diff options
Diffstat (limited to 'intern/cycles/util/math_int4.h')
-rw-r--r-- | intern/cycles/util/math_int4.h | 216 |
1 files changed, 185 insertions, 31 deletions
diff --git a/intern/cycles/util/math_int4.h b/intern/cycles/util/math_int4.h index fbdada223cb..c6d767d7587 100644 --- a/intern/cycles/util/math_int4.h +++ b/intern/cycles/util/math_int4.h @@ -1,4 +1,5 @@ /* SPDX-License-Identifier: Apache-2.0 + * Copyright 2011-2013 Intel Corporation * Copyright 2011-2022 Blender Foundation */ #ifndef __UTIL_MATH_INT4_H__ @@ -10,30 +11,8 @@ CCL_NAMESPACE_BEGIN -/******************************************************************************* - * Declaration. - */ - #ifndef __KERNEL_GPU__ -ccl_device_inline int4 operator+(const int4 &a, const int4 &b); -ccl_device_inline int4 operator+=(int4 &a, const int4 &b); -ccl_device_inline int4 operator>>(const int4 &a, int i); -ccl_device_inline int4 operator<<(const int4 &a, int i); -ccl_device_inline int4 operator<(const int4 &a, const int4 &b); -ccl_device_inline int4 operator>=(const int4 &a, const int4 &b); -ccl_device_inline int4 operator&(const int4 &a, const int4 &b); -ccl_device_inline int4 min(int4 a, int4 b); -ccl_device_inline int4 max(int4 a, int4 b); -ccl_device_inline int4 clamp(const int4 &a, const int4 &mn, const int4 &mx); -ccl_device_inline int4 select(const int4 &mask, const int4 &a, const int4 &b); -#endif /* __KERNEL_GPU__ */ - -/******************************************************************************* - * Definition. - */ - -#ifndef __KERNEL_GPU__ -ccl_device_inline int4 operator+(const int4 &a, const int4 &b) +ccl_device_inline int4 operator+(const int4 a, const int4 b) { # ifdef __KERNEL_SSE__ return int4(_mm_add_epi32(a.m128, b.m128)); @@ -42,12 +21,26 @@ ccl_device_inline int4 operator+(const int4 &a, const int4 &b) # endif } -ccl_device_inline int4 operator+=(int4 &a, const int4 &b) +ccl_device_inline int4 operator+=(int4 &a, const int4 b) { return a = a + b; } -ccl_device_inline int4 operator>>(const int4 &a, int i) +ccl_device_inline int4 operator-(const int4 a, const int4 b) +{ +# ifdef __KERNEL_SSE__ + return int4(_mm_sub_epi32(a.m128, b.m128)); +# else + return make_int4(a.x - b.x, a.y - b.y, a.z - b.z, a.w - b.w); +# endif +} + +ccl_device_inline int4 operator-=(int4 &a, const int4 b) +{ + return a = a - b; +} + +ccl_device_inline int4 operator>>(const int4 a, int i) { # ifdef __KERNEL_SSE__ return int4(_mm_srai_epi32(a.m128, i)); @@ -56,7 +49,7 @@ ccl_device_inline int4 operator>>(const int4 &a, int i) # endif } -ccl_device_inline int4 operator<<(const int4 &a, int i) +ccl_device_inline int4 operator<<(const int4 a, int i) { # ifdef __KERNEL_SSE__ return int4(_mm_slli_epi32(a.m128, i)); @@ -65,7 +58,7 @@ ccl_device_inline int4 operator<<(const int4 &a, int i) # endif } -ccl_device_inline int4 operator<(const int4 &a, const int4 &b) +ccl_device_inline int4 operator<(const int4 a, const int4 b) { # ifdef __KERNEL_SSE__ return int4(_mm_cmplt_epi32(a.m128, b.m128)); @@ -74,7 +67,26 @@ ccl_device_inline int4 operator<(const int4 &a, const int4 &b) # endif } -ccl_device_inline int4 operator>=(const int4 &a, const int4 &b) +ccl_device_inline int4 operator<(const int4 a, const int b) +{ + return a < make_int4(b); +} + +ccl_device_inline int4 operator==(const int4 a, const int4 b) +{ +# ifdef __KERNEL_SSE__ + return int4(_mm_cmpeq_epi32(a.m128, b.m128)); +# else + return make_int4(a.x == b.x, a.y == b.y, a.z == b.z, a.w == b.w); +# endif +} + +ccl_device_inline int4 operator==(const int4 a, const int b) +{ + return a == make_int4(b); +} + +ccl_device_inline int4 operator>=(const int4 a, const int4 b) { # ifdef __KERNEL_SSE__ return int4(_mm_xor_si128(_mm_set1_epi32(0xffffffff), _mm_cmplt_epi32(a.m128, b.m128))); @@ -83,7 +95,12 @@ ccl_device_inline int4 operator>=(const int4 &a, const int4 &b) # endif } -ccl_device_inline int4 operator&(const int4 &a, const int4 &b) +ccl_device_inline int4 operator>=(const int4 a, const int b) +{ + return a >= make_int4(b); +} + +ccl_device_inline int4 operator&(const int4 a, const int4 b) { # ifdef __KERNEL_SSE__ return