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
Unified · Split
| @@ -21,12 +21,15 @@ · 20 unmodified lines | |||
| 21 | #include "config.h" | 21 | #include "config.h" |
| 22 | 22 | ||
| 23 | #include <stdint.h> | 23 | #include <stdint.h> |
| 24 | #include <string.h> | ||
| 24 | 25 | ||
| 25 | #include "libavutil/attributes.h" | 26 | #include "libavutil/attributes.h" |
| 26 | #include "libavutil/cpu.h" | 27 | #include "libavutil/cpu.h" |
| 27 | #include "libavutil/riscv/cpu.h" | 28 | #include "libavutil/riscv/cpu.h" |
| 28 | #include "libavcodec/h264dsp.h" | 29 | #include "libavcodec/h264dsp.h" |
| 29 | 30 | ||
| 31 | extern const h264_weight_func ff_h264_weight_funcs_8_rvv[]; | ||
| 32 | |||
| 30 | void ff_h264_v_loop_filter_luma_8_rvv(uint8_t *pix, ptrdiff_t stride, | 33 | void ff_h264_v_loop_filter_luma_8_rvv(uint8_t *pix, ptrdiff_t stride, |
| 31 | int alpha, int beta, int8_t *tc0); | 34 | int alpha, int beta, int8_t *tc0); |
| 32 | void ff_h264_h_loop_filter_luma_8_rvv(uint8_t *pix, ptrdiff_t stride, | 35 | void ff_h264_h_loop_filter_luma_8_rvv(uint8_t *pix, ptrdiff_t stride, |
| @@ -60,6 +63,10 @@ av_cold void ff_h264dsp_init_riscv(H264DSPContext *dsp, const int bit_depth, · 27 unmodified lines | |||
| 60 | # if HAVE_RVV | 63 | # if HAVE_RVV |
| 61 | if (flags & AV_CPU_FLAG_RVV_I32) { | 64 | if (flags & AV_CPU_FLAG_RVV_I32) { |
| 62 | if (bit_depth == 8 && ff_rv_vlen_least(128)) { | 65 | if (bit_depth == 8 && ff_rv_vlen_least(128)) { |
| 66 | memcpy(dsp->weight_h264_pixels_tab, | ||
| 67 | ff_h264_weight_funcs_8_rvv, | ||
| 68 | sizeof (dsp->weight_h264_pixels_tab)); | ||
| 69 | |||
| 63 | dsp->h264_v_loop_filter_luma = ff_h264_v_loop_filter_luma_8_rvv; | 70 | dsp->h264_v_loop_filter_luma = ff_h264_v_loop_filter_luma_8_rvv; |
| 64 | dsp->h264_h_loop_filter_luma = ff_h264_h_loop_filter_luma_8_rvv; | 71 | dsp->h264_h_loop_filter_luma = ff_h264_h_loop_filter_luma_8_rvv; |
| 65 | dsp->h264_h_loop_filter_luma_mbaff = | 72 | dsp->h264_h_loop_filter_luma_mbaff = |
| @@ -26,6 +26,89 @@ · 25 unmodified lines | |||
| 26 | 26 | ||
| 27 | #include "libavutil/riscv/asm.S" | 27 | #include "libavutil/riscv/asm.S" |
| 28 | 28 | ||
| 29 | func ff_h264_weight_pixels_simple_8_rvv, zve32x | ||
| 30 | csrwi vxrm, 0 | ||
| 31 | sll a5, a5, a3 | ||
| 32 | 1: | ||
| 33 | vsetvli zero, a6, e32, m4, ta, ma | ||
| 34 | vle8.v v8, (a0) | ||
| 35 | addi a2, a2, -1 | ||
| 36 | vmv.v.x v16, a5 | ||
| 37 | vsetvli zero, zero, e16, m2, ta, ma | ||
| 38 | vzext.vf2 v24, v8 | ||
