Mineplay/Hallocy
1
0
Fork 0
Code Issues Pull requests Projects 1 Releases 2 Packages Wiki Activity Actions

Compare commits

base: Mineplay:9cef673a7e0d750f9f8233a6ddac478684c1e12c
Branches Tags
Mineplay:main
Mineplay:V1.1.0
Mineplay:V1.0.0
..
compare: Mineplay:db9d974f5e787ce92c8ddccf1453d86b32ec641e
Branches Tags
Mineplay:main
Mineplay:V1.1.0
Mineplay:V1.0.0

No commits in common. "9cef673a7e0d750f9f8233a6ddac478684c1e12c" and "db9d974f5e787ce92c8ddccf1453d86b32ec641e" have entirely different histories.
9cef673a7e ... db9d974f5e

4 changed files with 120 additions and 102 deletions
Show stats Expand all files Collapse all files

12
CMakeLists.txt
View file

@ -16,13 +16,11 @@ target_link_libraries(HallocyTest Hallocy)
if (MSVC) if (MSVC)
target_compile_options(Hallocy PRIVATE /W4 /Zl) target_compile_options(Hallocy PRIVATE /W4 /Zl)
else() else()
if (CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64") target_compile_options(Hallocy PRIVATE -mavx512f -mavx512vl)
target_compile_options(Hallocy PRIVATE -mavx512f -mavx512vl -march=native) target_compile_options(HallocyTest PRIVATE -mavx512f -mavx512vl)
target_compile_options(HallocyTest PRIVATE -mavx512f -mavx512vl -march=native)
elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "arm") target_compile_options(Hallocy PRIVATE -march=native)
target_compile_options(Hallocy PRIVATE -mfpu=neon) target_compile_options(HallocyTest PRIVATE -march=native)
target_compile_options(HallocyTest PRIVATE -mfpu=neon)
endif()
target_compile_options(Hallocy PRIVATE -Wall -Wextra -pedantic) target_compile_options(Hallocy PRIVATE -Wall -Wextra -pedantic)
endif() endif()

24
Include/Hallocy/Utils/Simd.h
View file

@ -23,20 +23,20 @@
#ifndef HALLOCY_SIMD #ifndef HALLOCY_SIMD
#define HALLOCY_SIMD #define HALLOCY_SIMD
#if defined(_MSC_VER) && (defined(_M_ARM64) || defined(_M_ARM)) #if defined(_MSC_VER)
#define WIN_NEON #if defined(_M_ARM64)
#include <arm_neon.h> #include <arm64intr.h>
#elif defined(_MSC_VER) #else
#define WIN_SIMD
#include <intrin.h> #include <intrin.h>
#elif defined(__GNUC__) && (defined(__x86_64__) || defined(__i386__)) #endif
#define LIN_SIMD
#include <immintrin.h>
#elif defined(__GNUC__) && (defined(__ARM_NEON__) || defined(__aarch64__))
#define LIN_NEON
#include <arm_neon.h>
#else #else
#warning "SIMD is unsupported by this architecture or compiler (only x86/x64/ARM/ARM64 supported)." #if defined(__aarch64__)
#include <arm64intr.h>
#elif defined(__arm__)
#include <arm_neon.h>
#else
#include <immintrin.h>
#endif
#endif #endif
#define SIMD_64_WIDTH 8 #define SIMD_64_WIDTH 8

