arm: vp9mc: Minor adjustments from review of the aarch64 version

This work is sponsored by, and copyright, Google.

The speedup for the large horizontal filters is surprisingly
big on A7 and A53, while there's a minor slowdown (almost within
measurement noise) on A8 and A9.

                            Cortex    A7        A8        A9       A53
orig:
vp9_put_8tap_smooth_64h_neon:    20270.0   14447.3   19723.9   10910.9
new:
vp9_put_8tap_smooth_64h_neon:    20165.8   14466.5   19730.2   10668.8

Signed-off-by: Martin Storsjö <martin@martin.st>

2 files changed, 44 insertions, 91 deletions

diff --git a/libavcodec/arm/vp9dsp_init_arm.c b/libavcodec/arm/vp9dsp_init_arm.c
index 1b00177..839037a 100644
--- a/libavcodec/arm/vp9dsp_init_arm.c
+++ b/libavcodec/arm/vp9dsp_init_arm.c
@@ -43,7 +43,7 @@ static void op##_##filter##sz##_hv_neon(uint8_t *dst, ptrdiff_t dst_stride,
                                         const uint8_t *src, ptrdiff_t src_stride, \
                                         int h, int mx, int my)                    \
 {                                                                                 \
-    LOCAL_ALIGNED_16(uint8_t, temp, [((sz < 64 ? 2 * sz : 64) + 8) * sz]);        \
+    LOCAL_ALIGNED_16(uint8_t, temp, [((1 + (sz < 64)) * sz + 8) * sz]);           \
     /* We only need h + 7 lines, but the horizontal filter assumes an             \
      * even number of rows, so filter h + 8 lines here. */                        \
     ff_vp9_put_##filter##sz##_h_neon(temp, sz,                                    \
diff --git a/libavcodec/arm/vp9mc_neon.S b/libavcodec/arm/vp9mc_neon.S
index cc8f241..9deb656 100644
--- a/libavcodec/arm/vp9mc_neon.S
+++ b/libavcodec/arm/vp9mc_neon.S
@@ -20,60 +20,6 @@
 
 #include "libavutil/arm/asm.S"
 
-const regular_filter, align=4
-        .short  0,  1,  -5, 126,   8,  -3,  1,  0
-        .short -1,  3, -10, 122,  18,  -6,  2,  0
-        .short -1,  4, -13, 118,  27,  -9,  3, -1
-        .short -1,  4, -16, 112,  37, -11,  4, -1
-        .short -1,  5, -18, 105,  48, -14,  4, -1
-        .short -1,  5, -19,  97,  58, -16,  5, -1
-        .short -1,  6, -19,  88,  68, -18,  5, -1
-        .short -1,  6, -19,  78,  78, -19,  6, -1
-        .short -1,  5, -18,  68,  88, -19,  6, -1
-        .short -1,  5, -16,  58,  97, -19,  5, -1
-        .short -1,  4, -14,  48, 105, -18,  5, -1
-        .short -1,  4, -11,  37, 112, -16,  4, -1
-        .short -1,  3,  -9,  27, 118, -13,  4, -1
-        .short  0,  2,  -6,  18, 122, -10,  3, -1
-        .short  0,  1,  -3,   8, 126,  -5,  1,  0
-endconst
-
-const sharp_filter, align=4
-        .short -1,  3,  -7, 127,   8,  -3,  1,  0
-        .short -2,  5, -13, 125,  17,  -6,  3, -1
-        .short -3,  7, -17, 121,  27, -10,  5, -2
-        .short -4,  9, -20, 115,  37, -13,  6, -2
-        .short -4, 10, -23, 108,  48, -16,  8, -3
-        .short -4, 10, -24, 100,  59, -19,  9, -3
-        .short -4, 11, -24,  90,  70, -21, 10, -4
-        .short -4, 11, -23,  80,  80, -23, 11, -4
-        .short -4, 10, -21,  70,  90, -24, 11, -4
-        .short -3,  9, -19,  59, 100, -24, 10, -4
-        .short -3,  8, -16,  48, 108, -23, 10, -4
-        .short -2,  6, -13,  37, 115, -20,  9, -4
-        .short -2,  5, -10,  27, 121, -17,  7, -3
-        .short -1,  3,  -6,  17, 125, -13,  5, -2
-        .short  0,  1,  -3,   8, 127,  -7,  3, -1
-endconst
-
-const smooth_filter, align=4
-        .short -3, -1,  32,  64,  38,   1, -3,  0
-        .short -2, -2,  29,  63,  41,   2, -3,  0
-        .short -2, -2,  26,  63,  43,   4, -4,  0
-        .short -2, -3,  24,  62,  46,   5, -4,  0
-        .short -2, -3,  21,  60,  49,   7, -4,  0
-        .short -1, -4,  18,  59,  51,   9, -4,  0
-        .short -1, -4,  16,  57,  53,  12, -4, -1
-        .short -1, -4,  14,  55,  55,  14, -4, -1
-        .short -1, -4,  12,  53,  57,  16, -4, -1
-        .short  0, -4,   9,  51,  59,  18, -4, -1
-        .short  0, -4,   7,  49,  60,  21, -3, -2
-        .short  0, -4,   5,  46,  62,  24, -3, -2
-        .short  0, -4,   4,  43,  63,  26, -2, -2
-        .short  0, -3,   2,  41,  63,  29, -2, -2
-        .short  0, -3,   1,  38,  64,  32, -1, -3
-endconst
-
 @ All public functions in this file have the following signature:
 @ typedef void (*vp9_mc_func)(uint8_t *dst, ptrdiff_t dst_stride,
 @                            const uint8_t *ref, ptrdiff_t ref_stride,
@@ -156,20 +102,21 @@ function ff_vp9_copy16_neon, export=1
 endfunc
 
