#include "zeus/Math.hpp" #include "ANIM.hpp" #define DUMP_KEYS 0 namespace DataSpec::DNAANIM { size_t ComputeBitstreamSize(size_t keyFrameCount, const std::vector& channels) { size_t bitsPerKeyFrame = 0; for (const Channel& chan : channels) { switch (chan.type) { case Channel::Type::Rotation: bitsPerKeyFrame += 1; [[fallthrough]]; case Channel::Type::Translation: case Channel::Type::Scale: bitsPerKeyFrame += chan.q[0]; bitsPerKeyFrame += chan.q[1]; bitsPerKeyFrame += chan.q[2]; break; case Channel::Type::KfHead: bitsPerKeyFrame += 1; break; case Channel::Type::RotationMP3: bitsPerKeyFrame += chan.q[0]; bitsPerKeyFrame += chan.q[1]; bitsPerKeyFrame += chan.q[2]; bitsPerKeyFrame += chan.q[3]; break; default: break; } } return (bitsPerKeyFrame * keyFrameCount + 31) / 32 * 4; } static inline QuantizedRot QuantizeRotation(const Value& quat, atUint32 div) { float q = float(div) / (M_PIF / 2.0f); zeus::simd_floats f(quat.simd); assert(std::abs(f[1]) <= 1.f && "Out of range quat X component"); assert(std::abs(f[2]) <= 1.f && "Out of range quat Y component"); assert(std::abs(f[3]) <= 1.f && "Out of range quat Z component"); return {{ atInt32(std::asin(f[1]) * q), atInt32(std::asin(f[2]) * q), atInt32(std::asin(f[3]) * q), }, (f[0] < 0.f)}; } static inline Value DequantizeRotation(const QuantizedRot& v, atUint32 div) { float q = (M_PIF / 2.0f) / float(div); athena::simd_floats f = { 0.0f, std::sin(v.v[0] * q), std::sin(v.v[1] * q), std::sin(v.v[2] * q), }; f[0] = std::sqrt(std::max((1.0f - (f[1] * f[1] + f[2] * f[2] + f[3] * f[3])), 0.0f)); f[0] = v.w ? -f[0] : f[0]; Value retval; retval.simd.copy_from(f); return retval; } static inline Value DequantizeRotation_3(const QuantizedRot& v, atUint32 div) { float q = 1.0f / float(div); athena::simd_floats f = { 0.0f, v.v[0] * q, v.v[1] * q, v.v[2] * q, }; f[0] = std::sqrt(std::max((1.0f - (f[1] * f[1] + f[2] * f[2] + f[3] * f[3])), 0.0f)); f[0] = v.w ? -f[0] : f[0]; Value retval; retval.simd.copy_from(f); return retval; } bool BitstreamReader::dequantizeBit(const atUint8* data) { atUint32 byteCur = (m_bitCur / 32) * 4; atUint32 bitRem = m_bitCur % 32; /* Fill 32 bit buffer with region containing bits */ /* Make them least significant */ atUint32 tempBuf = hecl::SBig(*reinterpret_cast(data + byteCur)) >> bitRem; /* That's it */ m_bitCur += 1; return tempBuf & 0x1; } atInt32 BitstreamReader::dequantize(const atUint8* data, atUint8 q) { atUint32 byteCur = (m_bitCur / 32) * 4; atUint32 bitRem = m_bitCur % 32; /* Fill 32 bit buffer with region containing bits */ /* Make them least significant */ atUint32 tempBuf = hecl::SBig(*reinterpret_cast(data + byteCur)) >> bitRem; /* If this shift underflows the value, buffer the next 32 bits */ /* And tack onto shifted buffer */ if ((bitRem + q) > 32) { atUint32 tempBuf2 = hecl::SBig(*reinterpret_cast(data + byteCur + 4)); tempBuf |= (tempBuf2 << (32 - bitRem)); } /* Mask it */ atUint32 mask = (1 << q) - 1; tempBuf &= mask; /* Sign extend */ atUint32 sign = (tempBuf >> (q - 1)) & 0x1; if (sign) tempBuf |= ~0u << q; /* Return delta value */ m_bitCur += q; return atInt32(tempBuf); } std::vector> BitstreamReader::read(const