| 39 | vwmaccsu.vx v16, a4, v24 | ||
| 40 | vnclip.wx v16, v16, a3 | ||
| 41 | vmax.vx v16, v16, zero | ||
| 42 | vsetvli zero, zero, e8, m1, ta, ma | ||
| 43 | vnclipu.wi v8, v16, 0 | ||
| 44 | vse8.v v8, (a0) | ||
| 45 | add a0, a0, a1 | ||
| 46 | bnez a2, 1b | ||
| 47 | |||
| 48 | ret | ||
| 49 | endfunc | ||
| 50 | |||
| 51 | func ff_h264_weight_pixels_8_rvv, zve32x | ||
| 52 | csrwi vxrm, 0 | ||
| 53 | sll a5, a5, a3 | ||
| 54 | 1: | ||
| 55 | mv t0, a0 | ||
| 56 | mv t6, a6 | ||
| 57 | 2: | ||
| 58 | vsetvli t2, a2, e32, m8, ta, ma | ||
| 59 | vlsseg2e8.v v0, (t0), a1 | ||
| 60 | addi t6, t6, -2 | ||
| 61 | vmv.v.x v16, a5 | ||
| 62 | vmv.v.x v24, a5 | ||
| 63 | vsetvli zero, zero, e16, m4, ta, ma | ||
| 64 | vzext.vf2 v8, v0 | ||
| 65 | vzext.vf2 v12, v2 | ||
| 66 | vwmaccsu.vx v16, a4, v8 | ||
| 67 | vwmaccsu.vx v24, a4, v12 | ||
| 68 | vnclip.wx v8, v16, a3 | ||
| 69 | vnclip.wx v12, v24, a3 | ||
| 70 | vmax.vx v8, v8, zero | ||
| 71 | vmax.vx v12, v12, zero | ||
| 72 | vsetvli zero, zero, e8, m2, ta, ma | ||
| 73 | vnclipu.wi v0, v8, 0 | ||
| 74 | vnclipu.wi v2, v12, 0 | ||
| 75 | vssseg2e8.v v0, (t0), a1 | ||
| 76 | addi t0, t0, 2 | ||
| 77 | bnez t6, 2b | ||
| 78 | |||
| 79 | mul t3, a1, t2 | ||
| 80 | sub a2, a2, t2 | ||
| 81 | add a0, a0, t3 | ||
| 82 | bnez a2, 1b | ||
| 83 | |||
| 84 | ret | ||
| 85 | endfunc | ||
| 86 | |||
| 87 | .irp w, 16, 8, 4, 2 | ||
| 88 | func ff_h264_weight_pixels\w\()_8_rvv, zve32x | ||
| 89 | li a6, \w | ||
| 90 | .if \w == 16 | ||
| 91 | j ff_h264_weight_pixels_simple_8_rvv | ||
| 92 | .else | ||
| 93 | j ff_h264_weight_pixels_8_rvv | ||
| 94 | .endif | ||
| 95 | endfunc | ||
| 96 | .endr | ||
| 97 | |||
| 98 | .global ff_h264_weight_funcs_8_rvv | ||
| 99 | .hidden ff_h264_weight_funcs_8_rvv | ||
| 100 | const ff_h264_weight_funcs_8_rvv | ||
| 101 | .irp w, 16, 8, 4, 2 | ||
| 102 | #if __riscv_xlen == 32 | ||
| 103 | .word ff_h264_weight_pixels\w\()_8_rvv | ||
| 104 | #elif __riscv_xlen == 64 | ||
| 105 | .dword ff_h264_weight_pixels\w\()_8_rvv | ||
| 106 | #else | ||
| 107 | .qword ff_h264_weight_pixels\w\()_8_rvv | ||
| 108 | #endif | ||
| 109 | .endr | ||
| 110 | endconst | ||
| 111 | |||
| 29 | .variant_cc ff_h264_loop_filter_luma_8_rvv | 112 | .variant_cc ff_h264_loop_filter_luma_8_rvv |
| 30 | func ff_h264_loop_filter_luma_8_rvv, zve32x | 113 | func ff_h264_loop_filter_luma_8_rvv, zve32x |
| 31 | # p2: v8, p1: v9, p0: v10, q0: v11, q1: v12, q2: v13 | 114 | # p2: v8, p1: v9, p0: v10, q0: v11, q1: v12, q2: v13 |