# --- T2-COPYRIGHT-BEGIN --- # t2/package/*/firefox/hotfix-swgl-ia64-tier3-yuv-kernels.patch # Copyright (C) 2026 The T2 SDE Project # SPDX-License-Identifier: GPL-2.0 or patched project license # --- T2-COPYRIGHT-END --- # Tier III (dav1d's terminology, see package/multimedia/dav1d/ia64-simd.patch) # kernels for SWGL's YUV -> BGRA video rows, in the same style as # hotfix-swgl-ia64-tier3-r8-kernels.patch: a generated, modulo-scheduled # assembly file with its own register frame. Two entry points: # # swgl_yuv42_row_ia64 upscaleYUV42R8, 4:2:0 at <= 1 luma texel/px # swgl_yuv42_row_linear_ia64 linear_row_yuv's generic R8 loop (any scale: # downscaled video, and the edge chunks of the # upscale path) # # The C upscale loop repairs a sliding window with lane shuffles that GCC # lowers to serial extr/dep chains, 127 bundles per four pixels; the kernels # gather the taps instead, under index rules that reproduce the shuffles (and # the generic loop's edge clamps) exactly - checked against the C on random # spans before any assembly was written. Positions advance as 32-bit lane # pairs, row 1 is loaded through ld1's register post-increment, and the # registers are allocated by packing lifetimes on the rotating file. # # On a browsing session (news, Discord, YouTube) linear_row_yuv was 2.7% of # all CPU cycles, 19% of the SWGL thread. Itanium 2 (Montecito), cycles per # pixel, byte-identical output, CompositeYUV of 4:2:0 R8 planes into # 1024x768; x86-64 is the same benchmark on a Ryzen 5900HX: # C path Tier III x86-64 # YUV 4:2:0 1:1 45.7 11.7 46.8 # YUV 4:2:0 upscale (640x360) 45.5 11.6 46.8 # YUV 4:2:0 downscale (1920x1080) 97.1 14.9 109.5 # # The upscale kernel needs a minimum step (an upscale below ~16000x) for one # of its index rules; the caller checks it and otherwise keeps the C loop. --- firefox-158.0/gfx/wr/swgl/src/composite.h.vanilla +++ firefox-158.0/gfx/wr/swgl/src/composite.h @@ -873,6 +873,96 @@ // and 1 bit for sign, which leaves 8 bits left for extra precision. const int STEP_BITS = 8; +#ifdef __ia64__ +// Tier III kernel for the loop below (swgl_yuv42_ia64.S, generated by +// tools/ia64-gen/genyuv_s.py of the SWGL ia64 bench). Every field is one +// 64-bit register: lanes are little-endian, positions and fractions are +// pre-decremented by one step since the kernel advances them first. +struct IA64YUVRow { + uint64_t yu01, yu23; // luma positions, two int32 lanes each + uint64_t cu; // chroma positions (A, B) + uint64_t fxy, fxc; // ycFracX, low and high half + uint64_t ydu, cdu; // position steps, both lanes + uint64_t fxdy, fxdc; // ycFracDX, low and high half + const uint8_t *y, *u, *v; + int64_t ystride, cstride; + uint64_t fy, fc; // row fractions, four lanes + uint64_t yc, yb, uvb, brc, ggc, mask; // YUVMatrix, low half of each +}; +extern "C" void swgl_yuv42_row_ia64(uint32_t* dst, long n, + const IA64YUVRow* params); + +template +static ALWAYS_INLINE uint64_t ia64_low_u64(const V& v) { + uint64_t r; + memcpy(&r, &v, sizeof(r)); + return r; +} + +static ALWAYS_INLINE uint64_t ia64_lanes32(int32_t a, int32_t b) { + return uint64_t(uint32_t(a)) | uint64_t(uint32_t(b)) << 32; +} + +static ALWAYS_INLINE uint64_t ia64_lanes16(int32_t a, int32_t b, int32_t c, + int32_t d) { + return uint64_t(uint16_t(a)) | uint64_t(uint16_t(b)) << 16 | + uint64_t(uint16_t(c)) << 32 | uint64_t(uint16_t(d)) << 48; +} + +// The generic R8 loop of linear_row_yuv (textureLinearRowR8 + +// textureLinearRowPairedR8 + convert), same generator: positions are not +// fractions-split here, the kernel clamps them as those functions do. +struct IA64YUVRowLinear { + uint64_t yu01, yu23, cu01, cu23; // positions (yU, cU), two int32 lanes each + uint64_t ydu, cdu; // position steps, both lanes + const uint8_t *y, *u, *v; // row 0 of each plane + int64_t