T-Head C908 (cycles):
vc1dsp.vc1_inv_trans_4x4_c: 310.7
vc1dsp.vc1_inv_trans_4x4_rvv_i32: 120.0
We could use 1 `vlseg4e64.v` instead of 4 `vle16.v`, but that seems to
be about 7% slower.
This is almost the same story as vp7_idct_add4y. We just have to use
strided loads of 2 64-bit elements to account for the different data
layout in memory.
T-Head C908:
vp7_idct_dc_add4uv_c: 7.5
vp7_idct_dc_add4uv_rvv_i64: 2.0
vp8_idct_dc_add4uv_c: 6.2
vp8_idct_dc_add4uv_rvv_i32: 2.2 (before)
vp8_idct_dc_add4uv_rvv_i64: 2.0
SpacemiT X60:
vp7_idct_dc_add4uv_c: 6.7
vp7_idct_dc_add4uv_rvv_i64: 2.2
vp8_idct_dc_add4uv_c: 5.7
vp8_idct_dc_add4uv_rvv_i32: 2.5 (before)
vp8_idct_dc_add4uv_rvv_i64: 2.0
DCT-related FFmpeg functions often add an unsigned 8-bit sample to a
signed 16-bit coefficient, then clip the result back to an unsigned
8-bit value. RISC-V has no signed 16-bit to unsigned 8-bit clip, so
instead our most common sequence is:
VWADDU.WV
set SEW to 16 bits
VMAX.VV zero # clip negative values to 0
set SEW to 8 bits
VNCLIPU.WI # clip values over 255 to 255 and narrow
Here we use a different sequence which does not require toggling the
vector type. This assumes that the wide addend vector is biased by
-128:
VWADDU.WV
VNCLIP.WI # clip values to signed 8-bit and narrow
VXOR.VX 0x80 # flip sign bit (convert signed to unsigned)
Also the VMAX is effectively replaced by a VXOR of half-width. In this
function, this comes for free as we anyway add a constant to the wide
vector in the prologue.
On C908, this has no observable effects. On X60, this improves
microbenchmarks by about 20%.
As with idct_dc_add, most of the code is shared with, and replaces, the
previous VP8 function. To improve performance, we break down the 16x4
matrix into 4 rows, rather than 4 squares. Thus strided loads and
stores are avoided, and the 4 DC calculations are vectored.
Unfortunately this requires a vector gather to splat the DC values, but
overall this is still a win for performance:
T-Head C908:
vp7_idct_dc_add4y_c: 7.2
vp7_idct_dc_add4y_rvv_i32: 2.2
vp8_idct_dc_add4y_c: 6.2
vp8_idct_dc_add4y_rvv_i32: 2.2 (before)
vp8_idct_dc_add4y_rvv_i32: 1.7
SpacemiT X60:
vp7_idct_dc_add4y_c: 6.2
vp7_idct_dc_add4y_rvv_i32: 2.0
vp8_idct_dc_add4y_c: 5.5
vp8_idct_dc_add4y_rvv_i32: 2.5 (before)
vp8_idct_dc_add4y_rvv_i32: 1.7
I also tried to provision the DC values using indexed loads. It ends up
slower overall, especially for VP7, as we then have to compute 16 DC's
instead of just 4.
This just computes the direct coefficient and hands over to code shared
with VP8. Accordingly the bulk of changes are just rewriting the VP8
code to share.
Nothing to write home about:
vp7_idct_dc_add_c: 1.7
vp7_idct_dc_add_rvv_i32: 1.2
The 8x8 pixel arrays are not necessarily aligned to 64 bits, so the
current code leads to Bus error on real hardware. This reproducible
with FATE's vc1_ilaced_twomv test case.
The new "pessimist" code can trivially be shared for 16x16 pixel
arrays so we also do that. FWIW, this also nominally reduces the
hardware requirement from Zve64x to Zve32x.
