diff options
author | 2016-08-06 14:20:16 +0200 | |
---|---|---|
committer | 2016-08-06 14:20:16 +0200 | |
commit | 3917850b6197950e5671f4d3b9cb0b9c921b798d (patch) | |
tree | 503b9690868a7c238ba10d73b5c6bece7ca8c338 /3rdparty/bgfx/src/image.cpp | |
parent | 057474989c1a92e9e935dbc8cb9d233ba3a4ea3a (diff) |
Update BX and BGFX (nw)
Diffstat (limited to '3rdparty/bgfx/src/image.cpp')
-rw-r--r-- | 3rdparty/bgfx/src/image.cpp | 206 |
1 files changed, 127 insertions, 79 deletions
diff --git a/3rdparty/bgfx/src/image.cpp b/3rdparty/bgfx/src/image.cpp index 1d91f1e7627..4b31a7ba08a 100644 --- a/3rdparty/bgfx/src/image.cpp +++ b/3rdparty/bgfx/src/image.cpp @@ -387,68 +387,68 @@ namespace bgfx const uint8_t* src = (const uint8_t*)_src; using namespace bx; - const float4_t unpack = float4_ld(1.0f, 1.0f/256.0f, 1.0f/65536.0f, 1.0f/16777216.0f); - const float4_t pack = float4_ld(1.0f, 256.0f*0.5f, 65536.0f, 16777216.0f*0.5f); - const float4_t umask = float4_ild(0xff, 0xff00, 0xff0000, 0xff000000); - const float4_t pmask = float4_ild(0xff, 0x7f80, 0xff0000, 0x7f800000); - const float4_t wflip = float4_ild(0, 0, 0, 0x80000000); - const float4_t wadd = float4_ld(0.0f, 0.0f, 0.0f, 32768.0f*65536.0f); - const float4_t gamma = float4_ld(1.0f/2.2f, 1.0f/2.2f, 1.0f/2.2f, 1.0f); - const float4_t linear = float4_ld(2.2f, 2.2f, 2.2f, 1.0f); - const float4_t quater = float4_splat(0.25f); + const simd128_t unpack = simd_ld(1.0f, 1.0f/256.0f, 1.0f/65536.0f, 1.0f/16777216.0f); + const simd128_t pack = simd_ld(1.0f, 256.0f*0.5f, 65536.0f, 16777216.0f*0.5f); + const simd128_t umask = simd_ild(0xff, 0xff00, 0xff0000, 0xff000000); + const simd128_t pmask = simd_ild(0xff, 0x7f80, 0xff0000, 0x7f800000); + const simd128_t wflip = simd_ild(0, 0, 0, 0x80000000); + const simd128_t wadd = simd_ld(0.0f, 0.0f, 0.0f, 32768.0f*65536.0f); + const simd128_t gamma = simd_ld(1.0f/2.2f, 1.0f/2.2f, 1.0f/2.2f, 1.0f); + const simd128_t linear = simd_ld(2.2f, 2.2f, 2.2f, 1.0f); + const simd128_t quater = simd_splat(0.25f); for (uint32_t yy = 0, ystep = _pitch*2; yy < dstheight; ++yy, src += ystep) { const uint8_t* rgba = src; for (uint32_t xx = 0; xx < dstwidth; ++xx, rgba += 8, dst += 4) { - const float4_t abgr0 = float4_splat(rgba); - const float4_t abgr1 = float4_splat(rgba+4); - const float4_t abgr2 = float4_splat(rgba+_pitch); - const float4_t abgr3 = float4_splat(rgba+_pitch+4); - - const float4_t abgr0m = float4_and(abgr0, umask); - const float4_t abgr1m = float4_and(abgr1, umask); - const float4_t abgr2m = float4_and(abgr2, umask); - const float4_t abgr3m = float4_and(abgr3, umask); - const float4_t abgr0x = float4_xor(abgr0m, wflip); - const float4_t abgr1x = float4_xor(abgr1m, wflip); - const