Add very basic feature flagging for NEON

This commit is contained in:
Doridian 2022-03-27 10:19:34 -07:00
commit 62704db5ed
8 changed files with 210 additions and 26 deletions

View file

@ -98,6 +98,20 @@
#define COUNT_BITARRAY_AND2 count_bitarray_AND2_MMX
#define COUNT_BITARRAY_AND3 count_bitarray_AND3_MMX
#define COUNT_BITARRAY_AND4 count_bitarray_AND4_MMX
#elif defined (__ARM_NEON) && !defined (NOSIMD_BUILD)
#define MALLOC_BITARRAY malloc_bitarray_NEON
#define FREE_BITARRAY free_bitarray_NEON
#define BITCOUNT bitcount_NEON
#define COUNT_STATES count_states_NEON
#define BITARRAY_AND bitarray_AND_NEON
#define BITARRAY_LOW20_AND bitarray_low20_AND_NEON
#define COUNT_BITARRAY_AND count_bitarray_AND_NEON
#define COUNT_BITARRAY_LOW20_AND count_bitarray_low20_AND_NEON
#define BITARRAY_AND4 bitarray_AND4_NEON
#define BITARRAY_OR bitarray_OR_NEON
#define COUNT_BITARRAY_AND2 count_bitarray_AND2_NEON
#define COUNT_BITARRAY_AND3 count_bitarray_AND3_NEON
#define COUNT_BITARRAY_AND4 count_bitarray_AND4_NEON
#else
#define MALLOC_BITARRAY malloc_bitarray_NOSIMD
#define FREE_BITARRAY free_bitarray_NOSIMD
@ -117,31 +131,31 @@
// typedefs and declaration of functions:
typedef uint32_t *malloc_bitarray_t(uint32_t);
malloc_bitarray_t malloc_bitarray_AVX512, malloc_bitarray_AVX2, malloc_bitarray_AVX, malloc_bitarray_SSE2, malloc_bitarray_MMX, malloc_bitarray_NOSIMD, malloc_bitarray_dispatch;
malloc_bitarray_t malloc_bitarray_AVX512, malloc_bitarray_AVX2, malloc_bitarray_AVX, malloc_bitarray_SSE2, malloc_bitarray_MMX, malloc_bitarray_NOSIMD, malloc_bitarray_NEON, malloc_bitarray_dispatch;
typedef void free_bitarray_t(uint32_t *);
free_bitarray_t free_bitarray_AVX512, free_bitarray_AVX2, free_bitarray_AVX, free_bitarray_SSE2, free_bitarray_MMX, free_bitarray_NOSIMD, free_bitarray_dispatch;
free_bitarray_t free_bitarray_AVX512, free_bitarray_AVX2, free_bitarray_AVX, free_bitarray_SSE2, free_bitarray_MMX, free_bitarray_NOSIMD, free_bitarray_NEON, free_bitarray_dispatch;
typedef uint32_t bitcount_t(uint32_t);
bitcount_t bitcount_AVX512, bitcount_AVX2, bitcount_AVX, bitcount_SSE2, bitcount_MMX, bitcount_NOSIMD, bitcount_dispatch;
bitcount_t bitcount_AVX512, bitcount_AVX2, bitcount_AVX, bitcount_SSE2, bitcount_MMX, bitcount_NOSIMD, bitcount_NEON, bitcount_dispatch;
typedef uint32_t count_states_t(uint32_t *);
count_states_t count_states_AVX512, count_states_AVX2, count_states_AVX, count_states_SSE2, count_states_MMX, count_states_NOSIMD, count_states_dispatch;
count_states_t count_states_AVX512, count_states_AVX2, count_states_AVX, count_states_SSE2, count_states_MMX, count_states_NOSIMD, count_states_NEON, count_states_dispatch;
typedef void bitarray_AND_t(uint32_t[], uint32_t[]);
bitarray_AND_t bitarray_AND_AVX512, bitarray_AND_AVX2, bitarray_AND_AVX, bitarray_AND_SSE2, bitarray_AND_MMX, bitarray_AND_NOSIMD, bitarray_AND_dispatch;
bitarray_AND_t bitarray_AND_AVX512, bitarray_AND_AVX2, bitarray_AND_AVX, bitarray_AND_SSE2, bitarray_AND_MMX, bitarray_AND_NOSIMD, bitarray_AND_NEON, bitarray_AND_dispatch;
typedef void bitarray_low20_AND_t(uint32_t *, uint32_t *);
