 #define hc_sleep(x) sleep ((x));
 #endif
 
-#ifdef _CUDA
-#include <ext_cuda.h>
-#elif _OCL
 #include <ext_OpenCL.h>
-#endif
 
 /**
  * temperature management
  */
 
 #ifdef LINUX
-#ifdef _CUDA
 #include <ext_nvml.h>
-#elif _OCL
 #include <ext_ADL.h>
 #endif
-#endif
 
 #ifdef WIN
-#ifdef _CUDA
 #include <ext_nvapi.h>
-#elif _OCL
 #include <ext_ADL.h>
 #endif
-#endif
 
 #ifdef OSX
-#ifdef _CUDA
 #include <ext_smi.h>
-#elif _OCL
 #include <ext_dummy.h>
 #endif
-#endif
 
 /**
  * shared stuff
@@ -97,11 +84,16 @@
 #define DEVICES_MAX           128
 
 #define CL_PLATFORMS_MAX      16
+
 #define CL_VENDOR_NV          "NVIDIA Corporation"
 #define CL_VENDOR_AMD         "Advanced Micro Devices, Inc."
 #define CL_VENDOR_SDS         "Shiloh Distributed Solutions"
 #define CL_VENDOR_APPLE       "Apple"
 
+#define VENDOR_ID_AMD         4098
+#define VENDOR_ID_NV          4318
+#define VENDOR_ID_UNKNOWN     0
+
 #define BLOCK_SIZE            64
 
 #define CHARSIZ               0x100
@@ -118,9 +110,6 @@
 
 #define LOOPBACK_FILE         "loopback"
 
-#define VENDOR_ID_AMD         4098
-#define VENDOR_ID_NV          4318
-
 /**
  * types
  */
@@ -1769,7 +1758,6 @@ extern hc_thread_mutex_t mux_display;
 #define OPTI_TYPE_SINGLE_HASH       (1 << 11)
 #define OPTI_TYPE_SINGLE_SALT       (1 << 12)
 #define OPTI_TYPE_BRUTE_FORCE       (1 << 13)
-#define OPTI_TYPE_SCALAR_MODE       (1 << 14)
 #define OPTI_TYPE_RAW_HASH          (1 << 15)
 
 #define OPTI_STR_ZERO_BYTE          "Zero-Byte"
@@ -1785,7 +1773,6 @@ extern hc_thread_mutex_t mux_display;
 #define OPTI_STR_SINGLE_HASH        "Single-Hash"
 #define OPTI_STR_SINGLE_SALT        "Single-Salt"
 #define OPTI_STR_BRUTE_FORCE        "Brute-Force"
-#define OPTI_STR_SCALAR_MODE        "Scalar-Mode"
 #define OPTI_STR_RAW_HASH           "Raw-Hash"
 
 /**
@@ -2022,11 +2009,9 @@ void logfile_append (const char *fmt, ...);
 void fsync (int fd);
 #endif
 
-#ifdef _CUDA
+/*
 int hm_get_adapter_index (HM_ADAPTER nvGPUHandle[DEVICES_MAX]);
-#endif
 
-#ifdef _OCL
 int get_adapters_num  (HM_LIB hm_dll, int *iNumberAdapters);
 
 int hm_get_device_num (HM_LIB hm_dll, HM_ADAPTER hm_adapter_index, int *hm_device_num);
@@ -2045,13 +2030,13 @@ int hm_check_fanspeed_control (HM_LIB hm_dll, hm_attrs_t *hm_device, uint32_t *v
 void hm_close (HM_LIB hm_dll);
 
 HM_LIB hm_init ();
-#endif
 
 int hm_get_temperature_with_device_id (const uint device_id);
 int hm_get_fanspeed_with_device_id    (const uint device_id);
 int hm_get_utilization_with_device_id (const uint device_id);
 
 int hm_set_fanspeed_with_device_id (const uint device_id, const int fanspeed);
+*/
 
 void myabort ();
 void myquit ();

include/types.h

@@ -812,10 +812,12 @@ struct __hc_device_param
 {
   uint              device_id;
 
