From c654e37254efb6de5d3f1355dd7936979dd0dca3 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?R=C3=A9mi=20Denis-Courmont?= Date: Thu, 11 Jul 2024 22:01:25 +0300 Subject: [PATCH] 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 --- libavcodec/riscv/h264dsp_init.c | 32 ++++++--- libavcodec/riscv/h264idct_rvv.S | 123 ++++++++++++++++++++++++++++++-- 2 files changed, 140 insertions(+), 15 deletions(-) diff --git a/libavcodec/riscv/h264dsp_init.c b/libavcodec/riscv/h264dsp_init.c index 463ffe7202..1fb73f810e 100644 --- a/libavcodec/riscv/h264dsp_init.c +++ b/libavcodec/riscv/h264dsp_init.c @@ -53,9 +53,13 @@ void ff_h264_idct8_add4_8_rvv(uint8_t *dst, const int *blockoffset, const uint8_t nnzc[5 * 8]); void ff_h264_idct_add_9_rvv(uint8_t *dst, int16_t *block, int stride); +void ff_h264_idct8_add_9_rvv(uint8_t *dst, int16_t *block, int stride); void ff_h264_idct_add_10_rvv(uint8_t *dst, int16_t *block, int stride); +void ff_h264_idct8_add_10_rvv(uint8_t *dst, int16_t *block, int stride); void ff_h264_idct_add_12_rvv(uint8_t *dst, int16_t *block, int stride); +void ff_h264_idct8_add_12_rvv(uint8_t *dst, int16_t *block, int stride); void ff_h264_idct_add_14_rvv(uint8_t *dst, int16_t *block, int stride); +void ff_h264_idct8_add_14_rvv(uint8_t *dst, int16_t *block, int stride); extern int ff_startcode_find_candidate_rvb(const uint8_t *, int); extern int ff_startcode_find_candidate_rvv(const uint8_t *, int); @@ -94,14 +98,26 @@ av_cold void ff_h264dsp_init_riscv(H264DSPContext *dsp, const int bit_depth, # endif } - if (bit_depth == 9 && zvl128b) - dsp->h264_idct_add = ff_h264_idct_add_9_rvv; - if (bit_depth == 10 && zvl128b) - dsp->h264_idct_add = ff_h264_idct_add_10_rvv; - if (bit_depth == 12 && zvl128b) - dsp->h264_idct_add = ff_h264_idct_add_12_rvv; - if (bit_depth == 14 && zvl128b) - dsp->h264_idct_add = ff_h264_idct_add_14_rvv; + if (bit_depth == 9) { + if (zvl128b) + dsp->h264_idct_add = ff_h264_idct_add_9_rvv; + dsp->h264_idct8_add = ff_h264_idct8_add_9_rvv; + } + if (bit_depth == 10) { + if (zvl128b) + dsp->h264_idct_add = ff_h264_idct_add_10_rvv; + dsp->h264_idct8_add = ff_h264_idct8_add_10_rvv; + } + if (bit_depth == 12) { + if (zvl128b) + dsp->h264_idct_add = ff_h264_idct_add_12_rvv; + dsp->h264_idct8_add = ff_h264_idct8_add_12_rvv; + } + if (bit_depth == 14) { + if (zvl128b) + dsp->h264_idct_add = ff_h264_idct_add_14_rvv; + dsp->h264_idct8_add = ff_h264_idct8_add_14_rvv; + } dsp->startcode_find_candidate = ff_startcode_find_candidate_rvv; } diff --git a/libavcodec/riscv/h264idct_rvv.S b/libavcodec/riscv/h264idct_rvv.S index 001ce0a0f4..7dd0a524fe 100644 --- a/libavcodec/riscv/h264idct_rvv.S +++ b/libavcodec/riscv/h264idct_rvv.S @@ -161,13 +161,6 @@ func ff_h264_idct_add_16_rvv, zve32x ret endfunc -.irp depth, 