float4_t abgr2x = float4_xor(abgr2m, wflip); - const float4_t abgr3x = float4_xor(abgr3m, wflip); - const float4_t abgr0f = float4_itof(abgr0x); - const float4_t abgr1f = float4_itof(abgr1x); - const float4_t abgr2f = float4_itof(abgr2x); - const float4_t abgr3f = float4_itof(abgr3x); - const float4_t abgr0c = float4_add(abgr0f, wadd); - const float4_t abgr1c = float4_add(abgr1f, wadd); - const float4_t abgr2c = float4_add(abgr2f, wadd); - const float4_t abgr3c = float4_add(abgr3f, wadd); - const float4_t abgr0n = float4_mul(abgr0c, unpack); - const float4_t abgr1n = float4_mul(abgr1c, unpack); - const float4_t abgr2n = float4_mul(abgr2c, unpack); - const float4_t abgr3n = float4_mul(abgr3c, unpack); - - const float4_t abgr0l = float4_pow(abgr0n, linear); - const float4_t abgr1l = float4_pow(abgr1n, linear); - const float4_t abgr2l = float4_pow(abgr2n, linear); - const float4_t abgr3l = float4_pow(abgr3n, linear); - - const float4_t sum0 = float4_add(abgr0l, abgr1l); - const float4_t sum1 = float4_add(abgr2l, abgr3l); - const float4_t sum2 = float4_add(sum0, sum1); - const float4_t avg0 = float4_mul(sum2, quater); - const float4_t avg1 = float4_pow(avg0, gamma); - - const float4_t avg2 = float4_mul(avg1, pack); - const float4_t ftoi0 = float4_ftoi(avg2); - const float4_t ftoi1 = float4_and(ftoi0, pmask); - const float4_t zwxy = float4_swiz_zwxy(ftoi1); - const float4_t tmp0 = float4_or(ftoi1, zwxy); - const float4_t yyyy = float4_swiz_yyyy(tmp0); - const float4_t tmp1 = float4_iadd(yyyy, yyyy); - const float4_t result = float4_or(tmp0, tmp1); - - float4_stx(dst, result); + const simd128_t abgr0 = simd_splat(rgba); + const simd128_t abgr1 = simd_splat(rgba+4); + const simd128_t abgr2 = simd_splat(rgba+_pitch); + const simd128_t abgr3 = simd_splat(rgba+_pitch+4); + + const simd128_t abgr0m = simd_and(abgr0, umask); + const simd128_t abgr1m = simd_and(abgr1, umask); + const simd128_t abgr2m = simd_and(abgr2, umask); + const simd128_t abgr3m = simd_and(abgr3, umask); + const simd128_t abgr0x = simd_xor(abgr0m, wflip); + const simd128_t abgr1x = simd_xor(abgr1m, wflip); + const simd128_t abgr2x = simd_xor(abgr2m, wflip); + const simd128_t abgr3x = simd_xor(abgr3m, wflip); + const simd128_t abgr0f = simd_itof(abgr0x); + const simd128_t abgr1f = simd_itof(abgr1x); + const simd128_t abgr2f = simd_itof(abgr2x); + const simd128_t abgr3f = simd_itof(abgr3x); + const simd128_t abgr0c = simd_add(abgr0f, wadd); + const simd128_t abgr1c = simd_add(abgr1f, wadd); + const simd128_t abgr2c = simd_add(abgr2f, wadd); + const simd128_t abgr3c = simd_add(abgr3f, wadd); + const simd128_t abgr0n = simd_mul(abgr0c, unpack); + const simd128_t abgr1n = simd_mul(abgr1c, unpack); + const