29
Src/Core/Memory.c
View file

@ -23,6 +23,9 @@
#include "../../Include/Hallocy/Core/Memory.h" #include "../../Include/Hallocy/Core/Memory.h"
#include "../../Include/Hallocy/Utils/Simd.h" #include "../../Include/Hallocy/Utils/Simd.h"
#include <immintrin.h>
#include <stddef.h>
HallocyError hallocy_set_memory(void *destination, int value, const size_t size) { HallocyError hallocy_set_memory(void *destination, int value, const size_t size) {
if (destination == NULL) { if (destination == NULL) {
return HALLOCY_ERROR_INVALID_POINTER; return HALLOCY_ERROR_INVALID_POINTER;
@ -34,7 +37,7 @@ HallocyError hallocy_set_memory(void *destination, int value, const size_t size)
unsigned char value_bytes = (unsigned char)value; unsigned char value_bytes = (unsigned char)value;
switch (hallocy_is_simd_supported()) { switch (hallocy_is_simd_supported()) {
#if defined(LIN_NEON) || defined(WIN_NEON) #if defined(_M_ARM64) || defined(__aarch64__) || defined(__arm__)
case HALLOCY_SIMD_NEON: { case HALLOCY_SIMD_NEON: {
while (((size_t)destination_bytes % SIMD_128_WIDTH) != 0 && destination_bytes != end_address) { while (((size_t)destination_bytes % SIMD_128_WIDTH) != 0 && destination_bytes != end_address) {
*destination_bytes = value_bytes; *destination_bytes = value_bytes;
@ -48,7 +51,7 @@ HallocyError hallocy_set_memory(void *destination, int value, const size_t size)
} }
break; break;
} }
#elif defined(LIN_SIMD) || defined(WIN_SIMD) #else
case HALLOCY_SIMD_AVX512: { case HALLOCY_SIMD_AVX512: {
while (((size_t)destination_bytes % SIMD_512_WIDTH) != 0 && destination_bytes != end_address) { while (((size_t)destination_bytes % SIMD_512_WIDTH) != 0 && destination_bytes != end_address) {
*destination_bytes = value_bytes; *destination_bytes = value_bytes;
@ -138,7 +141,7 @@ HallocyError hallocy_copy_memory(void *destination, void *source, const size_t s
unsigned char *end_address = destination_bytes + size; unsigned char *end_address = destination_bytes + size;
switch (hallocy_is_simd_supported()) { switch (hallocy_is_simd_supported()) {
#if defined(LIN_NEON) || defined(WIN_NEON) #if defined(_M_ARM64) || defined(__aarch64__) || defined(__arm__)
case HALLOCY_SIMD_NEON: { case HALLOCY_SIMD_NEON: {
if ((size_t)destination_bytes % SIMD_128_WIDTH == (size_t)source_bytes % SIMD_128_WIDTH) { if ((size_t)destination_bytes % SIMD_128_WIDTH == (size_t)source_bytes % SIMD_128_WIDTH) {
while (((size_t)destination_bytes % SIMD_128_WIDTH) != 0 && destination_bytes != end_address) { while (((size_t)destination_bytes % SIMD_128_WIDTH) != 0 && destination_bytes != end_address) {
@ -150,14 +153,14 @@ HallocyError hallocy_copy_memory(void *destination, void *source, const size_t s
uint8x16_t simd_value; uint8x16_t simd_value;
while (destination_bytes - end_address >= SIMD_128_WIDTH) { while (destination_bytes - end_address >= SIMD_128_WIDTH) {
simd_value = vdupq_n_u8(*source_bytes); simd_value = vdupq_n_u8(source_bytes);
vst1q_u8(destination_bytes, simd_value); vst1q_u8(destination_bytes, simd_value);
destination_bytes += SIMD_128_WIDTH; destination_bytes += SIMD_128_WIDTH;
source_bytes += SIMD_128_WIDTH; source_bytes += SIMD_128_WIDTH;
} }
break; break;
} }
#elif defined(LIN_SIMD) || defined(WIN_SIMD) #else
case HALLOCY_SIMD_AVX512: { case HALLOCY_SIMD_AVX512: {
if ((size_t)destination_bytes % SIMD_512_WIDTH == (size_t)source_bytes % SIMD_512_WIDTH) { if ((size_t)destination_bytes % SIMD_512_WIDTH == (size_t)source_bytes % SIMD_512_WIDTH) {
while (((size_t)destination_bytes % SIMD_512_WIDTH) != 0 && destination_bytes != end_address) { while (((size_t)destination_bytes % SIMD_512_WIDTH) != 0 && destination_bytes != end_address) {
@ -296,7 +299,7 @@ HallocyError hallocy_move_memory(void *destination, void *source, const size_t s
