LineScalers.hh

Go to the documentation of this file.

#ifndef LINESCALERS_HH
#define LINESCALERS_HH
 
#include "PixelOperations.hh"
 
#include "ranges.hh"
#include "view.hh"
#include "xrange.hh"
 
#include <bit>
#include <cassert>
#include <cstddef>
#include <cstdint>
#include <span>
#ifdef __SSE2__
#include "emmintrin.h"
#endif
 
namespace openmsx {
 
using Pixel = uint32_t;
 
void scale_1on3(std::span<const Pixel> in, std::span<Pixel> out);
void scale_1on4(std::span<const Pixel> in, std::span<Pixel> out);
void scale_1on6(std::span<const Pixel> in, std::span<Pixel> out);
void scale_1on2(std::span<const Pixel> in, std::span<Pixel> out);
void scale_2on1(std::span<const Pixel> in, std::span<Pixel> out);
void scale_6on1(std::span<const Pixel> in, std::span<Pixel> out);
void scale_4on1(std::span<const Pixel> in, std::span<Pixel> out);
void scale_3on1(std::span<const Pixel> in, std::span<Pixel> out);
void scale_3on2(std::span<const Pixel> in, std::span<Pixel> out);
void scale_3on4(std::span<const Pixel> in, std::span<Pixel> out);
void scale_3on8(std::span<const Pixel> in, std::span<Pixel> out);
void scale_2on3(std::span<const Pixel> in, std::span<Pixel> out);
void scale_4on3(std::span<const Pixel> in, std::span<Pixel> out);
void scale_8on3(std::span<const Pixel> in, std::span<Pixel> out);
void scale_2on9(std::span<const Pixel> in, std::span<Pixel> out);
void scale_4on9(std::span<const Pixel> in, std::span<Pixel> out);
void scale_8on9(std::span<const Pixel> in, std::span<Pixel> out);
 
template<unsigned w1 = 1, unsigned w2 = 1>
void blendLines(std::span<const Pixel> in1, std::span<const Pixel> in2,
                std::span<Pixel> out);
 
void alphaBlendLines(std::span<const Pixel> in1, std::span<const Pixel> in2,
                     std::span<Pixel> out);
void alphaBlendLines(Pixel in1, std::span<const Pixel> in2,
                     std::span<Pixel> out);
 
 
// implementation
 
template<unsigned N>
static inline void scale_1onN(
        std::span<const Pixel> in, std::span<Pixel> out)
{
        auto outWidth = out.size();
        assert(in.size() == (outWidth / N));
 
        size_t i = 0, j = 0;
        for (/* */; i < (outWidth - (N - 1)); i += N, j += 1) {
                Pixel pix = in[j];
                for (auto k : xrange(N)) {
                        out[i + k] = pix;
                }
        }
        for (auto k : xrange(N - 1)) {
                if ((i + k) < outWidth) out[i + k] = 0;
        }
}
 
inline void scale_1on3(std::span<const Pixel> in, std::span<Pixel> out)
{
        scale_1onN<3>(in, out);
}
 
inline void scale_1on4(std::span<const Pixel> in, std::span<Pixel> out)
{
        scale_1onN<4>(in, out);
}
 
inline void scale_1on6(std::span<const Pixel> in, std::span<Pixel> out)
{
        scale_1onN<6>(in, out);
}
 
#ifdef __SSE2__
inline __m128i unpacklo(__m128i x, __m128i y)
{
        // 32bpp
        return _mm_unpacklo_epi32(x, y);
}
inline __m128i unpackhi(__m128i x, __m128i y)
{
        // 32bpp
        return _mm_unpackhi_epi32(x, y);
}
 
inline void scale_1on2_SSE(const Pixel* __restrict in_, Pixel* __restrict out_, size_t srcWidth)
{
        size_t bytes = srcWidth * sizeof(Pixel);
        assert((bytes % (4 * sizeof(__m128i))) == 0);
        assert(bytes != 0);
 
        const auto* in  = std::bit_cast<const char*>(in_)  +     bytes;
              auto* out = std::bit_cast<      char*>(out_) + 2 * bytes;
 
