2f77ecc6bc
avcodec/riscv: add h264 qpel
...
Benched on K230 for VLEN 128, SpaceMIT for VLEN 256. Variants for 4
width have no speedup for VLEN 256 vs VLEN 128 on available hardware,
so were disabled.
C RVV128 C RVV256
avg_h264_qpel_4_mc00_8 33.9 33.6 (1.01x)
avg_h264_qpel_4_mc01_8 218.8 89.1 (2.46x)
avg_h264_qpel_4_mc02_8 218.8 79.8 (2.74x)
avg_h264_qpel_4_mc03_8 218.8 89.1 (2.46x)
avg_h264_qpel_4_mc10_8 172.3 126.1 (1.37x)
avg_h264_qpel_4_mc11_8 339.1 190.8 (1.78x)
avg_h264_qpel_4_mc12_8 533.6 357.6 (1.49x)
avg_h264_qpel_4_mc13_8 348.4 190.8 (1.83x)
avg_h264_qpel_4_mc20_8 144.8 116.8 (1.24x)
avg_h264_qpel_4_mc21_8 478.1 385.6 (1.24x)
avg_h264_qpel_4_mc22_8 348.4 283.6 (1.23x)
avg_h264_qpel_4_mc23_8 478.1 394.6 (1.21x)
avg_h264_qpel_4_mc30_8 172.6 126.1 (1.37x)
avg_h264_qpel_4_mc31_8 339.4 191.1 (1.78x)
avg_h264_qpel_4_mc32_8 542.9 357.6 (1.52x)
avg_h264_qpel_4_mc33_8 339.4 191.1 (1.78x)
avg_h264_qpel_8_mc00_8 116.8 42.9 (2.72x) 123.6 50.6 (2.44x)
avg_h264_qpel_8_mc01_8 774.4 163.1 (4.75x) 779.8 165.1 (4.72x)
avg_h264_qpel_8_mc02_8 774.4 154.1 (5.03x) 779.8 144.3 (5.40x)
avg_h264_qpel_8_mc03_8 774.4 163.3 (4.74x) 779.8 165.3 (4.72x)
avg_h264_qpel_8_mc10_8 617.1 237.3 (2.60x) 613.1 227.6 (2.69x)
avg_h264_qpel_8_mc11_8 1209.3 376.4 (3.21x) 1206.8 363.1 (3.32x)
avg_h264_qpel_8_mc12_8 1913.3 598.6 (3.20x) 1894.3 561.1 (3.38x)
avg_h264_qpel_8_mc13_8 1218.6 376.4 (3.24x) 1217.1 363.1 (3.35x)
avg_h264_qpel_8_mc20_8 524.4 228.1 (2.30x) 519.3 227.6 (2.28x)
avg_h264_qpel_8_mc21_8 1709.6 681.9 (2.51x) 1707.1 644.3 (2.65x)
avg_h264_qpel_8_mc22_8 1274.3 459.6 (2.77x) 1279.8 436.1 (2.93x)
avg_h264_qpel_8_mc23_8 1700.3 672.6 (2.53x) 1706.8 644.6 (2.65x)
avg_h264_qpel_8_mc30_8 607.6 246.6 (2.46x) 623.6 238.1 (2.62x)
avg_h264_qpel_8_mc31_8 1209.6 376.4 (3.21x) 1206.8 363.1 (3.32x)
avg_h264_qpel_8_mc32_8 1904.1 607.9 (3.13x) 1894.3 571.3 (3.32x)
avg_h264_qpel_8_mc33_8 1209.6 376.1 (3.22x) 1206.8 363.1 (3.32x)
avg_h264_qpel_16_mc00_8 431.9 89.1 (4.85x) 436.1 71.3 (6.12x)
avg_h264_qpel_16_mc01_8 2894.6 376.1 (7.70x) 2842.3 300.6 (9.46x)
avg_h264_qpel_16_mc02_8 2987.3 348.4 (8.57x) 2967.3 290.1 (10.23x)
avg_h264_qpel_16_mc03_8 2885.3 376.4 (7.67x) 2842.3 300.6 (9.46x)
avg_h264_qpel_16_mc10_8 2404.1 524.4 (4.58x) 2404.8 456.8 (5.26x)
avg_h264_qpel_16_mc11_8 4709.4 811.6 (5.80x) 4675.6 706.8 (6.62x)
avg_h264_qpel_16_mc12_8 7477.9 1274.3 (5.87x) 7436.1 1061.1 (7.01x)
avg_h264_qpel_16_mc13_8 4718.6 820.6 (5.75x) 4655.1 706.8 (6.59x)
avg_h264_qpel_16_mc20_8 2052.1 487.1 (4.21x) 2071.3 446.3 (4.64x)
avg_h264_qpel_16_mc21_8 7440.6 1422.6 (5.23x) 6727.8 1217.3 (5.53x)
avg_h264_qpel_16_mc22_8 5051.9 950.4 (5.32x) 5071.6 790.3 (6.42x)
avg_h264_qpel_16_mc23_8 6764.9 1422.3 (4.76x) 6748.6 1217.3 (5.54x)
avg_h264_qpel_16_mc30_8 2413.1 524.4 (4.60x) 2415.1 467.3 (5.17x)
avg_h264_qpel_16_mc31_8 4681.6 839.1 (5.58x) 4675.6 727.6 (6.43x)
avg_h264_qpel_16_mc32_8 8579.6 1292.8 (6.64x) 7436.3 1071.3 (6.94x)
avg_h264_qpel_16_mc33_8 5375.9 829.9 (6.48x) 4665.3 717.3 (6.50x)
put_h264_qpel_4_mc00_8 24.4 24.4 (1.00x)
put_h264_qpel_4_mc01_8 987.4 79.8 (12.37x)
put_h264_qpel_4_mc02_8 190.8 79.8 (2.39x)
put_h264_qpel_4_mc03_8 209.6 89.1 (2.35x)
put_h264_qpel_4_mc10_8 163.3 117.1 (1.39x)
put_h264_qpel_4_mc11_8 339.4 181.6 (1.87x)
put_h264_qpel_4_mc12_8 533.6 348.4 (1.53x)
put_h264_qpel_4_mc13_8 339.4 190.8 (1.78x)
put_h264_qpel_4_mc20_8 126.3 116.8 (1.08x)
put_h264_qpel_4_mc21_8 468.9 376.1 (1.25x)
put_h264_qpel_4_mc22_8 330.1 274.4 (1.20x)
put_h264_qpel_4_mc23_8 468.9 376.1 (1.25x)
put_h264_qpel_4_mc30_8 163.3 126.3 (1.29x)
put_h264_qpel_4_mc31_8 339.1 191.1 (1.77x)
put_h264_qpel_4_mc32_8 533.6 348.4 (1.53x)
