Page MenuHome
Differential D13361 Diff 45286 intern/cycles/util/math.h

Changeset View

Standalone View

View Options

intern/cycles/util/math.h

Show All 24 Lines
#ifndef __KERNEL_GPU__ #ifndef __KERNEL_GPU__
# include <cmath> # include <cmath>
#endif #endif
#ifdef __HIP__ #ifdef __HIP__
# include <hip/hip_vector_types.h> # include <hip/hip_vector_types.h>
#endif #endif
#if !defined(__KERNEL_METAL__)
#include <float.h> # include <float.h>
#include <math.h> # include <math.h>
#include <stdio.h> # include <stdio.h>
#endif /* !defined(__KERNEL_METAL__) */
#include "util/types.h" #include "util/types.h"
CCL_NAMESPACE_BEGIN CCL_NAMESPACE_BEGIN
/* Float Pi variations */ /* Float Pi variations */
/* Division */ /* Division */
▲ Show 20 Lines • Show All 125 Lines • ▼ Show 20 Linesccl_device_inline float min4(float a, float b, float c, float d)
return min(min(a, b), min(c, d)); return min(min(a, b), min(c, d));
} }
ccl_device_inline float max4(float a, float b, float c, float d) ccl_device_inline float max4(float a, float b, float c, float d)
{ {
return max(max(a, b), max(c, d)); return max(max(a, b), max(c, d));
} }
#if !defined(__KERNEL_METAL__)
/* Int/Float conversion */ /* Int/Float conversion */
ccl_device_inline int as_int(uint i) ccl_device_inline int as_int(uint i)
{ {
union { union {
uint ui; uint ui;
int i; int i;
} u; } u;
Show All 16 Linesccl_device_inline uint as_uint(float f)
union { union {
uint i; uint i;
float f; float f;
} u; } u;
u.f = f; u.f = f;
return u.i; return u.i;
} }
#ifndef __HIP__ # ifndef __HIP__
ccl_device_inline int __float_as_int(float f) ccl_device_inline int __float_as_int(float f)
{ {
union { union {
int i; int i;
float f; float f;
} u; } u;
u.f = f; u.f = f;
return u.i; return u.i;
Show All 23 Lines
{ {
union { union {
uint i; uint i;
float f; float f;
} u; } u;
u.i = i; u.i = i;
return u.f; return u.f;
} }
#endif # endif
ccl_device_inline int4 __float4_as_int4(float4 f) ccl_device_inline int4 __float4_as_int4(float4 f)
{ {
#ifdef __KERNEL_SSE__ # ifdef __KERNEL_SSE__
return int4(_mm_castps_si128(f.m128)); return int4(_mm_castps_si128(f.m128));
#else # else
return make_int4( return make_int4(
__float_as_int(f.x), __float_as_int(f.y), __float_as_int(f.z), __float_as_int(f.w)); __float_as_int(f.x), __float_as_int(f.y), __float_as_int(f.z), __float_as_int(f.w));
#endif # endif
} }
ccl_device_inline float4 __int4_as_float4(int4 i) ccl_device_inline float4 __int4_as_float4(int4 i)
{ {
#ifdef __KERNEL_SSE__ # ifdef __KERNEL_SSE__
return float4(_mm_castsi128_ps(i.m128)); return float4(_mm_castsi128_ps(i.m128));
#else # else
return make_float4( return make_float4(
__int_as_float(i.x), __int_as_float(i.y), __int_as_float(i.z), __int_as_float(i.w)); __int_as_float(i.x), __int_as_float(i.y), __int_as_float(i.z), __int_as_float(i.w));
#endif # endif
} }
#endif /* !defined(__KERNEL_METAL__) */
#if defined(__KERNEL_METAL__)
# define isnan_safe(v) isnan(v)
# define isfinite_safe(v) isfinite(v)
#else
template<typename T> ccl_device_inline uint pointer_pack_to_uint_0(T *ptr) template<typename T> ccl_device_inline uint pointer_pack_to_uint_0(T *ptr)
{ {
return ((uint64_t)ptr) & 0xFFFFFFFF; return ((uint64_t)ptr) & 0xFFFFFFFF;
} }
template<typename T> ccl_device_inline uint pointer_pack_to_uint_1(T *ptr) template<typename T> ccl_device_inline uint pointer_pack_to_uint_1(T *ptr)
{ {
return (((uint64_t)ptr) >> 32) & 0xFFFFFFFF; return (((uint64_t)ptr) >> 32) & 0xFFFFFFFF;
Show All 27 Lines
} }
ccl_device_inline bool isfinite_safe(float f) ccl_device_inline bool isfinite_safe(float f)
{ {
/* By IEEE 754 rule, 2*Inf equals Inf */ /* By IEEE 754 rule, 2*Inf equals Inf */
unsigned int x = __float_as_uint(f); unsigned int x = __float_as_uint(f);