        auto x = -ptrdiff_t(bytes);
        do {
                __m128i a0 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + x +  0));
                __m128i a1 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + x + 16));
                __m128i a2 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + x + 32));
                __m128i a3 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + x + 48));
                __m128i l0 = unpacklo(a0, a0);
                __m128i h0 = unpackhi(a0, a0);
                __m128i l1 = unpacklo(a1, a1);
                __m128i h1 = unpackhi(a1, a1);
                __m128i l2 = unpacklo(a2, a2);
                __m128i h2 = unpackhi(a2, a2);
                __m128i l3 = unpacklo(a3, a3);
                __m128i h3 = unpackhi(a3, a3);
                _mm_storeu_si128(std::bit_cast<__m128i*>(out + 2*x +   0), l0);
                _mm_storeu_si128(std::bit_cast<__m128i*>(out + 2*x +  16), h0);
                _mm_storeu_si128(std::bit_cast<__m128i*>(out + 2*x +  32), l1);
                _mm_storeu_si128(std::bit_cast<__m128i*>(out + 2*x +  48), h1);
                _mm_storeu_si128(std::bit_cast<__m128i*>(out + 2*x +  64), l2);
                _mm_storeu_si128(std::bit_cast<__m128i*>(out + 2*x +  80), h2);
                _mm_storeu_si128(std::bit_cast<__m128i*>(out + 2*x +  96), l3);
                _mm_storeu_si128(std::bit_cast<__m128i*>(out + 2*x + 112), h3);
                x += 4 * sizeof(__m128i);
        } while (x < 0);
}
#endif
 
inline void scale_1on2(std::span<const Pixel> in, std::span<Pixel> out)
{
        // This is a fairly simple algorithm (output each input pixel twice).
        // An ideal compiler should generate optimal (vector) code for it.
        // I checked the 2013-05-29 dev snapshots of gcc-4.9 and clang-3.4:
        // - Clang is not able to vectorize this loop. My best tuned C version
        //   of this routine is a little over 4x slower than the tuned
        //   SSE-intrinsics version.
        // - Gcc can auto-vectorize this routine. Though my best tuned version
        //   (I mean tuned to further improve the auto-vectorization, including
        //   using the new __builtin_assume_aligned() intrinsic) still runs
        //   approx 40% slower than the intrinsics version.
        // Hopefully in some years the compilers have improved further so that
        // the intrinsic version is no longer needed.
        auto srcWidth = in.size();
        assert((out.size() / 2) == srcWidth);
 
#ifdef __SSE2__
        size_t chunk = 4 * sizeof(__m128i) / sizeof(Pixel);
        size_t srcWidth2 = srcWidth & ~(chunk - 1);
        scale_1on2_SSE(in.data(), out.data(), srcWidth2);
        in  = in .subspan(    srcWidth2);
        out = out.subspan(2 * srcWidth2);
        srcWidth -= srcWidth2;
#endif
 
        // C++ version. Used both on non-x86 machines and (possibly) on x86 for
        // the last few pixels of the line.
        for (auto x : xrange(srcWidth)) {
                out[x * 2] = out[x * 2 + 1] = in[x];
        }
}
 
#ifdef __SSE2__
template<int IMM8> static inline __m128i shuffle(__m128i x, __m128i y)
{
        return _mm_castps_si128(_mm_shuffle_ps(
                _mm_castsi128_ps(x), _mm_castsi128_ps(y), IMM8));
}
 
inline __m128i blend(__m128i x, __m128i y)
{
        // 32bpp
        __m128i p = shuffle<0x88>(x, y);
        __m128i q = shuffle<0xDD>(x, y);
        return _mm_avg_epu8(p, q);
}
 
inline void scale_2on1_SSE(
        const Pixel* __restrict in_, Pixel* __restrict out_, size_t dstBytes)
{
        assert((dstBytes % (4 * sizeof(__m128i))) == 0);
        assert(dstBytes != 0);
 
        const auto* in  = std::bit_cast<const char*>(in_)  + 2 * dstBytes;
              auto* out = std::bit_cast<      char*>(out_) +     dstBytes;
 
        auto x = -ptrdiff_t(dstBytes);
        do {
                __m128i a0 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + 2*x +   0));
                __m128i a1 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + 2*x +  16));
                __m128i a2 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + 2*x +  32));
                __m128i a3 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + 2*x +  48));
                __m128i a4 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + 2*x +  64));
                __m128i a5 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + 2*x +  80));
                __m128i a6 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + 2*x +  96));
                __m128i a7 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + 2*x + 112));
                __m128i b0 = blend(a0, a1);
                __m128i b1 = blend(a2, a3);
                __m128i b2 = blend(a4, a5);
                __m128i b3 = blend(a6, a7);
                _mm_storeu_si128(std::bit_cast<__m128i*>(out + x +  0), b0);
                _mm_storeu_si128(std::bit_cast<__m128i*>(out + x + 16), b1);
                _mm_storeu_si128(std::bit_cast<__m128i*>(out + x + 32), b2);
                _mm_storeu_si128(std::bit_cast<__m128i*>(out + x + 48), b3);
                x += 4 * sizeof(__m128i);
        } while (x < 0);
}
#endif
 