put_h264_qpel_4_mc33_8 339.4 181.8 (1.87x)
put_h264_qpel_8_mc00_8 98.6 33.6 (2.93x) 92.3 40.1 (2.30x)
put_h264_qpel_8_mc01_8 737.1 153.8 (4.79x) 738.1 144.3 (5.12x)
put_h264_qpel_8_mc02_8 663.1 135.3 (4.90x) 665.1 134.1 (4.96x)
put_h264_qpel_8_mc03_8 737.4 154.1 (4.79x) 1508.8 144.3 (10.46x)
put_h264_qpel_8_mc10_8 598.4 237.1 (2.52x) 592.3 227.6 (2.60x)
put_h264_qpel_8_mc11_8 1172.3 357.9 (3.28x) 1175.6 342.3 (3.43x)
put_h264_qpel_8_mc12_8 1867.1 589.1 (3.17x) 1863.1 561.1 (3.32x)
put_h264_qpel_8_mc13_8 1172.6 366.9 (3.20x) 1175.6 352.8 (3.33x)
put_h264_qpel_8_mc20_8 450.4 218.8 (2.06x) 446.3 206.8 (2.16x)
put_h264_qpel_8_mc21_8 1672.3 663.1 (2.52x) 1675.6 633.8 (2.64x)
put_h264_qpel_8_mc22_8 1144.6 1200.1 (0.95x) 1144.3 425.6 (2.69x)
put_h264_qpel_8_mc23_8 1672.6 672.4 (2.49x) 1665.3 634.1 (2.63x)
put_h264_qpel_8_mc30_8 598.6 237.3 (2.52x) 613.1 227.6 (2.69x)
put_h264_qpel_8_mc31_8 1172.3 376.1 (3.12x) 1175.6 352.6 (3.33x)
put_h264_qpel_8_mc32_8 1857.8 598.6 (3.10x) 1863.1 561.1 (3.32x)
put_h264_qpel_8_mc33_8 1172.3 376.1 (3.12x) 1175.6 352.8 (3.33x)
put_h264_qpel_16_mc00_8 320.6 61.4 (5.22x) 321.3 60.8 (5.28x)
put_h264_qpel_16_mc01_8 2774.3 339.1 (8.18x) 2759.1 279.8 (9.86x)
put_h264_qpel_16_mc02_8 2589.1 320.6 (8.08x) 2571.6 269.3 (9.55x)
put_h264_qpel_16_mc03_8 2774.3 339.4 (8.17x) 2738.1 290.1 (9.44x)
put_h264_qpel_16_mc10_8 2274.3 487.4 (4.67x) 2290.1 436.1 (5.25x)
put_h264_qpel_16_mc11_8 5237.1 792.9 (6.60x) 4529.8 685.8 (6.61x)
put_h264_qpel_16_mc12_8 7357.6 1255.8 (5.86x) 7352.8 1040.1 (7.07x)
put_h264_qpel_16_mc13_8 4579.9 792.9 (5.78x) 4571.6 686.1 (6.66x)
put_h264_qpel_16_mc20_8 1802.1 459.6 (3.92x) 1800.6 425.6 (4.23x)
put_h264_qpel_16_mc21_8 6644.6 2246.6 (2.96x) 6644.3 1196.6 (5.55x)
put_h264_qpel_16_mc22_8 4589.1 913.4 (5.02x) 4592.3 769.3 (5.97x)
put_h264_qpel_16_mc23_8 6644.6 1394.6 (4.76x) 6634.1 1196.6 (5.54x)
put_h264_qpel_16_mc30_8 2274.3 496.6 (4.58x) 2290.1 456.8 (5.01x)
put_h264_qpel_16_mc31_8 5255.6 802.1 (6.55x) 4550.8 706.8 (6.44x)
put_h264_qpel_16_mc32_8 7376.1 1265.1 (5.83x) 7352.8 1050.6 (7.00x)
put_h264_qpel_16_mc33_8 4579.9 802.1 (5.71x) 4561.1 696.3 (6.55x)
Signed-off-by: Niklas Haas <git@haasn.dev>
Signed-off-by: J. Dekker <jdek@itanimul.li>
3 months ago
6611bf5484
lavc/h264dsp: optimise R-V V biweight for shorter heights
...
T-Head C908:
h264_biweight2_8_c: 313.7 ( 1.00x)
h264_biweight2_8_rvv_i32: before 239.5 ( 1.23x)
h264_biweight2_8_rvv_i32: after 72.7 ( 4.31x)
h264_biweight4_8_c: 582.0 ( 1.00x)
h264_biweight4_8_rvv_i32: before 471.0 ( 1.16x)
h264_biweight4_8_rvv_i32: after 91.5 ( 6.36x)
h264_biweight8_8_c: 1110.0 ( 1.00x)
h264_biweight8_8_rvv_i32: before 943.3 ( 1.10x)
h264_biweight8_8_rvv_i64: after 147.0 ( 7.55x)
SpacemiT X60:
h264_biweight2_8_c: 311.4 ( 1.00x)
h264_biweight2_8_rvv_i32: before 363.1 ( 0.83x)
h264_biweight2_8_rvv_i32: after 103.1 ( 3.02x)
h264_biweight4_8_c: 571.9 ( 1.00x)
h264_biweight4_8_rvv_i32: before 717.4 ( 0.78x)
h264_biweight4_8_rvv_i32: after 71.8 ( 7.96x)
h264_biweight8_8_c: 1103.1 ( 1.00x)
h264_biweight8_8_rvv_i32: before 1415.2 ( 0.76x)
h264_biweight8_8_rvv_i64: ater 92.8 (11.88x)
3 months ago
459a1512f1
lavc/h264dsp: unroll R-V V weight16
...
As VLSE128.V does not exist, we have no other way to deal with latency.
T-Head C908:
h264_weight16_8_c: 989.4 ( 1.00x)
h264_weight16_8_rvv_i32: 193.2 ( 5.12x)
SpacemiT X60:
h264_weight16_8_c: 874.1 ( 1.00x)
h264_weight16_8_rvv_i32: 196.9 ( 4.44x)
3 months ago
4936bb2508
lavc/h264dsp: optimise R-V V weight for shorter heights
...
The height is a power of two of up to 16 rows. The current code was
optimised for large sample counts.
