aarch64: vp9itxfm: Skip empty slices in the first pass of idct_idct 16x16 and 32x32

This work is sponsored by, and copyright, Google.

Previously all subpartitions except the eob=1 (DC) case ran with
the same runtime:

vp9_inv_dct_dct_16x16_sub16_add_neon:   1373.2
vp9_inv_dct_dct_32x32_sub32_add_neon:   8089.0

By skipping individual 8x16 or 8x32 pixel slices in the first pass,
we reduce the runtime of these functions like this:

vp9_inv_dct_dct_16x16_sub1_add_neon:     235.3
vp9_inv_dct_dct_16x16_sub2_add_neon:    1036.7
vp9_inv_dct_dct_16x16_sub4_add_neon:    1036.7
vp9_inv_dct_dct_16x16_sub8_add_neon:    1036.7
vp9_inv_dct_dct_16x16_sub12_add_neon:   1372.1
vp9_inv_dct_dct_16x16_sub16_add_neon:   1372.1
vp9_inv_dct_dct_32x32_sub1_add_neon:     555.1
vp9_inv_dct_dct_32x32_sub2_add_neon:    5190.2
vp9_inv_dct_dct_32x32_sub4_add_neon:    5180.0
vp9_inv_dct_dct_32x32_sub8_add_neon:    5183.1
vp9_inv_dct_dct_32x32_sub12_add_neon:   6161.5
vp9_inv_dct_dct_32x32_sub16_add_neon:   6155.5
vp9_inv_dct_dct_32x32_sub20_add_neon:   7136.3
vp9_inv_dct_dct_32x32_sub24_add_neon:   7128.4
vp9_inv_dct_dct_32x32_sub28_add_neon:   8098.9
vp9_inv_dct_dct_32x32_sub32_add_neon:   8098.8

I.e. in general a very minor overhead for the full subpartition case due
to the additional cmps, but a significant speedup for the cases when we
only need to process a small part of the actual input data.

This is cherrypicked from libav commits
cad42fadcd and
a0c443a398.

Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>

libavcodec/aarch64/vp9itxfm_neon.S

@@ -588,6 +588,9 @@ endfunc
 .macro store i, dst, inc
         st1             {v\i\().8h},  [\dst], \inc
 .endm
+.macro movi_v i, size, imm
+        movi            v\i\()\size,  \imm
+.endm
 .macro load_clear i, src, inc
         ld1             {v\i\().8h}, [\src]
         st1             {v2.8h},  [\src], \inc
@@ -596,9 +599,8 @@ endfunc
 // Read a vertical 8x16 slice out of a 16x16 matrix, do a transform on it,
 // transpose into a horizontal 16x8 slice and store.
 // x0 = dst (temp buffer)
-// x1 = unused
+// x1 = slice offset
 // x2 = src
-// x3 = slice offset
 // x9 = input stride
 .macro itxfm16_1d_funcs txfm
 function \txfm\()16_1d_8x16_pass1_neon
@@ -616,14 +618,14 @@ function \txfm\()16_1d_8x16_pass1_neon
         transpose_8x8H  v24, v25, v26, v27, v28, v29, v30, v31, v2, v3
 
         // Store the transposed 8x8 blocks horizontally.
-        cmp             x3,  #8
+        cmp             x1,  #8
         b.eq            1f
 .irp i, 16, 24, 17, 25, 18, 26, 19, 27, 20, 28, 21, 29, 22, 30, 23, 31
         store           \i,  x0,  #16
 .endr
         ret
 1:
-        // Special case: For the last input column (x3 == 8),
+        // Special case: For the last input column (x1 == 8),
         // which would be stored as the last row in the temp buffer,
         // don't store the first 8x8 block, but keep it in registers
         // for the first slice of the second pass (where it is the
@@ -751,13 +753,36 @@ function ff_vp9_\txfm1\()_\txfm2\()_16x16_add_neon, export=1
 
 .irp i, 0, 8
         add             x0,  sp,  #(\i*32)
+.ifc \txfm1\()_\txfm2,idct_idct
+.if \i == 8
+        cmp             w3,  #38
+        b.le            1f
+.endif
+.endif
+        mov             x1,  #\i
         add             x2,  x6,  #(\i*2)
-        mov             x3,  #\i
         bl              \txfm1\()16_1d_8x16_pass1_neon
 .endr
 .ifc \txfm1\()_\txfm2,iadst_idct
         ld1             {v0.8h,v1.8h}, [x10]
 .endif
+
+.ifc \txfm1\()_\txfm2,idct_idct
+        b               3f
+1:
+        // Set v24-v31 to zero, for the in-register passthrough of
+        // coefficients to pass 2. Since we only do two slices, this can
+        // only ever happen for the second slice. So we only need to store
+        // zeros to the temp buffer for the second half of the buffer.
+        // Move x0 to the second half, and use x9 == 32 as increment.
+        add             x0,  x0,  #16
+.irp i, 24, 25, 26, 27, 28, 29, 30, 31
+        movi_v          \i,  .16b, #0
+        st1             {v24.8h},  [x0], x9
+.endr
+3:
+.endif
+
 .irp i, 0, 8
         add             x0,  x4,  #(\i)
         mov             x1,  x5
@@ -1073,12 +1098,17 @@ function idct32_1d_8x32_pass2_neon
         ret
 endfunc
 
+const min_eob_idct_idct_32, align=4
+        .short  0, 34, 135, 336
+endconst
+
 function ff_vp9_idct_idct_32x32_add_neon, export=1
         cmp             w3,  #1
         b.eq            idct32x32_dc_add_neon
 
         movrel          x10, idct_coeffs
         add             x11, x10, #32
+        movrel          x12, min_eob_idct_idct_32, 2
 
         mov             x15, x30
 
@@ -1099,9 +1129,30 @@ function ff_vp9_idct_idct_32x32_add_neon, export=1
 
 .irp i, 0, 8, 16, 24
         add             x0,  sp,  #(\i*64)
+.if \i > 0
+        ldrh            w1,  [x12], #2
+        cmp             w3,  w1
+        mov             x1,  #(32 - \i)/4
+        b.le            1f
+.endif
         add             x2,  x6,  #(\i*2)
         bl              idct32_1d_8x32_pass1_neon
 .endr
+        b               3f
+
+1:
+        // Write zeros to the temp buffer for pass 2
+        movi            v16.8h,  #0
+        movi            v17.8h,  #0
+        movi            v18.8h,  #0
+        movi            v19.8h,  #0
+2:
+        subs            x1,  x1,  #1
+.rept 4
+        st1             {v16.8h-v19.8h},  [x0], #64
+.endr
+        b.ne            2b
+3:
 .irp i, 0, 8, 16, 24
         add             x0,  x4,  #(\i)
         mov             x1,  x5