282 lines
9.6 KiB
C
282 lines
9.6 KiB
C
#if !defined(SIMDE_X86_AVX512_DPBF16_H)
|
|
#define SIMDE_X86_AVX512_DPBF16_H
|
|
|
|
#include "types.h"
|
|
#include "mov.h"
|
|
|
|
HEDLEY_DIAGNOSTIC_PUSH
|
|
SIMDE_DISABLE_UNWANTED_DIAGNOSTICS
|
|
SIMDE_BEGIN_DECLS_
|
|
|
|
SIMDE_FUNCTION_ATTRIBUTES
|
|
simde__m128
|
|
simde_mm_dpbf16_ps (simde__m128 src, simde__m128bh a, simde__m128bh b) {
|
|
#if defined(SIMDE_X86_AVX512BF16_NATIVE) && defined(SIMDE_X86_AVX512VL_NATIVE)
|
|
return _mm_dpbf16_ps(src, a, b);
|
|
#else
|
|
simde__m128_private
|
|
src_ = simde__m128_to_private(src);
|
|
simde__m128bh_private
|
|
a_ = simde__m128bh_to_private(a),
|
|
b_ = simde__m128bh_to_private(b);
|
|
|
|
#if defined(SIMDE_VECTOR_SUBSCRIPT_SCALAR) && defined(SIMDE_CONVERT_VECTOR_) && defined(SIMDE_SHUFFLE_VECTOR_)
|
|
uint32_t x1 SIMDE_VECTOR(32);
|
|
uint32_t x2 SIMDE_VECTOR(32);
|
|
simde__m128_private
|
|
r1_[2],
|
|
r2_[2];
|
|
|
|
a_.u16 =
|
|
SIMDE_SHUFFLE_VECTOR_(
|
|
16, 16,
|
|
a_.u16, a_.u16,
|
|
0, 2, 4, 6,
|
|
1, 3, 5, 7
|
|
);
|
|
b_.u16 =
|
|
SIMDE_SHUFFLE_VECTOR_(
|
|
16, 16,
|
|
b_.u16, b_.u16,
|
|
0, 2, 4, 6,
|
|
1, 3, 5, 7
|
|
);
|
|
|
|
SIMDE_CONVERT_VECTOR_(x1, a_.u16);
|
|
SIMDE_CONVERT_VECTOR_(x2, b_.u16);
|
|
|
|
x1 <<= 16;
|
|
x2 <<= 16;
|
|
|
|
simde_memcpy(&r1_, &x1, sizeof(x1));
|
|
simde_memcpy(&r2_, &x2, sizeof(x2));
|
|
|
|
src_.f32 +=
|
|
HEDLEY_REINTERPRET_CAST(__typeof__(a_.f32), r1_[0].u32) * HEDLEY_REINTERPRET_CAST(__typeof__(a_.f32), r2_[0].u32) +
|
|
HEDLEY_REINTERPRET_CAST(__typeof__(a_.f32), r1_[1].u32) * HEDLEY_REINTERPRET_CAST(__typeof__(a_.f32), r2_[1].u32);
|
|
#else
|
|
SIMDE_VECTORIZE
|
|
for (size_t i = 0 ; i < (sizeof(a_.u16) / sizeof(a_.u16[0])) ; i++) {
|
|
src_.f32[i / 2] += (simde_uint32_as_float32(HEDLEY_STATIC_CAST(uint32_t, a_.u16[i]) << 16) * simde_uint32_as_float32(HEDLEY_STATIC_CAST(uint32_t, b_.u16[i]) << 16));
|
|
}
|
|
#endif
|
|
|
|
return simde__m128_from_private(src_);
|
|
#endif
|
|
}
|
|
#if defined(SIMDE_X86_AVX512BF16_ENABLE_NATIVE_ALIASES) && defined(SIMDE_X86_AVX512VL_ENABLE_NATIVE_ALIASES)
|
|
#undef _mm_dpbf16_ps
|
|
#define _mm_dpbf16_ps(src, a, b) simde_mm_dpbf16_ps(src, a, b)
|
|
#endif
|
|
|
|
SIMDE_FUNCTION_ATTRIBUTES
|
|
simde__m128
|
|
simde_mm_mask_dpbf16_ps (simde__m128 src, simde__mmask8 k, simde__m128bh a, simde__m128bh b) {
|
|
#if defined(SIMDE_X86_AVX512BF16_NATIVE) && defined(SIMDE_X86_AVX512VL_NATIVE)
|
|
return _mm_mask_dpbf16_ps(src, k, a, b);
|
|
#else
|
|