T-Head C908:
h264_weight2_8_c: 211.7 ( 1.00x)
h264_weight2_8_rvv_i32: before 184.0 ( 1.15x)
h264_weight2_8_rvv_i32: after 54.2 ( 3.90x)
h264_weight4_8_c: 285.7 ( 1.00x)
h264_weight4_8_rvv_i32: before 341.2 ( 0.86x)
h264_weight4_8_rvv_i32: after 82.2 ( 3.47x)
h264_weight8_8_c: 498.7 ( 1.00x)
h264_weight8_8_rvv_i32: before 683.7 ( 0.73x)
h264_weight8_8_rvv_i64: after 128.5 ( 3.95x)
h264_weight16_8_c: 878.2 ( 1.00x)
h264_weight16_8_rvv_i32: unchanged 239.5 ( 3.67x)
SpacemiT X60:
h264_weight2_8_c: 207.2 ( 1.00x)
h264_weight2_8_rvv_i32: before 259.6 ( 0.80x)
h264_weight2_8_rvv_i32: after 82.2 ( 2.52x)
h264_weight4_8_c: 290.8 ( 1.00x)
h264_weight4_8_rvv_i32: before 509.6 ( 0.57x)
h264_weight4_8_rvv_i32: after 61.5 ( 4.73x)
h264_weight8_8_c: 498.8 ( 1.00x)
h264_weight8_8_rvv_i32: before 1019.8 ( 0.49x)
h264_weight8_8_rvv_i64: after 71.8 ( 6.95x)
h264_weight16_8_c: 874.0 ( 1.00x)
h264_weight16_8_rvv_i32: unchanged 249.0 ( 3.51x)
3 months ago
ba7d0d5fc3
lavc/vvc_mc: R-V V avg w_avg
...
C908 X60
avg_8_2x2_c : 1.2 1.0
avg_8_2x2_rvv_i32 : 0.7 0.7
avg_8_2x4_c : 2.0 2.2
avg_8_2x4_rvv_i32 : 1.2 1.2
avg_8_2x8_c : 3.7 4.0
avg_8_2x8_rvv_i32 : 1.7 1.5
avg_8_2x16_c : 7.2 7.7
avg_8_2x16_rvv_i32 : 3.0 2.7
avg_8_2x32_c : 14.2 15.2
avg_8_2x32_rvv_i32 : 5.5 5.0
avg_8_2x64_c : 51.0 43.7
avg_8_2x64_rvv_i32 : 39.2 29.7
avg_8_2x128_c : 100.5 79.2
avg_8_2x128_rvv_i32 : 79.7 68.2
avg_8_4x2_c : 1.7 2.0
avg_8_4x2_rvv_i32 : 1.0 0.7
avg_8_4x4_c : 3.5 3.7
avg_8_4x4_rvv_i32 : 1.2 1.2
avg_8_4x8_c : 6.7 7.0
avg_8_4x8_rvv_i32 : 1.7 1.5
avg_8_4x16_c : 13.5 14.0
avg_8_4x16_rvv_i32 : 3.0 2.7
avg_8_4x32_c : 26.2 27.7
avg_8_4x32_rvv_i32 : 5.5 4.7
avg_8_4x64_c : 73.0 73.7
avg_8_4x64_rvv_i32 : 39.0 32.5
avg_8_4x128_c : 143.0 137.2
avg_8_4x128_rvv_i32 : 72.7 68.0
avg_8_8x2_c : 3.5 3.5
avg_8_8x2_rvv_i32 : 1.0 0.7
avg_8_8x4_c : 6.2 6.5
avg_8_8x4_rvv_i32 : 1.5 1.0
avg_8_8x8_c : 12.7 13.2
avg_8_8x8_rvv_i32 : 2.0 1.5
avg_8_8x16_c : 25.0 26.5
avg_8_8x16_rvv_i32 : 3.2 2.7
avg_8_8x32_c : 50.0 52.7
avg_8_8x32_rvv_i32 : 6.2 5.0
avg_8_8x64_c : 118.7 122.5
avg_8_8x64_rvv_i32 : 40.2 31.5
avg_8_8x128_c : 236.7 220.2
avg_8_8x128_rvv_i32 : 85.2 67.7
avg_8_16x2_c : 6.2 6.7
avg_8_16x2_rvv_i32 : 1.2 0.7
avg_8_16x4_c : 12.5 13.0
avg_8_16x4_rvv_i32 : 1.7 1.0
avg_8_16x8_c : 24.5 26.0
avg_8_16x8_rvv_i32 : 3.0 1.7
avg_8_16x16_c : 49.0 51.5
avg_8_16x16_rvv_i32 : 5.5 3.0
avg_8_16x32_c : 97.5 102.5
avg_8_16x32_rvv_i32 : 10.5 5.5
avg_8_16x64_c : 213.7 222.0
avg_8_16x64_rvv_i32 : 48.5 34.2
avg_8_16x128_c : 434.7 420.0
avg_8_16x128_rvv_i32 : 97.7 74.0
avg_8_32x2_c : 12.2 12.7
avg_8_32x2_rvv_i32 : 1.5 1.0
avg_8_32x4_c : 24.5 25.5
avg_8_32x4_rvv_i32 : 3.0 1.7
avg_8_32x8_c : 48.5 50.7
avg_8_32x8_rvv_i32 : 5.2 2.7
avg_8_32x16_c : 96.7 101.2
avg_8_32x16_rvv_i32 : 10.2 5.0
avg_8_32x32_c : 192.7 202.2
avg_8_32x32_rvv_i32 : 19.7 9.5
avg_8_32x64_c : 427.5 426.5
avg_8_32x64_rvv_i32 : 64.2 18.2
avg_8_32x128_c : 816.5 821.0
avg_8_32x128_rvv_i32 : 135.2 75.5
avg_8_64x2_c : 24.0 25.2
avg_8_64x2_rvv_i32 : 2.7 1.5
avg_8_64x4_c : 48.2 50.5
avg_8_64x4_rvv_i32 : 5.0 2.7
avg_8_64x8_c : 96.0 100.7
avg_8_64x8_rvv_i32 : 9.7 4.5
avg_8_64x16_c : 207.7 201.2
avg_8_64x16_rvv_i32 : 19.0 9.0
avg_8_64x32_c : 383.2 402.0
avg_8_64x32_rvv_i32 : 37.5 17.5
avg_8_64x64_c : 837.2 828.7
avg_8_64x64_rvv_i32 : 84.7 35.5
avg_8_64x128_c : 1640.7 1640.2
avg_8_64x128_rvv_i32 : 206.0 153.0
avg_8_128x2_c : 48.7 51.0
avg_8_128x2_rvv_i32 : 5.2 2.7
avg_8_128x4_c : 96.7 101.5
avg_8_128x4_rvv_i32 : 10.2 5.0
avg_8_128x8_c : 192.2 202.0
avg_8_128x8_rvv_i32 : 19.7 9.2