simd128_t abgr2n = simd_mul(abgr2c, unpack); + const simd128_t abgr3n = simd_mul(abgr3c, unpack); + + const simd128_t abgr0l = simd_pow(abgr0n, linear); + const simd128_t abgr1l = simd_pow(abgr1n, linear); + const simd128_t abgr2l = simd_pow(abgr2n, linear); + const simd128_t abgr3l = simd_pow(abgr3n, linear); + + const simd128_t sum0 = simd_add(abgr0l, abgr1l); + const simd128_t sum1 = simd_add(abgr2l, abgr3l); + const simd128_t sum2 = simd_add(sum0, sum1); + const simd128_t avg0 = simd_mul(sum2, quater); + const simd128_t avg1 = simd_pow(avg0, gamma); + + const simd128_t avg2 = simd_mul(avg1, pack); + const simd128_t ftoi0 = simd_ftoi(avg2); + const simd128_t ftoi1 = simd_and(ftoi0, pmask); + const simd128_t zwxy = simd_swiz_zwxy(ftoi1); + const simd128_t tmp0 = simd_or(ftoi1, zwxy); + const simd128_t yyyy = simd_swiz_yyyy(tmp0); + const simd128_t tmp1 = simd_iadd(yyyy, yyyy); + const simd128_t result = simd_or(tmp0, tmp1); + + simd_stx(dst, result); } } } @@ -630,8 +630,8 @@ namespace bgfx using namespace bx; - const float4_t mf0f0 = float4_isplat(0xff00ff00); - const float4_t m0f0f = float4_isplat(0x00ff00ff); + const simd128_t mf0f0 = simd_isplat(0xff00ff00); + const simd128_t m0f0f = simd_isplat(0x00ff00ff); const uint8_t* src = (uint8_t*) _src; const uint8_t* next = src + _pitch; uint8_t* dst = (uint8_t*)_dst; @@ -642,14 +642,14 @@ namespace bgfx { for (uint32_t xx = 0; xx < width; ++xx, src += 16, dst += 16) { - const float4_t tabgr = float4_ld(src); - const float4_t t00ab = float4_srl(tabgr, 16); - const float4_t tgr00 = float4_sll(tabgr, 16); - const float4_t tgrab = float4_or(t00ab, tgr00); - const float4_t ta0g0 = float4_and(tabgr, mf0f0); - const float4_t t0r0b = float4_and(tgrab, m0f0f); - const float4_t targb = float4_or(ta0g0, t0r0b); - float4_st(dst, targb); + const simd128_t tabgr = simd_ld(src); + const simd128_t t00ab = simd_srl(tabgr, 16); + const simd128_t tgr00 = simd_sll(tabgr, 16); + const simd128_t tgrab = simd_or(t00ab, tgr00); + const simd128_t ta0g0 = simd_and(tabgr, mf0f0); + const simd128_t t0r0b = simd_and(tgrab, m0f0f); + const simd128_t targb = simd_or(ta0g0, t0r0b); + simd_st(dst, targb); } } } @@ -707,6 +707,9 @@ namespace bgfx { const uint8_t* src = (const uint8_t*)_src; _dst[0] = fromUnorm(src[0], 255.0f); + _dst[1] = 0.0f; + _dst[2] = 0.0f; + _dst[3] = 1.0f; } // R8S @@ -720,6 +723,9 @@ namespace bgfx { const int8_t* src = (const int8_t*)_src; _dst[0] = fromSnorm(src[0], 127.0f); + _dst[1] = 0.0f; + _dst[2] = 0.0f; + _dst[3] = 1.0f; } // R8I @@ -733,6 +739,9 @@ namespace bgfx { const int8_t* src = (const int8_t*)_src; _dst[0] = float(src[0]); + _dst[1] = 0.0f; + _dst[2] = 0.0f; + _dst[3] = 1.0f; } // R8U @@ -746,6 +755,9 @@ namespace bgfx { const uint8_t* src = (const uint8_t*)_src; _dst[0] = float(src[0]); + _dst[1] = 0.0f; + _dst[2] = 0.0f; + _dst[3] = 1.0f; } // RG8 @@ -761,6 +773,8 @@ namespace bgfx const uint8_t* src = (const uint8_t*)_src; _dst[0] = fromUnorm(src[0], 255.0f); _dst[1] = fromUnorm(src[1], 255.0f); + _dst[2] = 0.0f; + _dst[3] = 1.0f; } // RG8S @@ -776,6 +790,8 @@ namespace bgfx const int8_t* src = (const int8_t*)_src; _dst[0] = fromSnorm(src[0], 127.0f); _dst[1] = fromSnorm(src[1], 127.0f); + _dst[2] = 0.0f; + _dst[3] = 1.0f; } // RG8I @@ -791,6 +807,8 @@ namespace bgfx const int8_t* src = (const int8_t*)_src; _dst[0] = float(src[0]); _dst[1] = float(src[1]); + _dst[2] = 0.0f; + _dst[3] = 1.0f; } // RG8U @@ -806,6 +824,8 @@ namespace bgfx const uint8_t* src = (const uint8_t*)_src; _dst[0] = float(src[0]); _dst[1] = float(src[1]); + _dst[2] = 0.0f; + _dst[3] = 1.0f; } // RGB8 @@ -823,6 +843,7 @@ namespace bgfx _dst[0] = fromUnorm(src[0], 255.0f); _dst[1] = fromUnorm(src[1], 255.0f); _dst[2] = fromUnorm(src[2], 255.0f); + _dst[3] = 1.0f; } // RGB8S @@ -840,6 +861,7 @@ namespace bgfx _dst[0] = fromSnorm(src[0], 127.0f); _dst[1] = fromSnorm(src[1], 127.0f); _dst[2] = fromSnorm(src[2], 127.0f); + _dst[3] = 1.0f; } // RGB8I @@ -857,6 +879,7 @@ namespace bgfx _dst[0] = float(src[0]); _dst[1] = float(src[1]); _dst[2] = float(src[2]); + _dst[3] = 1.0f; } // RGB8U @@ -874,6 +897,7 @@ namespace bgfx _dst[0] = float(src[0]); _dst[1] = float(src[1]); _dst[2] = float(src[2]); + _dst[3] = 1.0f; } // BGRA8 @@ -982,6 +1006,9 @@ namespace bgfx { const uint16_t* src = (const uint16_t*)_src; _dst[0] = fromUnorm(src[0], 65535.0f); + _dst[1] = 0.0f; + _dst[2] = 0.0f; + _dst[3] = 1.0f; } // R16S @@ -995,6 +1022,9 @@ namespace bgfx { const int16_t* src = (const int16_t*)_src; _dst[0] = fromSnorm(src[0], 32767.0f); + _dst[1] = 0.0f; + _dst[2] = 0.0f; + _dst[3] = 1.0f; } // R16I @@ -1008,6 +1038,9 @@ namespace bgfx { const int16_t* src = (const int16_t*)_src; _dst[0] = float(src[0]); + _dst[1] = 0.0f; + _dst[2] = 0.0f; + _dst[3] = 1.0f; } // R16U @@ -1034,6 +1067,9 @@ namespace bgfx { const uint16_t* src = (const uint16_t*)_src; _dst[0] = bx::halfToFloat(src[0]); + _dst[1] = 0.0f; + _dst[2] = 0.0f; + _dst[3] = 1.0f; } // RG16 @@ -1049,6 +1085,8 @@ namespace bgfx const uint16_t* src = (const uint16_t*)_src; _dst[0] = fromUnorm(src[0], 65535.0f); _dst[1] = fromUnorm(src[1], 65535.0f); + _dst[2] = 0.0f; + _dst[3] = 1.0f; } // RG16S @@ -1064,6 +1102,8 @@ namespace bgfx const int16_t* src = (const int16_t*)_src; _dst[0] = fromSnorm(src[0], 32767.0f); _dst[1] = fromSnorm(src[1], 32767.0f); + _dst[2] = 0.0f; + _dst[3] = 1.0f; } // RG16I @@ -1079,6 +1119,8 @@ namespace bgfx const