[FFmpeg-devel] [PATCH 6/6] avcodec/hevcdsp: Add NEON optimization for idct16x16

Shengbin Meng shengbinmeng at gmail.com
Wed Nov 22 13:12:06 EET 2017


From: Meng Wang <wangmeng.kids at bytedance.com>

Signed-off-by: Meng Wang <wangmeng.kids at bytedance.com>
---
 libavcodec/arm/hevcdsp_idct_neon.S | 241 +++++++++++++++++++++++++++++++++++++
 libavcodec/arm/hevcdsp_init_neon.c |   2 +
 2 files changed, 243 insertions(+)

diff --git a/libavcodec/arm/hevcdsp_idct_neon.S b/libavcodec/arm/hevcdsp_idct_neon.S
index e39d00634b..272abf279c 100644
--- a/libavcodec/arm/hevcdsp_idct_neon.S
+++ b/libavcodec/arm/hevcdsp_idct_neon.S
@@ -451,6 +451,247 @@ function ff_hevc_transform_8x8_neon_8, export=1
         bx lr
 endfunc
 
+/* 16x16 even line combine, input: q3-q10  output: q8-q15 */
+.macro tr8_combine
+        vsub.s32   q12, q3, q10  // e_8[3] - o_8[3], dst[4]
+        vadd.s32   q11, q3, q10  // e_8[3] + o_8[3], dst[3]
+
+        vsub.s32   q13, q6, q9   // e_8[2] - o_8[2], dst[5]
+        vadd.s32   q10, q6, q9   // e_8[2] + o_8[2], dst[2]
+
+        vsub.s32   q14, q5, q8   // e_8[1] - o_8[1], dst[6]
+        vadd.s32   q9,  q5, q8   // e_8[1] + o_8[1], dst[1]
+
+        vsub.s32   q15, q4, q7   // e_8[0] - o_8[0], dst[7]
+        vadd.s32   q8,  q4, q7   // e_8[0] + o_8[0], dst[0]
+.endm
+
+.macro tr16_begin in0, in1, in2, in3, in4, in5, in6, in7
+        vmull.s16  q2, \in0, d2[1]   // 90 * src1
+        vmull.s16  q3, \in0, d2[0]   // 87 * src1
+        vmull.s16  q4, \in0, d2[3]   // 80 * src1
+        vmull.s16  q5, \in0, d2[2]   // 70 * src1
+        vmull.s16  q6, \in0, d3[1]   // 57 * src1
+        vmull.s16  q7, \in0, d3[0]   // 43 * src1
+        vmull.s16  q8, \in0, d3[3]   // 25 * src1
+        vmull.s16  q9, \in0, d3[2]   //  9 * src1
+
+        vmlal.s16  q2, \in1, d2[0]   // 87 * src3
+        vmlal.s16  q3, \in1, d3[1]   // 57 * src3
+        vmlal.s16  q4, \in1, d3[2]   //  9 * src3
+        vmlsl.s16  q5, \in1, d3[0]   //-43 * src3
+        vmlsl.s16  q6, \in1, d2[3]   //-80 * src3
+        vmlsl.s16  q7, \in1, d2[1]   //-90 * src3
+        vmlsl.s16  q8, \in1, d2[2]   //-70 * src3
+        vmlsl.s16  q9, \in1, d3[3]   //-25 * src3
+
+        vmlal.s16  q2, \in2, d2[3]   // 80 * src5
+        vmlal.s16  q3, \in2, d3[2]   //  9 * src5
+        vmlsl.s16  q4, \in2, d2[2]   //-70 * src5
+        vmlsl.s16  q5, \in2, d2[0]   //-87 * src5
+        vmlsl.s16  q6, \in2, d3[3]   //-25 * src5
+        vmlal.s16  q7, \in2, d3[1]   // 57 * src5
+        vmlal.s16  q8, \in2, d2[1]   // 90 * src5
+        vmlal.s16  q9, \in2, d3[0]   // 43 * src5
+
+        vmlal.s16  q2, \in3, d2[2]   // 70 * src7
+        vmlsl.s16  q3, \in3, d3[0]   //-43 * src7
+        vmlsl.s16  q4, \in3, d2[0]   //-87 * src7
+        vmlal.s16  q5, \in3, d3[2]   //  9 * src7
+        vmlal.s16  q6, \in3, d2[1]   // 90 * src7
+        vmlal.s16  q7, \in3, d3[3]   // 25 * src7
+        vmlsl.s16  q8, \in3, d2[3]   //-80 * src7
+        vmlsl.s16  q9, \in3, d3[1]   //-57 * src7
+
+        vmlal.s16  q2, \in4, d3[1]   // 57 * src9
+        vmlsl.s16  q3, \in4, d2[3]   //-80 * src9
+        vmlsl.s16  q4, \in4, d3[3]   //-25 * src9
+        vmlal.s16  q5, \in4, d2[1]   // 90 * src9
+        vmlsl.s16  q6, \in4, d3[2]   // -9 * src9
+        vmlsl.s16  q7, \in4, d2[0]   //-87 * src9
+        vmlal.s16  q8, \in4, d3[0]   // 43 * src9
+        vmlal.s16  q9, \in4, d2[2]   // 70 * src9
+
+        vmlal.s16  q2, \in5, d3[0]   // 43 * src11
+        vmlsl.s16  q3, \in5, d2[1]   //-90 * src11
+        vmlal.s16  q4, \in5, d3[1]   // 57 * src11
+        vmlal.s16  q5, \in5, d3[3]   // 25 * src11