avg_8_128x16_c : 400.7 403.2
avg_8_128x16_rvv_i32 : 38.7 18.5
avg_8_128x32_c : 786.7 805.7
avg_8_128x32_rvv_i32 : 77.0 36.2
avg_8_128x64_c : 1615.5 1655.5
avg_8_128x64_rvv_i32 : 189.7 80.7
avg_8_128x128_c : 3182.0 3238.0
avg_8_128x128_rvv_i32 : 397.5 308.5
w_avg_8_2x2_c : 1.7 1.2
w_avg_8_2x2_rvv_i32 : 1.2 1.0
w_avg_8_2x4_c : 2.7 2.7
w_avg_8_2x4_rvv_i32 : 1.7 1.5
w_avg_8_2x8_c : 21.7 4.7
w_avg_8_2x8_rvv_i32 : 2.7 2.5
w_avg_8_2x16_c : 9.5 9.2
w_avg_8_2x16_rvv_i32 : 4.7 4.2
w_avg_8_2x32_c : 19.0 18.7
w_avg_8_2x32_rvv_i32 : 9.0 8.0
w_avg_8_2x64_c : 62.0 50.2
w_avg_8_2x64_rvv_i32 : 47.7 33.5
w_avg_8_2x128_c : 116.7 87.7
w_avg_8_2x128_rvv_i32 : 80.0 69.5
w_avg_8_4x2_c : 2.5 2.5
w_avg_8_4x2_rvv_i32 : 1.2 1.0
w_avg_8_4x4_c : 4.7 4.5
w_avg_8_4x4_rvv_i32 : 1.7 1.7
w_avg_8_4x8_c : 9.0 8.7
w_avg_8_4x8_rvv_i32 : 2.7 2.5
w_avg_8_4x16_c : 17.7 17.5
w_avg_8_4x16_rvv_i32 : 4.7 4.2
w_avg_8_4x32_c : 35.0 35.0
w_avg_8_4x32_rvv_i32 : 9.0 8.0
w_avg_8_4x64_c : 100.5 84.5
w_avg_8_4x64_rvv_i32 : 42.2 33.7
w_avg_8_4x128_c : 203.5 151.2
w_avg_8_4x128_rvv_i32 : 83.0 69.5
w_avg_8_8x2_c : 4.5 4.5
w_avg_8_8x2_rvv_i32 : 1.2 1.2
w_avg_8_8x4_c : 8.7 8.7
w_avg_8_8x4_rvv_i32 : 2.0 1.7
w_avg_8_8x8_c : 17.0 17.0
w_avg_8_8x8_rvv_i32 : 3.2 2.5
w_avg_8_8x16_c : 34.0 33.5
w_avg_8_8x16_rvv_i32 : 5.5 4.2
w_avg_8_8x32_c : 86.0 67.5
w_avg_8_8x32_rvv_i32 : 10.5 8.0
w_avg_8_8x64_c : 187.2 149.5
w_avg_8_8x64_rvv_i32 : 45.0 35.5
w_avg_8_8x128_c : 342.7 290.0
w_avg_8_8x128_rvv_i32 : 108.7 70.2
w_avg_8_16x2_c : 8.5 8.2
w_avg_8_16x2_rvv_i32 : 2.0 1.2
w_avg_8_16x4_c : 16.7 16.7
w_avg_8_16x4_rvv_i32 : 3.0 1.7
w_avg_8_16x8_c : 33.2 33.5
w_avg_8_16x8_rvv_i32 : 5.5 3.0
w_avg_8_16x16_c : 66.2 66.7
w_avg_8_16x16_rvv_i32 : 10.5 5.0
w_avg_8_16x32_c : 132.5 131.0
w_avg_8_16x32_rvv_i32 : 20.0 9.7
w_avg_8_16x64_c : 340.0 283.5
w_avg_8_16x64_rvv_i32 : 60.5 37.2
w_avg_8_16x128_c : 641.2 597.5
w_avg_8_16x128_rvv_i32 : 118.7 77.7
w_avg_8_32x2_c : 16.5 16.7
w_avg_8_32x2_rvv_i32 : 3.2 1.7
w_avg_8_32x4_c : 33.2 33.2
w_avg_8_32x4_rvv_i32 : 5.5 2.7
w_avg_8_32x8_c : 66.0 62.5
w_avg_8_32x8_rvv_i32 : 10.5 5.0
w_avg_8_32x16_c : 131.5 132.0
w_avg_8_32x16_rvv_i32 : 20.2 9.5
w_avg_8_32x32_c : 261.7 272.0
w_avg_8_32x32_rvv_i32 : 39.7 18.0
w_avg_8_32x64_c : 575.2 545.5
w_avg_8_32x64_rvv_i32 : 105.5 58.7
w_avg_8_32x128_c : 1154.2 1088.0
w_avg_8_32x128_rvv_i32 : 207.0 98.0
w_avg_8_64x2_c : 33.0 33.0
w_avg_8_64x2_rvv_i32 : 6.2 2.7
w_avg_8_64x4_c : 65.5 66.0
w_avg_8_64x4_rvv_i32 : 11.5 5.0
w_avg_8_64x8_c : 131.2 132.5
w_avg_8_64x8_rvv_i32 : 22.5 9.5
w_avg_8_64x16_c : 268.2 262.5
w_avg_8_64x16_rvv_i32 : 44.2 18.0
w_avg_8_64x32_c : 561.5 528.7
w_avg_8_64x32_rvv_i32 : 88.0 35.2
w_avg_8_64x64_c : 1136.2 1124.0
w_avg_8_64x64_rvv_i32 : 222.0 82.2
w_avg_8_64x128_c : 2345.0 2312.7
w_avg_8_64x128_rvv_i32 : 423.0 190.5
w_avg_8_128x2_c : 65.7 66.5
w_avg_8_128x2_rvv_i32 : 11.2 5.5
w_avg_8_128x4_c : 131.2 132.2
w_avg_8_128x4_rvv_i32 : 22.0 10.2
w_avg_8_128x8_c : 263.5 312.0
w_avg_8_128x8_rvv_i32 : 43.2 19.7
w_avg_8_128x16_c : 528.7 526.2
w_avg_8_128x16_rvv_i32 : 85.5 39.5
w_avg_8_128x32_c : 1067.7 1062.7
w_avg_8_128x32_rvv_i32 : 171.7 78.2
w_avg_8_128x64_c : 2234.7 2168.7
w_avg_8_128x64_rvv_i32 : 400.0 159.0
w_avg_8_128x128_c : 4752.5 4295.0
w_avg_8_128x128_rvv_i32 : 757.7 365.5
Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
3 months ago
3f9ca51015
lavc/opus*: move to opus/ subdir
4 months ago
6aafe61285
avcodec/mpegvideoencdsp: convert stride parameters from int to ptrdiff_t
4 months ago
7d1dda4892
lavc/h264dsp: R-V V loop_filter_chroma
...
