Deflicker.cc

Go to the documentation of this file.

#include "Deflicker.hh"
 
#include "PixelOperations.hh"
#include "RawFrame.hh"
 
#include "vla.hh"
#include "xrange.hh"
 
#include <bit>
#ifdef __SSE2__
#include <emmintrin.h>
#endif
 
namespace openmsx {
 
using Pixel = uint32_t;
 
Deflicker::Deflicker(std::span<std::unique_ptr<RawFrame>, 4> lastFrames_)
        : lastFrames(lastFrames_)
{
}
 
Deflicker::~Deflicker() = default;
 
void Deflicker::init()
{
        FrameSource::init(FieldType::NONINTERLACED);
        setHeight(lastFrames[0]->getHeight());
}
 
unsigned Deflicker::getLineWidth(unsigned line) const
{
        return lastFrames[0]->getLineWidthDirect(line);
}
 
#ifdef __SSE2__
static __m128i blend(__m128i x, __m128i y)
{
        // 32bpp
        return _mm_avg_epu8(x, y);
}
 
static __m128i uload(const Pixel* ptr, ptrdiff_t byteOffst)
{
        const auto* p8   = std::bit_cast<const   char *>(ptr);
        const auto* p128 = std::bit_cast<const __m128i*>(p8 + byteOffst);
        return _mm_loadu_si128(p128);
}
 
static void ustore(Pixel* ptr, ptrdiff_t byteOffst, __m128i val)
{
        auto* p8   = std::bit_cast<  char *>(ptr);
        auto* p128 = std::bit_cast<__m128i*>(p8 + byteOffst);
        return _mm_storeu_si128(p128, val);
}
 
static __m128i compare(__m128i x, __m128i y)
{
        // 32bpp
        return _mm_cmpeq_epi32(x, y);
}
#endif
 
std::span<const Pixel> Deflicker::getUnscaledLine(
        unsigned line, std::span<Pixel> helpBuf) const
{
        unsigned width0 = lastFrames[0]->getLineWidthDirect(line);
        unsigned width1 = lastFrames[1]->getLineWidthDirect(line);
        unsigned width2 = lastFrames[2]->getLineWidthDirect(line);
        unsigned width3 = lastFrames[3]->getLineWidthDirect(line);
        const Pixel* line0 = lastFrames[0]->getLineDirect(line).data();
        const Pixel* line1 = lastFrames[1]->getLineDirect(line).data();
        const Pixel* line2 = lastFrames[2]->getLineDirect(line).data();
        const Pixel* line3 = lastFrames[3]->getLineDirect(line).data();
        if ((width0 != width3) || (width0 != width2) || (width0 != width1)) {
                // Not all the same width.
                return std::span{line0, width0};
        }
 
        // Prefer to write directly to the output buffer, if that's not
        // possible store the intermediate result in a temp buffer.
        VLA_SSE_ALIGNED(Pixel, buf2, width0);
        auto* buf = helpBuf.data();
        Pixel* out = (width0 <= helpBuf.size()) ? buf : buf2.data();
 
        // Detect pixels that alternate between two different color values and
        // replace those with the average color. We search for an alternating
        // sequence with length (at least) 4. Or IOW we look for "A B A B".
        // The implementation below also detects a constant pixel value
        // "A A A A" as alternating between "A" and "A", but that's fine.
        Pixel* dst = out;
        size_t remaining = width0;
#ifdef __SSE2__
        size_t pixelsPerSSE = sizeof(__m128i) / sizeof(Pixel);
        size_t widthSSE = remaining & ~(pixelsPerSSE - 1); // rounded down to a multiple of pixels in a SSE register
        line0 += widthSSE;
        line1 += widthSSE;
        line2 += widthSSE;
        line3 += widthSSE;
        dst   += widthSSE;
        auto byteOffst = -ptrdiff_t(widthSSE * sizeof(Pixel));
 
        while (byteOffst < 0) {
                __m128i a0 = uload(line0, byteOffst);
                __m128i a1 = uload(line1, byteOffst);
                __m128i a2 = uload(line2, byteOffst);
                __m128i a3 = uload(line3, byteOffst);
 
                __m128i e02 = compare(a0, a2); // a0 == a2
                __m128i e13 = compare(a1, a3); // a1 == a3
                __m128i cnd = _mm_and_si128(e02, e13); // (a0==a2) && (a1==a3)
 
                __m128i a01 = blend(a0, a1);
                __m128i p = _mm_xor_si128(a0, a01);
                __m128i q = _mm_and_si128(p, cnd);
                __m128i r = _mm_xor_si128(q, a0); // select(a0, a01, cnd)
 
                ustore(dst, byteOffst, r);
                byteOffst += sizeof(__m128i);
        }
        remaining &= pixelsPerSSE - 1;
#endif
        PixelOperations pixelOps;
        for (auto x : xrange(remaining)) {
                dst[x] = ((line0[x] == line2[x]) && (line1[x] == line3[x]))
                       ? pixelOps.template blend<1, 1>(line0[x], line1[x])
                       : line0[x];
        }
 
        if (width0 <= helpBuf.size()) {
                // If it already fits, we're done
                return std::span{buf, width0};
        } else {
                // Otherwise scale so that it does fit.
                scaleLine(std::span{out, width0}, helpBuf);
                return helpBuf;
        }
}
 
} // namespace openmsx