Git Inbox Mirror of the ffmpeg-devel mailing list - see https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
 help / color / mirror / Atom feed
* [FFmpeg-devel] [PATCH 0/3] sw_scale: Provide neon implementation for hscale
@ 2022-10-28 11:34 Hubert Mazur
  2022-10-28 11:34 ` [FFmpeg-devel] [PATCH 1/3] sw_scale: Add specializations for hscale 8 to 19 Hubert Mazur
                   ` (3 more replies)
  0 siblings, 4 replies; 5+ messages in thread
From: Hubert Mazur @ 2022-10-28 11:34 UTC (permalink / raw)
  To: ffmpeg-devel; +Cc: gjb, upstream, jswinney, Hubert Mazur, martin, mw, spop

This patchset contains arm64 neon implementation of hscale functions.
Fixed minor style issues and declared C function wrappers as static.
This patchset do not contain the patch for checkasm tool, as the
previous one did. The reason behind it was failing tests on x86 arch
but not on aarch64 or loongarch. Probably the hscale functions on x86
have some bugs. Currently the checkasm tool does not check
the validity of hscale functions on x86 at all. Implementation of hscale
for x86 should be fixed anyway. As it comes to aarch64 the tests were
passing. Attaching a link to fate result in patchwork. After fixing x86
the patch for checkasm could be merged.

https://patchwork.ffmpeg.org/project/ffmpeg/patch/20221017130715.30896-3-hum@semihalf.com/

Hubert Mazur (3):
  sw_scale: Add specializations for hscale 8 to 19
  sw_scale: Add specializations for hscale 16 to 15
  sw_scale: Add specializations for hscale 16 to 19

 libswscale/aarch64/hscale.S  | 1100 ++++++++++++++++++++++++++++++++++
 libswscale/aarch64/swscale.c |  140 ++++-
 libswscale/swscale.c         |    1 -
 3 files changed, 1236 insertions(+), 5 deletions(-)

-- 
2.37.1

_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".

^ permalink raw reply	[flat|nested] 5+ messages in thread

* [FFmpeg-devel] [PATCH 1/3] sw_scale: Add specializations for hscale 8 to 19
  2022-10-28 11:34 [FFmpeg-devel] [PATCH 0/3] sw_scale: Provide neon implementation for hscale Hubert Mazur
@ 2022-10-28 11:34 ` Hubert Mazur
  2022-10-28 11:34 ` [FFmpeg-devel] [PATCH 2/3] sw_scale: Add specializations for hscale 16 to 15 Hubert Mazur
                   ` (2 subsequent siblings)
  3 siblings, 0 replies; 5+ messages in thread
From: Hubert Mazur @ 2022-10-28 11:34 UTC (permalink / raw)
  To: ffmpeg-devel; +Cc: gjb, upstream, jswinney, Hubert Mazur, martin, mw, spop

Add arm64 neon implementations for hscale 8 to 19 with filter
sizes 4, 4X and 8. Both implementations are based on very similar ones
dedicated to hscale 8 to 15. The major changes refer to saving
the data - instead of writing the result as int16_t it is done
with int32_t.

These functions are heavily inspired on patches provided by J. Swinney
and M. Storsjö for hscale8to15 which were slightly adapted for
hscale8to19.

The tests and benchmarks run on AWS Graviton 2 instances. The results
from a checkasm tool shown below.

hscale_8_to_19__fs_4_dstW_512_c: 5663.2
hscale_8_to_19__fs_4_dstW_512_neon: 1259.7
hscale_8_to_19__fs_8_dstW_512_c: 9306.0
hscale_8_to_19__fs_8_dstW_512_neon: 2020.2
hscale_8_to_19__fs_12_dstW_512_c: 12932.7
hscale_8_to_19__fs_12_dstW_512_neon: 2462.5
hscale_8_to_19__fs_16_dstW_512_c: 16844.2
hscale_8_to_19__fs_16_dstW_512_neon: 4671.2
hscale_8_to_19__fs_32_dstW_512_c: 32803.7
hscale_8_to_19__fs_32_dstW_512_neon: 5474.2
hscale_8_to_19__fs_40_dstW_512_c: 40948.0
hscale_8_to_19__fs_40_dstW_512_neon: 6669.7

Signed-off-by: Hubert Mazur <hum@semihalf.com>
---
 libswscale/aarch64/hscale.S  | 291 +++++++++++++++++++++++++++++++++++
 libswscale/aarch64/swscale.c |  13 +-
 2 files changed, 300 insertions(+), 4 deletions(-)