T-Head C908:
h264_v_loop_filter_chroma_8bpp_c: 137.4
h264_v_loop_filter_chroma_8bpp_rvv_i32: 54.2
4 months ago
3a53656837
lavc/h264dsp: do not write back unmodified rows in R-V V loop filter
4 months ago
d8fb44c0aa
lavc/mpegvideoencdsp: R-V V add_8x8basis
...
T-Head C908:
add_8x8basis_c: 440.6
add_8x8basis_rvv_i32: 70.3
SpacemiT X60:
add_8x8basis_c: 436.3
add_8x8basis_rvv_i32: 40.5
4 months ago
1907dd7f23
lavc/mpegvideoencdsp: R-V V try_8x8basis
...
T-Head C908:
try_8x8basis_c: 922.5
try_8x8basis_rvv_i32: 135.3
SpacemiT X60:
try_8x8basis_c: 926.1
try_8x8basis_rvv_i32: 103.1
4 months ago
0fd37c00d7
lavc/mpegvideoencdsp: R-V V pix_norm1
...
T-Head C908:
pix_norm1_c: 480.2
pix_norm1_rvv_i64: 146.9
SpacemiT X60:
pix_norm1_c: 478.2
pix_norm1_rvv_i64: 92.7
4 months ago
63d016aea5
lavc/mpegvideoencdsp: R-V V pix_sum
...
T-Head C908:
pix_sum_c: 332.2
pix_sum_rvv_i64: 91.2
SpacemiT X60:
pix_sum_c: 321.2
pix_sum_rvv_i64: 60.9
4 months ago
4e7b5ac48f
lavc/vp9dsp: R-V V mc bilin hv
...
C908 X60
vp9_avg_bilin_4hv_8bpp_c : 10.7 9.5
vp9_avg_bilin_4hv_8bpp_rvv_i32 : 4.0 3.5
vp9_avg_bilin_8hv_8bpp_c : 38.5 34.2
vp9_avg_bilin_8hv_8bpp_rvv_i32 : 7.2 6.5
vp9_avg_bilin_16hv_8bpp_c : 147.2 130.5
vp9_avg_bilin_16hv_8bpp_rvv_i32 : 14.5 12.7
vp9_avg_bilin_32hv_8bpp_c : 574.2 509.7
vp9_avg_bilin_32hv_8bpp_rvv_i32 : 42.5 38.0
vp9_avg_bilin_64hv_8bpp_c : 2321.2 2017.7
vp9_avg_bilin_64hv_8bpp_rvv_i32 : 163.5 131.0
vp9_put_bilin_4hv_8bpp_c : 10.0 8.7
vp9_put_bilin_4hv_8bpp_rvv_i32 : 3.5 3.0
vp9_put_bilin_8hv_8bpp_c : 35.2 31.2
vp9_put_bilin_8hv_8bpp_rvv_i32 : 6.5 5.7
vp9_put_bilin_16hv_8bpp_c : 134.0 119.0
vp9_put_bilin_16hv_8bpp_rvv_i32 : 12.7 11.5
vp9_put_bilin_32hv_8bpp_c : 538.5 464.2
vp9_put_bilin_32hv_8bpp_rvv_i32 : 39.7 35.2
vp9_put_bilin_64hv_8bpp_c : 2111.7 1833.2
vp9_put_bilin_64hv_8bpp_rvv_i32 : 138.5 122.5
Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
4 months ago
9edd2e723b
lavc/vp9dsp: R-V V mc bilin h v
...
C908 X60
vp9_avg_bilin_4h_8bpp_c : 5.5 4.7
vp9_avg_bilin_4h_8bpp_rvv_i32 : 1.7 1.5
vp9_avg_bilin_4v_8bpp_c : 5.5 4.7
vp9_avg_bilin_4v_8bpp_rvv_i32 : 1.5 1.2
vp9_avg_bilin_8h_8bpp_c : 20.0 17.7
vp9_avg_bilin_8h_8bpp_rvv_i32 : 3.0 2.7
vp9_avg_bilin_8v_8bpp_c : 20.7 18.7
vp9_avg_bilin_8v_8bpp_rvv_i32 : 3.0 2.7
vp9_avg_bilin_16h_8bpp_c : 78.2 69.7
vp9_avg_bilin_16h_8bpp_rvv_i32 : 7.0 6.2
vp9_avg_bilin_16v_8bpp_c : 98.5 73.2
vp9_avg_bilin_16v_8bpp_rvv_i32 : 7.0 6.0
vp9_avg_bilin_32h_8bpp_c : 325.5 275.5
vp9_avg_bilin_32h_8bpp_rvv_i32 : 23.0 20.5
vp9_avg_bilin_32v_8bpp_c : 342.2 290.0
vp9_avg_bilin_32v_8bpp_rvv_i32 : 21.7 19.5
vp9_avg_bilin_64h_8bpp_c : 1263.7 1095.7
vp9_avg_bilin_64h_8bpp_rvv_i32 : 91.2 81.2
vp9_avg_bilin_64v_8bpp_c : 1331.7 1155.2
vp9_avg_bilin_64v_8bpp_rvv_i32 : 91.2 81.0
vp9_put_bilin_4h_8bpp_c : 4.5 4.0
vp9_put_bilin_4h_8bpp_rvv_i32 : 1.0 1.0
vp9_put_bilin_4v_8bpp_c : 4.7 4.2
vp9_put_bilin_4v_8bpp_rvv_i32 : 1.0 1.0
vp9_put_bilin_8h_8bpp_c : 16.7 15.0
vp9_put_bilin_8h_8bpp_rvv_i32 : 2.2 2.0
vp9_put_bilin_8v_8bpp_c : 17.5 15.7
vp9_put_bilin_8v_8bpp_rvv_i32 : 2.2 2.0
vp9_put_bilin_16h_8bpp_c : 65.2 58.0
vp9_put_bilin_16h_8bpp_rvv_i32 : 6.0 5.5
vp9_put_bilin_16v_8bpp_c : 69.2 61.7
vp9_put_bilin_16v_8bpp_rvv_i32 : 5.7 5.2
vp9_put_bilin_32h_8bpp_c : 273.2 229.0
vp9_put_bilin_32h_8bpp_rvv_i32 : 19.7 17.7
vp9_put_bilin_32v_8bpp_c : 290.5 243.7
vp9_put_bilin_32v_8bpp_rvv_i32 : 18.7 16.7
vp9_put_bilin_64h_8bpp_c : 1040.5 910.5
vp9_put_bilin_64h_8bpp_rvv_i32 : 82.5 73.0
vp9_put_bilin_64v_8bpp_c : 1108.5 971.0
vp9_put_bilin_64v_8bpp_rvv_i32 : 82.2 73.2
Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
4 months ago
616fdeaea3
lavc/riscv: depend on RVB and simplify accordingly
...