ystride, cstride; // to row 1 (0 at the last row) + uint64_t fy, fc; // row fractions, four lanes + uint64_t yw2, cw2, c127; // width - 2 and 127, both lanes + uint64_t yc, yb, uvb, brc, ggc, mask; // YUVMatrix, low half of each +}; +extern "C" void swgl_yuv42_row_linear_ia64(uint32_t* dst, long n, + const IA64YUVRowLinear* params); + +// n chunks of the generic loop from positions yU, cU; returns nothing, the +// caller advances its own state as the C loop would have. +static void ia64_yuv_linear_chunks(uint32_t* dest, int n, sampler2DRect samplerY, + I32 yU, int32_t yDU, int32_t yOffsetV, + int32_t yStrideV, int16_t yFracV, + sampler2DRect samplerU, + sampler2DRect samplerV, I32 cU, int32_t cDU, + int32_t cOffsetV, int32_t cStrideV, + int16_t cFracV, + const YUVMatrix& colorSpace) { + IA64YUVRowLinear p; + p.yu01 = ia64_lanes32(yU.x - yDU, yU.y - yDU); + p.yu23 = ia64_lanes32(yU.z - yDU, yU.w - yDU); + p.cu01 = ia64_lanes32(cU.x - cDU, cU.y - cDU); + p.cu23 = ia64_lanes32(cU.z - cDU, cU.w - cDU); + p.ydu = ia64_lanes32(yDU, yDU); + p.cdu = ia64_lanes32(cDU, cDU); + p.y = (const uint8_t*)samplerY->buf + yOffsetV; + p.u = (const uint8_t*)samplerU->buf + cOffsetV; + p.v = (const uint8_t*)samplerV->buf + cOffsetV; + p.ystride = yStrideV; + p.cstride = cStrideV; + p.fy = ia64_lanes16(yFracV, yFracV, yFracV, yFracV); + p.fc = ia64_lanes16(cFracV, cFracV, cFracV, cFracV); + p.yw2 = ia64_lanes32(int32_t(samplerY->width) - 2, + int32_t(samplerY->width) - 2); + p.cw2 = ia64_lanes32(int32_t(samplerU->width) - 2, + int32_t(samplerU->width) - 2); + p.c127 = ia64_lanes32(127, 127); + p.yc = ia64_low_u64(colorSpace.yCoeffs); + p.yb = ia64_low_u64(colorSpace.yBias); + p.uvb = ia64_low_u64(colorSpace.uvBias); + p.brc = ia64_low_u64(colorSpace.br_uvCoeffs); + p.ggc = ia64_low_u64(colorSpace.gg_uvCoeffs); + p.mask = ia64_low_u64(colorSpace.br_yMask); + swgl_yuv42_row_linear_ia64(dest, n, &p); +} +#endif + // Optimized version of textureLinearPackedR8 for Y R8 texture with // half-resolution paired U/V R8 textures. This allows us to more efficiently // pack YUV samples into vectors to substantially reduce math operations even @@ -884,6 +974,44 @@ uint8_t* cRow2, I32 cU, int32_t cDU, int32_t cStrideV, int16_t cFracV, const YUVMatrix& colorSpace) { +#ifdef __ia64__ + // The kernel turns the shuffles' equality tests into ordered compares, so + // the next step's first index must not fall behind this step's last: true + // unless the upscale exceeds ~16000x, which a minimum step rules out. + if constexpr (!BLEND) { + if (yDU >= 64 && cDU >= 64) { + int32_t cA = (cU.x + cU.y) >> 1, cB = (cU.z + cU.w) >> 1; + int32_t fdy = yDU << 1, fdc = cDU << 1; + IA64YUVRow p; + p.yu01 = ia64_lanes32(yU.x - yDU, yU.y - yDU); + p.yu23 = ia64_lanes32(yU.z - yDU, yU.w - yDU); + p.cu = ia64_lanes32(cA - cDU, cB - cDU); + p.fxy = ia64_lanes16((yU.x << 1) - fdy, (yU.y << 1) - fdy, + (yU.z << 1) - fdy, (yU.w << 1) - fdy); + p.fxc = ia64_lanes16((cA << 1) - fdc, (cB << 1) - fdc, (cA << 1) - fdc, + (cB << 1) - fdc); + p.ydu = ia64_lanes32(yDU, yDU); + p.cdu = ia64_lanes32(cDU, cDU); + p.fxdy = ia64_lanes16(fdy, fdy, fdy, fdy); + p.fxdc = ia64_lanes16(fdc, fdc, fdc, fdc); + p.y = yRow; + p.u = cRow1; + p.v = cRow2; + p.ystride = yStrideV; + p.cstride = cStrideV; + p.fy = ia64_lanes16(yFracV, yFracV, yFracV, yFracV); + p.fc = ia64_lanes16(cFracV, cFracV, cFracV, cFracV); + p.yc = ia64_low_u64(colorSpace.yCoeffs); + p.yb = ia64_low_u64(colorSpace.yBias); + p.uvb = ia64_low_u64(colorSpace.uvBias); + p.brc = ia64_low_u64(colorSpace.br_uvCoeffs); + p.ggc = ia64_low_u64(colorSpace.gg_uvCoeffs); + p.mask = ia64_low_u64(colorSpace.br_yMask); + swgl_yuv42_row_ia64(dest, span / 4, &p); + return; + } + } +#endif // As much as possible try to utilize the fact that we're only using half // the UV samples