inline void scale_2on1(std::span<const Pixel> in, std::span<Pixel> out)
{
        assert((in.size() / 2) == out.size());
        auto outWidth = out.size();
#ifdef __SSE2__
        auto n64 = (outWidth * sizeof(Pixel)) & ~63;
        scale_2on1_SSE(in.data(), out.data(), n64); // process 64 byte chunks
        outWidth &= ((64 / sizeof(Pixel)) - 1); // remaining pixels (if any)
        if (outWidth == 0) [[likely]] return;
        in  = in .subspan(2 * n64 / sizeof(Pixel));
        out = out.subspan(    n64 / sizeof(Pixel));
        // fallthrough to c++ version
#endif
        // pure C++ version
        PixelOperations pixelOps;
        for (auto i : xrange(outWidth)) {
                out[i] = pixelOps.template blend<1, 1>(
                        in[2 * i + 0], in[2 * i + 1]);
        }
}
 
inline void scale_6on1(std::span<const Pixel> in, std::span<Pixel> out)
{
        assert((in.size() / 6) == out.size());
        PixelOperations pixelOps;
        for (auto i : xrange(out.size())) {
                out[i] = pixelOps.template blend<1, 1, 1, 1, 1, 1>(subspan<6>(in, 6 * i));
        }
}
 
inline void scale_4on1(std::span<const Pixel> in, std::span<Pixel> out)
{
        assert((in.size() / 4) == out.size());
        PixelOperations pixelOps;
        for (auto i : xrange(out.size())) {
                out[i] = pixelOps.template blend<1, 1, 1, 1>(subspan<4>(in, 4 * i));
        }
}
 
inline void scale_3on1(std::span<const Pixel> in, std::span<Pixel> out)
{
        assert((in.size() / 3) == out.size());
        PixelOperations pixelOps;
        for (auto i : xrange(out.size())) {
                out[i] = pixelOps.template blend<1, 1, 1>(subspan<3>(in, 3 * i));
        }
}
 
inline void scale_3on2(std::span<const Pixel> in, std::span<Pixel> out)
{
        assert((in.size() / 3) == (out.size() / 2));
        PixelOperations pixelOps;
        size_t n = out.size();
        size_t i = 0, j = 0;
        for (/* */; i < (n - 1); i += 2, j += 3) {
                out[i + 0] = pixelOps.template blend<2, 1>(subspan<2>(in, j + 0));
                out[i + 1] = pixelOps.template blend<1, 2>(subspan<2>(in, j + 1));
        }
        if (i < n) out[i] = 0;
}
 
inline void scale_3on4(std::span<const Pixel> in, std::span<Pixel> out)
{
        assert((in.size() / 3) == (out.size() / 4));
        PixelOperations pixelOps;
        size_t n = out.size();
        size_t i = 0, j = 0;
        for (/* */; i < (n - 3); i += 4, j += 3) {
                out[i + 0] = in[j + 0];
                out[i + 1] = pixelOps.template blend<1, 2>(subspan<2>(in, j + 0));
                out[i + 2] = pixelOps.template blend<2, 1>(subspan<2>(in, j + 1));
                out[i + 3] = in[j + 2];
        }
        for (auto k : xrange(4 - 1)) {
                if ((i + k) < n) out[i + k] = 0;
        }
}
 
inline void scale_3on8(std::span<const Pixel> in, std::span<Pixel> out)
{
        assert((in.size() / 3) == (out.size() / 8));
        PixelOperations pixelOps;
        size_t n = out.size();
        size_t i = 0, j = 0;
        for (/* */; i < (n - 7); i += 8, j += 3) {
                out[i + 0] = in[j + 0];
                out[i + 1] = in[j + 0];
                out[i + 2] = pixelOps.template blend<2, 1>(subspan<2>(in, j + 0));
                out[i + 3] = in[j + 1];
                out[i + 4] = in[j + 1];
                out[i + 5] = pixelOps.template blend<1, 2>(subspan<2>(in, j + 1));
                out[i + 6] = in[j + 2];
                out[i + 7] = in[j + 2];
        }
        for (auto k : xrange(8 - 1)) {
                if ((i + k) < n) out[i + k] = 0;
        }
}
 