There is no known (real) hardware with V and without the complete B
extension. B was indeed required in the RISC-V application profile from
2022, earlier than V. There should not be any relevant hardware in the
future either.
In practice, different R-V Vector optimisations in FFmpeg already depend on
every constituent of the B extension anyhow, so it would not work well.
5 months ago
4edfc11a28
lavc/h264dsp: R-V V idct4_add8 (all depths)
...
These are really just wrappers for idct4_add16intra functions, which are in
turn mostly wrappers for idct4_add and idct4_dc_add functions.
For benchmarks refer to the later two sets.
5 months ago
de7f999481
lavc/videodsp: work-around LLVM-as
...
For some reason, it can't handle the normal syntax for an address operand
without an offset, so add a dummy zero offset.
5 months ago
677f28b310
lavc/h264dsp: stick R-V V weight to 16-bit precision
...
T-Head C908 (ns):
h264_weight2_8_c: 1607.8
h264_weight2_8_rvv_i32: 515.0 (before)
h264_weight2_8_rvv_i32: 348.5 (after)
h264_weight4_8_c: 2255.8
h264_weight4_8_rvv_i32: 1015.0 (before)
h264_weight4_8_rvv_i32: 691.0 (after)
h264_weight8_8_c: 3857.5
h264_weight8_8_rvv_i32: 2218.8 (before)
h264_weight8_8_rvv_i32: 1561.3 (after)
h264_weight16_8_c: 7431.5
h264_weight16_8_rvv_i32: 2737.3 (before)
h264_weight16_8_rvv_i32: 1848.3 (after)
SpacemiT X60 (ns):
h264_weight2_8_c: 1624.1
h264_weight2_8_rvv_i32: 352.6 (before)
h264_weight2_8_rvv_i32: 259.3 (after)
h264_weight4_8_c: 2259.3
h264_weight4_8_rvv_i32: 685.8 (before)
h264_weight4_8_rvv_i32: 530.3 (after)
h264_weight8_8_c: 4103.3
h264_weight8_8_rvv_i32: 1581.8 (before)
h264_weight8_8_rvv_i32: 1238.6 (after)
h264_weight16_8_c: 7624.3
h264_weight16_8_rvv_i32: 2738.1 (before)
h264_weight16_8_rvv_i32: 1853.3 (after)
5 months ago
afd45c7ff7
lavc/h264dsp: stick R-V V biweight to 16-bit
...
T-Head C908 (ns):
h264_biweight2_8_c: 2414.5
h264_biweight2_8_rvv_i32: 701.8 (before)
h264_biweight2_8_rvv_i32: 468.5 (after)
h264_biweight4_8_c: 4655.3
h264_biweight4_8_rvv_i32: 1377.5 (before)
h264_biweight4_8_rvv_i32: 931.8 (after)
h264_biweight8_8_c: 9701.5
h264_biweight8_8_rvv_i32: 2896.0 (before)
h264_biweight8_8_rvv_i32: 2070.5 (after)
h264_biweight16_8_c: 18025.0
h264_biweight16_8_rvv_i32: 3460.8 (before)
h264_biweight16_8_rvv_i32: 1978.0 (after)
SpacemiT X60 (ns):
h264_biweight2_8_c: 2415.5
h264_biweight2_8_rvv_i32: 478.2 (before)
h264_biweight2_8_rvv_i32: 362.8 (after)
h264_biweight4_8_c: 4655.3
h264_biweight4_8_rvv_i32: 946.7 (before)
h264_biweight4_8_rvv_i32: 727.3 (after)
h264_biweight8_8_c: 9061.8
h264_biweight8_8_rvv_i32: 2071.7 (before)
h264_biweight8_8_rvv_i32: 1685.8 (after)
h264_biweight16_8_c: 18020.5
h264_biweight16_8_rvv_i32: 3457.2 (before)
h264_biweight16_8_rvv_i32: 1935.8 (after)
5 months ago
2f083fd581
lavc/audiodsp: drop R-V F vector_clipf
...
This is now firmly slower than C.
SiFive-U74 (cycles):
audiodsp.vector_clipf_c: 31.2
audiodsp.vector_clipf_rvf: 39.5
5 months ago
54ae270213
lavc/rv34dsp: use saturating add/sub for R-V V DC add
...
T-Head C908 (cycles):
rv34_idct_dc_add_c: 113.2
rv34_idct_dc_add_rvv_i32: 48.5 (before)
rv34_idct_dc_add_rvv_i32: 39.5 (after)
5 months ago
952b426f3b
lavc/bswapdsp: add RV Zvbb bswap16 and bswap32
5 months ago
262168b04e
lavc/videodsp: RISC-V zicbop prefetch
...
There are currently no ways to run-time detect the CPU capability, so we
take it for granted (in the worst case, it will execute NOPs).
5 months ago
324eba69f7
lavc/vc1dsp: use saturating arithmetic for RVV inv_trans_dc
...