diff --git a/libswscale/aarch64/hscale.S b/libswscale/aarch64/hscale.S
index a16d3dca42a..35e25462ce8 100644
--- a/libswscale/aarch64/hscale.S
+++ b/libswscale/aarch64/hscale.S
@@ -347,3 +347,294 @@ function ff_hscale8to15_4_neon, export=1
         add                 sp, sp, #32                 // clean up stack
         ret
 endfunc
+
+function ff_hscale8to19_4_neon, export=1
+        // x0               SwsContext *c (unused)
+        // x1               int32_t *dst
+        // w2               int dstW
+        // x3               const uint8_t *src // treat it as uint16_t *src
+        // x4               const uint16_t *filter
+        // x5               const int32_t *filterPos
+        // w6               int filterSize
+
+        movi                v18.4s, #1
+        movi                v17.4s, #1
+        shl                 v18.4s, v18.4s, #19
+        sub                 v18.4s, v18.4s, v17.4s      // max allowed value
+
+        cmp                 w2, #16
+        b.lt                2f // move to last block
+
+        ldp                 w8, w9, [x5]                // filterPos[0], filterPos[1]
+        ldp                 w10, w11, [x5, #8]          // filterPos[2], filterPos[3]
+        ldp                 w12, w13, [x5, #16]         // filterPos[4], filterPos[5]
+        ldp                 w14, w15, [x5, #24]         // filterPos[6], filterPos[7]
+        add                 x5, x5, #32
+
+        // load data from
+        ldr                 w8, [x3, w8, UXTW]
+        ldr                 w9, [x3, w9, UXTW]
+        ldr                 w10, [x3, w10, UXTW]
+        ldr                 w11, [x3, w11, UXTW]
+        ldr                 w12, [x3, w12, UXTW]
+        ldr                 w13, [x3, w13, UXTW]
+        ldr                 w14, [x3, w14, UXTW]
+        ldr                 w15, [x3, w15, UXTW]
+
+        sub                 sp, sp, #32
+
+        stp                 w8, w9, [sp]
+        stp                 w10, w11, [sp, #8]
+        stp                 w12, w13, [sp, #16]
+        stp                 w14, w15, [sp, #24]
+
+1:
+        ld4                 {v0.8b, v1.8b, v2.8b, v3.8b}, [sp]
+        ld4                 {v28.8h, v29.8h, v30.8h, v31.8h}, [x4], #64 // filter[0..7]
+        // load filterPositions into registers for next iteration
+
+        ldp                 w8, w9, [x5]                // filterPos[0], filterPos[1]
+        ldp                 w10, w11, [x5, #8]          // filterPos[2], filterPos[3]
+        ldp                 w12, w13, [x5, #16]         // filterPos[4], filterPos[5]
+        ldp                 w14, w15, [x5, #24]         // filterPos[6], filterPos[7]
+        add                 x5, x5, #32
+        uxtl                v0.8h, v0.8b
+        ldr                 w8, [x3, w8, UXTW]
+        smull               v5.4s, v0.4h, v28.4h        // multiply first column of src
+        ldr                 w9, [x3, w9, UXTW]
+        smull2              v6.4s, v0.8h, v28.8h
+        stp                 w8, w9, [sp]
+
+        uxtl                v1.8h, v1.8b
+        ldr                 w10, [x3, w10, UXTW]
+        smlal               v5.4s, v1.4h, v29.4h        // multiply second column of src
+        ldr                 w11, [x3, w11, UXTW]
+        smlal2              v6.4s, v1.8h, v29.8h
+        stp                 w10, w11, [sp, #8]
+
+        uxtl                v2.8h, v2.8b
+        ldr                 w12, [x3, w12, UXTW]
+        smlal               v5.4s, v2.4h, v30.4h        // multiply third column of src
+        ldr                 w13, [x3, w13, UXTW]
+        smlal2              v6.4s, v2.8h, v30.8h
+        stp                 w12, w13, [sp, #16]
+
+        uxtl                v3.8h, v3.8b
+        ldr                 w14, [x3, w14, UXTW]
+        smlal               v5.4s, v3.4h, v31.4h        // multiply fourth column of src
+        ldr                 w15, [x3, w15, UXTW]
+        smlal2              v6.4s, v3.8h, v31.8h
+        stp                 w14, w15, [sp, #24]
+
+        sub                 w2, w2, #8
+        sshr                v5.4s, v5.4s, #3
+        sshr                v6.4s, v6.4s, #3
+        smin                v5.4s, v5.4s, v18.4s
+        smin                v6.4s, v6.4s, v18.4s
+
+        st1                 {v5.4s, v6.4s}, [x1], #32
+        cmp                 w2, #16
+        b.ge                1b
+
+        // here we make last iteration, without updating the registers
+        ld4                 {v0.8b, v1.8b, v2.8b, v3.8b}, [sp]
+        ld4                 {v28.8h, v29.8h, v30.8h, v31.8h}, [x4], #64 // filter[0..7]
+
+        uxtl                v0.8h, v0.8b
+        uxtl                v1.8h, v1.8b
+        smull               v5.4s, v0.4h, v28.4h
+        smull2              v6.4s, v0.8h, v28.8h
+        uxtl                v2.8h, v2.8b
+        smlal               v5.4s, v1.4h, v29.4H
+        smlal2              v6.4s, v1.8h, v29.8H
+        uxtl                v3.8h, v3.8b
+        smlal               v5.4s, v2.4h, v30.4H
+        smlal2              v6.4s, v2.8h, v30.8H
+        smlal               v5.4s, v3.4h, v31.4H
+        smlal2              v6.4s, v3.8h, v31.8h
+
+        sshr                v5.4s, v5.4s, #3
+        sshr                v6.4s, v6.4s, #3
+
+        smin                v5.4s, v5.4s, v18.4s
+        smin                v6.4s, v6.4s, v18.4s
+
+        sub                 w2, w2, #8
+        st1                 {v5.4s, v6.4s}, [x1], #32
+        add                 sp, sp, #32 // restore stack
+        cbnz                w2, 2f
+
+        ret
+
+2:
+        ldr                 w8, [x5], #4 // load filterPos
+        add                 x9, x3, w8, UXTW // src + filterPos
+        ld1                 {v0.s}[0], [x9] // load 4 * uint8_t* into one single
+        ld1                 {v31.4h}, [x4], #8
+        uxtl                v0.8h, v0.8b
+        smull               v5.4s, v0.4h, v31.4H
+        saddlv              d0, v5.4S
+        sqshrn              s0, d0, #3
+        smin                v0.4s, v0.4s, v18.4s
+        st1                 {v0.s}[0], [x1], #4
+        sub                 w2, w2, #1
+        cbnz                w2, 2b // if iterations remain jump to beginning
+
+        ret
+endfunc
+
+function ff_hscale8to19_X8_neon, export=1
+        movi                v20.4s, #1
+        movi                v17.4s, #1
+        shl                 v20.4s, v20.4s, #19
+        sub                 v20.4s, v20.4s, v17.4s
+
+        sbfiz               x7, x6, #1, #32             // filterSize*2 (*2 because int16)
+1:
+        mov                 x16, x4                     // filter0 = filter
+        ldr                 w8, [x5], #4                // filterPos[idx]
+        add                 x12, x16, x7                // filter1 = filter0 + filterSize*2
+        ldr                 w0, [x5], #4                // filterPos[idx + 1]
+        add                 x13, x12, x7                // filter2 = filter1 + filterSize*2
+        ldr                 w11, [x5], #4               // filterPos[idx + 2]
+        add                 x4, x13, x7                 // filter3 = filter2 + filterSize*2
+        ldr                 w9, [x5], #4                // filterPos[idx + 3]
+        movi                v0.2D, #0                   // val sum part 1 (for dst[0])
+        movi                v1.2D, #0                   // val sum part 2 (for dst[1])
+        movi                v2.2D, #0                   // val sum part 3 (for dst[2])
+        movi                v3.2D, #0                   // val sum part 4 (for dst[3])
+        add                 x17, x3, w8, UXTW           // srcp + filterPos[0]
+        add                 x8,  x3, w0, UXTW           // srcp + filterPos[1]
+        add                 x0, x3, w11, UXTW           // srcp + filterPos[2]
+        add                 x11, x3, w9, UXTW           // srcp + filterPos[3]
+        mov                 w15, w6                     // filterSize counter
+2:      ld1                 {v4.8B}, [x17], #8          // srcp[filterPos[0] + {0..7}]
+        ld1                 {v5.8H}, [x16], #16         // load 8x16-bit filter values, part 1
+        uxtl                v4.8H, v4.8B                // unpack part 1 to 16-bit
+        smlal               v0.4S, v4.4H, v5.4H         // v0 accumulates srcp[filterPos[0] + {0..3}] * filter[{0..3}]
+        ld1                 {v6.8B}, [x8], #8           // srcp[filterPos[1] + {0..7}]
+        smlal2              v0.4S, v4.8H, v5.8H         // v0 accumulates srcp[filterPos[0] + {4..7}] * filter[{4..7}]
+        ld1                 {v7.8H}, [x12], #16         // load 8x16-bit at filter+filterSize
+        ld1                 {v16.8B}, [x0], #8          // srcp[filterPos[2] + {0..7}]
+        uxtl                v6.8H, v6.8B                // unpack part 2 to 16-bit
+        ld1                 {v17.8H}, [x13], #16        // load 8x16-bit at filter+2*filterSize
+        uxtl                v16.8H, v16.8B              // unpack part 3 to 16-bit
+        smlal               v1.4S, v6.4H, v7.4H         // v1 accumulates srcp[filterPos[1] + {0..3}] * filter[{0..3}]
+        ld1                 {v18.8B}, [x11], #8         // srcp[filterPos[3] + {0..7}]
+        smlal               v2.4S, v16.4H, v17.4H       // v2 accumulates srcp[filterPos[2] + {0..3}] * filter[{0..3}]
+        ld1                 {v19.8H}, [x4], #16         // load 8x16-bit at filter+3*filterSize
+        smlal2              v2.4S, v16.8H, v17.8H       // v2 accumulates srcp[filterPos[2] + {4..7}] * filter[{4..7}]
+        uxtl                v18.8H, v18.8B              // unpack part 4 to 16-bit
+        smlal2              v1.4S, v6.8H, v7.8H         // v1 accumulates srcp[filterPos[1] + {4..7}] * filter[{4..7}]
+        smlal               v3.4S, v18.4H, v19.4H       // v3 accumulates srcp[filterPos[3] + {0..3}] * filter[{0..3}]
+        subs                w15, w15, #8                // j -= 8: processed 8/filterSize
+        smlal2              v3.4S, v18.8H, v19.8H       // v3 accumulates srcp[filterPos[3] + {4..7}] * filter[{4..7}]
+        b.gt                2b                          // inner loop if filterSize not consumed completely
+        addp                v0.4S, v0.4S, v1.4S         // part01 horizontal pair adding
+        addp                v2.4S, v2.4S, v3.4S         // part23 horizontal pair adding
+        addp                v0.4S, v0.4S, v2.4S         // part0123 horizontal pair adding
+        subs                w2, w2, #4                  // dstW -= 4
+        sshr                v0.4s, v0.4S, #3            // shift and clip the 2x16-bit final values
+        smin                v0.4s, v0.4s, v20.4s
+        st1                 {v0.4s}, [x1], #16           // write to destination part0123
+        b.gt                1b                          // loop until end of line
+        ret
+endfunc
+
+function ff_hscale8to19_X4_neon, export=1
+        // x0  SwsContext *c (not used)
+        // x1  int16_t *dst
+        // w2  int dstW
+        // x3  const uint8_t *src
+        // x4  const int16_t *filter
+        // x5  const int32_t *filterPos
+        // w6  int filterSize
+
+        movi                v20.4s, #1
+        movi                v17.4s, #1
+        shl                 v20.4s, v20.4s, #19
+        sub                 v20.4s, v20.4s, v17.4s
+
+        lsl                 w7, w6, #1
+1:
+        ldp                 w8, w9, [x5]
+        ldp                 w10, w11, [x5, #8]
+
+        movi                v16.2d, #0                  // initialize accumulator for idx + 0
+        movi                v17.2d, #0                  // initialize accumulator for idx + 1
+        movi                v18.2d, #0                  // initialize accumulator for idx + 2
+        movi                v19.2d, #0                  // initialize accumulator for idx + 3
+
+        mov                 x12, x4                     // filter + 0
+        add                 x13, x4, x7                 // filter + 1
+        add                 x8, x3, w8, UXTW            // srcp + filterPos 0
+        add                 x14, x13, x7                // filter + 2
+        add                 x9, x3, w9, UXTW            // srcp + filterPos 1
+        add                 x15, x14, x7                // filter + 3
+        add                 x10, x3, w10, UXTW          // srcp + filterPos 2
+        mov                 w0, w6                      // save the filterSize to temporary variable
+        add                 x11, x3, w11, UXTW          // srcp + filterPos 3
+        add                 x5, x5, #16                 // advance filter position
+        mov                 x16, xzr                    // clear the register x16 used for offsetting the filter values
+
+2:
+        ldr                 d4, [x8], #8                // load src values for idx 0
+        ldr                 q31, [x12, x16]             // load filter values for idx 0
+        uxtl                v4.8h, v4.8b                // extend type to match the filter' size
+        ldr                 d5, [x9], #8                // load src values for idx 1
+        smlal               v16.4s, v4.4h, v31.4h       // multiplication of lower half for idx 0
+        uxtl                v5.8h, v5.8b                // extend type to match the filter' size
+        ldr                 q30, [x13, x16]             // load filter values for idx 1
+        smlal2              v16.4s, v4.8h, v31.8h       // multiplication of upper half for idx 0
+        ldr                 d6, [x10], #8               // load src values for idx 2
+        ldr                 q29, [x14, x16]             // load filter values for idx 2
+        smlal               v17.4s, v5.4h, v30.4H       // multiplication of lower half for idx 1
+        ldr                 d7, [x11], #8               // load src values for idx 3
+        smlal2              v17.4s, v5.8h, v30.8H       // multiplication of upper half for idx 1
+        uxtl                v6.8h, v6.8B                // extend tpye to matchi the filter's size
+        ldr                 q28, [x15, x16]             // load filter values for idx 3
+        smlal               v18.4s, v6.4h, v29.4h       // multiplication of lower half for idx 2
+        uxtl                v7.8h, v7.8B
+        smlal2              v18.4s, v6.8h, v29.8H       // multiplication of upper half for idx 2
+        sub                 w0, w0, #8
+        smlal               v19.4s, v7.4h, v28.4H       // multiplication of lower half for idx 3
+        cmp                 w0, #8
+        smlal2              v19.4s, v7.8h, v28.8h       // multiplication of upper half for idx 3
+        add                 x16, x16, #16                // advance filter values indexing
+
+        b.ge                2b
+
+
+        // 4 iterations left
+
+        sub                 x17, x7, #8                 // step back to wrap up the filter pos for last 4 elements
+
+        ldr                 s4, [x8]                    // load src values for idx 0
+        ldr                 d31, [x12, x17]             // load filter values for idx 0
+        uxtl                v4.8h, v4.8b                // extend type to match the filter' size
+        ldr                 s5, [x9]                    // load src values for idx 1
+        smlal               v16.4s, v4.4h, v31.4h
+        ldr                 d30, [x13, x17]             // load filter values for idx 1
+        uxtl                v5.8h, v5.8b                // extend type to match the filter' size
+        ldr                 s6, [x10]                   // load src values for idx 2
+        smlal               v17.4s, v5.4h, v30.4h
+        uxtl                v6.8h, v6.8B                // extend type to match the filter's size
+        ldr                 d29, [x14, x17]             // load filter values for idx 2
+        ldr                 s7, [x11]                   // load src values for idx 3
+        addp                v16.4s, v16.4s, v17.4s
+        uxtl                v7.8h, v7.8B
+        ldr                 d28, [x15, x17]             // load filter values for idx 3
+        smlal               v18.4s, v6.4h, v29.4h
+        smlal               v19.4s, v7.4h, v28.4h
+        subs                w2, w2, #4
+        addp                v18.4s, v18.4s, v19.4s
+        addp                v16.4s, v16.4s, v18.4s
+        sshr                v16.4s, v16.4s, #3
+        smin                v16.4s, v16.4s, v20.4s
+
+        st1                 {v16.4s}, [x1], #16
+        add                 x4, x4, x7, lsl #2
+        b.gt                1b
+        ret
+
+endfunc
diff --git a/libswscale/aarch64/swscale.c b/libswscale/aarch64/swscale.c
index d1312c66587..171fb51329b 100644
--- a/libswscale/aarch64/swscale.c
+++ b/libswscale/aarch64/swscale.c
@@ -29,7 +29,8 @@ void ff_hscale ## from_bpc ## to ## to_bpc ## _ ## filter_n ## _ ## opt( \
                                                 const int16_t *filter, \
                                                 const int32_t *filterPos, int filterSize)
 #define SCALE_FUNCS(filter_n, opt) \
-    SCALE_FUNC(filter_n,  8, 15, opt);
+    SCALE_FUNC(filter_n, 8, 15, opt); \
+    SCALE_FUNC(filter_n, 8, 19, opt);
 #define ALL_SCALE_FUNCS(opt) \
     SCALE_FUNCS(4, opt); \
     SCALE_FUNCS(X8, opt); \
@@ -48,9 +49,13 @@ void ff_yuv2plane1_8_neon(
         int offset);
 
