v3.23.2

JayDDee · Sep 21, 2023 · be88afc · be88afc
1 parent d6b5750
commit be88afc
Show file tree

Hide file tree

Showing 113 changed files with 3,354 additions and 2,925 deletions.
diff --git a/Makefile.am b/Makefile.am
@@ -163,16 +163,13 @@ cpuminer_SOURCES = \
   algo/sha/sph_sha2big.c \
   algo/sha/sha256-hash-4way.c \
   algo/sha/sha512-hash-4way.c \
-  algo/sha/sha256-hash-opt.c \
-  algo/sha/sha256-hash-2way-ni.c \
   algo/sha/hmac-sha256-hash.c \
   algo/sha/hmac-sha256-hash-4way.c \
   algo/sha/sha256d.c \
   algo/sha/sha2.c \
   algo/sha/sha256d-4way.c \
   algo/sha/sha256t-gate.c \
   algo/sha/sha256t-4way.c \
-  algo/sha/sha256t.c \
   algo/sha/sha256q-4way.c \
   algo/sha/sha256q.c \
   algo/sha/sha512256d-4way.c \
@@ -294,10 +291,10 @@ disable_flags =
 if USE_ASM
    cpuminer_SOURCES += asm/neoscrypt_asm.S
 if ARCH_x86
-   cpuminer_SOURCES += asm/sha2-x86.S asm/scrypt-x86.S asm/aesb-x86.S
+   cpuminer_SOURCES += asm/sha2-x86.S asm/scrypt-x86.S
 endif
 if ARCH_x86_64
-   cpuminer_SOURCES += asm/sha2-x64.S asm/scrypt-x64.S asm/aesb-x64.S
+   cpuminer_SOURCES += asm/sha2-x64.S asm/scrypt-x64.S
 endif
 if ARCH_ARM
    cpuminer_SOURCES += asm/sha2-arm.S asm/scrypt-arm.S

diff --git a/RELEASE_NOTES b/RELEASE_NOTES
@@ -65,6 +65,11 @@ If not what makes it happen or not happen?
 Change Log
 ----------
 
+v3.23.2
+
+sha256dt, sha256t & sha256d +10% with SHA, small improvement with AVX2.
+Other small improvements and code cleanup.
+
 v3.23.1
 
 #349: Fix sha256t low difficulty shares and low effective hash rate.

diff --git a/algo-gate-api.c b/algo-gate-api.c
@@ -248,7 +248,7 @@ int null_hash()
    return 0;
 };
 
-void init_algo_gate( algo_gate_t* gate )
+static void init_algo_gate( algo_gate_t* gate )
 {
    gate->miner_thread_init       = (void*)&return_true;
    gate->scanhash                = (void*)&scanhash_generic;

diff --git a/algo-gate-api.h b/algo-gate-api.h
@@ -269,7 +269,7 @@ void std_get_new_work( struct work *work, struct work *g_work, int thr_id,
 void sha256d_gen_merkle_root( char *merkle_root, struct stratum_ctx *sctx );
 void sha256_gen_merkle_root ( char *merkle_root, struct stratum_ctx *sctx );
 // OpenSSL sha256 deprecated
-void SHA256_gen_merkle_root ( char *merkle_root, struct stratum_ctx *sctx );
+//void SHA256_gen_merkle_root ( char *merkle_root, struct stratum_ctx *sctx );
 
 bool std_le_work_decode( struct work *work );
 bool std_be_work_decode( struct work *work );

diff --git a/algo/argon2/argon2a/argon2a.c b/algo/argon2/argon2a/argon2a.c
@@ -77,7 +77,7 @@ bool register_argon2_algo( algo_gate_t* gate )
   gate->optimizations = SSE2_OPT | AVX_OPT | AVX2_OPT;
   gate->scanhash        = (void*)&scanhash_argon2;
   gate->hash            = (void*)&argon2hash;
-  gate->gen_merkle_root = (void*)&SHA256_gen_merkle_root;
+  gate->gen_merkle_root = (void*)&sha256_gen_merkle_root;
   opt_target_factor = 65536.0;
 
   return true;

diff --git a/algo/blake/sph-blake2s.c b/algo/blake/sph-blake2s.c
@@ -15,7 +15,7 @@
 #include <string.h>
 #include <stdio.h>
 
