yading@10: /* yading@10: * VC-1 and WMV3 - DSP functions MMX-optimized yading@10: * Copyright (c) 2007 Christophe GISQUET yading@10: * yading@10: * Permission is hereby granted, free of charge, to any person yading@10: * obtaining a copy of this software and associated documentation yading@10: * files (the "Software"), to deal in the Software without yading@10: * restriction, including without limitation the rights to use, yading@10: * copy, modify, merge, publish, distribute, sublicense, and/or sell yading@10: * copies of the Software, and to permit persons to whom the yading@10: * Software is furnished to do so, subject to the following yading@10: * conditions: yading@10: * yading@10: * The above copyright notice and this permission notice shall be yading@10: * included in all copies or substantial portions of the Software. yading@10: * yading@10: * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, yading@10: * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES yading@10: * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND yading@10: * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT yading@10: * HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, yading@10: * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING yading@10: * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR yading@10: * OTHER DEALINGS IN THE SOFTWARE. yading@10: */ yading@10: yading@10: #include "libavutil/cpu.h" yading@10: #include "libavutil/mem.h" yading@10: #include "libavutil/x86/asm.h" yading@10: #include "libavutil/x86/cpu.h" yading@10: #include "dsputil_mmx.h" yading@10: #include "libavcodec/vc1dsp.h" yading@10: #include "vc1dsp.h" yading@10: yading@10: #if HAVE_INLINE_ASM yading@10: yading@10: #define OP_PUT(S,D) yading@10: #define OP_AVG(S,D) "pavgb " #S ", " #D " \n\t" yading@10: yading@10: /** Add rounder from mm7 to mm3 and pack result at destination */ yading@10: #define NORMALIZE_MMX(SHIFT) \ yading@10: "paddw %%mm7, %%mm3 \n\t" /* +bias-r */ \ yading@10: "paddw %%mm7, %%mm4 \n\t" /* +bias-r */ \ yading@10: "psraw "SHIFT", %%mm3 \n\t" \ yading@10: "psraw "SHIFT", %%mm4 \n\t" yading@10: yading@10: #define TRANSFER_DO_PACK(OP) \ yading@10: "packuswb %%mm4, %%mm3 \n\t" \ yading@10: OP((%2), %%mm3) \ yading@10: "movq %%mm3, (%2) \n\t" yading@10: yading@10: #define TRANSFER_DONT_PACK(OP) \ yading@10: OP(0(%2), %%mm3) \ yading@10: OP(8(%2), %%mm4) \ yading@10: "movq %%mm3, 0(%2) \n\t" \ yading@10: "movq %%mm4, 8(%2) \n\t" yading@10: yading@10: /** @see MSPEL_FILTER13_CORE for use as UNPACK macro */ yading@10: #define DO_UNPACK(reg) "punpcklbw %%mm0, " reg "\n\t" yading@10: #define DONT_UNPACK(reg) yading@10: yading@10: /** Compute the rounder 32-r or 8-r and unpacks it to mm7 */ yading@10: #define LOAD_ROUNDER_MMX(ROUND) \ yading@10: "movd "ROUND", %%mm7 \n\t" \ yading@10: "punpcklwd %%mm7, %%mm7 \n\t" \ yading@10: "punpckldq %%mm7, %%mm7 \n\t" yading@10: yading@10: #define SHIFT2_LINE(OFF, R0,R1,R2,R3) \ yading@10: "paddw %%mm"#R2", %%mm"#R1" \n\t" \ yading@10: "movd (%0,%3), %%mm"#R0" \n\t" \ yading@10: "pmullw %%mm6, %%mm"#R1" \n\t" \ yading@10: "punpcklbw %%mm0, %%mm"#R0" \n\t" \ yading@10: "movd (%0,%2), %%mm"#R3" \n\t" \ yading@10: "psubw %%mm"#R0", %%mm"#R1" \n\t" \ yading@10: "punpcklbw %%mm0, %%mm"#R3" \n\t" \ yading@10: "paddw %%mm7, %%mm"#R1" \n\t" \ yading@10: "psubw %%mm"#R3", %%mm"#R1" \n\t" \ yading@10: "psraw %4, %%mm"#R1" \n\t" \ yading@10: "movq %%mm"#R1", "#OFF"(%1) \n\t" \ yading@10: "add %2, %0 \n\t" yading@10: yading@10: /** Sacrifying mm6 allows to pipeline loads from src */ yading@10: static void vc1_put_ver_16b_shift2_mmx(int16_t *dst, yading@10: const uint8_t *src, x86_reg stride, yading@10: int rnd, int64_t shift) yading@10: { yading@10: __asm__ volatile( yading@10: "mov $3, %%"REG_c" \n\t" yading@10: LOAD_ROUNDER_MMX("%5") yading@10: "movq "MANGLE(ff_pw_9)", %%mm6 \n\t" yading@10: "1: \n\t" yading@10: "movd (%0), %%mm2 \n\t" yading@10: "add %2, %0 \n\t" yading@10: "movd (%0), %%mm3 \n\t" yading@10: "punpcklbw %%mm0, %%mm2 \n\t" yading@10: "punpcklbw %%mm0, %%mm3 \n\t" yading@10: SHIFT2_LINE( 0, 1, 2, 3, 4) yading@10: SHIFT2_LINE( 24, 2, 3, 4, 1) yading@10: SHIFT2_LINE( 48, 3, 4, 1, 2) yading@10: SHIFT2_LINE( 72, 4, 1, 2, 3) yading@10: SHIFT2_LINE( 96, 1, 2, 3, 4) yading@10: SHIFT2_LINE(120, 2, 3, 4, 1) yading@10: SHIFT2_LINE(144, 3, 4, 1, 2) yading@10: SHIFT2_LINE(168, 4, 1, 2, 3) yading@10: "sub %6, %0 \n\t" yading@10: "add $8, %1 \n\t" yading@10: "dec %%"REG_c" \n\t" yading@10: "jnz 1b \n\t" yading@10: : "+r"(src), "+r"(dst) yading@10: : "r"(stride), "r"(-2*stride), yading@10: "m"(shift), "m"(rnd), "r"(9*stride-4) yading@10: : "%"REG_c, "memory" yading@10: ); yading@10: } yading@10: yading@10: /** yading@10: * Data is already unpacked, so some operations can directly be made from yading@10: * memory. yading@10: */ yading@10: #define VC1_HOR_16b_SHIFT2(OP, OPNAME)\ yading@10: static void OPNAME ## vc1_hor_16b_shift2_mmx(uint8_t *dst, x86_reg stride,\ yading@10: const int16_t *src, int rnd)\ yading@10: {\ yading@10: int h = 8;\ yading@10: \ yading@10: src -= 1;\ yading@10: rnd -= (-1+9+9-1)*1024; /* Add -1024 bias */\ yading@10: __asm__ volatile(\ yading@10: LOAD_ROUNDER_MMX("%4")\ yading@10: "movq "MANGLE(ff_pw_128)", %%mm6\n\t"\ yading@10: "movq "MANGLE(ff_pw_9)", %%mm5 \n\t"\ yading@10: "1: \n\t"\ yading@10: "movq 2*0+0(%1), %%mm1 \n\t"\ yading@10: "movq 2*0+8(%1), %%mm2 \n\t"\ yading@10: "movq 2*1+0(%1), %%mm3 \n\t"\ yading@10: "movq 2*1+8(%1), %%mm4 \n\t"\ yading@10: "paddw 2*3+0(%1), %%mm1 \n\t"\ yading@10: "paddw 2*3+8(%1), %%mm2 \n\t"\ yading@10: "paddw 2*2+0(%1), %%mm3 \n\t"\ yading@10: "paddw 2*2+8(%1), %%mm4 \n\t"\ yading@10: "pmullw %%mm5, %%mm3 \n\t"\ yading@10: "pmullw %%mm5, %%mm4 \n\t"\ yading@10: "psubw %%mm1, %%mm3 \n\t"\ yading@10: "psubw %%mm2, %%mm4 \n\t"\ yading@10: NORMALIZE_MMX("$7")\ yading@10: /* Remove bias */\ yading@10: "paddw %%mm6, %%mm3 \n\t"\ yading@10: "paddw %%mm6, %%mm4 \n\t"\ yading@10: TRANSFER_DO_PACK(OP)\ yading@10: "add $24, %1 \n\t"\ yading@10: "add %3, %2 \n\t"\ yading@10: "decl %0 \n\t"\ yading@10: "jnz 1b \n\t"\ yading@10: : "+r"(h), "+r" (src), "+r" (dst)\ yading@10: : "r"(stride), "m"(rnd)\ yading@10: : "memory"\ yading@10: );\ yading@10: } yading@10: yading@10: VC1_HOR_16b_SHIFT2(OP_PUT, put_) yading@10: VC1_HOR_16b_SHIFT2(OP_AVG, avg_) yading@10: yading@10: yading@10: /** yading@10: * Purely vertical or horizontal 1/2 shift interpolation. yading@10: * Sacrify mm6 for *9 factor. yading@10: */ yading@10: #define VC1_SHIFT2(OP, OPNAME)\ yading@10: static void OPNAME ## vc1_shift2_mmx(uint8_t *dst, const uint8_t *src,\ yading@10: x86_reg stride, int rnd, x86_reg offset)\ yading@10: {\ yading@10: rnd = 8-rnd;\ yading@10: __asm__ volatile(\ yading@10: "mov $8, %%"REG_c" \n\t"\ yading@10: LOAD_ROUNDER_MMX("%5")\ yading@10: "movq "MANGLE(ff_pw_9)", %%mm6\n\t"\ yading@10: "1: \n\t"\ yading@10: "movd 0(%0 ), %%mm3 \n\t"\ yading@10: "movd 4(%0 ), %%mm4 \n\t"\ yading@10: "movd 0(%0,%2), %%mm1 \n\t"\ yading@10: "movd 4(%0,%2), %%mm2 \n\t"\ yading@10: "add %2, %0 \n\t"\ yading@10: "punpcklbw %%mm0, %%mm3 \n\t"\ yading@10: "punpcklbw %%mm0, %%mm4 \n\t"\ yading@10: "punpcklbw %%mm0, %%mm1 \n\t"\ yading@10: "punpcklbw %%mm0, %%mm2 \n\t"\ yading@10: "paddw %%mm1, %%mm3 \n\t"\ yading@10: "paddw %%mm2, %%mm4 \n\t"\ yading@10: "movd 0(%0,%3), %%mm1 \n\t"\ yading@10: "movd 4(%0,%3), %%mm2 \n\t"\ yading@10: "pmullw %%mm6, %%mm3 \n\t" /* 0,9,9,0*/\ yading@10: "pmullw %%mm6, %%mm4 \n\t" /* 0,9,9,0*/\ yading@10: "punpcklbw %%mm0, %%mm1 \n\t"\ yading@10: "punpcklbw %%mm0, %%mm2 \n\t"\ yading@10: "psubw %%mm1, %%mm3 \n\t" /*-1,9,9,0*/\ yading@10: "psubw %%mm2, %%mm4 \n\t" /*-1,9,9,0*/\ yading@10: "movd 0(%0,%2), %%mm1 \n\t"\ yading@10: "movd 4(%0,%2), %%mm2 \n\t"\ yading@10: "punpcklbw %%mm0, %%mm1 \n\t"\ yading@10: "punpcklbw %%mm0, %%mm2 \n\t"\ yading@10: "psubw %%mm1, %%mm3 \n\t" /*-1,9,9,-1*/\ yading@10: "psubw %%mm2, %%mm4 \n\t" /*-1,9,9,-1*/\ yading@10: NORMALIZE_MMX("$4")\ yading@10: "packuswb %%mm4, %%mm3 \n\t"\ yading@10: OP((%1), %%mm3)\ yading@10: "movq %%mm3, (%1) \n\t"\ yading@10: "add %6, %0 \n\t"\ yading@10: "add %4, %1 \n\t"\ yading@10: "dec %%"REG_c" \n\t"\ yading@10: "jnz 1b \n\t"\ yading@10: : "+r"(src), "+r"(dst)\ yading@10: : "r"(offset), "r"(-2*offset), "g"(stride), "m"(rnd),\ yading@10: "g"(stride-offset)\ yading@10: : "%"REG_c, "memory"\ yading@10: );\ yading@10: } yading@10: yading@10: VC1_SHIFT2(OP_PUT, put_) yading@10: VC1_SHIFT2(OP_AVG, avg_) yading@10: yading@10: /** yading@10: * Core of the 1/4 and 3/4 shift bicubic interpolation. yading@10: * yading@10: * @param UNPACK Macro unpacking arguments from 8 to 16bits (can be empty). yading@10: * @param MOVQ "movd 1" or "movq 2", if data read is already unpacked. yading@10: * @param A1 Address of 1st tap (beware of unpacked/packed). yading@10: * @param A2 Address of 2nd tap yading@10: * @param A3 Address of 3rd tap yading@10: * @param A4 Address of 4th tap yading@10: */ yading@10: #define MSPEL_FILTER13_CORE(UNPACK, MOVQ, A1, A2, A3, A4) \ yading@10: MOVQ "*0+"A1", %%mm1 \n\t" \ yading@10: MOVQ "*4+"A1", %%mm2 \n\t" \ yading@10: UNPACK("%%mm1") \ yading@10: UNPACK("%%mm2") \ yading@10: "pmullw "MANGLE(ff_pw_3)", %%mm1\n\t" \ yading@10: "pmullw "MANGLE(ff_pw_3)", %%mm2\n\t" \ yading@10: MOVQ "*0+"A2", %%mm3 \n\t" \ yading@10: MOVQ "*4+"A2", %%mm4 \n\t" \ yading@10: UNPACK("%%mm3") \ yading@10: UNPACK("%%mm4") \ yading@10: "pmullw %%mm6, %%mm3 \n\t" /* *18 */ \ yading@10: "pmullw %%mm6, %%mm4 \n\t" /* *18 */ \ yading@10: "psubw %%mm1, %%mm3 \n\t" /* 18,-3 */ \ yading@10: "psubw %%mm2, %%mm4 \n\t" /* 18,-3 */ \ yading@10: MOVQ "*0+"A4", %%mm1 \n\t" \ yading@10: MOVQ "*4+"A4", %%mm2 \n\t" \ yading@10: UNPACK("%%mm1") \ yading@10: UNPACK("%%mm2") \ yading@10: "psllw $2, %%mm1 \n\t" /* 4* */ \ yading@10: "psllw $2, %%mm2 \n\t" /* 4* */ \ yading@10: "psubw %%mm1, %%mm3 \n\t" /* -4,18,-3 */ \ yading@10: "psubw %%mm2, %%mm4 \n\t" /* -4,18,-3 */ \ yading@10: MOVQ "*0+"A3", %%mm1 \n\t" \ yading@10: MOVQ "*4+"A3", %%mm2 \n\t" \ yading@10: UNPACK("%%mm1") \ yading@10: UNPACK("%%mm2") \ yading@10: "pmullw %%mm5, %%mm1 \n\t" /* *53 */ \ yading@10: "pmullw %%mm5, %%mm2 \n\t" /* *53 */ \ yading@10: "paddw %%mm1, %%mm3 \n\t" /* 4,53,18,-3 */ \ yading@10: "paddw %%mm2, %%mm4 \n\t" /* 4,53,18,-3 */ yading@10: yading@10: /** yading@10: * Macro to build the vertical 16bits version of vc1_put_shift[13]. yading@10: * Here, offset=src_stride. Parameters passed A1 to A4 must use yading@10: * %3 (src_stride) and %4 (3*src_stride). yading@10: * yading@10: * @param NAME Either 1 or 3 yading@10: * @see MSPEL_FILTER13_CORE for information on A1->A4 yading@10: */ yading@10: #define MSPEL_FILTER13_VER_16B(NAME, A1, A2, A3, A4) \ yading@10: static void \ yading@10: vc1_put_ver_16b_ ## NAME ## _mmx(int16_t *dst, const uint8_t *src, \ yading@10: x86_reg src_stride, \ yading@10: int rnd, int64_t shift) \ yading@10: { \ yading@10: int h = 8; \ yading@10: src -= src_stride; \ yading@10: __asm__ volatile( \ yading@10: LOAD_ROUNDER_MMX("%5") \ yading@10: "movq "MANGLE(ff_pw_53)", %%mm5\n\t" \ yading@10: "movq "MANGLE(ff_pw_18)", %%mm6\n\t" \ yading@10: ".p2align 3 \n\t" \ yading@10: "1: \n\t" \ yading@10: MSPEL_FILTER13_CORE(DO_UNPACK, "movd 1", A1, A2, A3, A4) \ yading@10: NORMALIZE_MMX("%6") \ yading@10: TRANSFER_DONT_PACK(OP_PUT) \ yading@10: /* Last 3 (in fact 4) bytes on the line */ \ yading@10: "movd 8+"A1", %%mm1 \n\t" \ yading@10: DO_UNPACK("%%mm1") \ yading@10: "movq %%mm1, %%mm3 \n\t" \ yading@10: "paddw %%mm1, %%mm1 \n\t" \ yading@10: "paddw %%mm3, %%mm1 \n\t" /* 3* */ \ yading@10: "movd 8+"A2", %%mm3 \n\t" \ yading@10: DO_UNPACK("%%mm3") \ yading@10: "pmullw %%mm6, %%mm3 \n\t" /* *18 */ \ yading@10: "psubw %%mm1, %%mm3 \n\t" /*18,-3 */ \ yading@10: "movd 8+"A3", %%mm1 \n\t" \ yading@10: DO_UNPACK("%%mm1") \ yading@10: "pmullw %%mm5, %%mm1 \n\t" /* *53 */ \ yading@10: "paddw %%mm1, %%mm3 \n\t" /*53,18,-3 */ \ yading@10: "movd 8+"A4", %%mm1 \n\t" \ yading@10: DO_UNPACK("%%mm1") \ yading@10: "psllw $2, %%mm1 \n\t" /* 4* */ \ yading@10: "psubw %%mm1, %%mm3 \n\t" \ yading@10: "paddw %%mm7, %%mm3 \n\t" \ yading@10: "psraw %6, %%mm3 \n\t" \ yading@10: "movq %%mm3, 16(%2) \n\t" \ yading@10: "add %3, %1 \n\t" \ yading@10: "add $24, %2 \n\t" \ yading@10: "decl %0 \n\t" \ yading@10: "jnz 1b \n\t" \ yading@10: : "+r"(h), "+r" (src), "+r" (dst) \ yading@10: : "r"(src_stride), "r"(3*src_stride), \ yading@10: "m"(rnd), "m"(shift) \ yading@10: : "memory" \ yading@10: ); \ yading@10: } yading@10: yading@10: /** yading@10: * Macro to build the horizontal 16bits version of vc1_put_shift[13]. yading@10: * Here, offset=16bits, so parameters passed A1 to A4 should be simple. yading@10: * yading@10: * @param NAME Either 1 or 3 yading@10: * @see MSPEL_FILTER13_CORE for information on A1->A4 yading@10: */ yading@10: #define MSPEL_FILTER13_HOR_16B(NAME, A1, A2, A3, A4, OP, OPNAME) \ yading@10: static void \ yading@10: OPNAME ## vc1_hor_16b_ ## NAME ## _mmx(uint8_t *dst, x86_reg stride, \ yading@10: const int16_t *src, int rnd) \ yading@10: { \ yading@10: int h = 8; \ yading@10: src -= 1; \ yading@10: rnd -= (-4+58+13-3)*256; /* Add -256 bias */ \ yading@10: __asm__ volatile( \ yading@10: LOAD_ROUNDER_MMX("%4") \ yading@10: "movq "MANGLE(ff_pw_18)", %%mm6 \n\t" \ yading@10: "movq "MANGLE(ff_pw_53)", %%mm5 \n\t" \ yading@10: ".p2align 3 \n\t" \ yading@10: "1: \n\t" \ yading@10: MSPEL_FILTER13_CORE(DONT_UNPACK, "movq 2", A1, A2, A3, A4) \ yading@10: NORMALIZE_MMX("$7") \ yading@10: /* Remove bias */ \ yading@10: "paddw "MANGLE(ff_pw_128)", %%mm3 \n\t" \ yading@10: "paddw "MANGLE(ff_pw_128)", %%mm4 \n\t" \ yading@10: TRANSFER_DO_PACK(OP) \ yading@10: "add $24, %1 \n\t" \ yading@10: "add %3, %2 \n\t" \ yading@10: "decl %0 \n\t" \ yading@10: "jnz 1b \n\t" \ yading@10: : "+r"(h), "+r" (src), "+r" (dst) \ yading@10: : "r"(stride), "m"(rnd) \ yading@10: : "memory" \ yading@10: ); \ yading@10: } yading@10: yading@10: /** yading@10: * Macro to build the 8bits, any direction, version of vc1_put_shift[13]. yading@10: * Here, offset=src_stride. Parameters passed A1 to A4 must use yading@10: * %3 (offset) and %4 (3*offset). yading@10: * yading@10: * @param NAME Either 1 or 3 yading@10: * @see MSPEL_FILTER13_CORE for information on A1->A4 yading@10: */ yading@10: #define MSPEL_FILTER13_8B(NAME, A1, A2, A3, A4, OP, OPNAME) \ yading@10: static void \ yading@10: OPNAME ## vc1_## NAME ## _mmx(uint8_t *dst, const uint8_t *src, \ yading@10: x86_reg stride, int rnd, x86_reg offset) \ yading@10: { \ yading@10: int h = 8; \ yading@10: src -= offset; \ yading@10: rnd = 32-rnd; \ yading@10: __asm__ volatile ( \ yading@10: LOAD_ROUNDER_MMX("%6") \ yading@10: "movq "MANGLE(ff_pw_53)", %%mm5 \n\t" \ yading@10: "movq "MANGLE(ff_pw_18)", %%mm6 \n\t" \ yading@10: ".p2align 3 \n\t" \ yading@10: "1: \n\t" \ yading@10: MSPEL_FILTER13_CORE(DO_UNPACK, "movd 1", A1, A2, A3, A4) \ yading@10: NORMALIZE_MMX("$6") \ yading@10: TRANSFER_DO_PACK(OP) \ yading@10: "add %5, %1 \n\t" \ yading@10: "add %5, %2 \n\t" \ yading@10: "decl %0 \n\t" \ yading@10: "jnz 1b \n\t" \ yading@10: : "+r"(h), "+r" (src), "+r" (dst) \ yading@10: : "r"(offset), "r"(3*offset), "g"(stride), "m"(rnd) \ yading@10: : "memory" \ yading@10: ); \ yading@10: } yading@10: yading@10: /** 1/4 shift bicubic interpolation */ yading@10: MSPEL_FILTER13_8B (shift1, "0(%1,%4 )", "0(%1,%3,2)", "0(%1,%3 )", "0(%1 )", OP_PUT, put_) yading@10: MSPEL_FILTER13_8B (shift1, "0(%1,%4 )", "0(%1,%3,2)", "0(%1,%3 )", "0(%1 )", OP_AVG, avg_) yading@10: MSPEL_FILTER13_VER_16B(shift1, "0(%1,%4 )", "0(%1,%3,2)", "0(%1,%3 )", "0(%1 )") yading@10: MSPEL_FILTER13_HOR_16B(shift1, "2*3(%1)", "2*2(%1)", "2*1(%1)", "2*0(%1)", OP_PUT, put_) yading@10: MSPEL_FILTER13_HOR_16B(shift1, "2*3(%1)", "2*2(%1)", "2*1(%1)", "2*0(%1)", OP_AVG, avg_) yading@10: yading@10: /** 3/4 shift bicubic interpolation */ yading@10: MSPEL_FILTER13_8B (shift3, "0(%1 )", "0(%1,%3 )", "0(%1,%3,2)", "0(%1,%4 )", OP_PUT, put_) yading@10: MSPEL_FILTER13_8B (shift3, "0(%1 )", "0(%1,%3 )", "0(%1,%3,2)", "0(%1,%4 )", OP_AVG, avg_) yading@10: MSPEL_FILTER13_VER_16B(shift3, "0(%1 )", "0(%1,%3 )", "0(%1,%3,2)", "0(%1,%4 )") yading@10: MSPEL_FILTER13_HOR_16B(shift3, "2*0(%1)", "2*1(%1)", "2*2(%1)", "2*3(%1)", OP_PUT, put_) yading@10: MSPEL_FILTER13_HOR_16B(shift3, "2*0(%1)", "2*1(%1)", "2*2(%1)", "2*3(%1)", OP_AVG, avg_) yading@10: yading@10: typedef void (*vc1_mspel_mc_filter_ver_16bits)(int16_t *dst, const uint8_t *src, x86_reg src_stride, int rnd, int64_t shift); yading@10: typedef void (*vc1_mspel_mc_filter_hor_16bits)(uint8_t *dst, x86_reg dst_stride, const int16_t *src, int rnd); yading@10: typedef void (*vc1_mspel_mc_filter_8bits)(uint8_t *dst, const uint8_t *src, x86_reg stride, int rnd, x86_reg offset); yading@10: yading@10: /** yading@10: * Interpolate fractional pel values by applying proper vertical then yading@10: * horizontal filter. yading@10: * yading@10: * @param dst Destination buffer for interpolated pels. yading@10: * @param src Source buffer. yading@10: * @param stride Stride for both src and dst buffers. yading@10: * @param hmode Horizontal filter (expressed in quarter pixels shift). yading@10: * @param hmode Vertical filter. yading@10: * @param rnd Rounding bias. yading@10: */ yading@10: #define VC1_MSPEL_MC(OP)\ yading@10: static void OP ## vc1_mspel_mc(uint8_t *dst, const uint8_t *src, int stride,\ yading@10: int hmode, int vmode, int rnd)\ yading@10: {\ yading@10: static const vc1_mspel_mc_filter_ver_16bits vc1_put_shift_ver_16bits[] =\ yading@10: { NULL, vc1_put_ver_16b_shift1_mmx, vc1_put_ver_16b_shift2_mmx, vc1_put_ver_16b_shift3_mmx };\ yading@10: static const vc1_mspel_mc_filter_hor_16bits vc1_put_shift_hor_16bits[] =\ yading@10: { NULL, OP ## vc1_hor_16b_shift1_mmx, OP ## vc1_hor_16b_shift2_mmx, OP ## vc1_hor_16b_shift3_mmx };\ yading@10: static const vc1_mspel_mc_filter_8bits vc1_put_shift_8bits[] =\ yading@10: { NULL, OP ## vc1_shift1_mmx, OP ## vc1_shift2_mmx, OP ## vc1_shift3_mmx };\ yading@10: \ yading@10: __asm__ volatile(\ yading@10: "pxor %%mm0, %%mm0 \n\t"\ yading@10: ::: "memory"\ yading@10: );\ yading@10: \ yading@10: if (vmode) { /* Vertical filter to apply */\ yading@10: if (hmode) { /* Horizontal filter to apply, output to tmp */\ yading@10: static const int shift_value[] = { 0, 5, 1, 5 };\ yading@10: int shift = (shift_value[hmode]+shift_value[vmode])>>1;\ yading@10: int r;\ yading@10: DECLARE_ALIGNED(16, int16_t, tmp)[12*8];\ yading@10: \ yading@10: r = (1<<(shift-1)) + rnd-1;\ yading@10: vc1_put_shift_ver_16bits[vmode](tmp, src-1, stride, r, shift);\ yading@10: \ yading@10: vc1_put_shift_hor_16bits[hmode](dst, stride, tmp+1, 64-rnd);\ yading@10: return;\ yading@10: }\ yading@10: else { /* No horizontal filter, output 8 lines to dst */\ yading@10: vc1_put_shift_8bits[vmode](dst, src, stride, 1-rnd, stride);\ yading@10: return;\ yading@10: }\ yading@10: }\ yading@10: \ yading@10: /* Horizontal mode with no vertical mode */\ yading@10: vc1_put_shift_8bits[hmode](dst, src, stride, rnd, 1);\ yading@10: } yading@10: yading@10: VC1_MSPEL_MC(put_) yading@10: VC1_MSPEL_MC(avg_) yading@10: yading@10: /** Macro to ease bicubic filter interpolation functions declarations */ yading@10: #define DECLARE_FUNCTION(a, b) \ yading@10: static void put_vc1_mspel_mc ## a ## b ## _mmx(uint8_t *dst, \ yading@10: const uint8_t *src, \ yading@10: ptrdiff_t stride, \ yading@10: int rnd) \ yading@10: { \ yading@10: put_vc1_mspel_mc(dst, src, stride, a, b, rnd); \ yading@10: }\ yading@10: static void avg_vc1_mspel_mc ## a ## b ## _mmxext(uint8_t *dst, \ yading@10: const uint8_t *src, \ yading@10: ptrdiff_t stride, \ yading@10: int rnd) \ yading@10: { \ yading@10: avg_vc1_mspel_mc(dst, src, stride, a, b, rnd); \ yading@10: } yading@10: yading@10: DECLARE_FUNCTION(0, 1) yading@10: DECLARE_FUNCTION(0, 2) yading@10: DECLARE_FUNCTION(0, 3) yading@10: yading@10: DECLARE_FUNCTION(1, 0) yading@10: DECLARE_FUNCTION(1, 1) yading@10: DECLARE_FUNCTION(1, 2) yading@10: DECLARE_FUNCTION(1, 3) yading@10: yading@10: DECLARE_FUNCTION(2, 0) yading@10: DECLARE_FUNCTION(2, 1) yading@10: DECLARE_FUNCTION(2, 2) yading@10: DECLARE_FUNCTION(2, 3) yading@10: yading@10: DECLARE_FUNCTION(3, 0) yading@10: DECLARE_FUNCTION(3, 1) yading@10: DECLARE_FUNCTION(3, 2) yading@10: DECLARE_FUNCTION(3, 3) yading@10: yading@10: static void vc1_inv_trans_4x4_dc_mmxext(uint8_t *dest, int linesize, yading@10: int16_t *block) yading@10: { yading@10: int dc = block[0]; yading@10: dc = (17 * dc + 4) >> 3; yading@10: dc = (17 * dc + 64) >> 7; yading@10: __asm__ volatile( yading@10: "movd %0, %%mm0 \n\t" yading@10: "pshufw $0, %%mm0, %%mm0 \n\t" yading@10: "pxor %%mm1, %%mm1 \n\t" yading@10: "psubw %%mm0, %%mm1 \n\t" yading@10: "packuswb %%mm0, %%mm0 \n\t" yading@10: "packuswb %%mm1, %%mm1 \n\t" yading@10: ::"r"(dc) yading@10: ); yading@10: __asm__ volatile( yading@10: "movd %0, %%mm2 \n\t" yading@10: "movd %1, %%mm3 \n\t" yading@10: "movd %2, %%mm4 \n\t" yading@10: "movd %3, %%mm5 \n\t" yading@10: "paddusb %%mm0, %%mm2 \n\t" yading@10: "paddusb %%mm0, %%mm3 \n\t" yading@10: "paddusb %%mm0, %%mm4 \n\t" yading@10: "paddusb %%mm0, %%mm5 \n\t" yading@10: "psubusb %%mm1, %%mm2 \n\t" yading@10: "psubusb %%mm1, %%mm3 \n\t" yading@10: "psubusb %%mm1, %%mm4 \n\t" yading@10: "psubusb %%mm1, %%mm5 \n\t" yading@10: "movd %%mm2, %0 \n\t" yading@10: "movd %%mm3, %1 \n\t" yading@10: "movd %%mm4, %2 \n\t" yading@10: "movd %%mm5, %3 \n\t" yading@10: :"+m"(*(uint32_t*)(dest+0*linesize)), yading@10: "+m"(*(uint32_t*)(dest+1*linesize)), yading@10: "+m"(*(uint32_t*)(dest+2*linesize)), yading@10: "+m"(*(uint32_t*)(dest+3*linesize)) yading@10: ); yading@10: } yading@10: yading@10: static void vc1_inv_trans_4x8_dc_mmxext(uint8_t *dest, int linesize, yading@10: int16_t *block) yading@10: { yading@10: int dc = block[0]; yading@10: dc = (17 * dc + 4) >> 3; yading@10: dc = (12 * dc + 64) >> 7; yading@10: __asm__ volatile( yading@10: "movd %0, %%mm0 \n\t" yading@10: "pshufw $0, %%mm0, %%mm0 \n\t" yading@10: "pxor %%mm1, %%mm1 \n\t" yading@10: "psubw %%mm0, %%mm1 \n\t" yading@10: "packuswb %%mm0, %%mm0 \n\t" yading@10: "packuswb %%mm1, %%mm1 \n\t" yading@10: ::"r"(dc) yading@10: ); yading@10: __asm__ volatile( yading@10: "movd %0, %%mm2 \n\t" yading@10: "movd %1, %%mm3 \n\t" yading@10: "movd %2, %%mm4 \n\t" yading@10: "movd %3, %%mm5 \n\t" yading@10: "paddusb %%mm0, %%mm2 \n\t" yading@10: "paddusb %%mm0, %%mm3 \n\t" yading@10: "paddusb %%mm0, %%mm4 \n\t" yading@10: "paddusb %%mm0, %%mm5 \n\t" yading@10: "psubusb %%mm1, %%mm2 \n\t" yading@10: "psubusb %%mm1, %%mm3 \n\t" yading@10: "psubusb %%mm1, %%mm4 \n\t" yading@10: "psubusb %%mm1, %%mm5 \n\t" yading@10: "movd %%mm2, %0 \n\t" yading@10: "movd %%mm3, %1 \n\t" yading@10: "movd %%mm4, %2 \n\t" yading@10: "movd %%mm5, %3 \n\t" yading@10: :"+m"(*(uint32_t*)(dest+0*linesize)), yading@10: "+m"(*(uint32_t*)(dest+1*linesize)), yading@10: "+m"(*(uint32_t*)(dest+2*linesize)), yading@10: "+m"(*(uint32_t*)(dest+3*linesize)) yading@10: ); yading@10: dest += 4*linesize; yading@10: __asm__ volatile( yading@10: "movd %0, %%mm2 \n\t" yading@10: "movd %1, %%mm3 \n\t" yading@10: "movd %2, %%mm4 \n\t" yading@10: "movd %3, %%mm5 \n\t" yading@10: "paddusb %%mm0, %%mm2 \n\t" yading@10: "paddusb %%mm0, %%mm3 \n\t" yading@10: "paddusb %%mm0, %%mm4 \n\t" yading@10: "paddusb %%mm0, %%mm5 \n\t" yading@10: "psubusb %%mm1, %%mm2 \n\t" yading@10: "psubusb %%mm1, %%mm3 \n\t" yading@10: "psubusb %%mm1, %%mm4 \n\t" yading@10: "psubusb %%mm1, %%mm5 \n\t" yading@10: "movd %%mm2, %0 \n\t" yading@10: "movd %%mm3, %1 \n\t" yading@10: "movd %%mm4, %2 \n\t" yading@10: "movd %%mm5, %3 \n\t" yading@10: :"+m"(*(uint32_t*)(dest+0*linesize)), yading@10: "+m"(*(uint32_t*)(dest+1*linesize)), yading@10: "+m"(*(uint32_t*)(dest+2*linesize)), yading@10: "+m"(*(uint32_t*)(dest+3*linesize)) yading@10: ); yading@10: } yading@10: yading@10: static void vc1_inv_trans_8x4_dc_mmxext(uint8_t *dest, int linesize, yading@10: int16_t *block) yading@10: { yading@10: int dc = block[0]; yading@10: dc = ( 3 * dc + 1) >> 1; yading@10: dc = (17 * dc + 64) >> 7; yading@10: __asm__ volatile( yading@10: "movd %0, %%mm0 \n\t" yading@10: "pshufw $0, %%mm0, %%mm0 \n\t" yading@10: "pxor %%mm1, %%mm1 \n\t" yading@10: "psubw %%mm0, %%mm1 \n\t" yading@10: "packuswb %%mm0, %%mm0 \n\t" yading@10: "packuswb %%mm1, %%mm1 \n\t" yading@10: ::"r"(dc) yading@10: ); yading@10: __asm__ volatile( yading@10: "movq %0, %%mm2 \n\t" yading@10: "movq %1, %%mm3 \n\t" yading@10: "movq %2, %%mm4 \n\t" yading@10: "movq %3, %%mm5 \n\t" yading@10: "paddusb %%mm0, %%mm2 \n\t" yading@10: "paddusb %%mm0, %%mm3 \n\t" yading@10: "paddusb %%mm0, %%mm4 \n\t" yading@10: "paddusb %%mm0, %%mm5 \n\t" yading@10: "psubusb %%mm1, %%mm2 \n\t" yading@10: "psubusb %%mm1, %%mm3 \n\t" yading@10: "psubusb %%mm1, %%mm4 \n\t" yading@10: "psubusb %%mm1, %%mm5 \n\t" yading@10: "movq %%mm2, %0 \n\t" yading@10: "movq %%mm3, %1 \n\t" yading@10: "movq %%mm4, %2 \n\t" yading@10: "movq %%mm5, %3 \n\t" yading@10: :"+m"(*(uint32_t*)(dest+0*linesize)), yading@10: "+m"(*(uint32_t*)(dest+1*linesize)), yading@10: "+m"(*(uint32_t*)(dest+2*linesize)), yading@10: "+m"(*(uint32_t*)(dest+3*linesize)) yading@10: ); yading@10: } yading@10: yading@10: static void vc1_inv_trans_8x8_dc_mmxext(uint8_t *dest, int linesize, yading@10: int16_t *block) yading@10: { yading@10: int dc = block[0]; yading@10: dc = (3 * dc + 1) >> 1; yading@10: dc = (3 * dc + 16) >> 5; yading@10: __asm__ volatile( yading@10: "movd %0, %%mm0 \n\t" yading@10: "pshufw $0, %%mm0, %%mm0 \n\t" yading@10: "pxor %%mm1, %%mm1 \n\t" yading@10: "psubw %%mm0, %%mm1 \n\t" yading@10: "packuswb %%mm0, %%mm0 \n\t" yading@10: "packuswb %%mm1, %%mm1 \n\t" yading@10: ::"r"(dc) yading@10: ); yading@10: __asm__ volatile( yading@10: "movq %0, %%mm2 \n\t" yading@10: "movq %1, %%mm3 \n\t" yading@10: "movq %2, %%mm4 \n\t" yading@10: "movq %3, %%mm5 \n\t" yading@10: "paddusb %%mm0, %%mm2 \n\t" yading@10: "paddusb %%mm0, %%mm3 \n\t" yading@10: "paddusb %%mm0, %%mm4 \n\t" yading@10: "paddusb %%mm0, %%mm5 \n\t" yading@10: "psubusb %%mm1, %%mm2 \n\t" yading@10: "psubusb %%mm1, %%mm3 \n\t" yading@10: "psubusb %%mm1, %%mm4 \n\t" yading@10: "psubusb %%mm1, %%mm5 \n\t" yading@10: "movq %%mm2, %0 \n\t" yading@10: "movq %%mm3, %1 \n\t" yading@10: "movq %%mm4, %2 \n\t" yading@10: "movq %%mm5, %3 \n\t" yading@10: :"+m"(*(uint32_t*)(dest+0*linesize)), yading@10: "+m"(*(uint32_t*)(dest+1*linesize)), yading@10: "+m"(*(uint32_t*)(dest+2*linesize)), yading@10: "+m"(*(uint32_t*)(dest+3*linesize)) yading@10: ); yading@10: dest += 4*linesize; yading@10: __asm__ volatile( yading@10: "movq %0, %%mm2 \n\t" yading@10: "movq %1, %%mm3 \n\t" yading@10: "movq %2, %%mm4 \n\t" yading@10: "movq %3, %%mm5 \n\t" yading@10: "paddusb %%mm0, %%mm2 \n\t" yading@10: "paddusb %%mm0, %%mm3 \n\t" yading@10: "paddusb %%mm0, %%mm4 \n\t" yading@10: "paddusb %%mm0, %%mm5 \n\t" yading@10: "psubusb %%mm1, %%mm2 \n\t" yading@10: "psubusb %%mm1, %%mm3 \n\t" yading@10: "psubusb %%mm1, %%mm4 \n\t" yading@10: "psubusb %%mm1, %%mm5 \n\t" yading@10: "movq %%mm2, %0 \n\t" yading@10: "movq %%mm3, %1 \n\t" yading@10: "movq %%mm4, %2 \n\t" yading@10: "movq %%mm5, %3 \n\t" yading@10: :"+m"(*(uint32_t*)(dest+0*linesize)), yading@10: "+m"(*(uint32_t*)(dest+1*linesize)), yading@10: "+m"(*(uint32_t*)(dest+2*linesize)), yading@10: "+m"(*(uint32_t*)(dest+3*linesize)) yading@10: ); yading@10: } yading@10: yading@10: av_cold void ff_vc1dsp_init_mmx(VC1DSPContext *dsp) yading@10: { yading@10: dsp->put_vc1_mspel_pixels_tab[ 0] = ff_put_vc1_mspel_mc00_mmx; yading@10: dsp->put_vc1_mspel_pixels_tab[ 4] = put_vc1_mspel_mc01_mmx; yading@10: dsp->put_vc1_mspel_pixels_tab[ 8] = put_vc1_mspel_mc02_mmx; yading@10: dsp->put_vc1_mspel_pixels_tab[12] = put_vc1_mspel_mc03_mmx; yading@10: yading@10: dsp->put_vc1_mspel_pixels_tab[ 1] = put_vc1_mspel_mc10_mmx; yading@10: dsp->put_vc1_mspel_pixels_tab[ 5] = put_vc1_mspel_mc11_mmx; yading@10: dsp->put_vc1_mspel_pixels_tab[ 9] = put_vc1_mspel_mc12_mmx; yading@10: dsp->put_vc1_mspel_pixels_tab[13] = put_vc1_mspel_mc13_mmx; yading@10: yading@10: dsp->put_vc1_mspel_pixels_tab[ 2] = put_vc1_mspel_mc20_mmx; yading@10: dsp->put_vc1_mspel_pixels_tab[ 6] = put_vc1_mspel_mc21_mmx; yading@10: dsp->put_vc1_mspel_pixels_tab[10] = put_vc1_mspel_mc22_mmx; yading@10: dsp->put_vc1_mspel_pixels_tab[14] = put_vc1_mspel_mc23_mmx; yading@10: yading@10: dsp->put_vc1_mspel_pixels_tab[ 3] = put_vc1_mspel_mc30_mmx; yading@10: dsp->put_vc1_mspel_pixels_tab[ 7] = put_vc1_mspel_mc31_mmx; yading@10: dsp->put_vc1_mspel_pixels_tab[11] = put_vc1_mspel_mc32_mmx; yading@10: dsp->put_vc1_mspel_pixels_tab[15] = put_vc1_mspel_mc33_mmx; yading@10: } yading@10: yading@10: av_cold void ff_vc1dsp_init_mmxext(VC1DSPContext *dsp) yading@10: { yading@10: dsp->avg_vc1_mspel_pixels_tab[ 4] = avg_vc1_mspel_mc01_mmxext; yading@10: dsp->avg_vc1_mspel_pixels_tab[ 8] = avg_vc1_mspel_mc02_mmxext; yading@10: dsp->avg_vc1_mspel_pixels_tab[12] = avg_vc1_mspel_mc03_mmxext; yading@10: yading@10: dsp->avg_vc1_mspel_pixels_tab[ 1] = avg_vc1_mspel_mc10_mmxext; yading@10: dsp->avg_vc1_mspel_pixels_tab[ 5] = avg_vc1_mspel_mc11_mmxext; yading@10: dsp->avg_vc1_mspel_pixels_tab[ 9] = avg_vc1_mspel_mc12_mmxext; yading@10: dsp->avg_vc1_mspel_pixels_tab[13] = avg_vc1_mspel_mc13_mmxext; yading@10: yading@10: dsp->avg_vc1_mspel_pixels_tab[ 2] = avg_vc1_mspel_mc20_mmxext; yading@10: dsp->avg_vc1_mspel_pixels_tab[ 6] = avg_vc1_mspel_mc21_mmxext; yading@10: dsp->avg_vc1_mspel_pixels_tab[10] = avg_vc1_mspel_mc22_mmxext; yading@10: dsp->avg_vc1_mspel_pixels_tab[14] = avg_vc1_mspel_mc23_mmxext; yading@10: yading@10: dsp->avg_vc1_mspel_pixels_tab[ 3] = avg_vc1_mspel_mc30_mmxext; yading@10: dsp->avg_vc1_mspel_pixels_tab[ 7] = avg_vc1_mspel_mc31_mmxext; yading@10: dsp->avg_vc1_mspel_pixels_tab[11] = avg_vc1_mspel_mc32_mmxext; yading@10: dsp->avg_vc1_mspel_pixels_tab[15] = avg_vc1_mspel_mc33_mmxext; yading@10: yading@10: dsp->vc1_inv_trans_8x8_dc = vc1_inv_trans_8x8_dc_mmxext; yading@10: dsp->vc1_inv_trans_4x8_dc = vc1_inv_trans_4x8_dc_mmxext; yading@10: dsp->vc1_inv_trans_8x4_dc = vc1_inv_trans_8x4_dc_mmxext; yading@10: dsp->vc1_inv_trans_4x4_dc = vc1_inv_trans_4x4_dc_mmxext; yading@10: } yading@10: #endif /* HAVE_INLINE_ASM */