From 4a768458d2335fc3594f7a7f09bb4066442bed3f Mon Sep 17 00:00:00 2001 From: merlokk Date: Tue, 2 Jan 2018 19:05:12 +0200 Subject: [PATCH] hardnested SIMD select --- client/cmdhfmfhard.c | 40 ++++++---- client/hardnested/hardnested_bf_core.c | 100 ++++++++++++++++++------- client/hardnested/hardnested_bf_core.h | 12 +++ 3 files changed, 109 insertions(+), 43 deletions(-) diff --git a/client/cmdhfmfhard.c b/client/cmdhfmfhard.c index 0153541e..cbdf3f60 100644 --- a/client/cmdhfmfhard.c +++ b/client/cmdhfmfhard.c @@ -32,6 +32,7 @@ #include "crapto1/crapto1.h" #include "parity.h" #include "hardnested/hardnested_bruteforce.h" +#include "hardnested/hardnested_bf_core.h" #include "hardnested/hardnested_bitarray_core.h" #include "zlib.h" @@ -71,27 +72,32 @@ static float brute_force_per_second; static void get_SIMD_instruction_set(char* instruction_set) { -#if defined (__i386__) || defined (__x86_64__) - #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8 || __clang_major__ == 8 && __clang_minor__ >= 1)) - #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2) - if (__builtin_cpu_supports("avx512f")) strcpy(instruction_set, "AVX512F"); - else if (__builtin_cpu_supports("avx2")) strcpy(instruction_set, "AVX2"); - #else - if (__builtin_cpu_supports("avx2")) strcpy(instruction_set, "AVX2"); - #endif - else if (__builtin_cpu_supports("avx")) strcpy(instruction_set, "AVX"); - else if (__builtin_cpu_supports("sse2")) strcpy(instruction_set, "SSE2"); - else if (__builtin_cpu_supports("mmx")) strcpy(instruction_set, "MMX"); - else - #endif -#endif - strcpy(instruction_set, "no"); + switch(GetSIMDInstrAuto()) { + case SIMD_AVX512: + strcpy(instruction_set, "AVX512F"); + break; + case SIMD_AVX2: + strcpy(instruction_set, "AVX2"); + break; + case SIMD_AVX: + strcpy(instruction_set, "AVX"); + break; + case SIMD_SSE2: + strcpy(instruction_set, "SSE2"); + break; + case SIMD_MMX: + strcpy(instruction_set, "MMX"); + break; + default: + strcpy(instruction_set, "no"); + break; + } } static void print_progress_header(void) { char progress_text[80]; - char instr_set[12] = ""; + char instr_set[12] = {0}; get_SIMD_instruction_set(instr_set); sprintf(progress_text, "Start using %d threads and %s SIMD core", num_CPUs(), instr_set); PrintAndLog("\n\n"); @@ -2528,6 +2534,8 @@ static void set_test_state(uint8_t byte) int mfnestedhard(uint8_t blockNo, uint8_t keyType, uint8_t *key, uint8_t trgBlockNo, uint8_t trgKeyType, uint8_t *trgkey, bool nonce_file_read, bool nonce_file_write, bool slow, int tests) { char progress_text[80]; + + SetSIMDInstr(SIMD_NONE); srand((unsigned) time(NULL)); brute_force_per_second = brute_force_benchmark(); diff --git a/client/hardnested/hardnested_bf_core.c b/client/hardnested/hardnested_bf_core.c index 3c0c044f..d716f702 100644 --- a/client/hardnested/hardnested_bf_core.c +++ b/client/hardnested/hardnested_bf_core.c @@ -544,48 +544,94 @@ out: #ifndef __MMX__ -// pointers to functions: -crack_states_bitsliced_t *crack_states_bitsliced_function_p = &crack_states_bitsliced_dispatch; -bitslice_test_nonces_t *bitslice_test_nonces_function_p = &bitslice_test_nonces_dispatch; +static SIMDExecInstr intSIMDInstr = SIMD_AUTO; -// determine the available instruction set at runtime and call the correct function -const uint64_t crack_states_bitsliced_dispatch(uint32_t cuid, uint8_t *best_first_bytes, statelist_t *p, uint32_t *keys_found, uint64_t *num_keys_tested, uint32_t nonces_to_bruteforce, uint8_t *bf_test_nonce_2nd_byte, noncelist_t *nonces) { +void SetSIMDInstr(SIMDExecInstr instr) { + intSIMDInstr = instr; +} + +SIMDExecInstr GetSIMDInstr() { + SIMDExecInstr instr = SIMD_NONE; + #if defined (__i386__) || defined (__x86_64__) #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8 || __clang_major__ == 8 && __clang_minor__ >= 1)) #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2) - if (__builtin_cpu_supports("avx512f")) crack_states_bitsliced_function_p = &crack_states_bitsliced_AVX512; - else if (__builtin_cpu_supports("avx2")) crack_states_bitsliced_function_p = &crack_states_bitsliced_AVX2; + if (__builtin_cpu_supports("avx512f")) instr = SIMD_AVX512; + else if (__builtin_cpu_supports("avx2")) instr = SIMD_AVX2; #else - if (__builtin_cpu_supports("avx2")) crack_states_bitsliced_function_p = &crack_states_bitsliced_AVX2; + if (__builtin_cpu_supports("avx2")) instr = SIMD_AVX2; #endif - else if (__builtin_cpu_supports("avx")) crack_states_bitsliced_function_p = &crack_states_bitsliced_AVX; - else if (__builtin_cpu_supports("sse2")) crack_states_bitsliced_function_p = &crack_states_bitsliced_SSE2; - else