to combine Y and UV samples into single vectors. Here we // need to initialize several useful vector quantities for stepping fractional @@ -1099,6 +1227,24 @@ cDU <= (2 << (STEP_BITS + 7))) { // Ensure that samples don't fall outside of the valid bounds of each // planar texture. Step until the initial X coordinates are positive. +#ifdef __ia64__ + if constexpr (!BLEND) { + int n = 0; + for (I32 y = yU, c = cU; (y.x < 0 || c.x < 0) && n < span / 4; n++) { + y += yDU; + c += cDU; + } + if (n > 0) { + ia64_yuv_linear_chunks(dest, n, samplerY, yU, yDU, yOffsetV, + yStrideV, yFracV, samplerU, samplerV, cU, + cDU, cOffsetV, cStrideV, cFracV, colorSpace); + span -= n * 4; + dest += n * 4; + yU += n * yDU; + cU += n * cDU; + } + } +#endif for (; (yU.x < 0 || cU.x < 0) && span >= 4; span -= 4) { auto yPx = textureLinearRowR8(samplerY, yU >> STEP_BITS, yOffsetV, yStrideV, yFracV); @@ -1131,6 +1277,19 @@ // If there are any remaining chunks that weren't inside, handle them // below. } +#ifdef __ia64__ + if constexpr (!BLEND) { + if (int n = span / 4) { + ia64_yuv_linear_chunks(dest, n, samplerY, yU, yDU, yOffsetV, yStrideV, + yFracV, samplerU, samplerV, cU, cDU, cOffsetV, + cStrideV, cFracV, colorSpace); + span -= n * 4; + dest += n * 4; + yU += n * yDU; + cU += n * cDU; + } + } +#endif for (; span >= 4; span -= 4) { // Sample each YUV plane and then transform them by the appropriate // color space. --- firefox-158.0/gfx/wr/swgl/build.rs.vanilla +++ firefox-158.0/gfx/wr/swgl/build.rs @@ -271,6 +271,9 @@ if std::env::var("CARGO_CFG_TARGET_ARCH").as_deref() == Ok("ia64") { println!("cargo:rerun-if-changed=src/swgl_r8row_ia64.S"); build.file("src/swgl_r8row_ia64.S"); + // and the YUV row kernels, from the same generator family + println!("cargo:rerun-if-changed=src/swgl_yuv42_ia64.S"); + build.file("src/swgl_yuv42_ia64.S"); } build --- firefox-158.0/gfx/wr/swgl/src/swgl_yuv42_ia64.S.vanilla +++ firefox-158.0/gfx/wr/swgl/src/swgl_yuv42_ia64.S @@ -0,0 +1,1218 @@ +/* + * Tier III: SWGL 4:2:0 YUV -> BGRA row kernels, modulo-scheduled. + * + * Generated by genyuv_s.py, which lives in tools/ia64-gen/ of the T2 SWGL + * ia64 benchmark tree (not in Firefox): regenerate rather than hand-edit, the + * schedule (stages, rotating register names, bundle pairs) is derived. + * + * void swgl_yuv42_row_ia64(uint32_t* dst, long n, const IA64YUVRow* p); + * void swgl_yuv42_row_linear_ia64(uint32_t* dst, long n, + * const IA64YUVRowLinear* p); + * + * Each runs n four-pixel steps, bit-exact with the C: the first of + * upscaleYUV42R8() (composite.h), whose lane shuffles the gather formulation + * in the generator reproduces, the second of linear_row_yuv's generic R8 loop + * (textureLinearRowR8 + textureLinearRowPairedR8 + convert, any scale). The + * positions (and fractions) in *p are pre-decremented by one step; *p is not + * written. + * + * .explicit is mandatory: in gas's default auto mode the templates and stops + * below are ignored and re-derived. + */ + + .explicit + + .text + .align 32 + .global swgl_yuv42_row_ia64 + .hidden swgl_yuv42_row_ia64 + .proc swgl_yuv42_row_ia64 +swgl_yuv42_row_ia64: + .prologue + .save ar.pfs, r120 + alloc r120 = ar.pfs, 3, 88, 0, 56 + .save ar.lc, r121 + mov r121 = ar.lc + .save pr, r122 + mov r122 = pr + .body + /* the ins are r32-r34, inside the rotating window: copy them out + before the first br.ctop */ + mov r114 = r32 + mov r118 = r33 + mov r119 = r34 + ;; + ld8 r88 = [r119], 8 + ;; + ld8 r89 = [r119], 8 + ;; + ld8 r90 = [r119], 8 + ;; + ld8 r91 = [r119], 8 + ;; + ld8 r92 = [r119], 8 + ;; + ld8 r93 = [r119], 8 + ;; + ld8 r94 = [r119], 8 + ;; + ld8 r95 = [r119], 8 + ;; + ld8 r96 = [r119], 8 + ;; + ld8 r97 = [r119], 8 + ;; + ld8 r98 = [r119], 8 + ;; + ld8 r99 = [r119], 8 + ;; + ld8 r100 = [r119], 8 + ;; + ld8 r101 = [r119], 8 + ;; + ld8 r102 = [r119], 8 + ;; + ld8 r103 = [r119], 8 + ;; + ld8 r104 = [r119], 8 + ;; + ld8 r105 = [r119], 8 + ;; + ld8 r106 = [r119], 8 + ;; + ld8 r107 = [r119], 8 + ;; + ld8 r108 = [r119], 8 + ;; + ld8 r109 = [r119], 8 + ;; + adds r115 = 4, r114 + adds r116 = 8, r114 + adds r117 = 12, r114 + movl r113 = 0xff000000ff000000 + ;; + adds r110 = 1, r97 + adds r111 = 1, r98 + adds r112 = 1, r99 + cmp.ge p6, p0 = r0, r118 + adds r118 = -1, r118 + ;; + (p6) br.cond.dpnt.few .Ldone_yuv42 + ;; + mov ar.lc = r118 + mov r119 = 5 + ;; + mov ar.ec = r119 + mov pr.rot = 0x10000 + ;; +.Lloop_yuv42: + { .mii + (p16) padd4 r88 = r88, r93 + (p17) pshr2 r50 = r43, 7 + (p16) padd4 r90 = r90, r94 + } + { .mii + (p16) padd4 r89 = r89, r93 + (p16) padd2 r91 = r91, r95 + (p16) padd2 r92 = r92, r96 + } + ;; + { .mii + (p20) st4 [r117] = r56, 16 + (p17) pshr2 r40 = r52, 7 + (p19) pshr2 r54 = r55, 6 + } + { .mmf + nop.m 0 + nop.m 0 + nop.f 0 + } + ;; + { .mii + (p20) st4 [r114] = r47, 16 + (p17) pshr2 r51 = r36, 7 + (p19) pshr2 r55 = r53, 6 + } + { .mmi + nop.m 0 + nop.m 0 + (p18) psub2 r52 = r37, r105 + } + ;; + { .mii + nop.m 0 + (p16) extr.u r53 = r90, 15, 17 + (p19) pack2.uss r36 = r49, r38 + } + { .mmf + nop.m 0 + nop.m 0 + nop.f 0 + } + ;; + { .mii + (p16) padd4 r32 = r88, r93 + (p16) extr.u r56 = r88, 47, 17 + (p16) extr.u r58 = r90, 47, 17 + } + { .mii + nop.m 0 + (p18) padd2 r60 = r46, r41 + (p18) padd2 r61 = r48, r41 + } + ;; + { .mii + nop.m 0 + (p16) extr.u r57 = r89, 15, 17 + (p18) unpack2.h r37 = r52, r52 + } + { .mmi + nop.m 0 + nop.m 0 + (p16) padd4 r46 = r90, r94 + } + ;; + { .mii + nop.m 0 + (p16) extr.u r65 = r88, 15, 17 + (p16) extr.u r59 = r89, 47, 17 + } + { .mii + nop.m 0 + (p17) padd2 r43 = r44, r50 + (p17) padd2 r41 = r34, r39 + } + ;; + { .mii + (p16) pcmp4.gt r49 = r58, r53 + (p16) extr.u r48 = r32, 15, 17 + (p16) pshr2.u r34 = r91, 9 + } + { .mii + (p17) padd2 r39 = r45, r51 + (p18) psub2 r38 = r60, r106 + (p18) psub2 r64 = r61, r106 + } + ;; + { .mii + (p16) add r67 = r98, r53 + (p16) extr.u r66 = r46, 15, 17 + (p16) pshr2.u r32 = r92, 9 + } + { .mii + (p16) add r69 = r111, r53 + (p16) add r68 = r99, r53 + (p16) add r50 = r112, r53 + } + ;; + { .mii + (p16) ld1 r45 = [r69], r101 + (p16) pcmp4.gt r51 = r56, r65 + (p16) add r73 = r97, r65 + } + { .mii + (p16) ld1 r62 = [r68], r101 + (p16) add r60 = r110, r65 + (p17) padd2 r74 = r42, r40 + } + ;; + { .mii + (p16) ld1 r61 = [r73], r100 + (p18) pmpyshr2 r44 = r64, r108, 0 + (p19) pack2.uss r40 = r55, r54 + } + { .mii + (p16) ld1 r63 = [r68] + (p16) psub4 r70 = r53, r49 + (p17) psub2 r75 = r39, r43 + } + ;; + { .mii + (p16) ld1 r54 = [r73] + (p18) pmpyshr2 r39 = r64, r107, 0 + (p20) shr.u r68 = r47, 32 + } + { .mii + (p16) ld1 r46 = [r69] + (p16) pcmp4.gt r71 = r66, r58 + (p16) adds r72 = 1, r66 + } + ;; + { .mii + (p16) ld1 r66 = [r60], r100 + (p16) psub4 r73 = r65, r51 + (p16) pcmp4.gt r55 = r57, r56 + } + { .mii + (p16) ld1 r69 = [r67], r101 + (p16) adds r64 = 1, r53 + (p17) psub2 r76 = r74, r41 + } + ;; + { .mii + (p16) ld1 r53 = [r67] + (p17) pmpyshr2 r58 = r75, r35, 0 + (p18) unpack2.l r42 = r52, r52 + } + { .mii + (p16) pcmp4.gt r56 = r59, r57 + (p16) add r77 = r98, r70 + (p16) add r78 = r99, r70 + } + ;; + { .mii + (p16) ld1 r35 = [r77], r101 + (p18) pmpyshr2 r47 = r38, r108, 0 + (p19) unpack1.h r74 = r36, r40 + } + { .mii + (p16) ld1 r75 = [r78], r101 + (p16) pcmp4.gt r57 = r48, r59 + (p16) padd4 r67 = r72, r71 + } + ;; + { .mii + (p16) ld1 r80 = [r77] + (p17) pmpyshr2 r65 = r76, r33, 0 + (p16) psub4 r52 = r73, r55 + } + { .mii + (p16) ld1 r81 = [r78] + (p16) add r59 = r97, r73 + (p16) add r79 = r110, r73 + } + ;; + { .mii + (p16) ld1 r33 = [r59], r100 + (p16) dep r71 = r35, r69, 16, 16 + (p16) dep r72 = r75, r62, 16, 16 + } + { .mmi + (p16) ld1 r73 = [r79], r100 + (p20) st4 [r115] = r68, 16 + (p16) adds r70 = 1, r48 + } + ;; + { .mii + (p16) ld1 r75 = [r60] + (p16) dep r68 = r80, r53, 16, 16 + (p16) dep r69 = r81, r63, 16, 16 + } + { .mii + (p16) ld1 r76 = [r79] + (p16) xor r78 = r67, r64 + (p19) or r35 = r74, r113 + } + ;; + { .mii + (p16) ld1 r74 = [r59] + (p16) dep r53 = r33, r61, 16, 16 + (p16) dep r48 = r73, r66, 16, 16 + } + { .mii + (p16) psub4 r62 = r52, r56 + (p16) add r77 = r97, r52 + (p16) and r60 = r78, r49 + } + ;; + { .mii + (p16) ld1 r59 = [r77], r100 + (p16) dep r49 = r76, r75, 16, 16 + (p16) mix4.r r33 = r72, r71 + } + { .mii + (p16) ld1 r63 = [r50], r101 + (p16) padd4 r67 = r70, r57 + (p16) xor r61 = r60, r64 + } + ;; + { .mii + (p16) ld1 r60 = [r77] + (p16) dep r57 = r74, r54, 16, 16 + (p16) mix4.r r66 = r69, r68 + } + { .mii + (p16) ld1 r64 = [r50] + (p16) add r70 = r98, r61 + (p16) add r71 = r99, r61 + } + ;; + { .mii + (p16) ld1 r50 = [r70], r101 + (p17) pshr2 r69 = r65, 7 + (p18) pshr4.u r54 = r44, 16 + } + { .mii + (p16) ld1 r72 = [r71], r101 + (p16) and r73 = r51, r55 + (p16) adds r68 = 1, r62 + } + ;; + { .mii + (p16) ld1 r65 = [r71] + (p17) pshr2 r61 = r58, 7 + (p18) pshr4.u r51 = r47, 16 + } + { .mii + (p16) and r74 = r73, r56 + (p16) xor r75 = r67, r68 + (p16) add r55 = r97, r62 + } + ;; + { .mii + (p16) ld1 r67 = [r55], r100 + (p16) dep r56 = r50, r45, 16, 16 + (p16) dep r58 = r72, r63, 16, 16 + } + { .mmi + (p16) ld1 r71 = [r70] + nop.m 0 + (p16) and r73 = r75, r74 + } + ;; + { .mii + (p16) ld1 r70 = [r55] + (p16) dep r45 = r65, r64, 16, 16 + (p16) xor r72 = r73, r68 + } + { .mii + (p19) st4 [r116] = r35, 16 + (p16) psub2 r62 = r66, r33 + (p17) padd2 r50 = r41, r69 + } + ;; + { .mii + nop.m 0 + (p16) dep r63 = r67, r59, 16, 16 + (p16) dep r55 = r71, r46, 16, 16 + } + { .mii + nop.m 0 + (p16) add r65 = r110, r52 + (p16) add r66 = r97, r72 + } + ;; + { .mii + (p16) ld1 r52 = [r65], r100 + (p16) dep r59 = r70, r60, 16, 16 + (p16) mix4.r r41 = r58, r56 + } + { .mmf + (p16) ld1 r67 = [r66], r100 + nop.m 0 + nop.f 0 + } + ;; + { .mii + (p16) ld1 r68 = [r65] + (p16) pmpyshr2 r46 = r62, r103, 0 + (p17) mux2 r64 = r50, 0xb1 + } + { .mmi + (p16) ld1 r69 = [r66] + nop.m 0 + (p17) padd2 r60 = r43, r61 + } + ;; + { .mii + nop.m 0 + (p16) mix4.r r43 = r63, r53 + (p16) dep r56 = r67, r52, 16, 16 + } + { .mmi + nop.m 0 + nop.m 0 + (p18) padd2.sss r61 = r47, r51 + } + ;; + { .mii + nop.m 0 + (p16) mix4.r r52 = r59, r57 + (p16) dep r58 = r69, r68, 16, 16 + } + { .mmi + nop.m 0 + nop.m 0 + (p18) padd2.sss r62 = r44, r54 + } + ;; + { .mii + nop.m 0 + (p17) pmpyshr2.u r51 = r60, r104, 0 + (p16) mix4.r r53 = r45, r55 + } + { .mmi + nop.m 0 + nop.m 0 + (p17) psub2 r57 = r50, r64 + } + ;; + { .mii + nop.m 0 + (p16) mix4.r r44 = r56, r48 + (p17) mux2 r45 = r50, 0x88 + } + { .mmi + nop.m 0 + nop.m 0 + (p18) padd2.sss r59 = r42, r61 + } + ;; + { .mii + nop.m 0 + (p16) mix4.r r54 = r58, r49 + (p17) mux2 r47 = r50, 0xdd + } + { .mmi + nop.m 0 + nop.m 0 + (p18) padd2.sss r56 = r37, r62 + } + ;; + { .mii + nop.m 0 + (p18) pmpyshr2 r50 = r38, r107, 0 + (p17) pshr2 r49 = r57, 2 + } + { .mmi + nop.m 0 + nop.m 0 + (p16) psub2 r55 = r52, r43 + } + ;; + { .mii + nop.m 0 + (p16) pshr2 r38 = r46, 7 + (p18) pshr2 r48 = r59, 6 + } + { .mii + nop.m 0 + (p16) psub2 r52 = r53, r41 + (p18) and r57 = r37, r109 + } + ;; + { .mii + nop.m 0 + (p18) pshr2 r37 = r56, 6 + (p19) unpack1.l r59 = r36, r40 + } + { .mii + nop.m 0 + (p16) psub2 r53 = r54, r44 + (p18) and r58 = r42, r109 + } + ;; + { .mii + nop.m 0 + (p16) pmpyshr2 r42 = r55, r102, 0 + (p17) pshr2.u r36 = r51, 1 + } + { .mmf + nop.m 0 + nop.m 0 + nop.f 0 + } + ;; + { .mii + nop.m 0 + (p16) pmpyshr2 r51 = r52, r103, 0 + (p19) shr.u r55 = r35, 32 + } + { .mmi + nop.m 0 + nop.m 0 + (p18) padd2.sss r54 = r57, r39 + } + ;; + { .mii + (p18) padd2.sss r52 = r58, r50 + (p16) pmpyshr2 