+        vmlsl.s16  q6, \in5, d2[0]   //-87 * src11
+        vmlal.s16  q7, \in5, d2[2]   // 70 * src11
+        vmlal.s16  q8, \in5, d3[2]   //  9 * src11
+        vmlsl.s16  q9, \in5, d2[3]   //-80 * src11
+
+        vmlal.s16  q2, \in6, d3[3]   // 25 * src13
+        vmlsl.s16  q3, \in6, d2[2]   //-70 * src13
+        vmlal.s16  q4, \in6, d2[1]   // 90 * src13
+        vmlsl.s16  q5, \in6, d2[3]   //-80 * src13
+        vmlal.s16  q6, \in6, d3[0]   // 43 * src13
+        vmlal.s16  q7, \in6, d3[2]   //  9 * src13
+        vmlsl.s16  q8, \in6, d3[1]   //-57 * src13
+        vmlal.s16  q9, \in6, d2[0]   // 87 * src13
+
+
+        vmlal.s16  q2, \in7, d3[2]   //  9 * src15
+        vmlsl.s16  q3, \in7, d3[3]   //-25 * src15
+        vmlal.s16  q4, \in7, d3[0]   // 43 * src15
+        vmlsl.s16  q5, \in7, d3[1]   //-57 * src15
+        vmlal.s16  q6, \in7, d2[2]   // 70 * src15
+        vmlsl.s16  q7, \in7, d2[3]   //-80 * src15
+        vmlal.s16  q8, \in7, d2[0]   // 87 * src15
+        vmlsl.s16  q9, \in7, d2[1]   //-90 * src15
+.endm
+
+.macro tr16_end shift
+        vpop       {q2-q3}
+        vadd.s32   q4, q8,  q2
+        vsub.s32   q5, q8,  q2
+        vqrshrn.s32    d12, q4, \shift
+        vqrshrn.s32    d15, q5, \shift
+
+        vadd.s32   q4, q9,  q3
+        vsub.s32   q5, q9,  q3
+        vqrshrn.s32    d13, q4, \shift
+        vqrshrn.s32    d14, q5, \shift
+
+        vpop       {q2-q3}
+        vadd.s32   q4, q10, q2
+        vsub.s32   q5, q10, q2
+        vqrshrn.s32    d16, q4, \shift
+        vqrshrn.s32    d19, q5, \shift
+
+        vadd.s32   q4, q11, q3
+        vsub.s32   q5, q11, q3
+        vqrshrn.s32    d17, q4, \shift
+        vqrshrn.s32    d18, q5, \shift
+
+        vpop       {q2-q3}
+        vadd.s32   q4, q12, q2
+        vsub.s32   q5, q12, q2
+        vqrshrn.s32    d20, q4, \shift
+        vqrshrn.s32    d23, q5, \shift
+
+        vadd.s32   q4, q13, q3
+        vsub.s32   q5, q13, q3
+        vqrshrn.s32    d21, q4, \shift
+        vqrshrn.s32    d22, q5, \shift
+
+        vpop       {q2-q3}
+        vadd.s32   q4, q14, q2
+        vsub.s32   q5, q14, q2
+        vqrshrn.s32    d24, q4, \shift
+        vqrshrn.s32    d27, q5, \shift
+
+        vadd.s32   q4, q15, q3
+        vsub.s32   q5, q15, q3
+        vqrshrn.s32    d25, q4, \shift
+        vqrshrn.s32    d26, q5, \shift
+.endm
+
+function ff_hevc_transform_16x16_neon_8, export=1
+        push   {r4-r8}
+        vpush  {d8-d15}
+        mov    r5, #64
+        mov    r6, #32
+        mov    r7, #0
+        adr       r3, tr4f
+        vld1.16   {d0, d1, d2, d3}, [r3]
+        mov    r8, r0
+0:
+        add      r7, #4
+        add      r0, #32
+        // odd line
+        vld1.16 {d24}, [r0], r5
+        vld1.16 {d25}, [r0], r5
+        vld1.16 {d26}, [r0], r5
+        vld1.16 {d27}, [r0], r5
+        vld1.16 {d28}, [r0], r5
+        vld1.16 {d29}, [r0], r5
+        vld1.16 {d30}, [r0], r5
+        vld1.16 {d31}, [r0], r5
+        sub      r0, #544
+
+        tr16_begin d24, d25, d26, d27, d28, d29, d30, d31
+        vpush  {q2-q9}
+
+        // even line
+        vld1.16 {d24}, [r0], r5
+        vld1.16 {d25}, [r0], r5
+        vld1.16 {d26}, [r0], r5
+        vld1.16 {d27}, [r0], r5
+        vld1.16 {d28}, [r0], r5
+        vld1.16 {d29}, [r0], r5
+        vld1.16 {d30}, [r0], r5
+        vld1.16 {d31}, [r0], r5
+        sub      r0, #512
+
+        tr8_begin d25, d27, d29, d31
+        tr4       d24, d26, d28, d30
+        tr8_combine
+
+        // combine
+        tr16_end        #7
+
+        // store
+        vst1.16 {d12}, [r0], r6