if (__builtin_cpu_supports("mmx")) crack_states_bitsliced_function_p = &crack_states_bitsliced_MMX; + else if (__builtin_cpu_supports("avx")) instr = SIMD_AVX; + else if (__builtin_cpu_supports("sse2")) instr = SIMD_SSE2; + else if (__builtin_cpu_supports("mmx")) instr = SIMD_MMX; else #endif #endif - crack_states_bitsliced_function_p = &crack_states_bitsliced_NOSIMD; + instr = SIMD_NONE; + + return instr; +} + +SIMDExecInstr GetSIMDInstrAuto() { + SIMDExecInstr instr = intSIMDInstr; + if (instr == SIMD_AUTO) + return GetSIMDInstr(); + + return instr; +} + +// pointers to functions: +crack_states_bitsliced_t *crack_states_bitsliced_function_p = &crack_states_bitsliced_dispatch; +bitslice_test_nonces_t *bitslice_test_nonces_function_p = &bitslice_test_nonces_dispatch; + +// determine the available instruction set at runtime and call the correct function +const uint64_t crack_states_bitsliced_dispatch(uint32_t cuid, uint8_t *best_first_bytes, statelist_t *p, uint32_t *keys_found, uint64_t *num_keys_tested, uint32_t nonces_to_bruteforce, uint8_t *bf_test_nonce_2nd_byte, noncelist_t *nonces) { + switch(GetSIMDInstrAuto()) { + case SIMD_AVX512: + crack_states_bitsliced_function_p = &crack_states_bitsliced_AVX512; + break; + case SIMD_AVX2: + crack_states_bitsliced_function_p = &crack_states_bitsliced_AVX2; + break; + case SIMD_AVX: + crack_states_bitsliced_function_p = &crack_states_bitsliced_AVX; + break; + case SIMD_SSE2: + crack_states_bitsliced_function_p = &crack_states_bitsliced_SSE2; + break; + case SIMD_MMX: + crack_states_bitsliced_function_p = &crack_states_bitsliced_MMX; + break; + default: + crack_states_bitsliced_function_p = &crack_states_bitsliced_NOSIMD; + break; + } // call the most optimized function for this CPU return (*crack_states_bitsliced_function_p)(cuid, best_first_bytes, p, keys_found, num_keys_tested, nonces_to_bruteforce, bf_test_nonce_2nd_byte, nonces); } void bitslice_test_nonces_dispatch(uint32_t nonces_to_bruteforce, uint32_t *bf_test_nonce, uint8_t *bf_test_nonce_par) { -#if defined (__i386__) || defined (__x86_64__) - #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8 || __clang_major__ == 8 && __clang_minor__ >= 1)) - #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2) - if (__builtin_cpu_supports("avx512f")) bitslice_test_nonces_function_p = &bitslice_test_nonces_AVX512; - else if (__builtin_cpu_supports("avx2")) bitslice_test_nonces_function_p = &bitslice_test_nonces_AVX2; - #else - if (__builtin_cpu_supports("avx2")) bitslice_test_nonces_function_p = &bitslice_test_nonces_AVX2; - #endif - else if (__builtin_cpu_supports("avx")) bitslice_test_nonces_function_p = &bitslice_test_nonces_AVX; - else if (__builtin_cpu_supports("sse2")) bitslice_test_nonces_function_p = &bitslice_test_nonces_SSE2; - else if (__builtin_cpu_supports("mmx")) bitslice_test_nonces_function_p = &bitslice_test_nonces_MMX; - else - #endif -#endif - bitslice_test_nonces_function_p = &bitslice_test_nonces_NOSIMD; + switch(GetSIMDInstrAuto()) { + case SIMD_AVX512: + bitslice_test_nonces_function_p = &bitslice_test_nonces_AVX512; + break; + case SIMD_AVX2: + bitslice_test_nonces_function_p = &bitslice_test_nonces_AVX2; + break; + case SIMD_AVX: + bitslice_test_nonces_function_p = &bitslice_test_nonces_AVX; + break; + case SIMD_SSE2: + bitslice_test_nonces_function_p = &bitslice_test_nonces_SSE2; + break; + case SIMD_MMX: + bitslice_test_nonces_function_p = &bitslice_test_nonces_MMX; + break; + default: + bitslice_test_nonces_function_p = &bitslice_test_nonces_NOSIMD; + break; + } // call the most optimized function for this CPU (*bitslice_test_nonces_function_p)(nonces_to_bruteforce, bf_test_nonce, bf_test_nonce_par); diff --git a/client/hardnested/hardnested_bf_core.h b/client/hardnested/hardnested_bf_core.h index 7a445993..b3df0547 100644 --- a/client/hardnested/hardnested_bf_core.h +++ b/client/hardnested/hardnested_bf_core.h @@ -52,6 +52,18 @@ THE SOFTWARE. #include "hardnested_bruteforce.h" // statelist_t +typedef enum { + SIMD_AUTO, + SIMD_AVX512, + SIMD_AVX2, + SIMD_AVX, + SIMD_SSE2, + SIMD_MMX, + SIMD_NONE, +} SIMDExecInstr; +extern void SetSIMDInstr(SIMDExecInstr instr); +extern SIMDExecInstr GetSIMDInstrAuto(); + extern const uint64_t crack_states_bitsliced(uint32_t cuid, uint8_t *best_first_bytes, statelist_t *p, uint32_t *keys_found, uint64_t *num_keys_tested, uint32_t nonces_to_bruteforce, uint8_t *bf_test_nonces_2nd_byte, noncelist_t *nonces); extern void bitslice_test_nonces(uint32_t nonces_to_bruteforce, uint32_t *bf_test_nonces, uint8_t *bf_test_nonce_par); -- 2.39.2