-#include "algo/sha/sph_types.h"
+#include "compat/sph_types.h"
 #include "sph-blake2s.h"
 
 static const uint32_t blake2s_IV[8] =

diff --git a/algo/blake/sph_blake.h b/algo/blake/sph_blake.h
@@ -42,7 +42,7 @@ extern "C"{
 #endif
 
 #include <stddef.h>
-#include "algo/sha/sph_types.h"
+#include "compat/sph_types.h"
 
 /**
  * Output size (in bits) for BLAKE-224.

diff --git a/algo/blake/sph_blake2b.c b/algo/blake/sph_blake2b.c
@@ -31,7 +31,7 @@
 #include <stdint.h>
 #include <string.h>
 #include "simd-utils.h"
-#include "algo/sha/sph_types.h"
+#include "compat/sph_types.h"
 #include "sph_blake2b.h"
 
 // Little-endian byte access.

diff --git a/algo/bmw/bmw-hash-4way.h b/algo/bmw/bmw-hash-4way.h
@@ -41,8 +41,6 @@ extern "C"{
 #endif
 
 #include <stddef.h>
-
-#include "algo/sha/sph_types.h"
 #include "simd-utils.h"
 
 #define SPH_SIZE_bmw256   256
@@ -57,7 +55,7 @@ typedef struct {
    __m128i buf[64];
    __m128i H[16];
    size_t ptr;
-   sph_u32 bit_count;  // assume bit_count fits in 32 bits
+   uint32_t bit_count;  // assume bit_count fits in 32 bits
 } bmw_4way_small_context;
 
 typedef bmw_4way_small_context bmw256_4way_context;
@@ -144,7 +142,7 @@ typedef struct {
    __m256i buf[16];
    __m256i H[16];
    size_t ptr;
-   sph_u64 bit_count;
+   uint64_t bit_count;
 } bmw_4way_big_context __attribute__((aligned(128)));
 
 typedef bmw_4way_big_context bmw512_4way_context;

diff --git a/algo/bmw/bmw256-hash-4way.c b/algo/bmw/bmw256-hash-4way.c
@@ -109,7 +109,7 @@ static const uint32_t IV256[] = {
              _mm_sub_epi32( _mm_add_epi32( rol_off_32( M, j, 0 ), \
                                            rol_off_32( M, j, 3 ) ), \
                             rol_off_32( M, j, 10 ) ), \
-       _mm_set1_epi32( ( (j)+16 ) * SPH_C32(0x05555555UL) ) ), \
+       _mm_set1_epi32( ( (j)+16 ) * 0x05555555UL ) ), \
    H[ ( (j)+7 ) & 0xF ] )
 
 
@@ -485,7 +485,7 @@ bmw32_4way(bmw_4way_small_context *sc, const void *data, size_t len)
    size_t ptr;
    const int buf_size = 64;  // bytes of one lane, compatible with len
 
-   sc->bit_count += (sph_u32)len << 3;
+   sc->bit_count += (uint32_t)len << 3;
    buf = sc->buf;
    ptr = sc->ptr;
    h1 = sc->H;

diff --git a/algo/bmw/bmw512-hash-4way.c b/algo/bmw/bmw512-hash-4way.c
@@ -45,15 +45,15 @@ extern "C"{
 
 #define LPAR   (
 
-static const sph_u64 IV512[] = {
-        SPH_C64(0x8081828384858687), SPH_C64(0x88898A8B8C8D8E8F),
-        SPH_C64(0x9091929394959697), SPH_C64(0x98999A9B9C9D9E9F),
-        SPH_C64(0xA0A1A2A3A4A5A6A7), SPH_C64(0xA8A9AAABACADAEAF),
-        SPH_C64(0xB0B1B2B3B4B5B6B7), SPH_C64(0xB8B9BABBBCBDBEBF),
-        SPH_C64(0xC0C1C2C3C4C5C6C7), SPH_C64(0xC8C9CACBCCCDCECF),
-        SPH_C64(0xD0D1D2D3D4D5D6D7), SPH_C64(0xD8D9DADBDCDDDEDF),
-        SPH_C64(0xE0E1E2E3E4E5E6E7), SPH_C64(0xE8E9EAEBECEDEEEF),
-        SPH_C64(0xF0F1F2F3F4F5F6F7), SPH_C64(0xF8F9FAFBFCFDFEFF)
+static const uint64_t IV512[] = {
+        0x8081828384858687, 0x88898A8B8C8D8E8F,
+        0x9091929394959697, 0x98999A9B9C9D9E9F,
+        0xA0A1A2A3A4A5A6A7, 0xA8A9AAABACADAEAF,
+        0xB0B1B2B3B4B5B6B7, 0xB8B9BABBBCBDBEBF,
+        0xC0C1C2C3C4C5C6C7, 0xC8C9CACBCCCDCECF,
+        0xD0D1D2D3D4D5D6D7, 0xD8D9DADBDCDDDEDF,
+        0xE0E1E2E3E4E5E6E7, 0xE8E9EAEBECEDEEEF,
+        0xF0F1F2F3F4F5F6F7, 0xF8F9FAFBFCFDFEFF
 };
 