 #define ASSIGN_SCALE_FUNC2(hscalefn, filtersize, opt) do {              \
-    if (c->srcBpc == 8 && c->dstBpc <= 14) {                            \
-      hscalefn =                                                        \
-        ff_hscale8to15_ ## filtersize ## _ ## opt;                      \
+    if (c->srcBpc == 8) {                                               \
+        if(c->dstBpc <= 14) {                                           \
+            hscalefn =                                                  \
+                ff_hscale8to15_ ## filtersize ## _ ## opt;              \
+        } else                                                          \
+            hscalefn =                                                  \
+                ff_hscale8to19_ ## filtersize ## _ ## opt;              \
     }                                                                   \
 } while (0)
 
-- 
2.37.1

_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".

^ permalink raw reply	[flat|nested] 5+ messages in thread

* [FFmpeg-devel] [PATCH 2/3] sw_scale: Add specializations for hscale 16 to 15
  2022-10-28 11:34 [FFmpeg-devel] [PATCH 0/3] sw_scale: Provide neon implementation for hscale Hubert Mazur
  2022-10-28 11:34 ` [FFmpeg-devel] [PATCH 1/3] sw_scale: Add specializations for hscale 8 to 19 Hubert Mazur
@ 2022-10-28 11:34 ` Hubert Mazur
  2022-10-28 11:34 ` [FFmpeg-devel] [PATCH 3/3] sw_scale: Add specializations for hscale 16 to 19 Hubert Mazur
  2022-11-01 13:26 ` [FFmpeg-devel] [PATCH 0/3] sw_scale: Provide neon implementation for hscale Martin Storsjö
  3 siblings, 0 replies; 5+ messages in thread
From: Hubert Mazur @ 2022-10-28 11:34 UTC (permalink / raw)
  To: ffmpeg-devel; +Cc: gjb, upstream, jswinney, Hubert Mazur, martin, mw, spop

Add arm64 neon implementations for hscale 16 to 15 with filter
sizes 4, 8 and X4.

The tests and benchmarks run on AWS Graviton 2 instances.
The results from a checkasm tool are shown below.

hscale_16_to_15__fs_4_dstW_512_c: 6703.5
hscale_16_to_15__fs_4_dstW_512_neon: 2298.0
hscale_16_to_15__fs_8_dstW_512_c: 10983.0
hscale_16_to_15__fs_8_dstW_512_neon: 3216.5
hscale_16_to_15__fs_12_dstW_512_c: 15526.0
hscale_16_to_15__fs_12_dstW_512_neon: 3993.0
hscale_16_to_15__fs_16_dstW_512_c: 20183.5
hscale_16_to_15__fs_16_dstW_512_neon: 5369.7
hscale_16_to_15__fs_32_dstW_512_c: 39315.2
hscale_16_to_15__fs_32_dstW_512_neon: 9511.2
hscale_16_to_15__fs_40_dstW_512_c: 48995.7
hscale_16_to_15__fs_40_dstW_512_neon: 11570.0

Signed-off-by: Hubert Mazur <hum@semihalf.com>
---
 libswscale/aarch64/hscale.S  | 407 +++++++++++++++++++++++++++++++++++
 libswscale/aarch64/swscale.c |  61 ++++++
 libswscale/swscale.c         |   1 -
 3 files changed, 468 insertions(+), 1 deletion(-)

