Partially based on patches by clsid2 in ffdshow-tryout. ff_float_interleave6() x86 improvements by Loren Merrit.
Unified · Split
| @@ -56,11 +56,31 @@ static void float_to_int16_interleave_c(int16_t *dst, const float **src, · 55 unmodified lines | |||
| 56 | } | 56 | } |
| 57 | } | 57 | } |
| 58 | 58 | ||
| 59 | void ff_float_interleave_c(float *dst, const float **src, unsigned int len, | ||
| 60 | int channels) | ||
| 61 | { | ||
| 62 | int j, c; | ||
| 63 | unsigned int i; | ||
| 64 | if (channels == 2) { | ||
| 65 | for (i = 0; i < len; i++) { | ||
| 66 | dst[2*i] = src[0][i]; | ||
| 67 | dst[2*i+1] = src[1][i]; | ||
| 68 | } | ||
| 69 | } else if (channels == 1 && len < INT_MAX / sizeof(float)) { | ||
| 70 | memcpy(dst, src[0], len * sizeof(float)); | ||
| 71 | } else { | ||
| 72 | for (c = 0; c < channels; c++) | ||
| 73 | for (i = 0, j = c; i < len; i++, j += channels) | ||
| 74 | dst[j] = src[c][i]; | ||
| 75 | } | ||
| 76 | } | ||
| 77 | |||
| 59 | av_cold void ff_fmt_convert_init(FmtConvertContext *c, AVCodecContext *avctx) | 78 | av_cold void ff_fmt_convert_init(FmtConvertContext *c, AVCodecContext *avctx) |
| 60 | { | 79 | { |
| 61 | c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_c; | 80 | c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_c; |
| 62 | c->float_to_int16 = float_to_int16_c; | 81 | c->float_to_int16 = float_to_int16_c; |
| 63 | c->float_to_int16_interleave = float_to_int16_interleave_c; | 82 | c->float_to_int16_interleave = float_to_int16_interleave_c; |
| 83 | c->float_interleave = ff_float_interleave_c; | ||
| 64 | 84 | ||
| 65 | if (ARCH_ARM) ff_fmt_convert_init_arm(c, avctx); | 85 | if (ARCH_ARM) ff_fmt_convert_init_arm(c, avctx); |
| 66 | if (HAVE_ALTIVEC) ff_fmt_convert_init_altivec(c, avctx); | 86 | if (HAVE_ALTIVEC) ff_fmt_convert_init_altivec(c, avctx); |
| @@ -68,8 +68,17 @@ typedef struct FmtConvertContext { · 67 unmodified lines | |||
| 68 | */ | 68 | */ |
| 69 | void (*float_to_int16_interleave)(int16_t *dst, const float **src, | 69 | void (*float_to_int16_interleave)(int16_t *dst, const float **src, |
| 70 | long len, int channels); | 70 | long len, int channels); |
| 71 | |||
| 72 | /** | ||
| 73 | * Convert an array of interleaved float to multiple arrays of float. | ||
| 74 | */ | ||
| 75 | void (*float_interleave)(float *dst, const float **src, unsigned int len, | ||
| 76 | int channels); | ||
| 71 | } FmtConvertContext; | 77 | } FmtConvertContext; |
| 72 | 78 | ||
| 79 | void ff_float_interleave_c(float *dst, const float **src, unsigned int len, | ||
| 80 | int channels); | ||
| 81 | |||
| 73 | void ff_fmt_convert_init(FmtConvertContext *c, AVCodecContext *avctx); | 82 | void ff_fmt_convert_init(FmtConvertContext *c, AVCodecContext *avctx); |
| 74 | 83 | ||
