|
| 1 | +// Copyright (C) 2018-2026 - DevSH Graphics Programming Sp. z O.O. |
| 2 | +// This file is part of the "Nabla Engine". |
| 3 | +// For conditions of distribution and use, see copyright notice in nabla.h |
| 4 | +#ifndef _NBL_BUILTIN_HLSL_SAMPLING_SOBOL_INCLUDED_ |
| 5 | +#define _NBL_BUILTIN_HLSL_SAMPLING_SOBOL_INCLUDED_ |
| 6 | + |
| 7 | +#include <nbl/builtin/hlsl/cpp_compat.hlsl> |
| 8 | +#include <nbl/builtin/hlsl/macros.h> |
| 9 | + |
| 10 | +namespace nbl |
| 11 | +{ |
| 12 | +namespace hlsl |
| 13 | +{ |
| 14 | +namespace sampling |
| 15 | +{ |
| 16 | + |
| 17 | +// without acceleration |
| 18 | +struct RowMajorSobolMatrix |
| 19 | +{ |
| 20 | + inline uint16_t operator()(const uint16_t sampleIx) |
| 21 | + { |
| 22 | + // max number bits set is 16, if want to mask then need to be 5 apart |
| 23 | + const uint16_t mask = 0b10000100001u; |
| 24 | + uint16_t val = _static_cast<uint16_t>(spirv::bitCount<uint32_t>(rows[0] & sampleIx)); |
| 25 | + val |= _static_cast<uint16_t>(spirv::bitCount<uint32_t>(rows[5] & sampleIx)) << 5; |
| 26 | + val |= _static_cast<uint16_t>(spirv::bitCount<uint32_t>(rows[10] & sampleIx)) << 10; |
| 27 | + val |= _static_cast<uint16_t>(spirv::bitCount<uint32_t>(rows[15] & sampleIx)) << 15; |
| 28 | + val &= mask; |
| 29 | + NBL_UNROLL |
| 30 | + for (uint16_t i = 1; i < 5; i++) |
| 31 | + { |
| 32 | + uint16_t tmp = _static_cast<uint16_t>(spirv::bitCount<uint32_t>(rows[i] & sampleIx)); |
| 33 | + tmp |= _static_cast<uint16_t>(spirv::bitCount<uint32_t>(rows[5 + i] & sampleIx)) << 5; |
| 34 | + tmp |= _static_cast<uint16_t>(spirv::bitCount<uint32_t>(rows[10 + i] & sampleIx)) << 10; |
| 35 | + val |= (tmp & mask) << i; |
| 36 | + } |
| 37 | + return val; |
| 38 | + } |
| 39 | + |
| 40 | + uint16_t rows[16]; |
| 41 | +}; |
| 42 | + |
| 43 | +// bitcount is expensive, its full 32 bit and needs shifting and masking anyway |
| 44 | +struct ColMajorSobolMatrix |
| 45 | +{ |
| 46 | + inline uint16_t operator()(const uint16_t sampleIx) |
| 47 | + { |
| 48 | + // abuse 2s completement to get a 0xffffu mask when `sampleIx` bit is on |
| 49 | + uint16_t val = cols[0] & ((sampleIx & _static_cast<uint16_t>(0x1u)) - _static_cast<uint16_t>(2)); |
| 50 | + NBL_UNROLL |
| 51 | + for (uint16_t i = 1; i < 16; i++) |
| 52 | + { |
| 53 | + const uint16_t mask = _static_cast<uint16_t>(0x1u) << i; |
| 54 | + val |= cols[i] & ((sampleIx & mask) - _static_cast<uint16_t>(mask + 1)); |
| 55 | + } |
| 56 | + return val; |
| 57 | + } |
| 58 | + |
| 59 | + uint16_t cols[16]; |
| 60 | +}; |
| 61 | + |
| 62 | +} // namespace sampling |
| 63 | +} // namespace hlsl |
| 64 | +} // namespace nbl |
| 65 | + |
| 66 | +#endif |
0 commit comments