From 7ed753b10a61d0be95f683289dfb925b800b0676 Mon Sep 17 00:00:00 2001
From: Xiwei Gu <guxiwei-hf@loongson.cn>
Date: Tue, 5 Mar 2024 14:35:43 +0800
Subject: [PATCH] loongarch: Enhance ultrafast encoding performance
Using the following command, ultrafast encoding
has improved from 182fps to 189fps:
./x264 --preset ultrafast -o out.mkv yuv_1920x1080.yuv
---
common/loongarch/quant-a.S | 245 +++++++++++++++++++++++++++++++++++++
common/loongarch/quant.h | 12 ++
common/quant.c | 12 +-
3 files changed, 267 insertions(+), 2 deletions(-)
diff --git a/common/loongarch/quant-a.S b/common/loongarch/quant-a.S
index 279c83b6..a2f099d9 100644
--- a/common/loongarch/quant-a.S
+++ b/common/loongarch/quant-a.S
@@ -984,3 +984,248 @@ function_x264 decimate_score64_lsx
jirl $r0, $r1, 0x0
.END_SCORE_64_LSX:
endfunc_x264
+
+/*
+ * int coeff_level_run16( dctcoef *dct, x264_run_level_t *runlevel )
+ */
+function_x264 coeff_level_run16_lasx
+ addi.w t0, zero, 15
+
+ xvld xr0, a0, 0
+ xvldi xr2, 1
+
+ xvssrlni.bu.h xr0, xr0, 0
+ xvpermi.d xr1, xr0, 0xd8
+ xvsle.bu xr3, xr2, xr1
+ xvsrlni.b.h xr3, xr3, 4
+ xvpickve2gr.du t8, xr3, 0
+ clz.d t1, t8
+
+ srai.w t1, t1, 2
+ sub.w t0, t0, t1 // Index of the first non-zero element starting from the highest bit
+ st.w t0, a1, 0x00 // Store runlevel->last
+ addi.d t3, a1, 23
+ nor t2, zero, zero
+ addi.d t2, t2, -15
+ and t3, t3, t2 // runlevel->level
+ xor t4, t4, t4 // mask
+ xor t5, t5, t5 // total: number of non-zero elements
+ addi.w t6, zero, 1 // const 1
+.LOOP_COEFF_LEVEL_RUN16_LASX:
+ slli.w t7, t0, 1
+ ldx.h t2, a0, t7
+ st.h t2, t3, 0
+ addi.d t3, t3, 2
+
+ addi.w t5, t5, 1
+ sll.w t2, t6, t0
+ or t4, t4, t2
+ bge zero, t4, .END_COEFF_LEVEL_RUN16_LASX
+
+ addi.w t0, t0, -1
+ slli.w t1, t1, 2
+ addi.w t1, t1, 4
+ sll.d t8, t8, t1
+ clz.d t1, t8
+ srai.w t1, t1, 2
+ sub.w t0, t0, t1 // Index of the first non-zero element starting from the highest bit
+ bge t0, zero, .LOOP_COEFF_LEVEL_RUN16_LASX
+.END_COEFF_LEVEL_RUN16_LASX:
+ st.w t4, a1, 4
+ move a0, t5
+endfunc_x264
+
+function_x264 coeff_level_run15_lasx
+ addi.w t0, zero, 15
+
+ vld vr0, a0, 0
+ vld vr1, a0, 16
+ xvldi xr3, 1
+
+ vinsgr2vr.h vr1, zero, 7
+ xvpermi.q xr1, xr0, 0x20
+
+ xvssrlni.bu.h xr1, xr1, 0
+ xvpermi.d xr2, xr1, 0xd8
+ xvsle.bu xr4, xr3, xr2
+ xvsrlni.b.h xr4, xr4, 4
+ xvpickve2gr.du t8, xr4, 0
+ clz.d t1, t8
+
+ srai.w t1, t1, 2
+ sub.w t0, t0, t1 // Index of the first non-zero element starting from the highest bit
+ st.w t0, a1, 0x00 // Store runlevel->last
+ addi.d t3, a1, 23
+ nor t2, zero, zero
+ addi.d t2, t2, -15
+ and t3, t3, t2 // runlevel->level
+ xor t4, t4, t4 // mask
+ xor t5, t5, t5 // total: number of non-zero elements
+ addi.w t6, zero, 1 // const 1
+.LOOP_COEFF_LEVEL_RUN15_LASX:
+ slli.w t7, t0, 1
+ ldx.h t2, a0, t7
+ st.h t2, t3, 0
+ addi.d t3, t3, 2
+
+ addi.w t5, t5, 1
+ sll.w t2, t6, t0
+ or t4, t4, t2
+ bge zero, t4, .END_COEFF_LEVEL_RUN15_LASX
+
+ addi.w t0, t0, -1
+ slli.w t1, t1, 2
+ addi.w t1, t1, 4
+ sll.d t8, t8, t1
+ clz.d t1, t8
+ srai.w t1, t1, 2
+ sub.w t0, t0, t1 // Index of the first non-zero element starting from the highest bit