int4(_mm_and_si128(a.m128, b.m128)); @@ -92,6 +109,97 @@ ccl_device_inline int4 operator&(const int4 &a, const int4 &b) # endif } +ccl_device_inline int4 operator|(const int4 a, const int4 b) +{ +# ifdef __KERNEL_SSE__ + return int4(_mm_or_si128(a.m128, b.m128)); +# else + return make_int4(a.x | b.x, a.y | b.y, a.z | b.z, a.w | b.w); +# endif +} + +ccl_device_inline int4 operator^(const int4 a, const int4 b) +{ +# ifdef __KERNEL_SSE__ + return int4(_mm_xor_si128(a.m128, b.m128)); +# else + return make_int4(a.x ^ b.x, a.y ^ b.y, a.z ^ b.z, a.w ^ b.w); +# endif +} + +ccl_device_inline int4 operator&(const int32_t a, const int4 b) +{ + return make_int4(a) & b; +} + +ccl_device_inline int4 operator&(const int4 a, const int32_t b) +{ + return a & make_int4(b); +} + +ccl_device_inline int4 operator|(const int32_t a, const int4 b) +{ + return make_int4(a) | b; +} + +ccl_device_inline int4 operator|(const int4 a, const int32_t b) +{ + return a | make_int4(b); +} + +ccl_device_inline int4 operator^(const int32_t a, const int4 b) +{ + return make_int4(a) ^ b; +} + +ccl_device_inline int4 operator^(const int4 a, const int32_t b) +{ + return a ^ make_int4(b); +} + +ccl_device_inline int4 &operator&=(int4 &a, const int4 b) +{ + return a = a & b; +} +ccl_device_inline int4 &operator&=(int4 &a, const int32_t b) +{ + return a = a & b; +} + +ccl_device_inline int4 &operator|=(int4 &a, const int4 b) +{ + return a = a | b; +} +ccl_device_inline int4 &operator|=(int4 &a, const int32_t b) +{ + return a = a | b; +} + +ccl_device_inline int4 &operator^=(int4 &a, const int4 b) +{ + return a = a ^ b; +} +ccl_device_inline int4 &operator^=(int4 &a, const int32_t b) +{ + return a = a ^ b; +} + +ccl_device_inline int4 &operator<<=(int4 &a, const int32_t b) +{ + return a = a << b; +} +ccl_device_inline int4 &operator>>=(int4 &a, const int32_t b) +{ + return a = a >> b; +} + +# ifdef __KERNEL_SSE__ +ccl_device_forceinline const int4 srl(const int4 a, const int32_t b) +{ + return int4(_mm_srli_epi32(a.m128, b)); +} +# endif + ccl_device_inline int4 min(int4 a, int4 b) { # if defined(__KERNEL_SSE__) && defined(__KERNEL_SSE41__) @@ -110,12 +218,12 @@ ccl_device_inline int4 max(int4 a, int4 b) # endif } -ccl_device_inline int4 clamp(const int4 &a, const int4 &mn, const int4 &mx) +ccl_device_inline int4 clamp(const int4 a, const int4 mn, const int4 mx) { return min(max(a, mn), mx); } -ccl_device_inline int4 select(const int4 &mask, const int4 &a, const int4 &b) +ccl_device_inline int4 select(const int4 mask, const int4 a, const int4 b) { # ifdef __KERNEL_SSE__ return int4(_mm_or_si128(_mm_and_si128(mask, a), _mm_andnot_si128(mask, b))); @@ -135,6 +243,52 @@ ccl_device_inline int4 load_int4(const int *v) } #endif /* __KERNEL_GPU__ */ +ccl_device_inline float4 cast(const int4 a) +{ +#ifdef __KERNEL_SSE__ + return float4(_mm_castsi128_ps(a)); +#else + return make_float4( + __int_as_float(a.x), __int_as_float(a.y), __int_as_float(a.z), __int_as_float(a.w)); +#endif +} + +#ifdef __KERNEL_SSE__ +ccl_device_forceinline int4 andnot(const int4 a, const int4 b) +{ + return int4(_mm_andnot_si128(a.m128, b.m128)); +} + +template<size_t i0, size_t i1, size_t i2, size_t i3> +ccl_device_forceinline int4 shuffle(const int4 a) +{ +# ifdef __KERNEL_NEON__ + int32x4_t result = shuffle_neon<int32x4_t, i0, i1, i2, i3>(vreinterpretq_s32_m128i(a)); + return int4(vreinterpretq_m128i_s32(result)); +# else + return int4(_mm_shuffle_epi32(a, _MM_SHUFFLE(i3, i2, i1, i0))); +# endif +} + +template<size_t i0, size_t i1, size_t i2, size_t i3> +ccl_device_forceinline int4 shuffle(const int4 a, const int4 b) +{ +# ifdef __KERNEL_NEON__ + int32x4_t result = shuffle_neon<int32x4_t, i0, i1, i2, i3>(vreinterpretq_s32_m128i(a), + vreinterpretq_s32_m128i(b)); + return int4(vreinterpretq_m128i_s32(result)); +# else + return int4(_mm_castps_si128( + _mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), _MM_SHUFFLE(i3, i2, i1, i0)))); +# endif +} + +template<size_t i0> ccl_device_forceinline int4 shuffle(const int4 b) +{ + return shuffle<i0, i0, i0, i0>(b); +} +#endif + CCL_NAMESPACE_END #endif /* __UTIL_MATH_INT4_H__ */ |