 function ff_vp9_avg16_neon, export=1
-        ldr             r12, [sp]
+        push            {lr}
+        ldr             r12, [sp, #4]
+        mov             lr,  r0
 1:
         vld1.8          {q2},  [r2], r3
         vld1.8          {q0},  [r0, :128], r1
         vld1.8          {q3},  [r2], r3
         vrhadd.u8       q0,  q0,  q2
-        vld1.8          {q1},  [r0, :128]
-        sub             r0,  r0,  r1
+        vld1.8          {q1},  [r0, :128], r1
         vrhadd.u8       q1,  q1,  q3
         subs            r12, r12, #2
-        vst1.8          {q0},  [r0, :128], r1
-        vst1.8          {q1},  [r0, :128], r1
+        vst1.8          {q0},  [lr, :128], r1
+        vst1.8          {q1},  [lr, :128], r1
         bne             1b
-        bx              lr
+        pop             {pc}
 endfunc
 
 function ff_vp9_copy8_neon, export=1
@@ -218,7 +165,9 @@ function ff_vp9_copy4_neon, export=1
 endfunc
 
 function ff_vp9_avg4_neon, export=1
-        ldr             r12, [sp]
+        push            {lr}
+        ldr             r12, [sp, #4]
+        mov             lr,  r0
 1:
         vld1.32         {d4[]},   [r2], r3
         vld1.32         {d0[]},   [r0, :32], r1
@@ -231,15 +180,14 @@ function ff_vp9_avg4_neon, export=1
         vld1.32         {d7[]},   [r2], r3
         vrhadd.u8       d2,  d2,  d6
         vld1.32         {d3[]},   [r0, :32], r1
-        sub             r0,  r0,  r1, lsl #2
         subs            r12, r12, #4
-        vst1.32         {d0[0]},  [r0, :32], r1
+        vst1.32         {d0[0]},  [lr, :32], r1
         vrhadd.u8       d3,  d3,  d7
-        vst1.32         {d1[0]},  [r0, :32], r1
-        vst1.32         {d2[0]},  [r0, :32], r1
-        vst1.32         {d3[0]},  [r0, :32], r1
+        vst1.32         {d1[0]},  [lr, :32], r1
+        vst1.32         {d2[0]},  [lr, :32], r1
+        vst1.32         {d3[0]},  [lr, :32], r1
         bne             1b
-        bx              lr
+        pop             {pc}
 endfunc
 
 @ Helper macros for vmul/vmla with a constant from either d0 or d1 depending on index
@@ -327,7 +275,8 @@ function \type\()_8tap_\size\()h_\idx1\idx2
         sub             r3,  r3,  #8
 .endif
         @ Load the filter vector
-        vld1.16         {q0},  [r12,:128]
+        vld1.8          {d0},  [r12,:64]
+        vmovl.s8        q0,  d0
 1:
 .if \size >= 16
         mov             r12, r5
@@ -397,12 +346,12 @@ function \type\()_8tap_\size\()h_\idx1\idx2
 .endif
         @ Store and loop horizontally (for size >= 16)
 .if \size >= 16
+        subs            r12, r12, #16
         vst1.8          {q1}, [r0,:128]!
         vst1.8          {q3}, [r6,:128]!
+        beq             3f
         vmov            q8,  q10
         vmov            q11, q13
-        subs            r12, r12, #16
-        beq             3f
         vld1.8          {q10}, [r2]!
         vld1.8          {q13}, [r7]!
         vmovl.u8        q9,  d20
@@ -444,7 +393,7 @@ do_8tap_h_size 4
 do_8tap_h_size 8
 do_8tap_h_size 16
 