r35 = r53, r102, 0 + (p17) mux2 r40 = r49, 0xd8 + } + { .mib + nop.m 0 + (p19) or r46 = r59, r113 + br.ctop.sptk.few .Lloop_yuv42 + } + ;; +.Ldone_yuv42: + mov ar.lc = r121 + mov pr = r122, -1 + mov ar.pfs = r120 + br.ret.sptk.many b0 + .endp swgl_yuv42_row_ia64 + + .text + .align 32 + .global swgl_yuv42_row_linear_ia64 + .hidden swgl_yuv42_row_linear_ia64 + .proc swgl_yuv42_row_linear_ia64 +swgl_yuv42_row_linear_ia64: + .prologue + .save ar.pfs, r121 + alloc r121 = ar.pfs, 3, 89, 0, 56 + .save ar.lc, r122 + mov r122 = ar.lc + .save pr, r123 + mov r123 = pr + .body + /* the ins are r32-r34, inside the rotating window: copy them out + before the first br.ctop */ + mov r115 = r32 + mov r119 = r33 + mov r120 = r34 + ;; + ld8 r88 = [r120], 8 + ;; + ld8 r89 = [r120], 8 + ;; + ld8 r90 = [r120], 8 + ;; + ld8 r91 = [r120], 8 + ;; + ld8 r92 = [r120], 8 + ;; + ld8 r93 = [r120], 8 + ;; + ld8 r94 = [r120], 8 + ;; + ld8 r95 = [r120], 8 + ;; + ld8 r96 = [r120], 8 + ;; + ld8 r97 = [r120], 8 + ;; + ld8 r98 = [r120], 8 + ;; + ld8 r99 = [r120], 8 + ;; + ld8 r100 = [r120], 8 + ;; + ld8 r101 = [r120], 8 + ;; + ld8 r102 = [r120], 8 + ;; + ld8 r103 = [r120], 8 + ;; + ld8 r104 = [r120], 8 + ;; + ld8 r105 = [r120], 8 + ;; + ld8 r106 = [r120], 8 + ;; + ld8 r107 = [r120], 8 + ;; + ld8 r108 = [r120], 8 + ;; + ld8 r109 = [r120], 8 + ;; + adds r116 = 4, r115 + adds r117 = 8, r115 + adds r118 = 12, r115 + movl r114 = 0xff000000ff000000 + movl r113 = 0xffffffff + ;; + adds r110 = 1, r94 + adds r111 = 1, r95 + adds r112 = 1, r96 + cmp.ge p6, p0 = r0, r119 + adds r119 = -1, r119 + ;; + (p6) br.cond.dpnt.few .Ldone_yuv42lin + ;; + mov ar.lc = r119 + mov r120 = 6 + ;; + mov ar.ec = r120 + mov pr.rot = 0x10000 + ;; +.Lloop_yuv42lin: + { .mii + (p16) padd4 r88 = r88, r92 + (p17) mix4.r r37 = r63, r67 + (p17) dep r45 = r74, r73, 16, 16 + } + { .mii + (p16) padd4 r89 = r89, r92 + (p16) padd4 r90 = r90, r93 + (p16) padd4 r91 = r91, r93 + } + ;; + { .mii + nop.m 0 + (p17) pmpyshr2 r40 = r50, r100, 0 + (p17) dep r63 = r69, r49, 16, 16 + } + { .mii + nop.m 0 + (p18) psub2 r56 = r70, r106 + (p20) or r43 = r71, r114 + } + ;; + { .mii + (p21) st4 [r116] = r52, 16 + (p18) unpack2.l r49 = r55, r55 + (p19) unpack1.h r69 = r36, r39 + } + { .mmf + (p20) st4 [r118] = r51, 16 + nop.m 0 + nop.f 0 + } + ;; + { .mii + nop.m 0 + (p18) pmpyshr2 r65 = r62, r107, 0 + (p18) pshr4.u r67 = r64, 16 + } + { .mii + nop.m 0 + (p17) padd2 r50 = r59, r68 + (p17) padd2 r66 = r57, r48 + } + ;; + { .mii + (p20) st4 [r115] = r43, 16 + (p18) pmpyshr2 r51 = r56, r108, 0 + (p17) mix4.r r52 = r61, r58 + } + { .mmf + nop.m 0 + nop.m 0 + nop.f 0 + } + ;; + { .mii + (p17) psub2 r48 = r37, r47 + nop.i 0 + nop.i 0 + } + { .mii + nop.m 0 + (p18) and r58 = r49, r109 + (p19) or r44 = r69, r114 + } + ;; + { .mii + nop.m 0 + (p16) pshr4 r37 = r90, 15 + nop.i 0 + } + { .mii + nop.m 0 + (p17) psub2 r57 = r66, r50 + (p18) padd2.sss r59 = r64, r67 + } + ;; + { .mii + nop.m 0 + (p16) pshr4 r32 = r88, 15 + (p17) mix4.r r62 = r63, r45 + } + { .mmf + nop.m 0 + nop.m 0 + nop.f 0 + } + ;; + { .mii + (p19) st4 [r117] = r44, 16 + (p17) pmpyshr2 r63 = r48, r100, 0 + (p16) pshr4 r45 = r88, 8 + } + { .mmi + nop.m 0 + nop.m 0 + (p18) padd2.sss r64 = r58, r65 + } + ;; + { .mii + nop.m 0 + (p17) pmpyshr2 r58 = r57, r34, 0 + (p16) pshr4 r61 = r89, 15 + } + { .mmi + nop.m 0 + nop.m 0 + (p18) padd2.sss r65 = r49, r59 + } + ;; + { .mii + nop.m 0 + (p16) pshr4 r66 = r89, 8 + (p17) pshr2 r34 = r60, 7 + } + { .mii + nop.m 0 + (p17) and r68 = r46, r103 + (p17) psub2 r59 = r62, r52 + } + ;; + { .mii + nop.m 0 + (p17) pshr2 r57 = r63, 7 + (p18) pshr2 r48 = r64, 6 + } + { .mmi + nop.m 0 + nop.m 0 + (p16) pcmp4.gt r60 = r0, r32 + } + ;; + { .mii + nop.m 0 + (p18) pshr2 r49 = r65, 6 + (p18) pshr4.u r46 = r51, 16 + } + { .mii + nop.m 0 + (p16) pcmp4.gt r62 = r32, r101 + (p16) pcmp4.gt r64 = r0, r61 + } + ;; + { .mii + nop.m 0 + (p16) pshr4 r63 = r91, 15 + (p17) pshr2 r67 = r40, 7 + } + { .mii + nop.m 0 + (p16) pcmp4.gt r65 = r61, r101 + (p17) andcm r71 = r68, r41 + } + ;; + { .mii + (p16) andcm r41 = r32, r60 + nop.i 0 + nop.i 0 + } + { .mii + (p16) and r69 = r45, r103 + (p16) and r70 = r66, r103 + (p17) or r68 = r71, r42 + } + ;; + { .mii + (p16) andcm r45 = r41, r62 + nop.i 0 + nop.i 0 + } + { .mii + (p16) and r66 = r101, r62 + (p16) andcm r71 = r61, r64 + (p16) pcmp4.gt r40 = r0, r37 + } + ;; + { .mii + (p16) or r61 = r45, r66 + (p16) and r73 = r101, r65 + (p16) pcmp4.gt r41 = r37, r102 + } + { .mii + (p16) andcm r72 = r71, r65 + (p16) pcmp4.gt r32 = r0, r63 + (p17) padd2 r42 = r53, r34 + } + ;; + { .mii + (p16) andcm r71 = r69, r60 + nop.i 0 + (p16) andcm r74 = r70, r64 + } + { .mii + (p16) or r53 = r72, r73 + (p16) pcmp4.gt r34 = r63, r102 + (p17) padd2 r66 = r54, r67 + } + ;; + { .mii + (p16) and r54 = r61, r113 + (p16) pshr4 r45 = r90, 8 + (p16) or r69 = r74, r65 + } + { .mii + (p16) or r67 = r71, r62 + (p16) andcm r70 = r37, r40 + (p17) and r64 = r68, r103 + } + ;; + { .mii + (p16) andcm r68 = r70, r41 + (p16) shr.u r60 = r61, 32 + (p16) pshr4 r37 = r91, 8 + } + { .mii + (p16) and r71 = r102, r41 + (p16) andcm r73 = r63, r32 + (p17) and r65 = r38, r103 + } + ;; + { .mii + (p16) or r61 = r68, r71 + (p17) pmpyshr2 r62 = r59, r100, 0 + (p17) pshr2 r38 = r58, 7 + } + { .mii + nop.m 0 + (p16) andcm r63 = r73, r34 + (p16) add r72 = r110, r54 + } + ;; + { .mii + (p16) ld1 r74 = [r72], r97 + (p16) shr.u r70 = r53, 32 + (p16) and r71 = r69, r103 + } + { .mii + (p16) and r68 = r67, r103 + (p16) add r58 = r110, r60 + (p17) psub2 r59 = r66, r42 + } + ;; + { .mii + (p16) ld1 r67 = [r72] + (p17) mux2 r78 = r64, 0xa0 + (p16) and r66 = r53, r113 + } + { .mii + (p16) ld1 r69 = [r58], r97 + (p16) and r73 = r102, r34 + (p17) andcm r75 = r65, r33 + } + ;; + { .mii + (p16) ld1 r80 = [r58] + (p16) shr.u r65 = r61, 32 + (p17) or r79 = r75, r35 + } + { .mii + (p16) or r53 = r63, r73 + (p16) add r76 = r110, r66 + (p16) add r77 = r110, r70 + } + ;; + { .mii + (p16) ld1 r63 = [r76], r97 + (p16) pack4.sss r33 = r68, r71 + (p16) dep r58 = r69, r74, 16, 16 + } + { .mii + (p16) ld1 r81 = [r77], r97 + (p16) add r72 = r94, r66 + (p16) add r73 = r94, r70 + } + ;; + { .mii + (p16) ld1 r74 = [r72], r97 + (p17) pmpyshr2 r69 = r59, r78, 0 + (p16) dep r64 = r80, r67, 16, 16 + } + { .mmi + (p16) ld1 r75 = [r73], r97 + nop.m 0 + (p16) add r66 = r94, r54 + } + ;; + { .mii + (p16) ld1 r80 = [r76] + (p16) shr.u r35 = r53, 32 + (p16) dep r71 = r81, r63, 16, 16 + } + { .mii + (p16) ld1 r83 = [r77] + (p17) and r82 = r79, r103 + (p16) add r70 = r94, r60 + } + ;; + { .mii + (p16) ld1 r78 = [r66], r97 + (p18) pmpyshr2 r54 = r56, r107, 0 + (p17) pshr2 r63 = r62, 7 + } + { .mmf + (p16) ld1 r79 = [r70], r97 + nop.m 0 + nop.f 0 + } + ;; + { .mii + (p16) ld1 r56 = [r72] + (p16) dep r76 = r83, r80, 16, 16 + (p18) unpack2.h r59 = r55, r55 + } + { .mii + (p16) ld1 r62 = [r73] + (p16) add r67 = r111, r65 + (p16) add r68 = r112, r65 + } + ;; + { .mii + (p16) ld1 r80 = [r67], r98 + (p17) mux2 r60 = r82, 0xa0 + (p16) dep r77 = r75, r74, 16, 16 + } + { .mii + (p16) ld1 r81 = [r68], r98 + (p16) add r72 = r95, r65 + (p16) add r73 = r96, r65 + } + ;; + { .mii + (p16) ld1 r82 = [r72], r98 + (p16) dep r84 = r79, r78, 16, 16 + (p16) dep r65 = r62, r56, 16, 16 + } + { .mmi + (p16) ld1 r83 = [r73], r98 + nop.m 0 + (p16) and r55 = r61, r113 + } + ;; + { .mii + (p16) ld1 r78 = [r66] + (p16) mix4.r r56 = r71, r58 + (p17) pshr2 r61 = r69, 7 + } + { .mii + (p16) ld1 r79 = [r70] + (p16) add r74 = r95, r55 + (p16) add r75 = r96, r55 + } + ;; + { .mii + (p16) ld1 r71 = [r74], r98 + (p16) mix4.r r66 = r76, r64 + (p16) dep r62 = r81, r80, 16, 16 + } + { .mmf + (p16) ld1 r85 = [r75], r98 + nop.m 0 + nop.f 0 + } + ;; + { .mii + (p16) ld1 r64 = [r72] + (p16) mix4.r r58 = r77, r84 + (p16) dep r69 = r83, r82, 16, 16 + } + { .mmf + (p16) ld1 r76 = [r73] + nop.m 0 + nop.f 0 + } + ;; + { .mii + (p16) ld1 r77 = [r67] + (p16) dep r72 = r79, r78, 16, 16 + (p16) dep r70 = r85, r71, 16, 16 + } + { .mmf + (p16) ld1 r80 = [r68] + nop.m 0 + nop.f 0 + } + ;; + { .mii + (p16) ld1 r78 = [r74] + (p16) dep r68 = r76, r64, 16, 16 + (p16) add r79 = r111, r55 + } + { .mii + (p16) ld1 r81 = [r75] + (p16) add r82 = r112, r55 + (p16) psub2 r67 = r66, r56 + } + ;; + { .mii + (p16) ld1 r75 = [r79], r98 + (p16) dep r64 = r80, r77, 16, 16 + (p16) and r71 = r53, r113 + } + { .mii + (p16) ld1 r76 = [r82], r98 + (p17) padd2 r55 = r47, r57 + (p17) padd2 r73 = r52, r63 + } + ;; + { .mii + (p16) ld1 r53 = [r79] + (p16) mix4.r r57 = r65, r72 + (p16) dep r47 = r81, r78, 16, 16 + } + { .mii + (p16) ld1 r77 = [r82] + (p16) add r63 = r95, r71 + (p16) add r66 = r96, r71 + } + ;; + { .mii + (p16) ld1 r78 = [r63], r98 + (p16) pmpyshr2 r65 = r67, r99, 0 + (p16) dep r74 = r76, r75, 16, 16 + } + { .mmf + (p16) ld1 r79 = [r66], r98 + nop.m 0 + nop.f 0 + } + ;; + { .mii + (p16) add r80 = r95, r35 + (p16) mix4.r r52 = r69, r70 + (p16) dep r72 = r77, r53, 16, 16 + } + { .mii + nop.m 0 + (p16) add r81 = r96, r35 + (p17) padd2 r75 = r50, r38 + } + ;; + { .mii + (p16) ld1 r82 = [r80], r98 + (p16) mix4.r r67 = r68, r47 + (p16) dep r38 = r79, r78, 16, 16 + } + { .mii + (p16) ld1 r83 = [r81], r98 + (p17) psub2 r70 = r73, r55 + (p18) padd2.sss r76 = r51, r46 + } + ;; + { .mii + (p16) ld1 r68 = [r80] + (p16) pshr2 r47 = r65, 7 + (p16) mix4.r r53 = r62, r74 + } + { .mii + (p16) ld1 r79 = [r81] + (p16) add r50 = r111, r71 + (p16) add r51 = r112, r71 + } + ;; + { .mii + (p16) ld1 r80 = [r50], r98 + (p17) pmpyshr2.u r73 = r75, r104, 0 + (p16) dep r69 = r83, r82, 16, 16 + } + { .mii + (p16) ld1 r81 = [r51], r98 + (p16) psub2 r46 = r57, r58 + (p18) and r77 = r59, r109 + } + ;; + { .mii + nop.m 0 + (p17) pmpyshr2 r75 = r70, r60, 0 + (p16) dep r62 = r79, r68, 16, 16 + } + { .mii + nop.m 0 + (p17) padd2 r74 = r42, r61 + (p18) padd2.sss r78 = r59, r76 + } + ;; + { .mii + (p16) ld1 r59 = [r63] + (p16) mix4.r r68 = r64, r72 + (p16) dep r57 = r81, r80, 16, 16 + } + { .mii + (p16) ld1 r71 = [r66] + (p16) add r42 = r111, r35 + (p16) add r65 = r112, r35 + } + ;; + { .mii + (p16) ld1 r63 = [r42], r98 + (p16) pmpyshr2 r70 = r46, r99, 0 + (p17) pshr2.u r64 = r73, 1 + } + { .mmi + (p16) ld1 r76 = [r65], r98 + nop.m 0 + (p18) padd2.sss r35 = r77, r54 + } + ;; + { .mii + nop.m 0 + (p17) pshr2 r54 = r75, 7 + (p18) pshr2 r73 = r78, 6 + } + { .mmi + nop.m 0 + nop.m 0 + (p17) psub2 r61 = r74, r106 + } + ;; + { .mii + nop.m 0 + (p16) dep r66 = r71, r59, 16, 16 + (p16) dep r60 = r76, r63, 16, 16 + } + { .mmi + nop.m 0 + nop.m 0 + (p16) psub2 r72 = r67, r52 + } + ;; + { .mii + nop.m 0 + (p16) mix4.r r46 = r69, r38 + (p18) pshr2 r71 = r35, 6 + } + { .mmf + nop.m 0 + nop.m 0 + nop.f 0 + } + ;; + { .mii + nop.m 0 + (p17) pmpyshr2 r63 = r61, r108, 0 + (p18) pack2.uss r35 = r49, r73 + } + { .mmf + nop.m 0 + nop.m 0 + nop.f 0 + } + ;; + { .mii + nop.m 0 + (p16) pmpyshr2 r59 = r72, r100, 0 + (p16) pshr2 r67 = r70, 7 + } + { .mmf + nop.m 0 + nop.m 0 + nop.f 0 + } + ;; + { .mii + (p16) ld1 r72 = [r50] + (p18) pack2.uss r38 = r48, r71 + (p19) unpack1.l r70 = r36, r39 + } + { .mii + (p16) ld1 r73 = [r51] + (p16) psub2 r49 = r68, r53 + (p17) padd2 r69 = r55, r54 + } + ;; + { .mii + (p16) ld1 r48 = [r42] + (p20) shr.u r51 = r43, 32 + (p19) shr.u r50 = r44, 32 + } + { .mib + (p16) ld1 r68 = [r65] + (p17) psub2 r54 = r64, r105 + br.ctop.sptk.few .Lloop_yuv42lin + } + ;; +.Ldone_yuv42lin: + mov ar.lc = r122 + mov pr = r123, -1 + mov ar.pfs = r121 + br.ret.sptk.many b0 + .endp swgl_yuv42_row_linear_ia64