diff --git a/libswscale/aarch64/hscale.S b/libswscale/aarch64/hscale.S
index 35e25462ce8..b571a5dc219 100644
--- a/libswscale/aarch64/hscale.S
+++ b/libswscale/aarch64/hscale.S
@@ -636,5 +636,412 @@ function ff_hscale8to19_X4_neon, export=1
         add                 x4, x4, x7, lsl #2
         b.gt                1b
         ret
+endfunc
+
+function ff_hscale16to15_4_neon_asm, export=1
+        // w0               int shift
+        // x1               int32_t *dst
+        // w2               int dstW
+        // x3               const uint8_t *src // treat it as uint16_t *src
+        // x4               const uint16_t *filter
+        // x5               const int32_t *filterPos
+        // w6               int filterSize
+
+        movi                v18.4s, #1
+        movi                v17.4s, #1
+        shl                 v18.4s, v18.4s, #15
+        sub                 v18.4s, v18.4s, v17.4s      // max allowed value
+        dup                 v17.4s, w0                  // read shift
+        neg                 v17.4s, v17.4s              // negate it, so it can be used in sshl (effectively shift right)
+
+        cmp                 w2, #16
+        b.lt                2f // move to last block
+
+        ldp                 w8, w9, [x5]                // filterPos[0], filterPos[1]
+        ldp                 w10, w11, [x5, #8]          // filterPos[2], filterPos[3]
+        ldp                 w12, w13, [x5, #16]         // filterPos[4], filterPos[5]
+        ldp                 w14, w15, [x5, #24]         // filterPos[6], filterPos[7]
+        add                 x5, x5, #32
+
+        // shift all filterPos left by one, as uint16_t will be read
+        lsl                 x8, x8, #1
+        lsl                 x9, x9, #1
+        lsl                 x10, x10, #1
+        lsl                 x11, x11, #1
+        lsl                 x12, x12, #1
+        lsl                 x13, x13, #1
+        lsl                 x14, x14, #1
+        lsl                 x15, x15, #1
+
+        // load src with given offset
+        ldr                 x8,  [x3, w8,  UXTW]
+        ldr                 x9,  [x3, w9,  UXTW]
+        ldr                 x10, [x3, w10, UXTW]
+        ldr                 x11, [x3, w11, UXTW]
+        ldr                 x12, [x3, w12, UXTW]
+        ldr                 x13, [x3, w13, UXTW]
+        ldr                 x14, [x3, w14, UXTW]
+        ldr                 x15, [x3, w15, UXTW]
+
+        sub                 sp, sp, #64
+        // push src on stack so it can be loaded into vectors later
+        stp                 x8, x9, [sp]
+        stp                 x10, x11, [sp, #16]
+        stp                 x12, x13, [sp, #32]
+        stp                 x14, x15, [sp, #48]
+
+1:
+        ld4                 {v0.8h, v1.8h, v2.8h, v3.8h}, [sp]
+        ld4                 {v28.8h, v29.8h, v30.8h, v31.8h}, [x4], #64 // filter[0..7]
+
+        // Each of blocks does the following:
+        // Extend src and filter to 32 bits with uxtl and sxtl
+        // multiply or multiply and accumulate results
+        // Extending to 32 bits is necessary, as unit16_t values can't
+        // be represented as int16_t without type promotion.
+        uxtl                v26.4s, v0.4h
+        sxtl                v27.4s, v28.4H
+        uxtl2               v0.4s, v0.8h
+        mul                 v5.4s, v26.4s, v27.4s
+        sxtl2               v28.4s, v28.8H
+        uxtl                v26.4s, v1.4h
+        mul                 v6.4s, v0.4s, v28.4s
+
+        sxtl                v27.4s, v29.4H
+        uxtl2               v0.4s, v1.8h
+        mla                 v5.4s, v27.4s, v26.4s
+        sxtl2               v28.4s, v29.8H
+        uxtl                v26.4s, v2.4h
+        mla                 v6.4s, v28.4s, v0.4s
+
+        sxtl                v27.4s, v30.4H
+        uxtl2               v0.4s, v2.8h
+        mla                 v5.4s, v27.4s, v26.4s
+        sxtl2               v28.4s, v30.8H
+        uxtl                v26.4s, v3.4h
+        mla                 v6.4s, v28.4s, v0.4s
+
+        sxtl                v27.4s, v31.4H
+        uxtl2               v0.4s, v3.8h
+        mla                 v5.4s, v27.4s, v26.4s
+        sxtl2               v28.4s, v31.8H
+        sub                 w2, w2, #8
+        mla                 v6.4s, v28.4s, v0.4s
+
+        sshl                v5.4s, v5.4s, v17.4s
+        sshl                v6.4s, v6.4s, v17.4s
+        smin                v5.4s, v5.4s, v18.4s
+        smin                v6.4s, v6.4s, v18.4s
+        xtn                 v5.4h, v5.4s
+        xtn2                v5.8h, v6.4s
+
+        st1                 {v5.8h}, [x1], #16
+        cmp                 w2, #16
+
+        // load filterPositions into registers for next iteration
+        ldp                 w8, w9, [x5]                // filterPos[0], filterPos[1]
+        ldp                 w10, w11, [x5, #8]          // filterPos[2], filterPos[3]
+        ldp                 w12, w13, [x5, #16]         // filterPos[4], filterPos[5]
+        ldp                 w14, w15, [x5, #24]         // filterPos[6], filterPos[7]
+        add                 x5, x5, #32
+
+        lsl                 x8, x8, #1
+        lsl                 x9, x9, #1
+        lsl                 x10, x10, #1
+        lsl                 x11, x11, #1
+        lsl                 x12, x12, #1
+        lsl                 x13, x13, #1
+        lsl                 x14, x14, #1
+        lsl                 x15, x15, #1
+
+        ldr                 x8,  [x3, w8,  UXTW]
+        ldr                 x9,  [x3, w9,  UXTW]
+        ldr                 x10, [x3, w10, UXTW]
+        ldr                 x11, [x3, w11, UXTW]
+        ldr                 x12, [x3, w12, UXTW]
+        ldr                 x13, [x3, w13, UXTW]
+        ldr                 x14, [x3, w14, UXTW]
+        ldr                 x15, [x3, w15, UXTW]
+
+        stp                 x8, x9, [sp]
+        stp                 x10, x11, [sp, #16]
+        stp                 x12, x13, [sp, #32]
+        stp                 x14, x15, [sp, #48]
+
+        b.ge                1b
+
+        // here we make last iteration, without updating the registers
+        ld4                 {v0.8h, v1.8h, v2.8h, v3.8h}, [sp]
+        ld4                 {v28.8h, v29.8h, v30.8h, v31.8h}, [x4], #64
+
+        uxtl                v26.4s, v0.4h
+        sxtl                v27.4s, v28.4H
+        uxtl2               v0.4s, v0.8h
+        mul                 v5.4s, v26.4s, v27.4s
+        sxtl2               v28.4s, v28.8H
+        uxtl                v26.4s, v1.4h
+        mul                 v6.4s, v0.4s, v28.4s
+
+        sxtl                v27.4s, v29.4H
+        uxtl2               v0.4s, v1.8h
+        mla                 v5.4s, v26.4s, v27.4s
+        sxtl2               v28.4s, v29.8H
+        uxtl                v26.4s, v2.4h
+        mla                 v6.4s, v0.4s, v28.4s
+
+        sxtl                v27.4s, v30.4H
+        uxtl2               v0.4s, v2.8h
+        mla                 v5.4s, v26.4s, v27.4s
+        sxtl2               v28.4s, v30.8H
+        uxtl                v26.4s, v3.4h
+        mla                 v6.4s, v0.4s, v28.4s
+
+        sxtl                v27.4s, v31.4H
+        uxtl2               v0.4s, v3.8h
+        mla                 v5.4s, v26.4s, v27.4s
+        sxtl2               v28.4s, v31.8H
+        subs                w2, w2, #8
+        mla                 v6.4s, v0.4s, v28.4s
+
+        sshl                v5.4s, v5.4s, v17.4s
+        sshl                v6.4s, v6.4s, v17.4s
+        smin                v5.4s, v5.4s, v18.4s
+        smin                v6.4s, v6.4s, v18.4s
+        xtn                 v5.4h, v5.4S
+        xtn2                v5.8h, v6.4s
+
+        st1                 {v5.8h}, [x1], #16
+        add                 sp, sp, #64                 // restore stack
+        cbnz                w2, 2f
+
+        ret
+
+2:
+        ldr                 w8, [x5], #4                // load filterPos
+        lsl                 w8, w8, #1
+        add                 x9, x3, w8, UXTW            // src + filterPos
+        ld1                 {v0.4h}, [x9]               // load 4 * uint16_t
+        ld1                 {v31.4h}, [x4], #8
+
+        uxtl                v0.4s, v0.4h
+        sxtl                v31.4s, v31.4h
+        mul                 v5.4s, v0.4s, v31.4s
+        addv                s0, v5.4S
+        sshl                v0.4s, v0.4s, v17.4s
+        smin                v0.4s, v0.4s, v18.4s
+        st1                 {v0.h}[0], [x1], #2
+        sub                 w2, w2, #1
+        cbnz                w2, 2b                      // if iterations remain jump to beginning
+
+        ret
+endfunc
+
+function ff_hscale16to15_X8_neon_asm, export=1
+        // w0               int shift
+        // x1               int32_t *dst
+        // w2               int dstW
+        // x3               const uint8_t *src // treat it as uint16_t *src
+        // x4               const uint16_t *filter
+        // x5               const int32_t *filterPos
+        // w6               int filterSize
+
+        movi                v20.4s, #1
+        movi                v21.4s, #1
+        shl                 v20.4s, v20.4s, #15
+        sub                 v20.4s, v20.4s, v21.4s
+        dup                 v21.4s, w0
+        neg                 v21.4s, v21.4s
+
+        sbfiz               x7, x6, #1, #32             // filterSize*2 (*2 because int16)
+1:      ldr                 w8, [x5], #4                // filterPos[idx]
+        lsl                 w8, w8, #1
+        ldr                 w10, [x5], #4               // filterPos[idx + 1]
+        lsl                 w10, w10, #1
+        ldr                 w11, [x5], #4               // filterPos[idx + 2]
+        lsl                 w11, w11, #1
+        ldr                 w9, [x5], #4                // filterPos[idx + 3]
+        lsl                 w9, w9, #1
+        mov                 x16, x4                     // filter0 = filter
+        add                 x12, x16, x7                // filter1 = filter0 + filterSize*2
+        add                 x13, x12, x7                // filter2 = filter1 + filterSize*2
+        add                 x4, x13, x7                 // filter3 = filter2 + filterSize*2
+        movi                v0.2D, #0                   // val sum part 1 (for dst[0])
+        movi                v1.2D, #0                   // val sum part 2 (for dst[1])
+        movi                v2.2D, #0                   // val sum part 3 (for dst[2])
+        movi                v3.2D, #0                   // val sum part 4 (for dst[3])
+        add                 x17, x3, w8, UXTW           // srcp + filterPos[0]
+        add                 x8,  x3, w10, UXTW          // srcp + filterPos[1]
+        add                 x10, x3, w11, UXTW          // srcp + filterPos[2]
+        add                 x11, x3, w9, UXTW           // srcp + filterPos[3]
+        mov                 w15, w6                     // filterSize counter
+2:      ld1                 {v4.8H}, [x17], #16         // srcp[filterPos[0] + {0..7}]
+        ld1                 {v5.8H}, [x16], #16         // load 8x16-bit filter values, part 1
+        ld1                 {v6.8H}, [x8], #16          // srcp[filterPos[1] + {0..7}]
+        ld1                 {v7.8H}, [x12], #16         // load 8x16-bit at filter+filterSize
+        uxtl                v24.4s, v4.4H               // extend srcp lower half to 32 bits to preserve sign
+        sxtl                v25.4s, v5.4H               // extend filter lower half to 32 bits to match srcp size
+        uxtl2               v4.4s, v4.8h                // extend srcp upper half to 32 bits
+        mla                 v0.4s, v24.4s, v25.4s       // multiply accumulate lower half of v4 * v5
+        sxtl2               v5.4s, v5.8h                // extend filter upper half to 32 bits
+        uxtl                v26.4s, v6.4h               // extend srcp lower half to 32 bits
+        mla                 v0.4S, v4.4s, v5.4s         // multiply accumulate upper half of v4 * v5
+        sxtl                v27.4s, v7.4H               // exted filter lower half
+        uxtl2               v6.4s, v6.8H                // extend srcp upper half
+        sxtl2               v7.4s, v7.8h                // extend filter upper half
+        ld1                 {v16.8H}, [x10], #16        // srcp[filterPos[2] + {0..7}]
+        mla                 v1.4S, v26.4s, v27.4s       // v1 accumulates srcp[filterPos[1] + {0..3}] * filter[{0..3}]
+        ld1                 {v17.8H}, [x13], #16        // load 8x16-bit at filter+2*filterSize
+        uxtl                v22.4s, v16.4H              // extend srcp lower half
+        sxtl                v23.4s, v17.4H              // extend filter lower half
+        uxtl2               v16.4s, v16.8H              // extend srcp upper half
+        sxtl2               v17.4s, v17.8h              // extend filter upper half
+        mla                 v2.4S, v22.4s, v23.4s       // v2 accumulates srcp[filterPos[2] + {0..3}] * filter[{0..3}]
+        mla                 v2.4S, v16.4s, v17.4s       // v2 accumulates srcp[filterPos[2] + {4..7}] * filter[{4..7}]
+        ld1                 {v18.8H}, [x11], #16        // srcp[filterPos[3] + {0..7}]
+        mla                 v1.4S, v6.4s, v7.4s         // v1 accumulates srcp[filterPos[1] + {4..7}] * filter[{4..7}]
+        ld1                 {v19.8H}, [x4], #16         // load 8x16-bit at filter+3*filterSize
+        subs                w15, w15, #8                // j -= 8: processed 8/filterSize
+        uxtl                v28.4s, v18.4H              // extend srcp lower half
+        sxtl                v29.4s, v19.4H              // extend filter lower half
+        uxtl2               v18.4s, v18.8H              // extend srcp upper half
+        sxtl2               v19.4s, v19.8h              // extend filter upper half
+        mla                 v3.4S, v28.4s, v29.4s       // v3 accumulates srcp[filterPos[3] + {0..3}] * filter[{0..3}]
+        mla                 v3.4S, v18.4s, v19.4s       // v3 accumulates srcp[filterPos[3] + {4..7}] * filter[{4..7}]
+        b.gt                2b                          // inner loop if filterSize not consumed completely
+        addp                v0.4S, v0.4S, v1.4S         // part01 horizontal pair adding
+        addp                v2.4S, v2.4S, v3.4S         // part23 horizontal pair adding
+        addp                v0.4S, v0.4S, v2.4S         // part0123 horizontal pair adding
+        subs                w2, w2, #4                  // dstW -= 4
+        sshl                v0.4s, v0.4s, v21.4s        // shift right (effectively rigth, as shift is negative); overflow expected
+        smin                v0.4s, v0.4s, v20.4s        // apply min (do not use sqshl)
+        xtn                 v0.4h, v0.4s                // narrow down to 16 bits
+
+        st1                 {v0.4H}, [x1], #8           // write to destination part0123
+        b.gt                1b                          // loop until end of line
+        ret
+endfunc
 