-.macro do_8tap_h_func type, filter, size
+.macro do_8tap_h_func type, filter, offset, size
 function ff_vp9_\type\()_\filter\()\size\()_h_neon, export=1
         push            {r4-r7}
 .if \size >= 16
@@ -455,9 +404,10 @@ function ff_vp9_\type\()_\filter\()\size\()_h_neon, export=1
         ldr             r4,  [sp, #16]
         ldr             r5,  [sp, #20]
 .endif
-        movrel          r12, \filter\()_filter-16
+        movrelx         r12, X(ff_vp9_subpel_filters)
+        add             r12, r12, 120*\offset - 8
         cmp             r5,  #8
-        add             r12,  r12, r5, lsl #4
+        add             r12, r12, r5, lsl #3
         mov             r5, #\size
 .if \size >= 16
         bge             \type\()_8tap_16h_34
@@ -470,12 +420,12 @@ endfunc
 .endm
 
 .macro do_8tap_h_filters size
-do_8tap_h_func put, regular, \size
-do_8tap_h_func avg, regular, \size
-do_8tap_h_func put, sharp,   \size
-do_8tap_h_func avg, sharp,   \size
-do_8tap_h_func put, smooth,  \size
-do_8tap_h_func avg, smooth,  \size
+do_8tap_h_func put, regular, 1, \size
+do_8tap_h_func avg, regular, 1, \size
+do_8tap_h_func put, sharp,   2, \size
+do_8tap_h_func avg, sharp,   2, \size
+do_8tap_h_func put, smooth,  0, \size
+do_8tap_h_func avg, smooth,  0, \size
 .endm
 
 do_8tap_h_filters 64
@@ -590,7 +540,8 @@ do_8tap_h_filters 4
 function \type\()_8tap_8v_\idx1\idx2
         sub             r2,  r2,  r3, lsl #1
         sub             r2,  r2,  r3
-        vld1.16         {q0},  [r12, :128]
+        vld1.8          {d0},  [r12, :64]
+        vmovl.s8        q0,  d0
 1:
         mov             r12,  r4
 
@@ -660,7 +611,8 @@ do_8tap_8v avg, 4, 3
 function \type\()_8tap_4v_\idx1\idx2
         sub             r2,  r2,  r3, lsl #1
         sub             r2,  r2,  r3
-        vld1.16         {q0},  [r12, :128]
+        vld1.8          {d0},  [r12, :64]
+        vmovl.s8        q0,  d0
 
         vld1.32         {d2[]},   [r2], r3
         vld1.32         {d3[]},   [r2], r3
@@ -723,14 +675,15 @@ do_8tap_4v put, 4, 3
 do_8tap_4v avg, 3, 4
 do_8tap_4v avg, 4, 3
 
-.macro do_8tap_v_func type, filter, size
+.macro do_8tap_v_func type, filter, offset, size
 function ff_vp9_\type\()_\filter\()\size\()_v_neon, export=1
         push            {r4-r5}
         vpush           {q4-q7}
         ldr             r4,  [sp, #72]
         ldr             r5,  [sp, #80]
-        movrel          r12, \filter\()_filter-16
-        add             r12,  r12, r5, lsl #4
+        movrelx         r12, X(ff_vp9_subpel_filters)
+        add             r12, r12, 120*\offset - 8
+        add             r12, r12, r5, lsl #3
         cmp             r5,  #8
         mov             r5,  #\size
 .if \size >= 8
@@ -744,12 +697,12 @@ endfunc
 .endm
 
 .macro do_8tap_v_filters size
-do_8tap_v_func put, regular, \size
-do_8tap_v_func avg, regular, \size
-do_8tap_v_func put, sharp,   \size
-do_8tap_v_func avg, sharp,   \size
-do_8tap_v_func put, smooth,  \size
-do_8tap_v_func avg, smooth,  \size
+do_8tap_v_func put, regular, 1, \size
+do_8tap_v_func avg, regular, 1, \size
+do_8tap_v_func put, sharp,   2, \size
+do_8tap_v_func avg, sharp,   2, \size
+do_8tap_v_func put, smooth,  0, \size
+do_8tap_v_func avg, smooth,  0, \size
 .endm
 
 do_8tap_v_filters 64