Merging these functions allows merging some loops, which makes the results (particularly after SIMD optimizations) much faster. (cherry picked from commit f8bed30d8b176fa030f6737765338bb4a2bcabc9)
Showing the changes to libavcodec/ppc (1 of 5 files). Show all
Unified · Split
| @@ -130,7 +130,8 @@ do { \ · 129 unmodified lines | ||
| 130 | 130 | |
| 131 | 131 | /** Do inverse transform on 8x8 block |
| 132 | 132 | */ |
| 133 | −static void vc1_inv_trans_8x8_altivec(DCTELEM block[64]) | |
| 133 | +static void vc1_inv_trans_8x8_altivec(DCTELEM block[64], | |
| 134 | + int sign, int rangered) | |
| 134 | 135 | { |
| 135 | 136 | vector signed short src0, src1, src2, src3, src4, src5, src6, src7; |
| 136 | 137 | vector signed int s0, s1, s2, s3, s4, s5, s6, s7; |
| @@ -144,7 +145,9 @@ static void vc1_inv_trans_8x8_altivec(DCTELEM block[64]) · 7 unmodified lines | ||
| 144 | 145 | const vector unsigned int vec_2 = vec_splat_u32(2); |
| 145 | 146 | const vector signed int vec_1s = vec_splat_s32(1); |
| 146 | 147 | const vector unsigned int vec_1 = vec_splat_u32(1); |
| 147 | − | |
| 148 | + const vector unsigned short rangered_shift = vec_splat_u16(1); | |
| 149 | + const vector signed short signed_bias = vec_sl(vec_splat_u16(4), | |
| 150 | + vec_splat_u16(4)); | |
| 148 | 151 | |
| 149 | 152 | src0 = vec_ld( 0, block); |
| 150 | 153 | src1 = vec_ld( 16, block); |
| @@ -214,6 +217,27 @@ static void vc1_inv_trans_8x8_altivec(DCTELEM block[64]) · 63 unmodified lines | ||
| 214 | 217 | src6 = vec_pack(sE, s6); |
| 215 | 218 | src7 = vec_pack(sF, s7); |
| 216 | 219 | |
| 220 | + if (rangered) { | |
| 221 | + if (!sign) { | |
| 222 | + vec_sub(src0, signed_bias); | |
| 223 | + vec_sub(src1, signed_bias); | |
| 224 | + vec_sub(src2, signed_bias); | |
| 225 | + vec_sub(src3, signed_bias); | |
| 226 | + vec_sub(src4, signed_bias); | |
| 227 | + vec_sub(src5, signed_bias); | |
| 228 | + vec_sub(src6, signed_bias); | |
| 229 | + vec_sub(src7, signed_bias); | |
| 230 | + } | |
| 231 | + vec_sl(src0, rangered_shift); | |
| 232 | + vec_sl(src1, rangered_shift); | |
| 233 | + vec_sl(src2, rangered_shift); | |
| 234 | + vec_sl(src3, rangered_shift); | |
| 235 | + vec_sl(src4, rangered_shift); | |
| 236 | + vec_sl(src5, rangered_shift); | |
| 237 | + vec_sl(src6, rangered_shift); | |
| 238 | + vec_sl(src7, rangered_shift); | |
| 239 | + } | |
| 240 | + | |
| 217 | 241 | vec_st(src0, 0, block); |
| 218 | 242 | vec_st(src1, 16, block); |
| 219 | 243 | vec_st(src2, 32, block); |
| @@ -224,6 +248,36 @@ static void vc1_inv_trans_8x8_altivec(DCTELEM block[64]) · 4 unmodified lines | ||
| 224 | 248 | vec_st(src7,112, block); |
| 225 | 249 | } |
| 226 | 250 | |
| 251 | +static void vc1_inv_trans_8x8_add_altivec(uint8_t *dest, int stride, DCTELEM *b) | |
| 252 | +{ | |
| 253 | + vc1_inv_trans_8x8_altivec(b, 0, 0); | |
| 254 | + ff_add_pixels_clamped_c(b, dest, stride); | |
| 255 | +} | |
| 256 | + | |
| 257 | +static void vc1_inv_trans_8x8_put_signed_altivec(uint8_t *dest, int stride, DCTELEM *b) | |
| 258 | +{ | |
| 259 | + vc1_inv_trans_8x8_altivec(b, 1, 0); | |
| 260 | + ff_put_signed_pixels_clamped_c(b, dest, stride); | |
| 261 | +} | |
| 262 | + | |
| 263 | +static void vc1_inv_trans_8x8_put_signed_rangered_altivec(uint8_t *dest, int stride, DCTELEM *b) | |
| 264 | +{ | |
| 265 | + vc1_inv_trans_8x8_altivec(b, 1, 1); | |
| 266 | + ff_put_signed_pixels_clamped_c(b, dest, stride); | |
| 267 | +} | |
| 268 | + | |
| 269 | +static void vc1_inv_trans_8x8_put_altivec(uint8_t *dest, int stride, DCTELEM *b) | |
| 270 | +{ | |
| 271 | + vc1_inv_trans_8x8_altivec(b, 0, 0); | |
| 272 | + ff_put_pixels_clamped_c(b, dest, stride); | |
| 273 | +} | |
| 274 | + | |
| 275 | +static void vc1_inv_trans_8x8_put_rangered_altivec(uint8_t *dest, int stride, DCTELEM *b) | |
| 276 | +{ | |
| 277 | + vc1_inv_trans_8x8_altivec(b, 0, 1); | |
| 278 | + ff_put_pixels_clamped_c(b, dest, stride); | |
| 279 | +} | |
| 280 | + | |
| 227 | 281 | /** Do inverse transform on 8x4 part of block |
| 228 | 282 | */ |
| 229 | 283 | static void vc1_inv_trans_8x4_altivec(uint8_t *dest, int stride, DCTELEM *block) |
| @@ -342,7 +396,11 @@ void ff_vc1dsp_init_altivec(VC1DSPContext* dsp) · 112 unmodified lines | ||
| 342 | 396 | if (!(av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC)) |
| 343 | 397 | return; |
| 344 | 398 | |
| 345 | − dsp->vc1_inv_trans_8x8 = vc1_inv_trans_8x8_altivec; | |
| 399 | + dsp->vc1_inv_trans_8x8_add = vc1_inv_trans_8x8_add_altivec; | |
| 400 | + dsp->vc1_inv_trans_8x8_put_signed[0] = vc1_inv_trans_8x8_put_signed_altivec; | |
| 401 | + dsp->vc1_inv_trans_8x8_put_signed[1] = vc1_inv_trans_8x8_put_signed_rangered_altivec; | |
| 402 | + dsp->vc1_inv_trans_8x8_put[0] = vc1_inv_trans_8x8_put_altivec; | |
| 403 | + dsp->vc1_inv_trans_8x8_put[1] = vc1_inv_trans_8x8_put_rangered_altivec; | |
| 346 | 404 | dsp->vc1_inv_trans_8x4 = vc1_inv_trans_8x4_altivec; |
| 347 | 405 | dsp->put_no_rnd_vc1_chroma_pixels_tab[0] = put_no_rnd_vc1_chroma_mc8_altivec; |
| 348 | 406 | dsp->avg_no_rnd_vc1_chroma_pixels_tab[0] = avg_no_rnd_vc1_chroma_mc8_altivec; |