+function ff_hscale16to15_X4_neon_asm, export=1
+        // w0  int shift
+        // x1  int16_t *dst
+        // w2  int dstW
+        // x3  const uint8_t *src
+        // x4  const int16_t *filter
+        // x5  const int32_t *filterPos
+        // w6  int filterSize
+
+        stp                 d8, d9, [sp, #-0x20]!
+        stp                 d10, d11, [sp, #0x10]
+
+        movi                v18.4s, #1
+        movi                v17.4s, #1
+        shl                 v18.4s, v18.4s, #15
+        sub                 v21.4s, v18.4s, v17.4s      // max allowed value
+        dup                 v17.4s, w0                  // read shift
+        neg                 v20.4s, v17.4s              // negate it, so it can be used in sshl (effectively shift right)
+
+        lsl                 w7, w6, #1
+1:
+        ldp                 w8, w9, [x5]
+        ldp                 w10, w11, [x5, #8]
+
+        movi                v16.2d, #0                  // initialize accumulator for idx + 0
+        movi                v17.2d, #0                  // initialize accumulator for idx + 1
+        movi                v18.2d, #0                  // initialize accumulator for idx + 2
+        movi                v19.2d, #0                  // initialize accumulator for idx + 3
+
+        mov                 x12, x4                     // filter + 0
+        add                 x13, x4, x7                 // filter + 1
+        add                 x8, x3, x8, lsl #1          // srcp + filterPos 0
+        add                 x14, x13, x7                // filter + 2
+        add                 x9, x3, x9, lsl #1          // srcp + filterPos 1
+        add                 x15, x14, x7                // filter + 3
+        add                 x10, x3, x10, lsl #1        // srcp + filterPos 2
+        mov                 w0, w6                      // save the filterSize to temporary variable
+        add                 x11, x3, x11, lsl #1        // srcp + filterPos 3
+        add                 x5, x5, #16                 // advance filter position
+        mov                 x16, xzr                    // clear the register x16 used for offsetting the filter values
+
+2:
+        ldr                 q4, [x8], #16               // load src values for idx 0
+        ldr                 q5, [x9], #16               // load src values for idx 1
+        uxtl                v26.4s, v4.4h
+        uxtl2               v4.4s, v4.8h
+        ldr                 q31, [x12, x16]             // load filter values for idx 0
+        ldr                 q6, [x10], #16              // load src values for idx 2
+        sxtl                v22.4s, v31.4h
+        sxtl2               v31.4s, v31.8h
+        mla                 v16.4s, v26.4s, v22.4s      // multiplication of lower half for idx 0
+        uxtl                v25.4s, v5.4h
+        uxtl2               v5.4s, v5.8h
+        ldr                 q30, [x13, x16]             // load filter values for idx 1
+        ldr                 q7, [x11], #16              // load src values for idx 3
+        mla                 v16.4s, v4.4s, v31.4s       // multiplication of upper half for idx 0
+        uxtl                v24.4s, v6.4h
+        sxtl                v8.4s, v30.4h
+        sxtl2               v30.4s, v30.8h
+        mla                 v17.4s, v25.4s, v8.4s       // multiplication of lower half for idx 1
+        ldr                 q29, [x14, x16]             // load filter values for idx 2
+        uxtl2               v6.4s, v6.8h
+        sxtl                v9.4s, v29.4h
+        sxtl2               v29.4s, v29.8h
+        mla                 v17.4s, v5.4s, v30.4s       // multiplication of upper half for idx 1
+        mla                 v18.4s, v24.4s, v9.4s       // multiplication of lower half for idx 2
+        ldr                 q28, [x15, x16]             // load filter values for idx 3
+        uxtl                v23.4s, v7.4h
+        sxtl                v10.4s, v28.4h
+        mla                 v18.4s, v6.4s, v29.4s       // multiplication of upper half for idx 2
+        uxtl2               v7.4s, v7.8h
+        sxtl2               v28.4s, v28.8h
+        mla                 v19.4s, v23.4s, v10.4s      // multiplication of lower half for idx 3
+        sub                 w0, w0, #8
+        cmp                 w0, #8
+        mla                 v19.4s, v7.4s, v28.4s       // multiplication of upper half for idx 3
+
+        add                 x16, x16, #16               // advance filter values indexing
+
+        b.ge                2b
+
+        // 4 iterations left
+
+        sub                 x17, x7, #8                 // step back to wrap up the filter pos for last 4 elements
+
+        ldr                 d4, [x8]                    // load src values for idx 0
+        ldr                 d31, [x12, x17]             // load filter values for idx 0
+        uxtl                v4.4s, v4.4h
+        sxtl                v31.4s, v31.4h
+        ldr                 d5, [x9]                    // load src values for idx 1
+        mla                 v16.4s, v4.4s, v31.4s       // multiplication of upper half for idx 0
+        ldr                 d30, [x13, x17]             // load filter values for idx 1
+        uxtl                v5.4s, v5.4h
+        sxtl                v30.4s, v30.4h
+        ldr                 d6, [x10]                   // load src values for idx 2
+        mla                 v17.4s, v5.4s, v30.4s       // multiplication of upper half for idx 1
+        ldr                 d29, [x14, x17]             // load filter values for idx 2
+        uxtl                v6.4s, v6.4h
+        sxtl                v29.4s, v29.4h
+        ldr                 d7, [x11]                   // load src values for idx 3
+        ldr                 d28, [x15, x17]             // load filter values for idx 3
+        mla                 v18.4s, v6.4s, v29.4s       // multiplication of upper half for idx 2
+        uxtl                v7.4s, v7.4h
+        sxtl                v28.4s, v28.4h
+        addp                v16.4s, v16.4s, v17.4s
+        mla                 v19.4s, v7.4s, v28.4s       // multiplication of upper half for idx 3
+        subs                w2, w2, #4
+        addp                v18.4s, v18.4s, v19.4s
+        addp                v16.4s, v16.4s, v18.4s
+        sshl                v16.4s, v16.4s, v20.4s
+        smin                v16.4s, v16.4s, v21.4s
+        xtn                 v16.4h, v16.4s
+
+        st1                 {v16.4h}, [x1], #8
+        add                 x4, x4, x7, lsl #2
+        b.gt                1b
+
+        ldp                 d8, d9, [sp]
+        ldp                 d10, d11, [sp, #0x10]
+
+        add                 sp, sp, #0x20
+
+        ret
 endfunc
diff --git a/libswscale/aarch64/swscale.c b/libswscale/aarch64/swscale.c
index 171fb51329b..5d0887cfe51 100644
--- a/libswscale/aarch64/swscale.c
+++ b/libswscale/aarch64/swscale.c
@@ -22,6 +22,63 @@
 #include "libswscale/swscale_internal.h"
 #include "libavutil/aarch64/cpu.h"
 