T-Head C908 (cycles):
vc1dsp.vc1_inv_trans_4x4_dc_c: 113.7
vc1dsp.vc1_inv_trans_4x4_dc_rvv_i32: 46.5 (before)
vc1dsp.vc1_inv_trans_4x4_dc_rvv_i32: 45.5 (after)
vc1dsp.vc1_inv_trans_4x8_dc_c: 230.7
vc1dsp.vc1_inv_trans_4x8_dc_rvv_i32: 65.7 (before)
vc1dsp.vc1_inv_trans_4x8_dc_rvv_i32: 52.5 (after)
vc1dsp.vc1_inv_trans_8x4_dc_c: 246.7
vc1dsp.vc1_inv_trans_8x4_dc_rvv_i64: 56.7 (before)
vc1dsp.vc1_inv_trans_8x4_dc_rvv_i64: 45.5 (after)
vc1dsp.vc1_inv_trans_8x8_dc_c: 419.7
vc1dsp.vc1_inv_trans_8x8_dc_rvv_i64: 81.2 (before)
vc1dsp.vc1_inv_trans_8x8_dc_rvv_i64: 53.5 (after)
5 months ago
784a72a116
lavc/vc1dsp: unify R-V V DC bypass functions
5 months ago
bd0c3edb13
lavu/riscv: count bytes rather than words for bswap32
...
This removes the dependency on Zba at essentially zero cost.
5 months ago
5171baa228
lavc/ac3dsp: fix R-V CPU requirements
...
It probably will not matter on any real hardware, but the Zbb optimisations
do not require Zba. And then, we need HAVE_RVV to build the RVV stuff.
5 months ago
7b24f96c87
lavc/vp9dsp: remove R-V I intra functions
...
At this point, they are identical to the C code, except for instruction
ordering. In fact, they are typically slower or no faster than the C code.
5 months ago
b0b3bea10b
lavc/h264dsp: use saturing add/sub for R-V V 8-bit DC add
...
T-Head C908 (cycles):
h264_idct4_dc_add_8bpp_c: 109.2
h264_idct4_dc_add_8bpp_rvv_i32: 34.5 (before)
h264_idct4_dc_add_8bpp_rvv_i32: 25.5 (after)
h264_idct8_dc_add_8bpp_c: 418.7
h264_idct8_dc_add_8bpp_rvv_i64: 69.5 (before)
h264_idct8_dc_add_8bpp_rvv_i64: 33.5 (after)
5 months ago
9b4655c3a1
lavc/vp8dsp: use saturating add/sub for R-V V DC add
...
T-Head C908 (cycles):
vp7_idct_dc_add_c: 108.5
vp7_idct_dc_add_rvv_i32: 56.2 (before)
vp7_idct_dc_add_rvv_i32: 47.2 (after)
vp8_idct_dc_add_c: 96.2
vp8_idct_dc_add_rvv_i32: 43.0 (before)
vp8_idct_dc_add_rvv_i32: 34.0 (after)
5 months ago
bbfc0ac9ca
lavc/riscv: don't set vxrm if unnecessary
...
While narrowing clip is nominally a rounding operation, the rounding mode
has no arithmetic consequence if the right shift is by zero bits.
5 months ago
f2c30fe15a
lavc/riscv: add forward-edge CFI landing pads
5 months ago
b62586e310
lavc/h264dsp: use RISC-V B extension
...
This saves one register and one instruction per transform.
add16 and add16intra thus become stack-less.
5 months ago
187d4d066a
lavc/riscv: require B or zba explicitly
5 months ago
896c22ef00
lavc/vp8dsp: fix RV32 stack alignment
...
SP must be a multiple of 16 bytes at all times on POSIX - even in leaf
functions - so that signal handlers have a properly aligned stack.
5 months ago
9135dffd17
lavc/h264dsp: reduce spills in R-V V idct_add16
5 months ago
245f76ad74
lavc/h264dsp: reuse the R-V V IDCT DC add functions
...
This reuses the DC bypass functions from the multiple IDCT functions, to
leverage vector code.
As an added bonus, the caller functions can now rely on the callee functions
to preserve their parameters, thus cutting down on stack spills.
5 months ago
0a5b5bae89
lavc/h264dsp: correct VL and LMUL in idct_dc_add
...
T-Head C908 (cycles):
h264_idct4_dc_add_8bpp_c: 94.7
h264_idct4_dc_add_8bpp_rvv_i32: 55.0 (before)
h264_idct4_dc_add_8bpp_rvv_i32: 34.5 (after)
h264_idct4_dc_add_9bpp_c: 94.7
h264_idct4_dc_add_9bpp_rvv_i32: 43.5 (before)
h264_idct4_dc_add_9bpp_rvv_i32: 38.2 (after)
h264_idct4_dc_add_10bpp_c: 94.7
h264_idct4_dc_add_10bpp_rvv_i32: 43.5 (before)
h264_idct4_dc_add_10bpp_rvv_i32: 38.2 (after)
h264_idct4_dc_add_12bpp_c: 94.7
h264_idct4_dc_add_12bpp_rvv_i32: 43.7 (before)
h264_idct4_dc_add_12bpp_rvv_i32: 38.5 (after)
h264_idct4_dc_add_14bpp_c: 94.7
h264_idct4_dc_add_14bpp_rvv_i32: 43.7 (before)
h264_idct4_dc_add_14bpp_rvv_i32: 38.5 (after)
5 months ago
c9dc2ad09b
lavc/h264dsp: move R-V V idct_dc_add
...
No functional changes. This just moves the assembler so that it can be
referenced by other functions in h264idct_rvv.S with local jumps.
Edited-by: Rémi Denis-Courmont <remi@remlab.net>
5 months ago
d15169c51f
lavc/h264dsp: factor some mostly identical R-V V code
5 months ago
483fd732ab
lavc/h264dsp: R-V V high-depth idct_add{,intra}16, idct8_add4
...
As with 8-bit, this tends to be faster, but results are all over the
place due to the variable distribution of non-zero coefficients.
6 months ago
fa5a605542
avcodec/riscv: add h264 dc idct rvv
...
checkasm: bench runs 131072 (1 << 17)
h264_idct4_add_dc_8bpp_c: 1.5
h264_idct4_add_dc_8bpp_rvv_i64: 0.7
h264_idct4_add_dc_9bpp_c: 1.5
h264_idct4_add_dc_9bpp_rvv_i64: 0.7
h264_idct4_add_dc_10bpp_c: 1.5
h264_idct4_add_dc_10bpp_rvv_i64: 0.7
h264_idct4_add_dc_12bpp_c: 1.2
h264_idct4_add_dc_12bpp_rvv_i64: 0.7
h264_idct4_add_dc_14bpp_c: 1.2
h264_idct4_add_dc_14bpp_rvv_i64: 0.7
h264_idct8_add_dc_8bpp_c: 5.2
h264_idct8_add_dc_8bpp_rvv_i64: 1.5
h264_idct8_add_dc_9bpp_c: 5.5
h264_idct8_add_dc_9bpp_rvv_i64: 1.2
h264_idct8_add_dc_10bpp_c: 5.5
h264_idct8_add_dc_10bpp_rvv_i64: 1.2
h264_idct8_add_dc_12bpp_c: 4.2
h264_idct8_add_dc_12bpp_rvv_i64: 1.2
h264_idct8_add_dc_14bpp_c: 4.2
h264_idct8_add_dc_14bpp_rvv_i64: 1.2
Signed-off-by: J. Dekker <jdek@itanimul.li>
6 months ago
3002310b70
lavc/h264dsp: R-V V high-depth add_pixels8
...