bitarray_low20_AND_t bitarray_low20_AND_AVX512, bitarray_low20_AND_AVX2, bitarray_low20_AND_AVX, bitarray_low20_AND_SSE2, bitarray_low20_AND_MMX, bitarray_low20_AND_NOSIMD, bitarray_low20_AND_dispatch;
bitarray_low20_AND_t bitarray_low20_AND_AVX512, bitarray_low20_AND_AVX2, bitarray_low20_AND_AVX, bitarray_low20_AND_SSE2, bitarray_low20_AND_MMX, bitarray_low20_AND_NOSIMD, bitarray_low20_AND_NEON, bitarray_low20_AND_dispatch;
typedef uint32_t count_bitarray_AND_t(uint32_t *, uint32_t *);
count_bitarray_AND_t count_bitarray_AND_AVX512, count_bitarray_AND_AVX2, count_bitarray_AND_AVX, count_bitarray_AND_SSE2, count_bitarray_AND_MMX, count_bitarray_AND_NOSIMD, count_bitarray_AND_dispatch;
count_bitarray_AND_t count_bitarray_AND_AVX512, count_bitarray_AND_AVX2, count_bitarray_AND_AVX, count_bitarray_AND_SSE2, count_bitarray_AND_MMX, count_bitarray_AND_NOSIMD, count_bitarray_AND_NEON, count_bitarray_AND_dispatch;
typedef uint32_t count_bitarray_low20_AND_t(uint32_t *, uint32_t *);
count_bitarray_low20_AND_t count_bitarray_low20_AND_AVX512, count_bitarray_low20_AND_AVX2, count_bitarray_low20_AND_AVX, count_bitarray_low20_AND_SSE2, count_bitarray_low20_AND_MMX, count_bitarray_low20_AND_NOSIMD, count_bitarray_low20_AND_dispatch;
count_bitarray_low20_AND_t count_bitarray_low20_AND_AVX512, count_bitarray_low20_AND_AVX2, count_bitarray_low20_AND_AVX, count_bitarray_low20_AND_SSE2, count_bitarray_low20_AND_MMX, count_bitarray_low20_AND_NOSIMD, count_bitarray_low20_AND_NEON, count_bitarray_low20_AND_dispatch;
typedef void bitarray_AND4_t(uint32_t *, uint32_t *, uint32_t *, uint32_t *);
bitarray_AND4_t bitarray_AND4_AVX512, bitarray_AND4_AVX2, bitarray_AND4_AVX, bitarray_AND4_SSE2, bitarray_AND4_MMX, bitarray_AND4_NOSIMD, bitarray_AND4_dispatch;
bitarray_AND4_t bitarray_AND4_AVX512, bitarray_AND4_AVX2, bitarray_AND4_AVX, bitarray_AND4_SSE2, bitarray_AND4_MMX, bitarray_AND4_NOSIMD, bitarray_AND4_NEON, bitarray_AND4_dispatch;
typedef void bitarray_OR_t(uint32_t[], uint32_t[]);
bitarray_OR_t bitarray_OR_AVX512, bitarray_OR_AVX2, bitarray_OR_AVX, bitarray_OR_SSE2, bitarray_OR_MMX, bitarray_OR_NOSIMD, bitarray_OR_dispatch;
bitarray_OR_t bitarray_OR_AVX512, bitarray_OR_AVX2, bitarray_OR_AVX, bitarray_OR_SSE2, bitarray_OR_MMX, bitarray_OR_NOSIMD, bitarray_OR_NEON, bitarray_OR_dispatch;
typedef uint32_t count_bitarray_AND2_t(uint32_t *, uint32_t *);
count_bitarray_AND2_t count_bitarray_AND2_AVX512, count_bitarray_AND2_AVX2, count_bitarray_AND2_AVX, count_bitarray_AND2_SSE2, count_bitarray_AND2_MMX, count_bitarray_AND2_NOSIMD, count_bitarray_AND2_dispatch;
count_bitarray_AND2_t count_bitarray_AND2_AVX512, count_bitarray_AND2_AVX2, count_bitarray_AND2_AVX, count_bitarray_AND2_SSE2, count_bitarray_AND2_MMX, count_bitarray_AND2_NOSIMD, count_bitarray_AND2_NEON, count_bitarray_AND2_dispatch;
typedef uint32_t count_bitarray_AND3_t(uint32_t *, uint32_t *, uint32_t *);
count_bitarray_AND3_t count_bitarray_AND3_AVX512, count_bitarray_AND3_AVX2, count_bitarray_AND3_AVX, count_bitarray_AND3_SSE2, count_bitarray_AND3_MMX, count_bitarray_AND3_NOSIMD, count_bitarray_AND3_dispatch;
