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

Changeset View

Standalone View

View Options

intern/cycles/util/half.h

Show All 22 Lines
#if !defined(__KERNEL_GPU__) && defined(__KERNEL_SSE2__) #if !defined(__KERNEL_GPU__) && defined(__KERNEL_SSE2__)
# include "util/simd.h" # include "util/simd.h"
#endif #endif
CCL_NAMESPACE_BEGIN CCL_NAMESPACE_BEGIN
/* Half Floats */ /* Half Floats */
#if defined(__KERNEL_METAL__)
ccl_device_inline float half_to_float(half h_in)
{
float f;
union {
half h;
uint16_t s;
} val;
val.h = h_in;
*((ccl_private int *)&f) = ((val.s & 0x8000) << 16) | (((val.s & 0x7c00) + 0x1C000) << 13) |
((val.s & 0x03FF) << 13);
return f;
}
#else
/* CUDA has its own half data type, no need to define then */ /* CUDA has its own half data type, no need to define then */
#if !defined(__KERNEL_CUDA__) && !defined(__KERNEL_HIP__) # if !defined(__KERNEL_CUDA__) && !defined(__KERNEL_HIP__)
/* Implementing this as a class rather than a typedef so that the compiler can tell it apart from /* Implementing this as a class rather than a typedef so that the compiler can tell it apart from
* unsigned shorts. */ * unsigned shorts. */
class half { class half {
public: public:
half() : v(0) half() : v(0)
{ {
} }
half(const unsigned short &i) : v(i) half(const unsigned short &i) : v(i)
{ {
} }
operator unsigned short() operator unsigned short()
{ {
return v; return v;
} }
half &operator=(const unsigned short &i) half &operator=(const unsigned short &i)
{ {
v = i; v = i;
return *this; return *this;
} }
private: private:
unsigned short v; unsigned short v;
}; };
#endif # endif
struct half4 { struct half4 {
half x, y, z, w; half x, y, z, w;
}; };
#endif
/* Conversion to/from half float for image textures /* Conversion to/from half float for image textures
* *
* Simplified float to half for fast sampling on processor without a native * Simplified float to half for fast sampling on processor without a native
* instruction, and eliminating any NaN and inf values. */ * instruction, and eliminating any NaN and inf values. */
ccl_device_inline half float_to_half_image(float f) ccl_device_inline half float_to_half_image(float f)
{ {
#if defined(__KERNEL_CUDA__) || defined(__KERNEL_HIP__) #if defined(__KERNEL_METAL__)
return half(f);
#elif defined(__KERNEL_CUDA__) || defined(__KERNEL_HIP__)
return __float2half(f); return __float2half(f);
#else #else
const uint u = __float_as_uint(f); const uint u = __float_as_uint(f);
/* Sign bit, shifted to its position. */ /* Sign bit, shifted to its position. */
uint sign_bit = u & 0x80000000; uint sign_bit = u & 0x80000000;
sign_bit >>= 16; sign_bit >>= 16;
/* Exponent. */ /* Exponent. */
uint exponent_bits = u & 0x7f800000; uint exponent_bits = u & 0x7f800000;
Show All 9 Lines#else
value_bits = (exponent_bits == 0 ? 0 : value_bits); value_bits = (exponent_bits == 0 ? 0 : value_bits);
/* Re-insert sign bit and return. */ /* Re-insert sign bit and return. */
return (value_bits | sign_bit); return (value_bits | sign_bit);
#endif #endif
} }
ccl_device_inline float half_to_float_image(half h) ccl_device_inline float half_to_float_image(half h)
{ {
#if defined(__KERNEL_CUDA__) || defined(__KERNEL_HIP__) #if defined(__KERNEL_METAL__)
return half_to_float(h);
#elif defined(__KERNEL_CUDA__) || defined(__KERNEL_HIP__)
return __half2float(h); return __half2float(h);
#else #else
const int x = ((h & 0x8000) << 16) | (((h & 0x7c00) + 0x1C000) << 13) | ((h & 0x03FF) << 13); const int x = ((h & 0x8000) << 16) | (((h & 0x7c00) + 0x1C000) << 13) | ((h & 0x03FF) << 13);
return __int_as_float(x); return __int_as_float(x);
#endif #endif
} }
ccl_device_inline float4 half4_to_float4_image(const half4 h) ccl_device_inline float4 half4_to_float4_image(const half4 h)
Show All 16 Lines
/* Conversion to half float texture for display. /* Conversion to half float texture for display.
* *
* Simplified float to half for fast display texture conversion on processors * Simplified float to half for fast display texture conversion on processors
* without a native instruction. Assumes no negative, no NaN, no inf, and sets * without a native instruction. Assumes no negative, no NaN, no inf, and sets
* denormal to 0. */ * denormal to 0. */
ccl_device_inline half float_to_half_display(const float f) ccl_device_inline half float_to_half_display(const float f)
{ {
#if defined(__KERNEL_CUDA__) || defined(__KERNEL_HIP__) #if defined(__KERNEL_METAL__)
return half(f);
#elif defined(__KERNEL_CUDA__) || defined(__KERNEL_HIP__)
return __float2half(f); return __float2half(f);
#else #else
const int x = __float_as_int((f > 0.0f) ? ((f < 65504.0f) ? f : 65504.0f) : 0.0f); const int x = __float_as_int((f > 0.0f) ? ((f < 65504.0f) ? f : 65504.0f) : 0.0f);
const int absolute = x & 0x7FFFFFFF; const int absolute = x & 0x7FFFFFFF;
const int Z = absolute + 0xC8000000; const int Z = absolute + 0xC8000000;
const int result = (absolute < 0x38800000) ? 0 : Z; const int result = (absolute < 0x38800000) ? 0 : Z;
const int rshift = (result >> 13); const int rshift = (result >> 13);
return (rshift & 0x7FFF); return (rshift & 0x7FFF);
Show All 33 Lines