atUint8* data, size_t keyFrameCount, const std::vector& channels, atUint32 rotDiv, float transMult, float scaleMult) { m_bitCur = 0; std::vector> chanKeys; std::vector chanAccum; chanKeys.reserve(channels.size()); chanAccum.reserve(channels.size()); for (const Channel& chan : channels) { chanAccum.push_back(chan.i); chanKeys.emplace_back(); std::vector& keys = chanKeys.back(); keys.reserve(keyFrameCount); switch (chan.type) { case Channel::Type::Rotation: { QuantizedRot qr = {{chan.i[0], chan.i[1], chan.i[2]}, false}; keys.emplace_back(DequantizeRotation(qr, rotDiv)); break; } case Channel::Type::Translation: { keys.push_back({chan.i[0] * transMult, chan.i[1] * transMult, chan.i[2] * transMult}); break; } case Channel::Type::Scale: { keys.push_back({chan.i[0] * scaleMult, chan.i[1] * scaleMult, chan.i[2] * scaleMult}); break; } case Channel::Type::KfHead: { break; } case Channel::Type::RotationMP3: { QuantizedRot qr = {{chan.i[1], chan.i[2], chan.i[3]}, bool(chan.i[0] & 0x1)}; keys.emplace_back(DequantizeRotation_3(qr, rotDiv)); break; } default: break; } } for (size_t f = 0; f < keyFrameCount; ++f) { #if DUMP_KEYS fprintf(stderr, "\nFRAME %" PRISize " %u %u\n", f, (m_bitCur / 32) * 4, m_bitCur % 32); int lastId = -1; #endif auto kit = chanKeys.begin(); auto ait = chanAccum.begin(); for (const Channel& chan : channels) { #if DUMP_KEYS if (chan.id != lastId) { lastId = chan.id; fprintf(stderr, "\n"); } #endif QuantizedValue& p = *ait; switch (chan.type) { case Channel::Type::Rotation: { bool wBit = dequantizeBit(data); p[0] += dequantize(data, chan.q[0]); p[1] += dequantize(data, chan.q[1]); p[2] += dequantize(data, chan.q[2]); QuantizedRot qr = {{p[0], p[1], p[2]}, wBit}; kit->emplace_back(DequantizeRotation(qr, rotDiv)); #if DUMP_KEYS fprintf(stderr, "%d R: %d %d %d %d\t", chan.id, wBit, p[0], p[1], p[2]); #endif break; } case Channel::Type::Translation: { atInt32 val1 = dequantize(data, chan.q[0]); p[0] += val1; atInt32 val2 = dequantize(data, chan.q[1]); p[1] += val2; atInt32 val3 = dequantize(data, chan.q[2]); p[2] += val3; kit->push_back({p[0] * transMult, p[1] * transMult, p[2] * transMult}); #if DUMP_KEYS fprintf(stderr, "%d T: %d %d %d\t", chan.id, p[0], p[1], p[2]); #endif break; } case Channel::Type::Scale: { p[0] += dequantize(data, chan.q[0]); p[1] += dequantize(data, chan.q[1]); p[2] += dequantize(data, chan.q[2]); kit->push_back({p[0] * scaleMult, p[1] * scaleMult, p[2] * scaleMult}); #if DUMP_KEYS fprintf(stderr, "%d S: %d %d %d\t", chan.id, p[0], p[1], p[2]); #endif break; } case Channel::Type::KfHead: { dequantizeBit(data); break; } case Channel::Type::RotationMP3: { atInt32 val1 = dequantize(data, chan.q[0]); p[0] += val1; atInt32 val2 = dequantize(data, chan.q[1]); p[1] += val2; atInt32 val3 = dequantize(data, chan.q[2]); p[2] += val3; atInt32 val4 = dequantize(data, chan.q[3]); p[3] += val4; QuantizedRot qr = {{p[1], p[2], p[3]}, bool(p[0] & 0x1)}; kit->emplace_back(DequantizeRotation_3(qr, rotDiv)); break; } default: break; } ++kit; ++ait; } #if DUMP_KEYS fprintf(stderr, "\n"); #endif } return chanKeys; } void BitstreamWriter::quantizeBit(atUint8* data, bool val) { atUint32 