count_bitarray_AND3_t count_bitarray_AND3_AVX512, count_bitarray_AND3_AVX2, count_bitarray_AND3_AVX, count_bitarray_AND3_SSE2, count_bitarray_AND3_MMX, count_bitarray_AND3_NOSIMD, count_bitarray_AND3_NEON, count_bitarray_AND3_dispatch;
typedef uint32_t count_bitarray_AND4_t(uint32_t *, uint32_t *, uint32_t *, uint32_t *);
count_bitarray_AND4_t count_bitarray_AND4_AVX512, count_bitarray_AND4_AVX2, count_bitarray_AND4_AVX, count_bitarray_AND4_SSE2, count_bitarray_AND4_MMX, count_bitarray_AND4_NOSIMD, count_bitarray_AND4_dispatch;
count_bitarray_AND4_t count_bitarray_AND4_AVX512, count_bitarray_AND4_AVX2, count_bitarray_AND4_AVX, count_bitarray_AND4_SSE2, count_bitarray_AND4_MMX, count_bitarray_AND4_NOSIMD, count_bitarray_AND4_NEON, count_bitarray_AND4_dispatch;
inline uint32_t *MALLOC_BITARRAY(uint32_t x) {
@ -287,7 +301,7 @@ inline uint32_t COUNT_BITARRAY_AND4(uint32_t *restrict A, uint32_t *restrict B,
}
#ifndef __MMX__
#ifdef NOSIMD_BUILD
// pointers to functions:
malloc_bitarray_t *malloc_bitarray_function_p = &malloc_bitarray_dispatch;
@ -306,6 +320,11 @@ count_bitarray_AND4_t *count_bitarray_AND4_function_p = &count_bitarray_AND4_dis
// determine the available instruction set at runtime and call the correct function
uint32_t *malloc_bitarray_dispatch(uint32_t x) {
#if defined(COMPILER_HAS_SIMD_NEON)
if (arm_has_neon()) malloc_bitarray_function_p = &malloc_bitarray_NEON;
else
#endif
#if defined(COMPILER_HAS_SIMD_AVX512)
if (__builtin_cpu_supports("avx512f")) malloc_bitarray_function_p = &malloc_bitarray_AVX512;
else
@ -324,6 +343,11 @@ uint32_t *malloc_bitarray_dispatch(uint32_t x) {
}
void free_bitarray_dispatch(uint32_t *x) {
#if defined(COMPILER_HAS_SIMD_NEON)
if (arm_has_neon()) free_bitarray_function_p = &free_bitarray_NEON;
else
#endif
#if defined(COMPILER_HAS_SIMD_AVX512)
if (__builtin_cpu_supports("avx512f")) free_bitarray_function_p = &free_bitarray_AVX512;
else
@ -342,6 +366,11 @@ void free_bitarray_dispatch(uint32_t *x) {
}
uint32_t bitcount_dispatch(uint32_t a) {
#if defined(COMPILER_HAS_SIMD_NEON)
if (arm_has_neon()) bitcount_function_p = &bitcount_NEON;
else
#endif
#if defined(COMPILER_HAS_SIMD_AVX512)
if (__builtin_cpu_supports("avx512f")) bitcount_function_p = &bitcount_AVX512;
else
@ -360,6 +389,11 @@ uint32_t bitcount_dispatch(uint32_t a) {
}
uint32_t count_states_dispatch(uint32_t *bitarray) {
#if defined(COMPILER_HAS_SIMD_NEON)
if (arm_has_neon()) count_states_function_p = &count_states_NEON;
else
#endif
#if defined(COMPILER_HAS_SIMD_AVX512)
if (__builtin_cpu_supports("avx512f")) count_states_function_p = &count_states_AVX512;
else
@ -378,6 +412,11 @@ uint32_t count_states_dispatch(uint32_t *bitarray) {
}
void bitarray_AND_dispatch(uint32_t *A, uint32_t *B) {
#if defined(COMPILER_HAS_SIMD_NEON)
if (arm_has_neon()) bitarray_AND_function_p = &bitarray_AND_NEON;
else
#endif
#if defined(COMPILER_HAS_SIMD_AVX512)
if (__builtin_cpu_supports("avx512f")) bitarray_AND_function_p = &bitarray_AND_AVX512;
else
@ -396,6 +435,11 @@ void bitarray_AND_dispatch(uint32_t *A, uint32_t *B) {
}
void bitarray_low20_AND_dispatch(uint32_t *A, uint32_t *B) {
#if defined(COMPILER_HAS_SIMD_NEON)
if (arm_has_neon()) bitarray_low20_AND_function_p = &bitarray_low20_AND_NEON;
else
#endif