 #if defined(__SSE2__)
@@ -894,7 +894,7 @@ static const __m256i final_b[16] =
 };
 
 static void
-bmw64_4way_init( bmw_4way_big_context *sc, const sph_u64 *iv )
+bmw64_4way_init( bmw_4way_big_context *sc, const uint64_t *iv )
 {
    sc->H[ 0] = _mm256_set1_epi64x( 0x8081828384858687 );
    sc->H[ 1] = _mm256_set1_epi64x( 0x88898A8B8C8D8E8F );
@@ -926,7 +926,7 @@ bmw64_4way( bmw_4way_big_context *sc, const void *data, size_t len )
    size_t ptr;
    const int buf_size = 128;  // bytes of one lane, compatible with len
 
-   sc->bit_count += (sph_u64)len << 3;
+   sc->bit_count += (uint64_t)len << 3;
    buf = sc->buf;
    ptr = sc->ptr;
    h1 = sc->H;
@@ -1377,7 +1377,7 @@ static const __m512i final_b8[16] =
 
 
 void bmw512_8way_init( bmw512_8way_context *ctx )
-//bmw64_4way_init( bmw_4way_big_context *sc, const sph_u64 *iv )
+//bmw64_4way_init( bmw_4way_big_context *sc, const uint64_t *iv )
 {
    ctx->H[ 0] = _mm512_set1_epi64( 0x8081828384858687 );
    ctx->H[ 1] = _mm512_set1_epi64( 0x88898A8B8C8D8E8F );

diff --git a/algo/bmw/sph_bmw.h b/algo/bmw/sph_bmw.h
@@ -41,7 +41,7 @@ extern "C"{
 #endif
 
 #include <stddef.h>
-#include "algo/sha/sph_types.h"
+#include "compat/sph_types.h"
 
 /**
  * Output size (in bits) for BMW-224.

diff --git a/algo/cubehash/cubehash_sse2.c b/algo/cubehash/cubehash_sse2.c
@@ -9,7 +9,6 @@
 #include <immintrin.h>
 #endif
 #include "cubehash_sse2.h"
-#include "algo/sha/sha3-defs.h"
 #include <stdbool.h>
 #include <unistd.h>
 #include <memory.h>

diff --git a/algo/cubehash/cubehash_sse2.h b/algo/cubehash/cubehash_sse2.h
@@ -3,7 +3,7 @@
 
 #include "compat.h"
 #include <stdint.h>
-#include "algo/sha/sha3-defs.h"
+#include "compat/sha3-defs.h"
 
 #define	OPTIMIZE_SSE2
 

diff --git a/algo/cubehash/sph_cubehash.h b/algo/cubehash/sph_cubehash.h
@@ -42,7 +42,7 @@ extern "C"{
 #endif
 
 #include <stddef.h>
-#include "algo/sha/sph_types.h"
+#include "compat/sph_types.h"
 
 /**
  * Output size (in bits) for CubeHash-224.

diff --git a/algo/echo/aes_ni/hash_api.h b/algo/echo/aes_ni/hash_api.h
@@ -22,7 +22,7 @@
 #endif
 
 
-#include "algo/sha/sha3_common.h"
+#include "compat/sha3_common.h"
 
 #include <emmintrin.h>
 

diff --git a/algo/echo/sph_echo.c b/algo/echo/sph_echo.c
@@ -73,7 +73,7 @@ extern "C"{
 #endif
 
 #define AES_BIG_ENDIAN   0
-#include "algo/sha/aes_helper.c"
+#include "compat/aes_helper.c"
 