byteCur = (m_bitCur / 32) * 4; atUint32 bitRem = m_bitCur % 32; /* Fill 32 bit buffer with region containing bits */ /* Make them least significant */ *(atUint32*)(data + byteCur) = hecl::SBig(hecl::SBig(*(atUint32*)(data + byteCur)) | (val << bitRem)); m_bitCur += 1; } void BitstreamWriter::quantize(atUint8* data, atUint8 q, atInt32 val) { atUint32 byteCur = (m_bitCur / 32) * 4; atUint32 bitRem = m_bitCur % 32; atUint32 masked = val & ((1 << q) - 1); assert(((((val >> 31) & 0x1) == 0x1) || (((masked >> (q - 1)) & 0x1) == 0)) && "Twos compliment fail"); /* Fill 32 bit buffer with region containing bits */ /* Make them least significant */ *(atUint32*)(data + byteCur) = hecl::SBig(hecl::SBig(*(atUint32*)(data + byteCur)) | (masked << bitRem)); /* If this shift underflows the value, buffer the next 32 bits */ /* And tack onto shifted buffer */ if ((bitRem + q) > 32) { *(atUint32*)(data + byteCur + 4) = hecl::SBig(hecl::SBig(*(atUint32*)(data + byteCur + 4)) | (masked >> (32 - bitRem))); } m_bitCur += q; } std::unique_ptr BitstreamWriter::write(const std::vector>& chanKeys, size_t keyFrameCount, std::vector& channels, atUint32 quantRange, atUint32& rotDivOut, float& transMultOut, float& scaleMultOut, size_t& sizeOut) { m_bitCur = 0; rotDivOut = quantRange; /* Normalized range of values */ float quantRangeF = float(quantRange); /* Pre-pass to calculate translation multiplier */ float maxTransDelta = 0.0f; float maxScaleDelta = 0.0f; auto kit = chanKeys.begin(); for (Channel& chan : channels) { switch (chan.type) { case Channel::Type::Translation: { zeus::simd lastVal = {}; for (auto it = kit->begin(); it != kit->end(); ++it) { const Value* key = &*it; zeus::simd_floats f(key->simd - lastVal); lastVal = key->simd; maxTransDelta = std::max(maxTransDelta, std::fabs(f[0])); maxTransDelta = std::max(maxTransDelta, std::fabs(f[1])); maxTransDelta = std::max(maxTransDelta, std::fabs(f[2])); } break; } case Channel::Type::Scale: { zeus::simd lastVal = {}; for (auto it = kit->begin(); it != kit->end(); ++it) { const Value* key = &*it; zeus::simd_floats f(key->simd - lastVal); lastVal = key->simd; maxScaleDelta = std::max(maxScaleDelta, std::fabs(f[0])); maxScaleDelta = std::max(maxScaleDelta, std::fabs(f[1])); maxScaleDelta = std::max(maxScaleDelta, std::fabs(f[2])); } break; } default: break; } ++kit; } transMultOut = maxTransDelta / quantRangeF + FLT_EPSILON; scaleMultOut = maxScaleDelta / quantRangeF + FLT_EPSILON; /* Output channel inits */ std::vector initVals; initVals.reserve(channels.size()); kit = chanKeys.begin(); for (Channel& chan : channels) { chan.q[0] = 1; chan.q[1] = 1; chan.q[2] = 1; switch (chan.type) { case Channel::Type::Rotation: { QuantizedRot qr = QuantizeRotation((*kit)[0], rotDivOut); chan.i = qr.v; initVals.push_back(chan.i); break; } case Channel::Type::Translation: { zeus::simd_floats f((*kit)[0].simd); chan.i = {atInt32(f[0] / transMultOut), atInt32(f[1] / transMultOut), atInt32(f[2] / transMultOut)}; initVals.push_back(chan.i); break; } case Channel::Type::Scale: { zeus::simd_floats f((*kit)[0].simd); chan.i = {atInt32(f[0] / scaleMultOut), atInt32(f[1] / scaleMultOut), atInt32(f[2] / scaleMultOut)}; initVals.push_back(chan.i); break; } default: break; } ++kit; } /* Pre-pass to analyze quantization factors for channels */ std::vector lastVals = initVals; kit = chanKeys.begin(); auto vit = lastVals.begin(); for (Channel& chan : channels) { QuantizedValue& last = *vit++; switch (chan.type) { case Channel::Type::Rotation: { for (auto it = kit->begin() + 1; it != kit->end(); ++it) { QuantizedRot qrCur = QuantizeRotation(*it, rotDivOut); chan.q[0] = std::max(chan.q[0], atUint8(qrCur.v.qFrom(last, 0))); chan.q[1] = std::max(chan.q[1], atUint8(qrCur.v.qFrom(last, 1))); chan.q[2] = std::max(chan.q[2], atUint8(qrCur.v.qFrom(last, 2))); last = qrCur.v; } break; } case Channel::Type::Translation: { for (auto it = kit->begin() + 1; it != kit->end(); ++it) { zeus::simd_floats f(it->simd); QuantizedValue cur = {atInt32(f[0] / transMultOut), atInt32(f[1] / transMultOut), atInt32(f[2] / transMultOut)}; chan.q[0] = std::max(chan.q[0], atUint8(cur.qFrom(last, 0))); chan.q[1] = std::max(chan.q[1], atUint8(cur.qFrom(last, 1))); chan.q[2] = std::max(chan.q[2], atUint8(cur.qFrom(last, 2))); last = cur; } break; } case Channel::Type::Scale: { for (auto it = kit->begin() + 1; it != kit->end(); ++it) { zeus::simd_floats f(it->simd); QuantizedValue cur = {atInt32(f[0] / scaleMultOut), atInt32(f[1] / scaleMultOut), atInt32(f[2] / scaleMultOut)}; chan.q[0] = std::max(chan.q[0], atUint8(cur.qFrom(last, 0))); chan.q[1] = std::max(chan.q[1], atUint8(cur.qFrom(last, 1))); chan.q[2] = std::max(chan.q[2], atUint8(cur.qFrom(last, 2))); last = cur; } break; } default: break; } ++kit; } /* Generate Bitstream */ sizeOut = ComputeBitstreamSize(keyFrameCount, channels); std::unique_ptr newData(new atUint8[sizeOut]); memset(newData.get(), 0, sizeOut); lastVals = initVals; for (size_t frame = 0; frame < keyFrameCount; ++frame) { kit = chanKeys.begin(); vit = lastVals.begin(); for (const Channel& chan : channels) { const Value& val = (*kit++)[frame + 1]; QuantizedValue& last = *vit++; switch (chan.type) { case Channel::Type::Rotation: { QuantizedRot qrCur = QuantizeRotation(val, rotDivOut); quantizeBit(newData.get(), qrCur.w); quantize(newData.get(), chan.q[0], qrCur.v[0] - last.v[0]); quantize(newData.get(), chan.q[1], qrCur.v[1] - last.v[1]); quantize(newData.get(), chan.q[2], qrCur.v[2] - last.v[2]); last = qrCur.v; break; } case Channel::Type::Translation: { zeus::simd_floats f(val.simd); QuantizedValue cur = {atInt32(f[0] / transMultOut), atInt32(f[1] / transMultOut), atInt32(f[2] / transMultOut)}; quantize(newData.get(), chan.q[0], cur[0] - last[0]); quantize(newData.get(), chan.q[1], cur[1] - last[1]); quantize(newData.get(), chan.q[2], cur[2] - last[2]); last = cur; break; } case Channel::Type::Scale: { zeus::simd_floats f(val.simd); QuantizedValue cur = {atInt32(f[0] / scaleMultOut), atInt32(f[1] / scaleMultOut), atInt32(f[2] / scaleMultOut)}; quantize(newData.get(), chan.q[0], cur[0] - last[0]); quantize(newData.get(), chan.q[1], cur[1] - last[1]); quantize(newData.get(), chan.q[2], cur[2] - last[2]); last = cur; break; } default: break; } } } return newData; } } // namespace DataSpec::DNAANIM