unsigned char *source_bytes = (unsigned char*)source + size; unsigned char *source_bytes = (unsigned char*)source + size;
switch (hallocy_is_simd_supported()) { switch (hallocy_is_simd_supported()) {
#if defined(LIN_NEON) || defined(WIN_NEON) #if defined(_M_ARM64) || defined(__aarch64__) || defined(__arm__)
case HALLOCY_SIMD_NEON: { case HALLOCY_SIMD_NEON: {
if ((size_t)destination_bytes % SIMD_128_WIDTH == (size_t)source_bytes % SIMD_128_WIDTH) { if ((size_t)destination_bytes % SIMD_128_WIDTH == (size_t)source_bytes % SIMD_128_WIDTH) {
while (((size_t)destination_bytes % SIMD_128_WIDTH) != 0 && destination_bytes != end_address) { while (((size_t)destination_bytes % SIMD_128_WIDTH) != 0 && destination_bytes != end_address) {
@ -311,12 +314,12 @@ HallocyError hallocy_move_memory(void *destination, void *source, const size_t s
destination_bytes -= SIMD_128_WIDTH; destination_bytes -= SIMD_128_WIDTH;
source_bytes -= SIMD_128_WIDTH; source_bytes -= SIMD_128_WIDTH;
simd_value = vdupq_n_u8(*source_bytes); simd_value = vdupq_n_u8(source_bytes);
vst1q_u8(destination_bytes, simd_value); vst1q_u8(destination_bytes, simd_value);
} }
break; break;
} }
#elif defined(LIN_SIMD) || defined(WIN_SIMD) #else
case HALLOCY_SIMD_AVX512: { case HALLOCY_SIMD_AVX512: {
if ((size_t)destination_bytes % SIMD_512_WIDTH == (size_t)source_bytes % SIMD_512_WIDTH) { if ((size_t)destination_bytes % SIMD_512_WIDTH == (size_t)source_bytes % SIMD_512_WIDTH) {
while (((size_t)destination_bytes % SIMD_512_WIDTH) != 0 && destination_bytes != end_address) { while (((size_t)destination_bytes % SIMD_512_WIDTH) != 0 && destination_bytes != end_address) {
@ -452,7 +455,7 @@ bool hallocy_compare_memory(void *left_side, void *right_side, const size_t size
unsigned char *end_address = left_side_bytes + size; unsigned char *end_address = left_side_bytes + size;
switch (hallocy_is_simd_supported()) { switch (hallocy_is_simd_supported()) {
#if defined(LIN_NEON) || defined(WIN_NEON) #if defined(_M_ARM64) || defined(__aarch64__) || defined(__arm__)
case HALLOCY_SIMD_NEON: { case HALLOCY_SIMD_NEON: {
if ((size_t)left_side_bytes % SIMD_128_WIDTH == (size_t)right_side_bytes % SIMD_128_WIDTH) { if ((size_t)left_side_bytes % SIMD_128_WIDTH == (size_t)right_side_bytes % SIMD_128_WIDTH) {
while (((size_t)left_side_bytes % SIMD_128_WIDTH) != 0 && left_side_bytes != end_address) { while (((size_t)left_side_bytes % SIMD_128_WIDTH) != 0 && left_side_bytes != end_address) {
@ -465,9 +468,9 @@ bool hallocy_compare_memory(void *left_side, void *right_side, const size_t size
} }
} }
while (left_side_bytes - end_address >= SIMD_128_WIDTH) { while (end_address - destination_bytes >= SIMD_128_WIDTH) {
uint8x16_t simd_left_side_value = vdupq_n_u8(*left_side_bytes); uint8x16_t simd_left_side_value = vdupq_n_u8(left_side_bytes);
uint8x16_t simd_right_side_value = vdupq_n_u8(*right_side_bytes); uint8x16_t simd_right_side_value = vdupq_n_u8(right_side_bytes);
uint8x16_t result = vceqq_u8(simd_left_side_value, simd_right_side_value); uint8x16_t result = vceqq_u8(simd_left_side_value, simd_right_side_value);
if (vmaxvq_u8(result) != 0xFF) { if (vmaxvq_u8(result) != 0xFF) {
@ -479,7 +482,7 @@ bool hallocy_compare_memory(void *left_side, void *right_side, const size_t size
} }
break; break;
} }
#elif defined(LIN_SIMD) || defined(WIN_SIMD) #else
case HALLOCY_SIMD_AVX512: { case HALLOCY_SIMD_AVX512: {
if ((size_t)left_side_bytes % SIMD_512_WIDTH == (size_t)right_side_bytes % SIMD_512_WIDTH) { if ((size_t)left_side_bytes % SIMD_512_WIDTH == (size_t)right_side_bytes % SIMD_512_WIDTH) {
while (((size_t)left_side_bytes % SIMD_512_WIDTH) != 0 && left_side_bytes != end_address) { while (((size_t)left_side_bytes % SIMD_512_WIDTH) != 0 && left_side_bytes != end_address) {