| 75 | void ff_fmt_convert_init_arm(FmtConvertContext *c, AVCodecContext *avctx); | 84 | void ff_fmt_convert_init_arm(FmtConvertContext *c, AVCodecContext *avctx); |
| @@ -20,6 +20,7 @@ · 19 unmodified lines | |||
| 20 | ;****************************************************************************** | 20 | ;****************************************************************************** |
| 21 | 21 | ||
| 22 | %include "x86inc.asm" | 22 | %include "x86inc.asm" |
| 23 | %include "x86util.asm" | ||
| 23 | 24 | ||
| 24 | section .text align=16 | 25 | section .text align=16 |
| 25 | 26 | ||
| @@ -89,3 +90,143 @@ FLOAT_TO_INT16_INTERLEAVE6 3dnow · 63 unmodified lines | |||
| 89 | %undef pswapd | 90 | %undef pswapd |
| 90 | FLOAT_TO_INT16_INTERLEAVE6 3dn2 | 91 | FLOAT_TO_INT16_INTERLEAVE6 3dn2 |
| 91 | %undef cvtps2pi | 92 | %undef cvtps2pi |
| 93 | |||
| 94 | ;----------------------------------------------------------------------------- | ||
| 95 | ; void ff_float_interleave6(float *dst, const float **src, unsigned int len); | ||
| 96 | ;----------------------------------------------------------------------------- | ||
| 97 | |||
| 98 | %macro BUTTERFLYPS 3 | ||
| 99 | movaps m%3, m%1 | ||
| 100 | unpcklps m%1, m%2 | ||
| 101 | unpckhps m%3, m%2 | ||
| 102 | SWAP %2, %3 | ||
| 103 | %endmacro | ||
| 104 | |||
| 105 | %macro FLOAT_INTERLEAVE6 2 | ||
| 106 | cglobal float_interleave6_%1, 2,7,%2, dst, src, src1, src2, src3, src4, src5 | ||
| 107 | %ifdef ARCH_X86_64 | ||
| 108 | %define lend r10d | ||
| 109 | mov lend, r2d | ||
| 110 | %else | ||
| 111 | %define lend dword r2m | ||
| 112 | %endif | ||
| 113 | mov src1q, [srcq+1*gprsize] | ||
| 114 | mov src2q, [srcq+2*gprsize] | ||
| 115 | mov src3q, [srcq+3*gprsize] | ||
| 116 | mov src4q, [srcq+4*gprsize] | ||
| 117 | mov src5q, [srcq+5*gprsize] | ||
| 118 | mov srcq, [srcq] | ||
| 119 | sub src1q, srcq | ||
| 120 | sub src2q, srcq | ||
| 121 | sub src3q, srcq | ||
| 122 | sub src4q, srcq | ||
| 123 | sub src5q, srcq | ||
| 124 | .loop: | ||
| 125 | %ifidn %1, sse | ||
| 126 | movaps m0, [srcq] | ||
| 127 | movaps m1, [srcq+src1q] | ||
| 128 | movaps m2, [srcq+src2q] | ||
| 129 | movaps m3, [srcq+src3q] | ||
| 130 | movaps m4, [srcq+src4q] | ||
| 131 | movaps m5, [srcq+src5q] | ||
| 132 | |||
| 133 | BUTTERFLYPS 0, 1, 6 | ||
| 134 | BUTTERFLYPS 2, 3, 6 | ||
| 135 | BUTTERFLYPS 4, 5, 6 | ||
| 136 | |||
| 137 | movaps m6, m4 | ||
| 138 | shufps m4, m0, 0xe4 | ||
| 139 | movlhps m0, m2 | ||
| 140 | movhlps m6, m2 | ||
| 141 | movaps [dstq ], m0 | ||
| 142 | movaps [dstq+16], m4 | ||
| 143 | movaps [dstq+32], m6 | ||
| 144 | |||
| 145 | movaps m6, m5 | ||
| 146 | shufps m5, m1, 0xe4 | ||
| 147 | movlhps m1, m3 | ||
| 148 | movhlps m6, m3 | ||
| 149 | movaps [dstq+48], m1 | ||
| 150 | movaps [dstq+64], m5 | ||
| 151 | movaps [dstq+80], m6 | ||