+void ff_hscale16to15_4_neon_asm(int shift, int16_t *_dst, int dstW,
+                      const uint8_t *_src, const int16_t *filter,
+                      const int32_t *filterPos, int filterSize);
+void ff_hscale16to15_X8_neon_asm(int shift, int16_t *_dst, int dstW,
+                      const uint8_t *_src, const int16_t *filter,
+                      const int32_t *filterPos, int filterSize);
+void ff_hscale16to15_X4_neon_asm(int shift, int16_t *_dst, int dstW,
+                      const uint8_t *_src, const int16_t *filter,
+                      const int32_t *filterPos, int filterSize);
+
+static void ff_hscale16to15_4_neon(SwsContext *c, int16_t *_dst, int dstW,
+                      const uint8_t *_src, const int16_t *filter,
+                      const int32_t *filterPos, int filterSize)
+{
+    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(c->srcFormat);
+    int sh              = desc->comp[0].depth - 1;
+
+    if (sh<15) {
+        sh = isAnyRGB(c->srcFormat) || c->srcFormat==AV_PIX_FMT_PAL8 ? 13 : (desc->comp[0].depth - 1);
+    } else if (desc->flags & AV_PIX_FMT_FLAG_FLOAT) { /* float input are process like uint 16bpc */
+        sh = 16 - 1;
+    }
+    ff_hscale16to15_4_neon_asm(sh, _dst, dstW, _src, filter, filterPos, filterSize);
+
+}
+
+static void ff_hscale16to15_X8_neon(SwsContext *c, int16_t *_dst, int dstW,
+                      const uint8_t *_src, const int16_t *filter,
+                      const int32_t *filterPos, int filterSize)
+{
+    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(c->srcFormat);
+    int sh              = desc->comp[0].depth - 1;
+
+    if (sh<15) {
+        sh = isAnyRGB(c->srcFormat) || c->srcFormat==AV_PIX_FMT_PAL8 ? 13 : (desc->comp[0].depth - 1);
+    } else if (desc->flags & AV_PIX_FMT_FLAG_FLOAT) { /* float input are process like uint 16bpc */
+        sh = 16 - 1;
+    }
+    ff_hscale16to15_X8_neon_asm(sh, _dst, dstW, _src, filter, filterPos, filterSize);
+
+}
+
+static void ff_hscale16to15_X4_neon(SwsContext *c, int16_t *_dst, int dstW,
+                      const uint8_t *_src, const int16_t *filter,
+                      const int32_t *filterPos, int filterSize)
+{
+    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(c->srcFormat);
+    int sh              = desc->comp[0].depth - 1;
+
+    if (sh<15) {
+        sh = isAnyRGB(c->srcFormat) || c->srcFormat==AV_PIX_FMT_PAL8 ? 13 : (desc->comp[0].depth - 1);
+    } else if (desc->flags & AV_PIX_FMT_FLAG_FLOAT) { /* float input are process like uint 16bpc */
+        sh = 16 - 1;
+    }
+    ff_hscale16to15_X4_neon_asm(sh, _dst, dstW, _src, filter, filterPos, filterSize);
+}
+
 #define SCALE_FUNC(filter_n, from_bpc, to_bpc, opt) \
 void ff_hscale ## from_bpc ## to ## to_bpc ## _ ## filter_n ## _ ## opt( \
                                                 SwsContext *c, int16_t *data, \
@@ -56,6 +113,10 @@ void ff_yuv2plane1_8_neon(
         } else                                                          \
             hscalefn =                                                  \
                 ff_hscale8to19_ ## filtersize ## _ ## opt;              \
+    } else {                                                            \
+        if (c->dstBpc <= 14)                                            \
+            hscalefn =                                                  \
+                ff_hscale16to15_ ## filtersize ## _ ## opt;             \
     }                                                                   \
 } while (0)
 
diff --git a/libswscale/swscale.c b/libswscale/swscale.c
index 367d045a024..ad0ddc07e0d 100644
--- a/libswscale/swscale.c
+++ b/libswscale/swscale.c
@@ -109,7 +109,6 @@ static void hScale16To15_c(SwsContext *c, int16_t *dst, int dstW,
         int j;
         int srcPos = filterPos[i];
         int val    = 0;
-
         for (j = 0; j < filterSize; j++) {
             val += src[srcPos + j] * filter[filterSize * i + j];
         }
-- 
2.37.1

_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".

^ permalink raw reply	[flat|nested] 5+ messages in thread

* [FFmpeg-devel] [PATCH 3/3] sw_scale: Add specializations for hscale 16 to 19
  2022-10-28 11:34 [FFmpeg-devel] [PATCH 0/3] sw_scale: Provide neon implementation for hscale Hubert Mazur
  2022-10-28 11:34 ` [FFmpeg-devel] [PATCH 1/3] sw_scale: Add specializations for hscale 8 to 19 Hubert Mazur
  2022-10-28 11:34 ` [FFmpeg-devel] [PATCH 2/3] sw_scale: Add specializations for hscale 16 to 15 Hubert Mazur
@ 2022-10-28 11:34 ` Hubert Mazur
  2022-11-01 13:26 ` [FFmpeg-devel] [PATCH 0/3] sw_scale: Provide neon implementation for hscale Martin Storsjö
  3 siblings, 0 replies; 5+ messages in thread
From: Hubert Mazur @ 2022-10-28 11:34 UTC (permalink / raw)
  To: ffmpeg-devel; +Cc: gjb, upstream, jswinney, Hubert Mazur, martin, mw, spop

Provide arm64 neon optimized implementations for hscale16To19 with
filter sizes 4, 8 and X4.