 #if SPH_ECHO_64
 

diff --git a/algo/echo/sph_echo.h b/algo/echo/sph_echo.h
@@ -43,7 +43,7 @@ extern "C"{
 #endif
 
 #include <stddef.h>
-#include "algo/sha/sph_types.h"
+#include "compat/sph_types.h"
 
 /**
  * Output size (in bits) for ECHO-224.

diff --git a/algo/fugue/fugue-aesni.h b/algo/fugue/fugue-aesni.h
@@ -20,7 +20,7 @@
 #error "Unsupported configuration, AES needs SSE4.1. Compile without AES."
 #endif
 
-#include "algo/sha/sha3_common.h"
+#include "compat/sha3_common.h"
 #include "simd-utils.h"
 
 

diff --git a/algo/fugue/sph_fugue.h b/algo/fugue/sph_fugue.h
@@ -2,7 +2,7 @@
 #define SPH_FUGUE_H__
 
 #include <stddef.h>
-#include "algo/sha/sph_types.h"
+#include "compat/sph_types.h"
 
 #ifdef __cplusplus
 extern "C"{

diff --git a/algo/gost/sph_gost.h b/algo/gost/sph_gost.h
@@ -41,7 +41,7 @@ extern "C"{
 #endif
 
 #include <stddef.h>
-#include "algo/sha/sph_types.h"
+#include "compat/sph_types.h"
 
 /**
  * Output size (in bits) for GOST-256.

diff --git a/algo/groestl/aes_ni/hash-groestl.h b/algo/groestl/aes_ni/hash-groestl.h
@@ -20,8 +20,8 @@
 #define LENGTH (512)
 
 #include "brg_endian.h"
-#define NEED_UINT_64T
-#include "algo/sha/brg_types.h"
+//#define NEED_UINT_64T
+#include "compat/brg_types.h"
 
 /* some sizes (number of bytes) */
 #define ROWS (8)

diff --git a/algo/groestl/aes_ni/hash-groestl256.h b/algo/groestl/aes_ni/hash-groestl256.h
@@ -34,8 +34,7 @@ typedef crypto_uint64 u64;
 //#define LENGTH (512)
 
 #include "brg_endian.h"
-#define NEED_UINT_64T
-#include "algo/sha/brg_types.h"
+#include "compat/brg_types.h"
 
 #ifdef IACA_TRACE
   #include IACA_MARKS

diff --git a/algo/groestl/groestl-gate.c b/algo/groestl/groestl-gate.c
@@ -17,7 +17,7 @@ bool register_dmd_gr_algo( algo_gate_t *gate )
 bool register_groestl_algo( algo_gate_t* gate )
 {
     register_dmd_gr_algo( gate );
-    gate->gen_merkle_root = (void*)&SHA256_gen_merkle_root;
+    gate->gen_merkle_root = (void*)&sha256_gen_merkle_root;
     return true;
 };
 
diff --git a/algo/groestl/groestl256-hash-4way.h b/algo/groestl/groestl256-hash-4way.h
@@ -22,10 +22,6 @@
 
 #define LENGTH (256)
 
-//#include "brg_endian.h"
-//#define NEED_UINT_64T
-//#include "algo/sha/brg_types.h"
-
 /* some sizes (number of bytes) */
 #define ROWS (8)
 #define LENGTHFIELDLEN (ROWS)

diff --git a/algo/groestl/myrgr-4way.c b/algo/groestl/myrgr-4way.c
@@ -4,7 +4,7 @@
 #include <stdint.h>
 #include <string.h>
 #include "aes_ni/hash-groestl.h"
-#include "algo/sha/sha-hash-4way.h"
+#include "algo/sha/sha256-hash.h"
 #if defined(__VAES__)
   #include "groestl512-hash-4way.h"
 #endif

diff --git a/algo/groestl/sph_groestl.h b/algo/groestl/sph_groestl.h
@@ -40,7 +40,7 @@ extern "C"{
 #endif
 
 #include <stddef.h>
-#include "algo/sha/sph_types.h"
+#include "compat/sph_types.h"
 
 #if !defined(__AES__)   
 /**
-Original file line number
+Diff line change
@@ Expand Up / @@ -22,7 +22,7 @@ @@
     #endif
-    #include "algo/sha/sha3_common.h"
+    #include "compat/sha3_common.h"
     #include <emmintrin.h>
@@ Expand Down @@