+ bge t0, zero, .LOOP_COEFF_LEVEL_RUN15_LASX
+.END_COEFF_LEVEL_RUN15_LASX:
+ st.w t4, a1, 4
+ move a0, t5
+endfunc_x264
+
+function_x264 coeff_level_run16_lsx
+ addi.w t0, zero, 15
+ vld vr0, a0, 0
+ vld vr1, a0, 16
+ vldi vr2, 1
+
+ vssrlni.bu.h vr0, vr0, 0
+ vssrlni.bu.h vr1, vr1, 0
+ vpermi.w vr1, vr0, 0x44
+ vsle.bu vr3, vr2, vr1
+ vsrlni.b.h vr3, vr3, 4
+ vpickve2gr.du t8, vr3, 0
+ clz.d t1, t8
+
+ srai.w t1, t1, 2
+ sub.w t0, t0, t1 // Index of the first non-zero element starting from the highest bit
+ st.w t0, a1, 0x00 // Store runlevel->last
+ addi.d t3, a1, 23
+ nor t2, zero, zero
+ addi.d t2, t2, -15
+ and t3, t3, t2 // runlevel->level
+ xor t4, t4, t4 // mask
+ xor t5, t5, t5 // total: number of non-zero elements
+ addi.w t6, zero, 1 // const 1
+.LOOP_COEFF_LEVEL_RUN16_LSX:
+ slli.w t7, t0, 1
+ ldx.h t2, a0, t7
+ st.h t2, t3, 0
+ addi.d t3, t3, 2
+
+ addi.w t5, t5, 1
+ sll.w t2, t6, t0
+ or t4, t4, t2
+ bge zero, t4, .END_COEFF_LEVEL_RUN16_LSX
+
+ addi.w t0, t0, -1
+ slli.w t1, t1, 2
+ addi.w t1, t1, 4
+ sll.d t8, t8, t1
+ clz.d t1, t8
+ srai.w t1, t1, 2
+ sub.w t0, t0, t1 // Index of the first non-zero element starting from the highest bit
+ bge t0, zero, .LOOP_COEFF_LEVEL_RUN16_LSX
+.END_COEFF_LEVEL_RUN16_LSX:
+ st.w t4, a1, 4
+ move a0, t5
+endfunc_x264
+
+function_x264 coeff_level_run15_lsx
+ addi.w t0, zero, 15
+ vld vr0, a0, 0
+ vld vr1, a0, 16
+ vldi vr2, 1
+ vinsgr2vr.h vr1, zero, 7
+
+ vssrlni.bu.h vr0, vr0, 0
+ vssrlni.bu.h vr1, vr1, 0
+ vpermi.w vr1, vr0, 0x44
+ vsle.bu vr3, vr2, vr1
+ vsrlni.b.h vr3, vr3, 4
+ vpickve2gr.du t8, vr3, 0
+ clz.d t1, t8
+
+ srai.w t1, t1, 2
+ sub.w t0, t0, t1 // Index of the first non-zero element starting from the highest bit
+ st.w t0, a1, 0x00 // Store runlevel->last
+ addi.d t3, a1, 23
+ nor t2, zero, zero
+ addi.d t2, t2, -15
+ and t3, t3, t2 // runlevel->level
+ xor t4, t4, t4 // mask
+ xor t5, t5, t5 // total: number of non-zero elements
+ addi.w t6, zero, 1 // const 1
+.LOOP_COEFF_LEVEL_RUN15_LSX:
+ slli.w t7, t0, 1
+ ldx.h t2, a0, t7
+ st.h t2, t3, 0
+ addi.d t3, t3, 2
+
+ addi.w t5, t5, 1
+ sll.w t2, t6, t0
+ or t4, t4, t2
+ bge zero, t4, .END_COEFF_LEVEL_RUN15_LSX
+
+ addi.w t0, t0, -1
+ slli.w t1, t1, 2
+ addi.w t1, t1, 4
+ sll.d t8, t8, t1
+ clz.d t1, t8
+ srai.w t1, t1, 2
+ sub.w t0, t0, t1 // Index of the first non-zero element starting from the highest bit
+ bge t0, zero, .LOOP_COEFF_LEVEL_RUN15_LSX
+.END_COEFF_LEVEL_RUN15_LSX:
+ st.w t4, a1, 4
+ move a0, t5
+endfunc_x264
+
+function_x264 coeff_level_run8_lsx
+ addi.w t0, zero, 15
+ vld vr0, a0, 0
+ vxor.v vr1, vr1, vr1
+ vldi vr2, 1
+
+ vssrlni.bu.h vr0, vr0, 0
+ vpermi.w vr1, vr0, 0x44
+ vsle.bu vr3, vr2, vr1
+ vsrlni.b.h vr3, vr3, 4
+ vpickve2gr.du t8, vr3, 0
+ clz.d t1, t8
+
+ srai.w t1, t1, 2
+ sub.w t0, t0, t1 // Index of the first non-zero element starting from the highest bit
+ st.w t0, a1, 0x00 // Store runlevel->last
+ addi.d t3, a1, 23
+ nor t2, zero, zero
+ addi.d t2, t2, -15
+ and t3, t3, t2 // runlevel->level
+ xor t4, t4, t4 // mask
+ xor t5, t5, t5 // total: number of non-zero elements
+ addi.w t6, zero, 1 // const 1