return (f == f) && (x == 0 || x == (1u << 31) || (f != 2.0f * f)) && !((x << 1) > 0xff000000u); return (f == f) && (x == 0 || x == (1u << 31) || (f != 2.0f * f)) && !((x << 1) > 0xff000000u);
} }
#endif
ccl_device_inline float ensure_finite(float v) ccl_device_inline float ensure_finite(float v)
{ {
return isfinite_safe(v) ? v : 0.0f; return isfinite_safe(v) ? v : 0.0f;
} }
#if !defined(__KERNEL_METAL__)
ccl_device_inline int clamp(int a, int mn, int mx) ccl_device_inline int clamp(int a, int mn, int mx)
{ {
return min(max(a, mn), mx); return min(max(a, mn), mx);
} }
ccl_device_inline float clamp(float a, float mn, float mx) ccl_device_inline float clamp(float a, float mn, float mx)
{ {
return min(max(a, mn), mx); return min(max(a, mn), mx);
Show All 13 Lines else if (x >= edge1)
result = 1.0f; result = 1.0f;
else { else {
float t = (x - edge0) / (edge1 - edge0); float t = (x - edge0) / (edge1 - edge0);
result = (3.0f - 2.0f * t) * (t * t); result = (3.0f - 2.0f * t) * (t * t);
} }
return result; return result;
} }
#ifndef __KERNEL_CUDA__ #endif /* !defined(__KERNEL_METAL__) */
#if defined(__KERNEL_CUDA__)
ccl_device_inline float saturatef(float a) ccl_device_inline float saturatef(float a)
{ {
return clamp(a, 0.0f, 1.0f); return __saturatef(a);
} }
#else #elif !defined(__KERNEL_METAL__)
ccl_device_inline float saturatef(float a) ccl_device_inline float saturatef(float a)
{ {
return __saturatef(a); return clamp(a, 0.0f, 1.0f);
} }
#endif /* __KERNEL_CUDA__ */ #endif /* __KERNEL_CUDA__ */
ccl_device_inline int float_to_int(float f) ccl_device_inline int float_to_int(float f)
{ {
return (int)f; return (int)f;
} }
▲ Show 20 Lines • Show All 120 Lines • ▼ Show 20 Lines
#include "util/math_float2.h" #include "util/math_float2.h"
#include "util/math_float3.h" #include "util/math_float3.h"
#include "util/math_float4.h" #include "util/math_float4.h"
#include "util/rect.h" #include "util/rect.h"
CCL_NAMESPACE_BEGIN CCL_NAMESPACE_BEGIN
#if !defined(__KERNEL_METAL__)
/* Interpolation */ /* Interpolation */
template<class A, class B> A lerp(const A &a, const A &b, const B &t) template<class A, class B> A lerp(const A &a, const A &b, const B &t)
{ {
return (A)(a * ((B)1 - t) + b * t); return (A)(a * ((B)1 - t) + b * t);
} }
#endif /* __KERNEL_METAL__ */
/* Triangle */ /* Triangle */
ccl_device_inline float triangle_area(ccl_private const float3 &v1, ccl_device_inline float triangle_area(ccl_private const float3 &v1,
ccl_private const float3 &v2, ccl_private const float3 &v2,
ccl_private const float3 &v3) ccl_private const float3 &v3)
{ {
return len(cross(v3 - v2, v1 - v2)) * 0.5f; return len(cross(v3 - v2, v1 - v2)) * 0.5f;
} }
▲ Show 20 Lines • Show All 113 Lines • ▼ Show 20 Lines
ccl_device_inline float safe_sqrtf(float f) ccl_device_inline float safe_sqrtf(float f)
{ {
return sqrtf(max(f, 0.0f)); return sqrtf(max(f, 0.0f));
} }
ccl_device_inline float inversesqrtf(float f) ccl_device_inline float inversesqrtf(float f)
{ {
#if defined(__KERNEL_METAL__)
return (f > 0.0f) ? rsqrt(f) : 0.0f;
#else
return (f > 0.0f) ? 1.0f / sqrtf(f) : 0.0f; return (f > 0.0f) ? 1.0f / sqrtf(f) : 0.0f;
#endif
} }
ccl_device float safe_asinf(float a) ccl_device float safe_asinf(float a)
{ {
return asinf(clamp(a, -1.0f, 1.0f)); return asinf(clamp(a, -1.0f, 1.0f));
} }
ccl_device float safe_acosf(float a) ccl_device float safe_acosf(float a)
▲ Show 20 Lines • Show All 71 Lines • ▼ Show 20 Linesccl_device_inline float xor_signmask(float x, int y)
return __int_as_float(__float_as_int(x) ^ y); return __int_as_float(__float_as_int(x) ^ y);
} }
ccl_device float bits_to_01(uint bits) ccl_device float bits_to_01(uint bits)
{ {