The tests and benchmarks run on AWS Graviton 2 instances.
The results from a checkasm tool are shown below.

hscale_16_to_19__fs_4_dstW_512_c: 6216.0
hscale_16_to_19__fs_4_dstW_512_neon: 2257.0
hscale_16_to_19__fs_8_dstW_512_c: 10417.7
hscale_16_to_19__fs_8_dstW_512_neon: 3112.5
hscale_16_to_19__fs_12_dstW_512_c: 14890.5
hscale_16_to_19__fs_12_dstW_512_neon: 3899.0
hscale_16_to_19__fs_16_dstW_512_c: 19006.5
hscale_16_to_19__fs_16_dstW_512_neon: 5341.2
hscale_16_to_19__fs_32_dstW_512_c: 36629.5
hscale_16_to_19__fs_32_dstW_512_neon: 9502.7
hscale_16_to_19__fs_40_dstW_512_c: 45477.5
hscale_16_to_19__fs_40_dstW_512_neon: 11552.0

Signed-off-by: Hubert Mazur <hum@semihalf.com>
---
 libswscale/aarch64/hscale.S  | 402 +++++++++++++++++++++++++++++++++++
 libswscale/aarch64/swscale.c |  66 ++++++
 2 files changed, 468 insertions(+)

diff --git a/libswscale/aarch64/hscale.S b/libswscale/aarch64/hscale.S
index b571a5dc219..a4fa0709f81 100644
--- a/libswscale/aarch64/hscale.S
+++ b/libswscale/aarch64/hscale.S
@@ -1045,3 +1045,405 @@ function ff_hscale16to15_X4_neon_asm, export=1
 