int16_t* src = (const int16_t*)_src; _dst[0] = float(src[0]); _dst[1] = float(src[1]); + _dst[2] = 0.0f; + _dst[3] = 1.0f; } // RG16U @@ -1094,6 +1136,8 @@ namespace bgfx const uint16_t* src = (const uint16_t*)_src; _dst[0] = float(src[0]); _dst[1] = float(src[1]); + _dst[2] = 0.0f; + _dst[3] = 1.0f; } // RG16F @@ -1109,6 +1153,8 @@ namespace bgfx const uint16_t* src = (const uint16_t*)_src; _dst[0] = bx::halfToFloat(src[0]); _dst[1] = bx::halfToFloat(src[1]); + _dst[2] = 0.0f; + _dst[3] = 1.0f; } // RGBA16 @@ -1389,6 +1435,7 @@ namespace bgfx _dst[0] = float( ( (packed>>11) & 0x1f) ) / 31.0f; _dst[1] = float( ( (packed>> 5) & 0x3f) ) / 63.0f; _dst[2] = float( ( (packed ) & 0x1f) ) / 31.0f; + _dst[3] = 1.0f; } // RGBA4 @@ -1507,6 +1554,7 @@ namespace bgfx _dst[0] = bx::halfToFloat( (packed<< 4) & 0x7ff0); _dst[1] = bx::halfToFloat( (packed>> 7) & 0x7ff0); _dst[2] = bx::halfToFloat( (packed>>17) & 0x7fe0); + _dst[3] = 1.0f; } struct PackUnpack @@ -3581,7 +3629,7 @@ namespace bgfx default: imageDecodeToBgra8(_dst, _src, _width, _height, _pitch, _format); - imageSwizzleBgra8(_width, _height, _pitch, _dst, _dst); + imageSwizzleBgra8(_width, _height, _width*4, _dst, _dst); break; } } @@ -3628,24 +3676,24 @@ namespace bgfx const uint8_t* src = (const uint8_t*)_src; using namespace bx; - const float4_t unpack = float4_ld(1.0f, 1.0f/256.0f, 1.0f/65536.0f, 1.0f/16777216.0f); - const float4_t umask = float4_ild(0xff, 0xff00, 0xff0000, 0xff000000); - const float4_t wflip = float4_ild(0, 0, 0, 0x80000000); - const float4_t wadd = float4_ld(0.0f, 0.0f, 0.0f, 32768.0f*65536.0f); + const simd128_t unpack = simd_ld(1.0f, 1.0f/256.0f, 1.0f/65536.0f, 1.0f/16777216.0f); + const simd128_t umask = simd_ild(0xff, 0xff00, 0xff0000, 0xff000000); + const simd128_t wflip = simd_ild(0, 0, 0, 0x80000000); + const simd128_t wadd = simd_ld(0.0f, 0.0f, 0.0f, 32768.0f*65536.0f); for (uint32_t yy = 0, ystep = _pitch; yy < dstheight; ++yy, src += ystep) { const uint8_t* rgba = src; for (uint32_t xx = 0; xx < dstwidth; ++xx, rgba += 4, dst += 4) { - const float4_t abgr0 = float4_splat(rgba); - const float4_t abgr0m = float4_and(abgr0, umask); - const float4_t abgr0x = float4_xor(abgr0m, wflip); - const float4_t abgr0f = float4_itof(abgr0x); - const float4_t abgr0c = float4_add(abgr0f, wadd); - const float4_t abgr0n = float4_mul(abgr0c, unpack); - - float4_st(dst, abgr0n); + const simd128_t abgr0 = simd_splat(rgba); + const simd128_t abgr0m = simd_and(abgr0, umask); + const simd128_t abgr0x = simd_xor(abgr0m, wflip); + const simd128_t abgr0f = simd_itof(abgr0x); + const simd128_t abgr0c = simd_add(abgr0f, wadd); + const simd128_t abgr0n = simd_mul(abgr0c, unpack); + + simd_st(dst, abgr0n); } } } |