+.LOOP_COEFF_LEVEL_RUN8_LSX:
+ slli.w t7, t0, 1
+ ldx.h t2, a0, t7
+ st.h t2, t3, 0
+ addi.d t3, t3, 2
+
+ addi.w t5, t5, 1
+ sll.w t2, t6, t0
+ or t4, t4, t2
+ bge zero, t4, .END_COEFF_LEVEL_RUN8_LSX
+
+ addi.w t0, t0, -1
+ slli.w t1, t1, 2
+ addi.w t1, t1, 4
+ sll.d t8, t8, t1
+ clz.d t1, t8
+ srai.w t1, t1, 2
+ sub.w t0, t0, t1 // Index of the first non-zero element starting from the highest bit
+ bge t0, zero, .LOOP_COEFF_LEVEL_RUN8_LSX
+.END_COEFF_LEVEL_RUN8_LSX:
+ st.w t4, a1, 4
+ move a0, t5
+endfunc_x264
diff --git a/common/loongarch/quant.h b/common/loongarch/quant.h
index 36d88b4b..cc3a5399 100644
--- a/common/loongarch/quant.h
+++ b/common/loongarch/quant.h
@@ -81,4 +81,16 @@ void x264_dequant_8x8_lasx( dctcoef dct[64], int dequant_mf[6][64], int i_qp );
#define x264_dequant_4x4_dc_lasx x264_template(dequant_4x4_dc_lasx)
void x264_dequant_4x4_dc_lasx( dctcoef dct[16], int dequant_mf[6][16], int i_qp );
+#define x264_coeff_level_run16_lasx x264_template(coeff_level_run16_lasx)
+int x264_coeff_level_run16_lasx( dctcoef *, x264_run_level_t * );
+#define x264_coeff_level_run15_lasx x264_template(coeff_level_run15_lasx)
+int x264_coeff_level_run15_lasx( dctcoef *, x264_run_level_t * );
+
+#define x264_coeff_level_run16_lsx x264_template(coeff_level_run16_lsx)
+int x264_coeff_level_run16_lsx( dctcoef *, x264_run_level_t * );
+#define x264_coeff_level_run15_lsx x264_template(coeff_level_run15_lsx)
+int x264_coeff_level_run15_lsx( dctcoef *, x264_run_level_t * );
+#define x264_coeff_level_run8_lsx x264_template(coeff_level_run8_lsx)
+int x264_coeff_level_run8_lsx( dctcoef *, x264_run_level_t * );
+
#endif/* X264_LOONGARCH_QUANT_H */
diff --git a/common/quant.c b/common/quant.c
index 64e33216..262c5c52 100644
--- a/common/quant.c
+++ b/common/quant.c
@@ -848,11 +848,17 @@ void x264_quant_init( x264_t *h, uint32_t cpu, x264_quant_function_t *pf )
pf->dequant_4x4 = x264_dequant_4x4_lsx;
pf->dequant_8x8 = x264_dequant_8x8_lsx;
pf->dequant_4x4_dc = x264_dequant_4x4_dc_lsx;
- pf->coeff_last4 = x264_coeff_last4_lsx;
- pf->coeff_last8 = x264_coeff_last8_lsx;
+ pf->decimate_score15 = x264_decimate_score15_lsx;
+ pf->decimate_score16 = x264_decimate_score16_lsx;
+ pf->decimate_score64 = x264_decimate_score64_lsx;
+ pf->coeff_last4 = x264_coeff_last4_lsx;
+ pf->coeff_last8 = x264_coeff_last8_lsx;
pf->coeff_last[ DCT_LUMA_AC] = x264_coeff_last15_lsx;
pf->coeff_last[DCT_LUMA_4x4] = x264_coeff_last16_lsx;
pf->coeff_last[DCT_LUMA_8x8] = x264_coeff_last64_lsx;
+ pf->coeff_level_run8 = x264_coeff_level_run8_lsx;
+ pf->coeff_level_run[ DCT_LUMA_AC] = x264_coeff_level_run15_lsx;
+ pf->coeff_level_run[DCT_LUMA_4x4] = x264_coeff_level_run16_lsx;
}
if( cpu&X264_CPU_LASX )
{
@@ -863,6 +869,8 @@ void x264_quant_init( x264_t *h, uint32_t cpu, x264_quant_function_t *pf )
pf->coeff_last[ DCT_LUMA_AC] = x264_coeff_last15_lasx;
pf->coeff_last[DCT_LUMA_4x4] = x264_coeff_last16_lasx;
pf->coeff_last[DCT_LUMA_8x8] = x264_coeff_last64_lasx;
+ pf->coeff_level_run[ DCT_LUMA_AC] = x264_coeff_level_run15_lasx;
+ pf->coeff_level_run[DCT_LUMA_4x4] = x264_coeff_level_run16_lasx;
}
#endif
--
GitLab