return bits * (1.0f / (float)0xFFFFFFFF); return bits * (1.0f / (float)0xFFFFFFFF);
} }
#if !defined(__KERNEL_GPU__)
# if defined(__GNUC__)
# define popcount(x) __builtin_popcount(x)
# else
ccl_device_inline uint popcount(uint x)
{
/* TODO(Stefan): pop-count intrinsic for Windows with fallback for older CPUs. */
uint i = x & 0xaaaaaaaa;
i = i - ((i >> 1) & 0x55555555);
i = (i & 0x33333333) + ((i >> 2) & 0x33333333);
i = (((i + (i >> 4)) & 0xF0F0F0F) * 0x1010101) >> 24;
return i & 1;
}
# endif
#elif !defined(__KERNEL_METAL__)
# define popcount(x) __popc(x)
#endif
ccl_device_inline uint count_leading_zeros(uint x) ccl_device_inline uint count_leading_zeros(uint x)
{ {
#if defined(__KERNEL_CUDA__) || defined(__KERNEL_OPTIX__) || defined(__KERNEL_HIP__) #if defined(__KERNEL_CUDA__) || defined(__KERNEL_OPTIX__) || defined(__KERNEL_HIP__)
return __clz(x); return __clz(x);
#elif defined(__KERNEL_METAL__)
return clz(x);
#else #else
assert(x != 0); assert(x != 0);
# ifdef _MSC_VER # ifdef _MSC_VER
unsigned long leading_zero = 0; unsigned long leading_zero = 0;
_BitScanReverse(&leading_zero, x); _BitScanReverse(&leading_zero, x);
return (31 - leading_zero); return (31 - leading_zero);
# else # else
return __builtin_clz(x); return __builtin_clz(x);
# endif # endif
#endif #endif
} }
ccl_device_inline uint count_trailing_zeros(uint x) ccl_device_inline uint count_trailing_zeros(uint x)
{ {
#if defined(__KERNEL_CUDA__) || defined(__KERNEL_OPTIX__) || defined(__KERNEL_HIP__) #if defined(__KERNEL_CUDA__) || defined(__KERNEL_OPTIX__) || defined(__KERNEL_HIP__)
return (__ffs(x) - 1); return (__ffs(x) - 1);
#elif defined(__KERNEL_METAL__)
return ctz(x);
#else #else
assert(x != 0); assert(x != 0);
# ifdef _MSC_VER # ifdef _MSC_VER
unsigned long ctz = 0; unsigned long ctz = 0;
_BitScanForward(&ctz, x); _BitScanForward(&ctz, x);
return ctz; return ctz;
# else # else
return __builtin_ctz(x); return __builtin_ctz(x);
# endif # endif
#endif #endif
} }
ccl_device_inline uint find_first_set(uint x) ccl_device_inline uint find_first_set(uint x)
{ {
#if defined(__KERNEL_CUDA__) || defined(__KERNEL_OPTIX__) || defined(__KERNEL_HIP__) #if defined(__KERNEL_CUDA__) || defined(__KERNEL_OPTIX__) || defined(__KERNEL_HIP__)
return __ffs(x); return __ffs(x);
#elif defined(__KERNEL_METAL__)
return (x != 0) ? ctz(x) + 1 : 0;
#else #else
# ifdef _MSC_VER # ifdef _MSC_VER
return (x != 0) ? (32 - count_leading_zeros(x & (-x))) : 0; return (x != 0) ? (32 - count_leading_zeros(x & (-x))) : 0;
# else # else
return __builtin_ffs(x); return __builtin_ffs(x);
# endif # endif
#endif #endif
} }
▲ Show 20 Lines • Show All 82 Lines • ▼ Show 20 Lines
ccl_device_inline uint32_t reverse_integer_bits(uint32_t x) ccl_device_inline uint32_t reverse_integer_bits(uint32_t x)
{ {
/* Use a native instruction if it exists. */ /* Use a native instruction if it exists. */
#if defined(__arm__) || defined(__aarch64__) #if defined(__arm__) || defined(__aarch64__)
__asm__("rbit %w0, %w1" : "=r"(x) : "r"(x)); __asm__("rbit %w0, %w1" : "=r"(x) : "r"(x));
return x; return x;
#elif defined(__KERNEL_CUDA__) #elif defined(__KERNEL_CUDA__)
return __brev(x); return __brev(x);
#elif defined(__KERNEL_METAL__)
return reverse_bits(x);
#elif __has_builtin(__builtin_bitreverse32) #elif __has_builtin(__builtin_bitreverse32)
return __builtin_bitreverse32(x); return __builtin_bitreverse32(x);
#else #else
/* Flip pairwise. */ /* Flip pairwise. */
x = ((x & 0x55555555) << 1) | ((x & 0xAAAAAAAA) >> 1); x = ((x & 0x55555555) << 1) | ((x & 0xAAAAAAAA) >> 1);
/* Flip pairs. */ /* Flip pairs. */
x = ((x & 0x33333333) << 2) | ((x & 0xCCCCCCCC) >> 2); x = ((x & 0x33333333) << 2) | ((x & 0xCCCCCCCC) >> 2);
/* Flip nibbles. */ /* Flip nibbles. */
Show All 16 Lines