T-Head C908:
vc1dsp.avg_vc1_mspel_pixels_tab[0][0]_c: 14.7
vc1dsp.avg_vc1_mspel_pixels_tab[0][0]_rvv_i32: 3.5
vc1dsp.avg_vc1_mspel_pixels_tab[1][0]_c: 3.7
vc1dsp.avg_vc1_mspel_pixels_tab[1][0]_rvv_i32: 1.5
SpacemiT X60:
vc1dsp.avg_vc1_mspel_pixels_tab[0][0]_c: 13.0
vc1dsp.avg_vc1_mspel_pixels_tab[0][0]_rvv_i32: 3.0
vc1dsp.avg_vc1_mspel_pixels_tab[1][0]_c: 3.2
vc1dsp.avg_vc1_mspel_pixels_tab[1][0]_rvv_i32: 1.2
hf_apply_noise_0_c: 35.7
hf_apply_noise_0_rvv_f32: 9.5
hf_apply_noise_1_c: 38.5
hf_apply_noise_1_rvv_f32: 10.0
hf_apply_noise_2_c: 35.5
hf_apply_noise_2_rvv_f32: 9.7
hf_apply_noise_3_c: 38.5
hf_apply_noise_3_rvv_f32: 10.0
Maybe extending the noise table manually is not such great idea, but I
not quite sure how to deal with that otherwise? Allocating the table
dynamically is possible but would require an ELF destructor to clean up.
This works out a bit more favourably than VP8's due to:
- additional multiplications that can be vectored,
- hardware-supported fixed-point rounding mode.
vp7_luma_dc_wht_c: 3.2
vp7_luma_dc_wht_rvv_i64: 2.0
This saves one instruction and frees up A5, which will be repurposed in
later changes. Unfortunately, we need to add quite a lot of alternative
code for this.
128-bit is the maximum, not the minimum here. Larger vector sizes can
result in reads past the end of the noise value table.
This partially reverts commit cdcb4b98b7.
This loop correctly assumes that VLMAX=16 (4x128-bit vectors
with 32-bit elements) and 32 >= pred_order > 16. We need to alternate
between VL=16 and VL=t2=pred_order-16 elements to add up to pred_order.
The current code requests AVL=a2=pred_order elements. In QEMU and on
thte K230 hardware, this sets VL=16 as we need. But the specification
merely guarantees that we get: ceil(AVL / 2) <= VL <= VLMAX. For
instance, if pred_order equals 27, we could end up with VL=14 or VL=15
instead of VL=16. So instead, request literally VLMAX=16.
Since the horizontal and vertical filters are identical except for a
transposition, this uses a common subprocedure with an ad-hoc ABI.
To preserve return-address stack prediction, a link register has to be
used (c.f. the "Control Transfer Instructions" from the
RISC-V ISA Manual). The alternate/temporary link register T0 is used
here, so that the normal RA is preserved (something Arm cannot do!).
To load the strength value based on `qscale`, the shortest possible
and PIC-compatible sequence is used: AUIPC; ADD; LBU. The classic
LLA; ADD; LBU sequence would add one more instruction since LLA is a
convenience alias for AUIPC; ADDI. To ensure that this trick works,
relocation relaxation is disabled.
To implement the two signed divisions by a power of two toward zero:
(x / (1 << SHIFT))
the code relies on the small range of integers involved, computing:
(x + (x >> (16 - SHIFT))) >> SHIFT
rather than the more general:
(x + ((x >> (16 - 1)) & ((1 << SHIFT) - 1))) >> SHIFT
Thus one ANDI instruction is avoided.
T-Head C908:
h263dsp.h_loop_filter_c: 228.2
h263dsp.h_loop_filter_rvv_i32: 144.0
h263dsp.v_loop_filter_c: 242.7
h263dsp.v_loop_filter_rvv_i32: 114.0
(C is probably worse in real use due to less predictible branches.)
While this function can easily be written with vectors, it just fails to
get any performance improvement.