| 152 | %else ; mmx | ||
| 153 | movq m0, [srcq] | ||
| 154 | movq m1, [srcq+src1q] | ||
| 155 | movq m2, [srcq+src2q] | ||
| 156 | movq m3, [srcq+src3q] | ||
| 157 | movq m4, [srcq+src4q] | ||
| 158 | movq m5, [srcq+src5q] | ||
| 159 | |||
| 160 | SBUTTERFLY dq, 0, 1, 6 | ||
| 161 | SBUTTERFLY dq, 2, 3, 6 | ||
| 162 | SBUTTERFLY dq, 4, 5, 6 | ||
| 163 | movq [dstq ], m0 | ||
| 164 | movq [dstq+ 8], m2 | ||
| 165 | movq [dstq+16], m4 | ||
| 166 | movq [dstq+24], m1 | ||
| 167 | movq [dstq+32], m3 | ||
| 168 | movq [dstq+40], m5 | ||
| 169 | %endif | ||
| 170 | add srcq, mmsize | ||
| 171 | add dstq, mmsize*6 | ||
| 172 | sub lend, mmsize/4 | ||
| 173 | jg .loop | ||
| 174 | %ifidn %1, mmx | ||
| 175 | emms | ||
| 176 | %endif | ||
| 177 | REP_RET | ||
| 178 | %endmacro | ||
| 179 | |||
| 180 | INIT_MMX | ||
| 181 | FLOAT_INTERLEAVE6 mmx, 0 | ||
| 182 | INIT_XMM | ||
| 183 | FLOAT_INTERLEAVE6 sse, 7 | ||
| 184 | |||
| 185 | ;----------------------------------------------------------------------------- | ||
| 186 | ; void ff_float_interleave2(float *dst, const float **src, unsigned int len); | ||
| 187 | ;----------------------------------------------------------------------------- | ||
| 188 | |||
| 189 | %macro FLOAT_INTERLEAVE2 2 | ||
| 190 | cglobal float_interleave2_%1, 3,4,%2, dst, src, len, src1 | ||
| 191 | mov src1q, [srcq+gprsize] | ||
| 192 | mov srcq, [srcq ] | ||
| 193 | sub src1q, srcq | ||
| 194 | .loop | ||
| 195 | MOVPS m0, [srcq ] | ||
| 196 | MOVPS m1, [srcq+src1q ] | ||
| 197 | MOVPS m3, [srcq +mmsize] | ||
| 198 | MOVPS m4, [srcq+src1q+mmsize] | ||
| 199 | |||
| 200 | MOVPS m2, m0 | ||
| 201 | PUNPCKLDQ m0, m1 | ||
| 202 | PUNPCKHDQ m2, m1 | ||
| 203 | |||
| 204 | MOVPS m1, m3 | ||
| 205 | PUNPCKLDQ m3, m4 | ||
| 206 | PUNPCKHDQ m1, m4 | ||
| 207 | |||
| 208 | MOVPS [dstq ], m0 | ||
| 209 | MOVPS [dstq+1*mmsize], m2 | ||
| 210 | MOVPS [dstq+2*mmsize], m3 | ||
| 211 | MOVPS [dstq+3*mmsize], m1 | ||
| 212 | |||
| 213 | add srcq, mmsize*2 | ||
| 214 | add dstq, mmsize*4 | ||
| 215 | sub lend, mmsize/2 | ||
| 216 | jg .loop | ||
| 217 | %ifidn %1, mmx | ||
| 218 | emms | ||
| 219 | %endif | ||
| 220 | REP_RET | ||
| 221 | %endmacro | ||
| 222 | |||
| 223 | INIT_MMX | ||
| 224 | %define MOVPS movq | ||
| 225 | %define PUNPCKLDQ punpckldq | ||
| 226 | %define PUNPCKHDQ punpckhdq | ||
| 227 | FLOAT_INTERLEAVE2 mmx, 0 | ||
| 228 | INIT_XMM | ||
| 229 | %define MOVPS movaps | ||
| 230 | %define PUNPCKLDQ unpcklps | ||
| 231 | %define PUNPCKHDQ unpckhps | ||
| 232 | FLOAT_INTERLEAVE2 sse, 5 | ||
| @@ -235,11 +235,40 @@ static void float_to_int16_interleave_3dn2(int16_t *dst, const float **src, long · 234 unmodified lines | |||
| 235 | float_to_int16_interleave_3dnow(dst, src, len, channels); | 235 | float_to_int16_interleave_3dnow(dst, src, len, channels); |