return simde_mm_mask_mov_ps(src, k, simde_mm_dpbf16_ps(src, a, b));
|
|
#endif
|
|
}
|
|
#if defined(SIMDE_X86_AVX512BF16_ENABLE_NATIVE_ALIASES) && defined(SIMDE_X86_AVX512VL_ENABLE_NATIVE_ALIASES)
|
|
#undef _mm_mask_dpbf16_ps
|
|
#define _mm_mask_dpbf16_ps(src, k, a, b) simde_mm_mask_dpbf16_ps(src, k, a, b)
|
|
#endif
|
|
|
|
SIMDE_FUNCTION_ATTRIBUTES
|
|
simde__m128
|
|
simde_mm_maskz_dpbf16_ps (simde__mmask8 k, simde__m128 src, simde__m128bh a, simde__m128bh b) {
|
|
#if defined(SIMDE_X86_AVX512BF16_NATIVE) && defined(SIMDE_X86_AVX512VL_NATIVE)
|
|
return _mm_maskz_dpbf16_ps(k, src, a, b);
|
|
#else
|
|
return simde_mm_maskz_mov_ps(k, simde_mm_dpbf16_ps(src, a, b));
|
|
#endif
|
|
}
|
|
#if defined(SIMDE_X86_AVX512BF16_ENABLE_NATIVE_ALIASES) && defined(SIMDE_X86_AVX512VL_ENABLE_NATIVE_ALIASES)
|
|
#undef _mm_maskz_dpbf16_ps
|
|
#define _mm_maskz_dpbf16_ps(k, src, a, b) simde_mm_maskz_dpbf16_ps(k, src, a, b)
|
|
#endif
|
|
|
|
SIMDE_FUNCTION_ATTRIBUTES
|
|
simde__m256
|
|
simde_mm256_dpbf16_ps (simde__m256 src, simde__m256bh a, simde__m256bh b) {
|
|
#if defined(SIMDE_X86_AVX512BF16_NATIVE) && defined(SIMDE_X86_AVX512VL_NATIVE)
|
|
return _mm256_dpbf16_ps(src, a, b);
|
|
#else
|
|
simde__m256_private
|
|
src_ = simde__m256_to_private(src);
|
|
simde__m256bh_private
|
|
a_ = simde__m256bh_to_private(a),
|
|
b_ = simde__m256bh_to_private(b);
|
|
|
|
#if defined(SIMDE_VECTOR_SUBSCRIPT_SCALAR) && defined(SIMDE_CONVERT_VECTOR_) && defined(SIMDE_SHUFFLE_VECTOR_)
|
|
uint32_t x1 SIMDE_VECTOR(64);
|
|
uint32_t x2 SIMDE_VECTOR(64);
|
|
simde__m256_private
|
|
r1_[2],
|
|
r2_[2];
|
|
|
|
a_.u16 =
|
|
SIMDE_SHUFFLE_VECTOR_(
|
|
16, 32,
|
|
a_.u16, a_.u16,
|
|
0, 2, 4, 6, 8, 10, 12, 14,
|
|
1, 3, 5, 7, 9, 11, 13, 15
|
|
);
|
|
b_.u16 =
|
|
SIMDE_SHUFFLE_VECTOR_(
|
|
16, 32,
|
|
b_.u16, b_.u16,
|
|
0, 2, 4, 6, 8, 10, 12, 14,
|
|
1, 3, 5, 7, 9, 11, 13, 15
|
|
);
|
|
|
|
SIMDE_CONVERT_VECTOR_(x1, a_.u16);
|
|
SIMDE_CONVERT_VECTOR_(x2, b_.u16);
|
|
|
|
x1 <<= 16;
|
|
x2 <<= 16;
|
|
|
|
simde_memcpy(&r1_, &x1, sizeof(x1));
|
|
simde_memcpy(&r2_, &x2, sizeof(x2));
|
|
|
|
src_.f32 +=
|
|
HEDLEY_REINTERPRET_CAST(__typeof__(a_.f32), r1_[0].u32) * HEDLEY_REINTERPRET_CAST(__typeof__(a_.f32), r2_[0].u32) +
|
|
HEDLEY_REINTERPRET_CAST(__typeof__(a_.f32), r1_[1].u32) * HEDLEY_REINTERPRET_CAST(__typeof__(a_.f32), r2_[1].u32);
|
|
#else
|
|
SIMDE_VECTORIZE
|
|
for (size_t i = 0 ; i < (sizeof(a_.u16) / sizeof(a_.u16[0])) ; i++) {
|
|