35
Src/Utils/Simd.c
View file

@ -21,22 +21,19 @@
*/ */
#include "../../Include/Hallocy/Utils/Simd.h" #include "../../Include/Hallocy/Utils/Simd.h"
#ifdef LIN_NEON
#include <sys/auxv.h>
#endif
HallocySimdType hallocy_is_simd_supported(void) { HallocySimdType hallocy_is_simd_supported(void) {
static HallocySimdType hallocy_supported_simd = HALLOCY_SIMD_UNDEFINED; static HallocySimdType hallocy_supported_simd = HALLOCY_SIMD_UNDEFINED;
if (hallocy_supported_simd != HALLOCY_SIMD_UNDEFINED) { if (hallocy_supported_simd != HALLOCY_SIMD_UNDEFINED) {
return hallocy_supported_simd; return hallocy_supported_simd;
} }
#if defined(WIN_NEON) #if defined(_MSC_VER)
#if defined(_M_ARM64)
if (isProcessorFeaturePresent(PF_ARM64_SVE)) { if (isProcessorFeaturePresent(PF_ARM64_SVE)) {
hallocy_supported_simd = HALLOCY_SIMD_NEON; hallocy_supported_simd = HALLOCY_SIMD_NEON;
return hallocy_supported_simd; return hallocy_supported_simd;
} }
#elif defined(WIN_SIMD) #else
int cpu_info[4] = { 0 }; int cpu_info[4] = { 0 };
__cpuid(cpu_info, 7); __cpuid(cpu_info, 7);
if ((cpu_info[1] & (1 << 16)) != 0) { if ((cpu_info[1] & (1 << 16)) != 0) {
@ -65,12 +62,31 @@ HallocySimdType hallocy_is_simd_supported(void) {
hallocy_supported_simd = HALLOCY_SIMD_SSE2; hallocy_supported_simd = HALLOCY_SIMD_SSE2;
return hallocy_supported_simd; return hallocy_supported_simd;
} }
#elif defined(LIN_NEON) #endif
if (getauxval(16) & (1 << 12)) { #else
#if defined(__aarch64__) || defined(__arm__)
int file_descriptor = open("/proc/cpuinfo", O_READONLY);
if (file_descriptor == -1) {
return hallocy_supported_simd;
}
char buffer[256];
int bytes_read = read(file_descriptor, buffer, sizeof(buffer));
while (bytes_read > 0) {
for (size_t i = 0; i < bytes_read - 4; i++) {
if (buffer[i] == 'n' && buffer[i + 1] == 'e' && buffer[i + 2] == 'o' && buffer[i + 3] == 'n') {
close(file_descriptor);
hallocy_supported_simd = HALLOCY_SIMD_NEON; hallocy_supported_simd = HALLOCY_SIMD_NEON;
return hallocy_supported_simd; return hallocy_supported_simd;
} }
#elif defined(LIN_SIMD) }
bytes_read = read(file_descriptor, buffer, sizeof(buffer));
}
close(file_descriptor);
#else
unsigned int a, b, c, d; unsigned int a, b, c, d;
__asm__ __volatile__ ( __asm__ __volatile__ (
"cpuid" "cpuid"
@ -109,6 +125,7 @@ HallocySimdType hallocy_is_simd_supported(void) {
return hallocy_supported_simd; return hallocy_supported_simd;
} }
#endif #endif
#endif
hallocy_supported_simd = HALLOCY_SIMD_NONE; hallocy_supported_simd = HALLOCY_SIMD_NONE;
return hallocy_supported_simd; return hallocy_supported_simd;