inline void scale_2on3(std::span<const Pixel> in, std::span<Pixel> out)
{
        assert((in.size() / 2) == (out.size() / 3));
        PixelOperations pixelOps;
        size_t n = out.size();
        size_t i = 0, j = 0;
        for (/* */; i < (n - 2); i += 3, j += 2) {
                out[i + 0] = in[j + 0];
                out[i + 1] = pixelOps.template blend<1, 1>(subspan<2>(in, j));
                out[i + 2] = in[j + 1];
        }
        if ((i + 0) < n) out[i + 0] = 0;
        if ((i + 1) < n) out[i + 1] = 0;
}
 
inline void scale_4on3(std::span<const Pixel> in, std::span<Pixel> out)
{
        assert((in.size() / 4) == (out.size() / 3));
        PixelOperations pixelOps;
        size_t n = out.size();
        size_t i = 0, j = 0;
        for (/* */; i < (n - 2); i += 3, j += 4) {
                out[i + 0] = pixelOps.template blend<3, 1>(subspan<2>(in, j + 0));
                out[i + 1] = pixelOps.template blend<1, 1>(subspan<2>(in, j + 1));
                out[i + 2] = pixelOps.template blend<1, 3>(subspan<2>(in, j + 2));
        }
        if ((i + 0) < n) out[i + 0] = 0;
        if ((i + 1) < n) out[i + 1] = 0;
}
 
inline void scale_8on3(std::span<const Pixel> in, std::span<Pixel> out)
{
        assert((in.size() / 8) == (out.size() / 3));
        PixelOperations pixelOps;
        size_t n = out.size();
        size_t i = 0, j = 0;
        for (/* */; i < (n - 2); i += 3, j += 8) {
                out[i + 0] = pixelOps.template blend<3, 3, 2>   (subspan<3>(in, j + 0));
                out[i + 1] = pixelOps.template blend<1, 3, 3, 1>(subspan<4>(in, j + 2));
                out[i + 2] = pixelOps.template blend<2, 3, 3>   (subspan<3>(in, j + 5));
        }
        if ((i + 0) < n) out[i + 0] = 0;
        if ((i + 1) < n) out[i + 1] = 0;
}
 
inline void scale_2on9(std::span<const Pixel> in, std::span<Pixel> out)
{
        assert((in.size() / 2) == (out.size() / 9));
        PixelOperations pixelOps;
        size_t n = out.size();
        size_t i = 0, j = 0;
        for (/* */; i < (n - 8); i += 9, j += 2) {
                out[i + 0] = in[j + 0];
                out[i + 1] = in[j + 0];
                out[i + 2] = in[j + 0];
                out[i + 3] = in[j + 0];
                out[i + 4] = pixelOps.template blend<1, 1>(subspan<2>(in, j));
                out[i + 5] = in[j + 1];
                out[i + 6] = in[j + 1];
                out[i + 7] = in[j + 1];
                out[i + 8] = in[j + 1];
        }
        if ((i + 0) < n) out[i + 0] = 0;
        if ((i + 1) < n) out[i + 1] = 0;
        if ((i + 2) < n) out[i + 2] = 0;
        if ((i + 3) < n) out[i + 3] = 0;
        if ((i + 4) < n) out[i + 4] = 0;
        if ((i + 5) < n) out[i + 5] = 0;
        if ((i + 6) < n) out[i + 6] = 0;
        if ((i + 7) < n) out[i + 7] = 0;
}
 