+        vst1.16 {d13}, [r0], r6
+        vst1.16 {d16}, [r0], r6
+        vst1.16 {d17}, [r0], r6
+        vst1.16 {d20}, [r0], r6
+        vst1.16 {d21}, [r0], r6
+        vst1.16 {d24}, [r0], r6
+        vst1.16 {d25}, [r0], r6
+        vst1.16 {d26}, [r0], r6
+        vst1.16 {d27}, [r0], r6
+        vst1.16 {d22}, [r0], r6
+        vst1.16 {d23}, [r0], r6
+        vst1.16 {d18}, [r0], r6
+        vst1.16 {d19}, [r0], r6
+        vst1.16 {d14}, [r0], r6
+        vst1.16 {d15}, [r0], r6
+        sub      r0,    #504  // 512 - 8
+
+        cmp      r1, r7
+        blt      1f
+
+        cmp      r7, #16
+        blt      0b
+
+1:      mov r0, r8
+        mov       r7, #4
+2:      subs      r7, #1
+        // 1st 4 line
+        vldm r0, {q8-q15}                                    // coeffs
+        transpose_16b_4x4 d16, d20, d24, d28
+        transpose_16b_4x4 d17, d21, d25, d29
+        transpose_16b_4x4 d18, d22, d26, d30
+        transpose_16b_4x4 d19, d23, d27, d31
+        vpush  {q12-q13}                                     // 16x16 even line (8x8 odd  line)
+        vpush  {q8-q9}                                       // 16x16 even line (8x8 even line)
+        tr16_begin d20, d28, d21, d29, d22, d30, d23, d31    // odd  line transform 2n+1
+        vpop   {q12-q15}                                     // pop  even line
+        vpush  {q2-q9}                                       // push results of 16x16 odd line
+        tr8_begin d28, d29, d30, d31                         // even line transform 2n
+        tr4       d24, d25, d26, d27
+        tr8_combine
+        tr16_end        #12
+        transpose_16b_4x4 d12, d13, d16, d17
+        transpose_16b_4x4 d20, d21, d24, d25
+        transpose_16b_4x4 d26, d27, d22, d23
+        transpose_16b_4x4 d18, d19, d14, d15
+        vswp     d13, d20
+        vswp     d14, d23
+        vswp     d17, d24
+        vswp     d18, d27
+        vswp     q8, q10
+        vswp     q7, q13
+        vstm r0!, {q6-q13}
+        bne      2b
+
+        vpop {d8-d15}
+        pop {r4-r8}
+        bx  lr
+endfunc
+
 .align 4
 tr4f:
 .word 0x00240053  // 36 and d1[0] = 83
diff --git a/libavcodec/arm/hevcdsp_init_neon.c b/libavcodec/arm/hevcdsp_init_neon.c
index 33cc44ef40..d846d01081 100644
--- a/libavcodec/arm/hevcdsp_init_neon.c
+++ b/libavcodec/arm/hevcdsp_init_neon.c
@@ -36,6 +36,7 @@ void ff_hevc_v_loop_filter_chroma_neon(uint8_t *_pix, ptrdiff_t _stride, int *_t
 void ff_hevc_h_loop_filter_chroma_neon(uint8_t *_pix, ptrdiff_t _stride, int *_tc, uint8_t *_no_p, uint8_t *_no_q);
 void ff_hevc_transform_4x4_neon_8(int16_t *coeffs, int col_limit);
 void ff_hevc_transform_8x8_neon_8(int16_t *coeffs, int col_limit);
+void ff_hevc_transform_16x16_neon_8(int16_t *coeffs, int col_limit);
 void ff_hevc_idct_4x4_dc_neon_8(int16_t *coeffs);
 void ff_hevc_idct_8x8_dc_neon_8(int16_t *coeffs);
 void ff_hevc_idct_16x16_dc_neon_8(int16_t *coeffs);
@@ -550,6 +551,7 @@ av_cold void ff_hevcdsp_init_neon(HEVCDSPContext *c, const int bit_depth)
         c->sao_edge_filter[4]          = ff_hevc_sao_edge_filter_neon_wrapper;
         c->idct[0]                     = ff_hevc_transform_4x4_neon_8;
         c->idct[1]                     = ff_hevc_transform_8x8_neon_8;
+        c->idct[2]                     = ff_hevc_transform_16x16_neon_8;
         c->idct_dc[0]                  = ff_hevc_idct_4x4_dc_neon_8;
         c->idct_dc[1]                  = ff_hevc_idct_8x8_dc_neon_8;
         c->idct_dc[2]                  = ff_hevc_idct_16x16_dc_neon_8;
-- 
2.13.6 (Apple Git-96)



More information about the ffmpeg-devel mailing list