#if defined(COMPILER_HAS_SIMD_AVX512)
if (__builtin_cpu_supports("avx512f")) bitarray_low20_AND_function_p = &bitarray_low20_AND_AVX512;
else
@ -414,6 +458,11 @@ void bitarray_low20_AND_dispatch(uint32_t *A, uint32_t *B) {
}
uint32_t count_bitarray_AND_dispatch(uint32_t *A, uint32_t *B) {
#if defined(COMPILER_HAS_SIMD_NEON)
if (arm_has_neon()) count_bitarray_AND_function_p = &count_bitarray_AND_NEON;
else
#endif
#if defined(COMPILER_HAS_SIMD_AVX512)
if (__builtin_cpu_supports("avx512f")) count_bitarray_AND_function_p = &count_bitarray_AND_AVX512;
else
@ -432,6 +481,11 @@ uint32_t count_bitarray_AND_dispatch(uint32_t *A, uint32_t *B) {
}
uint32_t count_bitarray_low20_AND_dispatch(uint32_t *A, uint32_t *B) {
#if defined(COMPILER_HAS_SIMD_NEON)
if (arm_has_neon()) count_bitarray_low20_AND_function_p = &count_bitarray_low20_AND_NEON;
else
#endif
#if defined(COMPILER_HAS_SIMD_AVX512)
if (__builtin_cpu_supports("avx512f")) count_bitarray_low20_AND_function_p = &count_bitarray_low20_AND_AVX512;
else
@ -450,6 +504,11 @@ uint32_t count_bitarray_low20_AND_dispatch(uint32_t *A, uint32_t *B) {
}
void bitarray_AND4_dispatch(uint32_t *A, uint32_t *B, uint32_t *C, uint32_t *D) {
#if defined(COMPILER_HAS_SIMD_NEON)
if (arm_has_neon()) bitarray_AND4_function_p = &bitarray_AND4_NEON;
else
#endif
#if defined(COMPILER_HAS_SIMD_AVX512)
if (__builtin_cpu_supports("avx512f")) bitarray_AND4_function_p = &bitarray_AND4_AVX512;
else
@ -468,6 +527,11 @@ void bitarray_AND4_dispatch(uint32_t *A, uint32_t *B, uint32_t *C, uint32_t *D)
}
void bitarray_OR_dispatch(uint32_t *A, uint32_t *B) {
#if defined(COMPILER_HAS_SIMD_NEON)
if (arm_has_neon()) bitarray_OR_function_p = &bitarray_OR_NEON;
else
#endif
#if defined(COMPILER_HAS_SIMD_AVX512)
if (__builtin_cpu_supports("avx512f")) bitarray_OR_function_p = &bitarray_OR_AVX512;
else
@ -486,6 +550,11 @@ void bitarray_OR_dispatch(uint32_t *A, uint32_t *B) {
}
uint32_t count_bitarray_AND2_dispatch(uint32_t *A, uint32_t *B) {
#if defined(COMPILER_HAS_SIMD_NEON)
if (arm_has_neon()) count_bitarray_AND2_function_p = &count_bitarray_AND2_NEON;
else
#endif
#if defined(COMPILER_HAS_SIMD_AVX512)
if (__builtin_cpu_supports("avx512f")) count_bitarray_AND2_function_p = &count_bitarray_AND2_AVX512;
else
@ -504,6 +573,11 @@ uint32_t count_bitarray_AND2_dispatch(uint32_t *A, uint32_t *B) {
}
uint32_t count_bitarray_AND3_dispatch(uint32_t *A, uint32_t *B, uint32_t *C) {
#if defined(COMPILER_HAS_SIMD_NEON)
if (arm_has_neon()) count_bitarray_AND3_function_p = &count_bitarray_AND3_NEON;
else
#endif
#if defined(COMPILER_HAS_SIMD_AVX512)
if (__builtin_cpu_supports("avx512f")) count_bitarray_AND3_function_p = &count_bitarray_AND3_AVX512;
else
@ -522,6 +596,11 @@ uint32_t count_bitarray_AND3_dispatch(uint32_t *A, uint32_t *B, uint32_t *C) {
}
uint32_t count_bitarray_AND4_dispatch(uint32_t *A, uint32_t *B, uint32_t *C, uint32_t *D) {
#if defined(COMPILER_HAS_SIMD_NEON)
if (arm_has_neon()) count_bitarray_AND4_function_p = &count_bitarray_AND4_NEON;
else
#endif
#if defined(COMPILER_HAS_SIMD_AVX512)
if (__builtin_cpu_supports("avx512f")) count_bitarray_AND4_function_p = &count_bitarray_AND4_AVX512;
else