src_.f32[i / 2] += (simde_uint32_as_float32(HEDLEY_STATIC_CAST(uint32_t, a_.u16[i]) << 16) * simde_uint32_as_float32(HEDLEY_STATIC_CAST(uint32_t, b_.u16[i]) << 16));
|
|
}
|
|
#endif
|
|
|
|
return simde__m256_from_private(src_);
|
|
#endif
|
|
}
|
|
#if defined(SIMDE_X86_AVX512BF16_ENABLE_NATIVE_ALIASES) && defined(SIMDE_X86_AVX512VL_ENABLE_NATIVE_ALIASES)
|
|
#undef _mm256_dpbf16_ps
|
|
#define _mm256_dpbf16_ps(src, a, b) simde_mm256_dpbf16_ps(src, a, b)
|
|
#endif
|
|
|
|
SIMDE_FUNCTION_ATTRIBUTES
|
|
simde__m256
|
|
simde_mm256_mask_dpbf16_ps (simde__m256 src, simde__mmask8 k, simde__m256bh a, simde__m256bh b) {
|
|
#if defined(SIMDE_X86_AVX512BF16_NATIVE) && defined(SIMDE_X86_AVX512VL_NATIVE)
|
|
return _mm256_mask_dpbf16_ps(src, k, a, b);
|
|
#else
|
|
return simde_mm256_mask_mov_ps(src, k, simde_mm256_dpbf16_ps(src, a, b));
|
|
#endif
|
|
}
|
|
#if defined(SIMDE_X86_AVX512BF16_ENABLE_NATIVE_ALIASES) && defined(SIMDE_X86_AVX512VL_ENABLE_NATIVE_ALIASES)
|
|
#undef _mm256_mask_dpbf16_ps
|
|
#define _mm256_mask_dpbf16_ps(src, k, a, b) simde_mm256_mask_dpbf16_ps(src, k, a, b)
|
|
#endif
|
|
|
|
SIMDE_FUNCTION_ATTRIBUTES
|
|
simde__m256
|
|
simde_mm256_maskz_dpbf16_ps (simde__mmask8 k, simde__m256 src, simde__m256bh a, simde__m256bh b) {
|
|
#if defined(SIMDE_X86_AVX512BF16_NATIVE) && defined(SIMDE_X86_AVX512VL_NATIVE)
|
|
return _mm256_maskz_dpbf16_ps(k, src, a, b);
|
|
#else
|
|
return simde_mm256_maskz_mov_ps(k, simde_mm256_dpbf16_ps(src, a, b));
|
|
#endif
|
|
}
|
|
#if defined(SIMDE_X86_AVX512BF16_ENABLE_NATIVE_ALIASES) && defined(SIMDE_X86_AVX512VL_ENABLE_NATIVE_ALIASES)
|
|
#undef _mm256_maskz_dpbf16_ps
|
|
#define _mm256_maskz_dpbf16_ps(k, src, a, b) simde_mm256_maskz_dpbf16_ps(k, src, a, b)
|
|
#endif
|
|
|
|
SIMDE_FUNCTION_ATTRIBUTES
|
|
simde__m512
|
|
simde_mm512_dpbf16_ps (simde__m512 src, simde__m512bh a, simde__m512bh b) {
|
|
#if defined(SIMDE_X86_AVX512BF16_NATIVE)
|
|
return _mm512_dpbf16_ps(src, a, b);
|
|
#else
|
|
simde__m512_private
|
|
src_ = simde__m512_to_private(src);
|
|
simde__m512bh_private
|
|
a_ = simde__m512bh_to_private(a),
|
|
b_ = simde__m512bh_to_private(b);
|
|
|
|
#if defined(SIMDE_VECTOR_SUBSCRIPT_SCALAR) && defined(SIMDE_CONVERT_VECTOR_) && defined(SIMDE_SHUFFLE_VECTOR_)
|
|
uint32_t x1 SIMDE_VECTOR(128);
|
|
uint32_t x2 SIMDE_VECTOR(128);
|
|
simde__m512_private
|
|
r1_[2],
|
|
r2_[2];
|
|
|
|
a_.u16 =
|
|
SIMDE_SHUFFLE_VECTOR_(
|
|
16, 64,
|
|
a_.u16, a_.u16,
|
|
0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30,