9, 10, 12, 14 -func ff_h264_idct_add_\depth\()_rvv, zve32x - li a3, (1 << \depth) - 1 - j ff_h264_idct_add_16_rvv -endfunc -.endr - .variant_cc ff_h264_idct8_rvv func ff_h264_idct8_rvv, zve32x vsra.vi v9, v7, 1 @@ -301,6 +294,122 @@ func ff_h264_idct8_add_8_rvv, zve32x ret endfunc +func ff_h264_idct8_add_16_rvv, zve32x + li a4, 8 + csrwi vxrm, 0 + vsetivli a5, 8, e32, m1, ta, ma +1: + addi t1, a1, 1 * 8 * 4 + vle32.v v0, (a1) + addi t2, a1, 2 * 8 * 4 + vle32.v v1, (t1) + addi t3, a1, 3 * 8 * 4 + vle32.v v2, (t2) + addi t4, a1, 4 * 8 * 4 + vle32.v v3, (t3) + addi t5, a1, 5 * 8 * 4 + vle32.v v4, (t4) + addi t6, a1, 6 * 8 * 4 + vle32.v v5, (t5) + addi a7, a1, 7 * 8 * 4 + vle32.v v6, (t6) + sub a4, a4, a5 + vle32.v v7, (a7) + jal t0, ff_h264_idct8_rvv + vse32.v v0, (a1) + sh2add a1, a5, a1 + vse32.v v1, (t1) + vse32.v v2, (t2) + vse32.v v3, (t3) + vse32.v v4, (t4) + vse32.v v5, (t5) + vse32.v v6, (t6) + vse32.v v7, (a7) + bnez a4, 1b + + addi a1, a1, -8 * 4 + li a4, 8 + slli a6, a5, 3 + 2 +2: + vsetvli zero, zero, e32, m1, ta, ma + vlseg8e32.v v0, (a1) + jal t0, ff_h264_idct8_rvv + add t1, a0, a2 + vle16.v v16, (a0) + add t2, t1, a2 + vle16.v v17, (t1) + add t3, t2, a2 + vle16.v v18, (t2) + add t4, t3, a2 + vle16.v v19, (t3) + add t5, t4, a2 + vle16.v v20, (t4) + add t6, t5, a2 + vle16.v v21, (t5) + add a7, t6, a2 + vle16.v v22, (t6) + sub a4, a4, a5 + vle16.v v23, (a7) + .irp n,0,1,2,3,4,5,6,7 + vssra.vi v\n, v\n, 6 + .endr + vsetvli zero, zero, e16, mf2, ta, ma + vwaddu.wv v0, v0, v16 + add a1, a6, a1 + vwaddu.wv v1, v1, v17 + vwaddu.wv v2, v2, v18 + vwaddu.wv v3, v3, v19 + vwaddu.wv v4, v4, v20 + vwaddu.wv v5, v5, v21 + vwaddu.wv v6, v6, v22 + vwaddu.wv v7, v7, v23 + vsetvli zero, zero, e32, m1, ta, ma + .irp n,0,1,2,3,4,5,6,7 + vmax.vx v\n, v\n, zero + .endr + .irp n,0,1,2,3,4,5,6,7 + vmin.vx v\n, v\n, a3 + .endr + vsetvli zero, zero, e16, mf2, ta, ma + vncvt.x.x.w v16, v0 + vncvt.x.x.w v17, v1 + vncvt.x.x.w v18, v2 + vncvt.x.x.w v19, v3 + vncvt.x.x.w v20, v4 + vncvt.x.x.w v21, v5 + vncvt.x.x.w v22, v6 + vncvt.x.x.w v23, v7 + vse16.v v16, (a0) + sh1add a0, a5, a0 + vse16.v v17, (t1) + vse16.v v18, (t2) + vse16.v v19, (t3) + vse16.v v20, (t4) + vse16.v v21, (t5) + vse16.v v22, (t6) + vse16.v v23, (a7) + bnez a4, 2b + + .equ offset, 0 + .rept 2048 / __riscv_xlen + sx zero, offset - 8 * 8 * 4(a1) + .equ offset, offset + (__riscv_xlen / 8) + .endr + ret +endfunc + +.irp depth, 9, 10, 12, 14 +func ff_h264_idct_add_\depth\()_rvv, zve32x + li a3, (1 << \depth) - 1 + j ff_h264_idct_add_16_rvv +endfunc + +func ff_h264_idct8_add_\depth\()_rvv, zve32x + li a3, (1 << \depth) - 1 + j ff_h264_idct8_add_16_rvv +endfunc +.endr + const ff_h264_scan8 .byte 014, 015, 024, 025, 016, 017, 026, 027 .byte 034, 035, 044, 045, 036, 037, 046, 047