T-Head C908 (cycles);
h264_add_pixels8_9bpp_c: 270.5
h264_add_pixels8_9bpp_rvv_i32: 164.2
h264_add_pixels8_10bpp_c: 270.5
h264_add_pixels8_10bpp_rvv_i32: 164.2
h264_add_pixels8_12bpp_c: 270.5
h264_add_pixels8_12bpp_rvv_i32: 164.2
h264_add_pixels8_14bpp_c: 270.5
h264_add_pixels8_14bpp_rvv_i32: 164.2
6 months ago
7744c08240
lavc/h264dsp: R-V V add_pixels4 and 8-bit add_pixels8
...
T-Head C908 (cycles):
h264_add_pixels4_8bpp_c: 93.5
h264_add_pixels4_8bpp_rvv_i32: 39.5
h264_add_pixels4_9bpp_c: 87.5
h264_add_pixels4_9bpp_rvv_i64: 50.5
h264_add_pixels4_10bpp_c: 87.5
h264_add_pixels4_10bpp_rvv_i64: 50.5
h264_add_pixels4_12bpp_c: 87.5
h264_add_pixels4_12bpp_rvv_i64: 50.5
h264_add_pixels4_14bpp_c: 87.5
h264_add_pixels4_14bpp_rvv_i64: 50.5
h264_add_pixels8_8bpp_c: 265.2
h264_add_pixels8_8bpp_rvv_i64: 84.5
6 months ago
c654e37254
lavc/h264dsp: R-V V high-depth h264_idct8_add
...
Unlike the 8-bit version, we need two iterations to process this within
128-bit vectors. This adds some extra complexity for pointer arithmetic
and counting down which is unnecessary in the 8-bit variant.
Accordingly the gain relative to C are just slight better than half as
good with 128-bit vectors as with 256-bit ones.
T-Head C908 (2 iterations):
h264_idct8_add_9bpp_c: 17.5
h264_idct8_add_9bpp_rvv_i32: 10.0
h264_idct8_add_10bpp_c: 17.5
h264_idct8_add_10bpp_rvv_i32: 9.7
h264_idct8_add_12bpp_c: 17.7
h264_idct8_add_12bpp_rvv_i32: 9.7
h264_idct8_add_14bpp_c: 17.7
h264_idct8_add_14bpp_rvv_i32: 9.7
SpacemiT X60 (single iteration):
h264_idct8_add_9bpp_c: 15.2
h264_idct8_add_9bpp_rvv_i32: 5.0
h264_idct8_add_10bpp_c: 15.2
h264_idct8_add_10bpp_rvv_i32: 5.0
h264_idct8_add_12bpp_c: 14.7
h264_idct8_add_12bpp_rvv_i32: 5.0
h264_idct8_add_14bpp_c: 14.7
h264_idct8_add_14bpp_rvv_i32: 4.7
6 months ago
4e0e872881
lavc/h264dsp: R-V V high-depth h264_idct_add
...
T-Head C908 (cycles):
h264_idct4_add_9bpp_c: 248.2
h264_idct4_add_9bpp_rvv_i32: 128.7
h264_idct4_add_10bpp_c: 256.7
h264_idct4_add_10bpp_rvv_i32: 128.7
h264_idct4_add_12bpp_c: 252.5
h264_idct4_add_12bpp_rvv_i32: 129.7
h264_idct4_add_14bpp_c: 258.0
h264_idct4_add_14bpp_rvv_i32: 129.7
6 months ago
d28a7e8eb7
lavc/h264dsp: avoid \+ expansion
...
This seems to be unsupported by LLVM-as.
6 months ago
f1ed351d3b
lavc/h264dsp: R-V V 8-bit h264_biweight_pixels
...
T-Head C908:
h264_biweight2_8_c: 58.0
h264_biweight2_8_rvv_i32: 11.2
h264_biweight4_8_c: 106.0
h264_biweight4_8_rvv_i32: 22.7
h264_biweight8_8_c: 205.7
h264_biweight8_8_rvv_i32: 50.0
h264_biweight16_8_c: 403.5
h264_biweight16_8_rvv_i32: 83.2
SpacemiT X60:
h264_weight2_8_c: 48.2
h264_weight2_8_rvv_i32: 8.2
h264_weight4_8_c: 90.5
h264_weight4_8_rvv_i32: 16.5
h264_weight8_8_c: 175.2
h264_weight8_8_rvv_i32: 38.0
h264_weight16_8_c: 342.2
h264_weight16_8_rvv_i32: 66.0
6 months ago
3606e592ea
lavc/h264dsp: R-V V 8-bit h264_weight_pixels
...
There are two implementations here:
- a generic scalable one processing two columns at a time,
- a specialised processing one (fixed-size) row at a time.
Unsurprisingly, the generic one works out better with smaller widths.
With larger widths, the gains from filling vectors are outweighed by
the extra cost of strided loads and stores. In other words, memory
accesses become the bottleneck.
T-Head C908:
h264_weight2_8_c: 54.5
h264_weight2_8_rvv_i32: 13.7
h264_weight4_8_c: 101.7
h264_weight4_8_rvv_i32: 27.5
h264_weight8_8_c: 197.0
h264_weight8_8_rvv_i32: 75.5
h264_weight16_8_c: 385.0
h264_weight16_8_rvv_i32: 74.2
SpacemiT X60:
h264_weight2_8_c: 48.5
h264_weight2_8_rvv_i32: 8.2
h264_weight4_8_c: 90.7
h264_weight4_8_rvv_i32: 16.5
h264_weight8_8_c: 175.0
h264_weight8_8_rvv_i32: 37.7
h264_weight16_8_c: 342.2
h264_weight16_8_rvv_i32: 66.0
6 months ago
Niklas Haas
Rémi Denis-Courmont
sunyuechi
Anton Khirnov
Ramiro Polla
J. Dekker