For reference, this is a simpler loop-free implementation that does get
better performance than the current one depending on hardware, but still
more or less the same metrics as the C code:
func ff_sbr_neg_odd_64_rvv, zve64x
li a1, 32
addi a0, a0, 7
li t0, 8
vsetvli zero, a1, e8, m2, ta, ma
li t1, 0x80
vlse8.v v8, (a0), t0
vxor.vx v8, v8, t1
vsse8.v v8, (a0), t0
ret
endfunc
This reverts commit d06fd18f8f.
Notes:
- The loop is biased toward no unescaped bytes as that should be most common.
- The input byte array is slid rather than the (8 times smaller) bit-mask,
as RISC-V V does not provide a bit-mask (or bit-wise) slide instruction.
- There are two comparisons with 0 per iteration, for the same reason.
- In case of match, bytes are copied until the first match, and the loop is
restarted after the escape byte. Vector compression (vcompress.vm) could
discard all escape bytes but that is slower if escape bytes are rare.
Further optimisations should be possible, e.g.:
- processing 2 bytes fewer per iteration to get rid of a 2 slides,
- taking a short cut if the input vector contains less than 2 zeroes.
But this is a good starting point:
T-Head C908:
vc1dsp.vc1_unescape_buffer_c: 12749.5
vc1dsp.vc1_unescape_buffer_rvv_i32: 6009.0
SpacemiT X60:
vc1dsp.vc1_unescape_buffer_c: 11038.0
vc1dsp.vc1_unescape_buffer_rvv_i32: 2061.0
This is pretty much the same as for lpc16, though it only improves half
as large prediction orders. With 128-bit vectors, this gives:
C V old V new
1 69.2 181.5 95.5
2 107.7 180.7 95.2
3 145.5 180.0 103.5
4 183.0 179.2 102.7
5 220.7 178.5 128.0
6 257.7 194.0 127.5
7 294.5 193.7 126.7
8 331.0 193.0 126.5
Larger prediction orders see no significant changes at that size.
This calculates the optimal vector type value at run-time based on the
hardware vector length and the FLAC LPC prediction order. In this
particular case, the additional computation is easily amortised over
the loop iterations:
T-Head C908:
C V before V after
1 48.0 214.7 95.2
2 64.7 214.2 94.7
3 79.7 213.5 94.5
4 96.2 196.5 94.2 #
5 111.0 195.7 118.5
6 127.0 211.2 102.0
7 143.7 194.2 101.5
8 175.7 193.2 101.2 #
9 176.2 224.2 126.0
10 191.5 192.0 125.5
11 224.5 191.2 124.7
12 223.0 190.2 124.2
13 239.2 189.5 123.7
14 253.7 188.7 139.5
15 286.2 188.0 122.7
16 284.0 187.0 122.5 #
17 300.2 186.5 186.5
18 314.0 185.5 185.7
19 329.7 184.7 185.0
20 343.0 184.2 184.2
21 358.7 199.2 183.7
22 371.7 182.7 182.7
23 387.5 181.7 182.0
24 400.7 181.0 181.2
25 431.5 180.2 196.5
26 443.7 195.5 196.0
27 459.0 178.7 196.2
28 470.7 177.7 194.2
29 470.0 177.0 193.5
30 481.2 176.2 176.5
31 496.2 175.5 175.7
32 507.2 174.7 191.0 #
# Power of two boundary.
With 128-bit vectors, improvements are expected for the first two
test cases only. For the other two, there is overhead but below noise.
Improvements should be better observable with prediction order of 8
and less, or on hardware with larger vector sizes.
The main loop processes 8 bytes in 5 instructions.
For comparison, the optimal plain strnlen() requires 4 instructions per
byte (6.4x worse): LBU; ADDI; BEQZ; BNE. The current libavcodec C code
involves 5 instructions per byte (8x worse). Actual benchmarks may be
slightly less favourable due to latency from ORC.B to BNE.
Rémi Denis-Courmont
sunyuechi