+  uint              sm_major;
+  uint              sm_minor;
+
   uint              gpu_processors;
   uint              gpu_threads;
   uint              gpu_accel;
-  uint              gpu_vector_width;
   uint64_t          gpu_maxmem_alloc;
   uint              gpu_power;          // these both are based on their _user counterpart
   uint              gpu_blocks;         // but are modified by autotuner and used inside crack loops
@@ -833,8 +835,6 @@ struct __hc_device_param
   uint              size_results;
   uint              size_plains;
 
-  uint              vect_size;
-
   uint (*pw_add)    (struct __hc_device_param *, const uint8_t *, const uint);
 
   void (*pw_transpose) (const pw_t *, pw_t *);
@@ -870,68 +870,6 @@ struct __hc_device_param
 
   // device specific attributes starting
 
-  #ifdef _CUDA
-
-  int               sm_major;
-  int               sm_minor;
-
-  CUdevice          device;
-
-  CUfunction        function1;
-  CUfunction        function12;
-  CUfunction        function2;
-  CUfunction        function23;
-  CUfunction        function3;
-  CUfunction        function_mp;
-  CUfunction        function_mp_l;
-  CUfunction        function_mp_r;
-  CUfunction        function_amp;
-  CUfunction        function_tb;
-  CUfunction        function_tm;
-
-  CUcontext         context;
-  CUmodule          module;
-  CUmodule          module_mp;
-  CUmodule          module_amp;
-  CUstream          stream;
-
-  CUdeviceptr       d_pws_buf;
-  CUdeviceptr       d_pws_amp_buf;
-  CUdeviceptr       d_words_buf_l;
-  CUdeviceptr       d_words_buf_r;
-  CUdeviceptr       c_words_buf_r;
-  CUdeviceptr       d_rules;
-  CUdeviceptr       c_rules;
-  CUdeviceptr       d_combs;
-  CUdeviceptr       c_combs;
-  CUdeviceptr       d_bfs;
-  CUdeviceptr       c_bfs;
-  CUdeviceptr       d_tm;
-  CUdeviceptr       c_tm;
-  size_t            c_bytes;
-  CUdeviceptr       d_bitmap_s1_a;
-  CUdeviceptr       d_bitmap_s1_b;
-  CUdeviceptr       d_bitmap_s1_c;
-  CUdeviceptr       d_bitmap_s1_d;
-  CUdeviceptr       d_bitmap_s2_a;
-  CUdeviceptr       d_bitmap_s2_b;
-  CUdeviceptr       d_bitmap_s2_c;
-  CUdeviceptr       d_bitmap_s2_d;
-  CUdeviceptr       d_plain_bufs;
-  CUdeviceptr       d_digests_buf;
-  CUdeviceptr       d_digests_shown;
-  CUdeviceptr       d_salt_bufs;
-  CUdeviceptr       d_esalt_bufs;
-  CUdeviceptr       d_bcrypt_bufs;
-  CUdeviceptr       d_tmps;
-  CUdeviceptr       d_hooks;
-  CUdeviceptr       d_result;
-  CUdeviceptr       d_scryptV_buf;
-  CUdeviceptr       d_root_css_buf;
-  CUdeviceptr       d_markov_css_buf;
-
-  #elif _OCL
-
   char             *device_name;
   char             *device_version;
   char             *driver_version;
@@ -990,8 +928,6 @@ struct __hc_device_param
   cl_mem            d_root_css_buf;
   cl_mem            d_markov_css_buf;
 
-  #endif
-
   #define PARAMCNT 32
 
   void             *kernel_params[PARAMCNT];
@@ -1021,11 +957,9 @@ typedef struct __hc_device_param hc_device_param_t;
 
 typedef struct
 {
-  HM_ADAPTER adapter_index;
+//  HM_ADAPTER adapter_index;
 
-  #ifdef _OCL
   int od_version;
-  #endif
 
   int fan_supported;
 
@@ -1040,6 +974,8 @@ typedef struct
    * threads
    */
 
+  uint                vendor_id;
+
   uint                devices_status;
   uint                devices_cnt;
   hc_device_param_t  *devices_param;