| 236 | } | 236 | } |
| 237 | 237 | ||
| 238 | void ff_float_interleave2_mmx(float *dst, const float **src, unsigned int len); | ||
| 239 | void ff_float_interleave2_sse(float *dst, const float **src, unsigned int len); | ||
| 240 | |||
| 241 | void ff_float_interleave6_mmx(float *dst, const float **src, unsigned int len); | ||
| 242 | void ff_float_interleave6_sse(float *dst, const float **src, unsigned int len); | ||
| 243 | |||
| 244 | static void float_interleave_mmx(float *dst, const float **src, | ||
| 245 | unsigned int len, int channels) | ||
| 246 | { | ||
| 247 | if (channels == 2) { | ||
| 248 | ff_float_interleave2_mmx(dst, src, len); | ||
| 249 | } else if (channels == 6) | ||
| 250 | ff_float_interleave6_mmx(dst, src, len); | ||
| 251 | else | ||
| 252 | ff_float_interleave_c(dst, src, len, channels); | ||
| 253 | } | ||
| 254 | |||
| 255 | static void float_interleave_sse(float *dst, const float **src, | ||
| 256 | unsigned int len, int channels) | ||
| 257 | { | ||
| 258 | if (channels == 2) { | ||
| 259 | ff_float_interleave2_sse(dst, src, len); | ||
| 260 | } else if (channels == 6) | ||
| 261 | ff_float_interleave6_sse(dst, src, len); | ||
| 262 | else | ||
| 263 | ff_float_interleave_c(dst, src, len, channels); | ||
| 264 | } | ||
| 265 | |||
| 238 | void ff_fmt_convert_init_x86(FmtConvertContext *c, AVCodecContext *avctx) | 266 | void ff_fmt_convert_init_x86(FmtConvertContext *c, AVCodecContext *avctx) |
| 239 | { | 267 | { |
| 240 | int mm_flags = av_get_cpu_flags(); | 268 | int mm_flags = av_get_cpu_flags(); |
| 241 | 269 | ||
| 242 | if (mm_flags & AV_CPU_FLAG_MMX) { | 270 | if (mm_flags & AV_CPU_FLAG_MMX) { |
| 271 | c->float_interleave = float_interleave_mmx; | ||
| 243 | 272 | ||
| 244 | if(mm_flags & AV_CPU_FLAG_3DNOW){ | 273 | if(mm_flags & AV_CPU_FLAG_3DNOW){ |
| 245 | if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ | 274 | if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ |
| @@ -256,6 +285,7 @@ void ff_fmt_convert_init_x86(FmtConvertContext *c, AVCodecContext *avctx) · 10 unmodified lines | |||
| 256 | c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_sse; | 285 | c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_sse; |
| 257 | c->float_to_int16 = float_to_int16_sse; | 286 | c->float_to_int16 = float_to_int16_sse; |
| 258 | c->float_to_int16_interleave = float_to_int16_interleave_sse; | 287 | c->float_to_int16_interleave = float_to_int16_interleave_sse; |
| 288 | c->float_interleave = float_interleave_sse; | ||
| 259 | } | 289 | } |
| 260 | if(mm_flags & AV_CPU_FLAG_SSE2){ | 290 | if(mm_flags & AV_CPU_FLAG_SSE2){ |
| 261 | c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_sse2; | 291 | c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_sse2; |