|
|
1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31
|
|
);
|
|
b_.u16 =
|
|
SIMDE_SHUFFLE_VECTOR_(
|
|
16, 64,
|
|
b_.u16, b_.u16,
|
|
0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30,
|
|
1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31
|
|
);
|
|
|
|
SIMDE_CONVERT_VECTOR_(x1, a_.u16);
|
|
SIMDE_CONVERT_VECTOR_(x2, b_.u16);
|
|
|
|
x1 <<= 16;
|
|
x2 <<= 16;
|
|
|
|
simde_memcpy(&r1_, &x1, sizeof(x1));
|
|
simde_memcpy(&r2_, &x2, sizeof(x2));
|
|
|
|
src_.f32 +=
|
|
HEDLEY_REINTERPRET_CAST(__typeof__(a_.f32), r1_[0].u32) * HEDLEY_REINTERPRET_CAST(__typeof__(a_.f32), r2_[0].u32) +
|
|
HEDLEY_REINTERPRET_CAST(__typeof__(a_.f32), r1_[1].u32) * HEDLEY_REINTERPRET_CAST(__typeof__(a_.f32), r2_[1].u32);
|
|
#else
|
|
SIMDE_VECTORIZE
|
|
for (size_t i = 0 ; i < (sizeof(a_.u16) / sizeof(a_.u16[0])) ; i++) {
|
|
src_.f32[i / 2] += (simde_uint32_as_float32(HEDLEY_STATIC_CAST(uint32_t, a_.u16[i]) << 16) * simde_uint32_as_float32(HEDLEY_STATIC_CAST(uint32_t, b_.u16[i]) << 16));
|
|
}
|
|
#endif
|
|
|
|
return simde__m512_from_private(src_);
|
|
#endif
|
|
}
|
|
#if defined(SIMDE_X86_AVX512BF16_ENABLE_NATIVE_ALIASES)
|
|
#undef _mm512_dpbf16_ps
|
|
#define _mm512_dpbf16_ps(src, a, b) simde_mm512_dpbf16_ps(src, a, b)
|
|
#endif
|
|
|
|
SIMDE_FUNCTION_ATTRIBUTES
|
|
simde__m512
|
|
simde_mm512_mask_dpbf16_ps (simde__m512 src, simde__mmask16 k, simde__m512bh a, simde__m512bh b) {
|
|
#if defined(SIMDE_X86_AVX512BF16_NATIVE)
|
|
return _mm512_mask_dpbf16_ps(src, k, a, b);
|
|
#else
|
|
return simde_mm512_mask_mov_ps(src, k, simde_mm512_dpbf16_ps(src, a, b));
|
|
#endif
|
|
}
|
|
#if defined(SIMDE_X86_AVX512BF16_ENABLE_NATIVE_ALIASES)
|
|
#undef _mm512_mask_dpbf16_ps
|
|
#define _mm512_mask_dpbf16_ps(src, k, a, b) simde_mm512_mask_dpbf16_ps(src, k, a, b)
|
|
#endif
|
|
|
|
SIMDE_FUNCTION_ATTRIBUTES
|
|
simde__m512
|
|
simde_mm512_maskz_dpbf16_ps (simde__mmask16 k, simde__m512 src, simde__m512bh a, simde__m512bh b) {
|
|
#if defined(SIMDE_X86_AVX512BF16_NATIVE)
|
|
return _mm512_maskz_dpbf16_ps(k, src, a, b);
|
|
#else
|
|
return simde_mm512_maskz_mov_ps(k, simde_mm512_dpbf16_ps(src, a, b));
|
|
#endif
|
|
}
|
|
#if defined(SIMDE_X86_AVX512BF16_ENABLE_NATIVE_ALIASES)
|
|
#undef _mm512_maskz_dpbf16_ps
|
|
#define _mm512_maskz_dpbf16_ps(k, src, a, b) simde_mm512_maskz_dpbf16_ps(k, src, a, b)
|
|
#endif
|
|
|
|
SIMDE_END_DECLS_
|
|
HEDLEY_DIAGNOSTIC_POP
|
|
|
|
#endif /* !defined(SIMDE_X86_AVX512_DPBF16_H) */
|