inline void scale_4on9(std::span<const Pixel> in, std::span<Pixel> out)
{
        assert((in.size() / 4) == (out.size() / 9));
        PixelOperations pixelOps;
        size_t n = out.size();
        size_t i = 0, j = 0;
        for (/* */; i < (n - 8); i += 9, j += 4) {
                out[i + 0] = in[j + 0];
                out[i + 1] = in[j + 0];
                out[i + 2] = pixelOps.template blend<1, 3>(subspan<2>(in, j + 0));
                out[i + 3] = in[j + 1];
                out[i + 4] = pixelOps.template blend<1, 1>(subspan<2>(in, j + 1));
                out[i + 5] = in[j + 2];
                out[i + 6] = pixelOps.template blend<3, 1>(subspan<2>(in, j + 2));
                out[i + 7] = in[j + 3];
                out[i + 8] = in[j + 3];
        }
        if ((i + 0) < n) out[i + 0] = 0;
        if ((i + 1) < n) out[i + 1] = 0;
        if ((i + 2) < n) out[i + 2] = 0;
        if ((i + 3) < n) out[i + 3] = 0;
        if ((i + 4) < n) out[i + 4] = 0;
        if ((i + 5) < n) out[i + 5] = 0;
        if ((i + 6) < n) out[i + 6] = 0;
        if ((i + 7) < n) out[i + 7] = 0;
}
 
inline void scale_8on9(std::span<const Pixel> in, std::span<Pixel> out)
{
        assert((in.size() / 8) == (out.size() / 9));
        PixelOperations pixelOps;
        size_t n = out.size();
        size_t i = 0, j = 0;
        for (/* */; i < (n - 8); i += 9, j += 8) {
                out[i + 0] = in[j + 0];
                out[i + 1] = pixelOps.template blend<1, 7>(subspan<2>(in, j + 0));
                out[i + 2] = pixelOps.template blend<1, 3>(subspan<2>(in, j + 1));
                out[i + 3] = pixelOps.template blend<3, 5>(subspan<2>(in, j + 2));
                out[i + 4] = pixelOps.template blend<1, 1>(subspan<2>(in, j + 3));
                out[i + 5] = pixelOps.template blend<5, 3>(subspan<2>(in, j + 4));
                out[i + 6] = pixelOps.template blend<3, 1>(subspan<2>(in, j + 5));
                out[i + 7] = pixelOps.template blend<7, 1>(subspan<2>(in, j + 6));
                out[i + 8] = in[j + 7];
        }
        if ((i + 0) < n) out[i + 0] = 0;
        if ((i + 1) < n) out[i + 1] = 0;
        if ((i + 2) < n) out[i + 2] = 0;
        if ((i + 3) < n) out[i + 3] = 0;
        if ((i + 4) < n) out[i + 4] = 0;
        if ((i + 5) < n) out[i + 5] = 0;
        if ((i + 6) < n) out[i + 6] = 0;
        if ((i + 7) < n) out[i + 7] = 0;
}
 
template<unsigned w1, unsigned w2>
void blendLines(std::span<const Pixel> in1, std::span<const Pixel> in2, std::span<Pixel> out)
{
        // It _IS_ allowed that the output is the same as one of the inputs.
        // TODO SSE optimizations
        // pure C++ version
        assert(in1.size() == in2.size());
        assert(in1.size() == out.size());
        PixelOperations pixelOps;
        for (auto [i1, i2, o] : view::zip_equal(in1, in2, out)) {
                o = pixelOps.template blend<w1, w2>(i1, i2);
        }
}
 
inline void alphaBlendLines(
        std::span<const Pixel> in1, std::span<const Pixel> in2, std::span<Pixel> out)
{
        // It _IS_ allowed that the output is the same as one of the inputs.
        assert(in1.size() == in2.size());
        assert(in1.size() == out.size());
        PixelOperations pixelOps;
        for (auto [i1, i2, o] : view::zip_equal(in1, in2, out)) {
                o = pixelOps.alphaBlend(i1, i2);
        }
}
 
inline void alphaBlendLines(
        Pixel in1, std::span<const Pixel> in2, std::span<Pixel> out)
{
        // It _IS_ allowed that the output is the same as the input.
 
        // ATM this routine is only called when 'in1' is not fully opaque nor
        // fully transparent.
        assert(in2.size() == out.size());
 
        PixelOperations pixelOps;
        unsigned alpha = pixelOps.alpha(in1);
 
        // When one of the two colors is loop-invariant, using the
        // pre-multiplied-alpha-blending equation is a tiny bit more efficient
        // than using alphaBlend() or even lerp().
        //    for (auto i : xrange(width)) {
        //        out[i] = pixelOps.lerp(in1, in2[i], alpha);
        //    }
        Pixel in1M = pixelOps.multiply(in1, alpha);
        unsigned alpha2 = 256 - alpha;
        for (auto [i2, o] : view::zip_equal(in2, out)) {
                o = in1M + pixelOps.multiply(i2, alpha2);
        }
}
 
} // namespace openmsx
 
#endif