         ret
 endfunc
+
+function ff_hscale16to19_4_neon_asm, export=1
+        // w0               int shift
+        // x1               int32_t *dst
+        // w2               int dstW
+        // x3               const uint8_t *src // treat it as uint16_t *src
+        // x4               const uint16_t *filter
+        // x5               const int32_t *filterPos
+        // w6               int filterSize
+
+        movi                v18.4s, #1
+        movi                v17.4s, #1
+        shl                 v18.4s, v18.4s, #19
+        sub                 v18.4s, v18.4s, v17.4s      // max allowed value
+        dup                 v17.4s, w0                  // read shift
+        neg                 v17.4s, v17.4s              // negate it, so it can be used in sshl (effectively shift right)
+
+        cmp                 w2, #16
+        b.lt                2f // move to last block
+
+        ldp                 w8, w9, [x5]                // filterPos[0], filterPos[1]
+        ldp                 w10, w11, [x5, #8]          // filterPos[2], filterPos[3]
+        ldp                 w12, w13, [x5, #16]         // filterPos[4], filterPos[5]
+        ldp                 w14, w15, [x5, #24]         // filterPos[6], filterPos[7]
+        add                 x5, x5, #32
+
+        // shift all filterPos left by one, as uint16_t will be read
+        lsl                 x8, x8, #1
+        lsl                 x9, x9, #1
+        lsl                 x10, x10, #1
+        lsl                 x11, x11, #1
+        lsl                 x12, x12, #1
+        lsl                 x13, x13, #1
+        lsl                 x14, x14, #1
+        lsl                 x15, x15, #1
+
+        // load src with given offset
+        ldr                 x8,  [x3, w8,  UXTW]
+        ldr                 x9,  [x3, w9,  UXTW]
+        ldr                 x10, [x3, w10, UXTW]
+        ldr                 x11, [x3, w11, UXTW]
+        ldr                 x12, [x3, w12, UXTW]
+        ldr                 x13, [x3, w13, UXTW]
+        ldr                 x14, [x3, w14, UXTW]
+        ldr                 x15, [x3, w15, UXTW]
+
+        sub                 sp, sp, #64
+        // push src on stack so it can be loaded into vectors later
+        stp                 x8, x9, [sp]
+        stp                 x10, x11, [sp, #16]
+        stp                 x12, x13, [sp, #32]
+        stp                 x14, x15, [sp, #48]
+
+1:
+        ld4                 {v0.8h, v1.8h, v2.8h, v3.8h}, [sp]
+        ld4                 {v28.8h, v29.8h, v30.8h, v31.8h}, [x4], #64 // filter[0..7]
+
+        // Each of blocks does the following:
+        // Extend src and filter to 32 bits with uxtl and sxtl
+        // multiply or multiply and accumulate results
+        // Extending to 32 bits is necessary, as unit16_t values can't
+        // be represented as int16_t without type promotion.
+        uxtl                v26.4s, v0.4h
+        sxtl                v27.4s, v28.4H
+        uxtl2               v0.4s, v0.8h
+        mul                 v5.4s, v26.4s, v27.4s
+        sxtl2               v28.4s, v28.8H
+        uxtl                v26.4s, v1.4h
+        mul                 v6.4s, v0.4s, v28.4s
+
+        sxtl                v27.4s, v29.4H
+        uxtl2               v0.4s, v1.8h
+        mla                 v5.4s, v27.4s, v26.4s
+        sxtl2               v28.4s, v29.8H
+        uxtl                v26.4s, v2.4h
+        mla                 v6.4s, v28.4s, v0.4s
+
+        sxtl                v27.4s, v30.4H
+        uxtl2               v0.4s, v2.8h
+        mla                 v5.4s, v27.4s, v26.4s
+        sxtl2               v28.4s, v30.8H
+        uxtl                v26.4s, v3.4h
+        mla                 v6.4s, v28.4s, v0.4s
+
+        sxtl                v27.4s, v31.4H
+        uxtl2               v0.4s, v3.8h
+        mla                 v5.4s, v27.4s, v26.4s
+        sxtl2               v28.4s, v31.8H
+        sub                 w2, w2, #8
+        mla                 v6.4s, v28.4s, v0.4s
+
+        sshl                v5.4s, v5.4s, v17.4s
+        sshl                v6.4s, v6.4s, v17.4s
+        smin                v5.4s, v5.4s, v18.4s
+        smin                v6.4s, v6.4s, v18.4s
+
+        st1                 {v5.4s, v6.4s}, [x1], #32
+        cmp                 w2, #16
+
+        // load filterPositions into registers for next iteration
+        ldp                 w8, w9, [x5]                // filterPos[0], filterPos[1]
+        ldp                 w10, w11, [x5, #8]          // filterPos[2], filterPos[3]
+        ldp                 w12, w13, [x5, #16]         // filterPos[4], filterPos[5]
+        ldp                 w14, w15, [x5, #24]         // filterPos[6], filterPos[7]
+        add                 x5, x5, #32
+
+        lsl                 x8, x8, #1
+        lsl                 x9, x9, #1
+        lsl                 x10, x10, #1
+        lsl                 x11, x11, #1
+        lsl                 x12, x12, #1
+        lsl                 x13, x13, #1
+        lsl                 x14, x14, #1
+        lsl                 x15, x15, #1
+
+        ldr                 x8,  [x3, w8,  UXTW]
+        ldr                 x9,  [x3, w9,  UXTW]
+        ldr                 x10, [x3, w10, UXTW]
+        ldr                 x11, [x3, w11, UXTW]
+        ldr                 x12, [x3, w12, UXTW]
+        ldr                 x13, [x3, w13, UXTW]
+        ldr                 x14, [x3, w14, UXTW]
+        ldr                 x15, [x3, w15, UXTW]
+
+        stp                 x8, x9, [sp]
+        stp                 x10, x11, [sp, #16]
+        stp                 x12, x13, [sp, #32]
+        stp                 x14, x15, [sp, #48]
+
+        b.ge                1b
+
+        // here we make last iteration, without updating the registers
+        ld4                 {v0.8h, v1.8h, v2.8h, v3.8h}, [sp]
+        ld4                 {v28.8h, v29.8h, v30.8h, v31.8h}, [x4], #64
+
+        uxtl                v26.4s, v0.4h
+        sxtl                v27.4s, v28.4H
+        uxtl2               v0.4s, v0.8h
+        mul                 v5.4s, v26.4s, v27.4s
+        sxtl2               v28.4s, v28.8H
+        uxtl                v26.4s, v1.4h
+        mul                 v6.4s, v0.4s, v28.4s
+
+        sxtl                v27.4s, v29.4H
+        uxtl2               v0.4s, v1.8h
+        mla                 v5.4s, v26.4s, v27.4s
+        sxtl2               v28.4s, v29.8H
+        uxtl                v26.4s, v2.4h
+        mla                 v6.4s, v0.4s, v28.4s
+
+        sxtl                v27.4s, v30.4H
+        uxtl2               v0.4s, v2.8h
+        mla                 v5.4s, v26.4s, v27.4s
+        sxtl2               v28.4s, v30.8H
+        uxtl                v26.4s, v3.4h
+        mla                 v6.4s, v0.4s, v28.4s
+
+        sxtl                v27.4s, v31.4H
+        uxtl2               v0.4s, v3.8h
+        mla                 v5.4s, v26.4s, v27.4s
+        sxtl2               v28.4s, v31.8H
+        subs                w2, w2, #8
+        mla                 v6.4s, v0.4s, v28.4s
+
+        sshl                v5.4s, v5.4s, v17.4s
+        sshl                v6.4s, v6.4s, v17.4s
+
+        smin                v5.4s, v5.4s, v18.4s
+        smin                v6.4s, v6.4s, v18.4s
+
+        st1                 {v5.4s, v6.4s}, [x1], #32
+        add                 sp, sp, #64                 // restore stack
+        cbnz                w2, 2f
+
+        ret
+
+2:
+        ldr                 w8, [x5], #4                // load filterPos
+        lsl                 w8, w8, #1
+        add                 x9, x3, w8, UXTW            // src + filterPos
+        ld1                 {v0.4h}, [x9]               // load 4 * uint16_t
+        ld1                 {v31.4h}, [x4], #8
+
+        uxtl                v0.4s, v0.4h
+        sxtl                v31.4s, v31.4h
+        subs                w2, w2, #1
+        mul                 v5.4s, v0.4s, v31.4s
+        addv                s0, v5.4S
+        sshl                v0.4s, v0.4s, v17.4s
+        smin                v0.4s, v0.4s, v18.4s
+        st1                 {v0.s}[0], [x1], #4
+        cbnz                w2, 2b                      // if iterations remain jump to beginning
+
+        ret
+endfunc
+
+function ff_hscale16to19_X8_neon_asm, export=1
+        // w0               int shift
+        // x1               int32_t *dst
+        // w2               int dstW
+        // x3               const uint8_t *src // treat it as uint16_t *src
+        // x4               const uint16_t *filter
+        // x5               const int32_t *filterPos
+        // w6               int filterSize
+
+        movi                v20.4s, #1
+        movi                v21.4s, #1
+        shl                 v20.4s, v20.4s, #19
+        sub                 v20.4s, v20.4s, v21.4s
+        dup                 v21.4s, w0
+        neg                 v21.4s, v21.4s
+
+        sbfiz               x7, x6, #1, #32             // filterSize*2 (*2 because int16)
+1:      ldr                 w8, [x5], #4                // filterPos[idx]
+        ldr                 w10, [x5], #4               // filterPos[idx + 1]
+        lsl                 w8, w8, #1
+        ldr                 w11, [x5], #4               // filterPos[idx + 2]
+        ldr                 w9, [x5], #4                // filterPos[idx + 3]
+        mov                 x16, x4                     // filter0 = filter
+        lsl                 w11, w11, #1
+        add                 x12, x16, x7                // filter1 = filter0 + filterSize*2
+        lsl                 w9, w9, #1
+        add                 x13, x12, x7                // filter2 = filter1 + filterSize*2
+        lsl                 w10, w10, #1
+        add                 x4, x13, x7                 // filter3 = filter2 + filterSize*2
+        movi                v0.2D, #0                   // val sum part 1 (for dst[0])
+        movi                v1.2D, #0                   // val sum part 2 (for dst[1])
+        movi                v2.2D, #0                   // val sum part 3 (for dst[2])
+        movi                v3.2D, #0                   // val sum part 4 (for dst[3])
+        add                 x17, x3, w8, UXTW           // srcp + filterPos[0]
+        add                 x8,  x3, w10, UXTW          // srcp + filterPos[1]
+        add                 x10, x3, w11, UXTW          // srcp + filterPos[2]
+        add                 x11, x3, w9, UXTW           // srcp + filterPos[3]
+        mov                 w15, w6                     // filterSize counter
+2:      ld1                 {v4.8H}, [x17], #16         // srcp[filterPos[0] + {0..7}]
+        ld1                 {v5.8H}, [x16], #16         // load 8x16-bit filter values, part 1
+        ld1                 {v6.8H}, [x8], #16          // srcp[filterPos[1] + {0..7}]
+        ld1                 {v7.8H}, [x12], #16         // load 8x16-bit at filter+filterSize
+        uxtl                v24.4s, v4.4H               // extend srcp lower half to 32 bits to preserve sign
+        sxtl                v25.4s, v5.4H               // extend filter lower half to 32 bits to match srcp size
+        uxtl2               v4.4s, v4.8h                // extend srcp upper half to 32 bits
+        mla                 v0.4s, v24.4s, v25.4s       // multiply accumulate lower half of v4 * v5
+        sxtl2               v5.4s, v5.8h                // extend filter upper half to 32 bits
+        uxtl                v26.4s, v6.4h               // extend srcp lower half to 32 bits
+        mla                 v0.4S, v4.4s, v5.4s         // multiply accumulate upper half of v4 * v5
+        sxtl                v27.4s, v7.4H               // exted filter lower half
+        uxtl2               v6.4s, v6.8H                // extend srcp upper half
+        sxtl2               v7.4s, v7.8h                // extend filter upper half
+        ld1                 {v16.8H}, [x10], #16        // srcp[filterPos[2] + {0..7}]
+        mla                 v1.4S, v26.4s, v27.4s       // v1 accumulates srcp[filterPos[1] + {0..3}] * filter[{0..3}]
+        ld1                 {v17.8H}, [x13], #16        // load 8x16-bit at filter+2*filterSize
+        uxtl                v22.4s, v16.4H              // extend srcp lower half
+        sxtl                v23.4s, v17.4H              // extend filter lower half
+        uxtl2               v16.4s, v16.8H              // extend srcp upper half
+        sxtl2               v17.4s, v17.8h              // extend filter upper half
+        mla                 v2.4S, v22.4s, v23.4s       // v2 accumulates srcp[filterPos[2] + {0..3}] * filter[{0..3}]
+        mla                 v2.4S, v16.4s, v17.4s       // v2 accumulates srcp[filterPos[2] + {4..7}] * filter[{4..7}]
+        ld1                 {v18.8H}, [x11], #16        // srcp[filterPos[3] + {0..7}]
+        mla                 v1.4S, v6.4s, v7.4s         // v1 accumulates srcp[filterPos[1] + {4..7}] * filter[{4..7}]
+        ld1                 {v19.8H}, [x4], #16         // load 8x16-bit at filter+3*filterSize
+        subs                w15, w15, #8                // j -= 8: processed 8/filterSize
+        uxtl                v28.4s, v18.4H              // extend srcp lower half
+        sxtl                v29.4s, v19.4H              // extend filter lower half
+        uxtl2               v18.4s, v18.8H              // extend srcp upper half
+        sxtl2               v19.4s, v19.8h              // extend filter upper half
+        mla                 v3.4S, v28.4s, v29.4s       // v3 accumulates srcp[filterPos[3] + {0..3}] * filter[{0..3}]
+        mla                 v3.4S, v18.4s, v19.4s       // v3 accumulates srcp[filterPos[3] + {4..7}] * filter[{4..7}]
+        b.gt                2b                          // inner loop if filterSize not consumed completely
+        addp                v0.4S, v0.4S, v1.4S         // part01 horizontal pair adding
+        addp                v2.4S, v2.4S, v3.4S         // part23 horizontal pair adding
+        addp                v0.4S, v0.4S, v2.4S         // part0123 horizontal pair adding
+        subs                w2, w2, #4                  // dstW -= 4
+        sshl                v0.4s, v0.4s, v21.4s        // shift right (effectively rigth, as shift is negative); overflow expected
+        smin                v0.4s, v0.4s, v20.4s        // apply min (do not use sqshl)
+        st1                 {v0.4s}, [x1], #16          // write to destination part0123
+        b.gt                1b                          // loop until end of line
+        ret
+endfunc
+
+function ff_hscale16to19_X4_neon_asm, export=1
+        // w0  int shift
+        // x1  int16_t *dst
+        // w2  int dstW
+        // x3  const uint8_t *src
+        // x4  const int16_t *filter
+        // x5  const int32_t *filterPos
+        // w6  int filterSize
+
+        stp                 d8, d9, [sp, #-0x20]!
+        stp                 d10, d11, [sp, #0x10]
+
+        movi                v18.4s, #1
+        movi                v17.4s, #1
+        shl                 v18.4s, v18.4s, #19
+        sub                 v21.4s, v18.4s, v17.4s      // max allowed value
+        dup                 v17.4s, w0                  // read shift
+        neg                 v20.4s, v17.4s              // negate it, so it can be used in sshl (effectively shift right)
+
+        lsl                 w7, w6, #1
+1:
+        ldp                 w8, w9, [x5]
+        ldp                 w10, w11, [x5, #8]
+
+        movi                v16.2d, #0                  // initialize accumulator for idx + 0
+        movi                v17.2d, #0                  // initialize accumulator for idx + 1
+        movi                v18.2d, #0                  // initialize accumulator for idx + 2
+        movi                v19.2d, #0                  // initialize accumulator for idx + 3
+
+        mov                 x12, x4                     // filter + 0
+        add                 x13, x4, x7                 // filter + 1
+        add                 x8, x3, x8, lsl #1          // srcp + filterPos 0
+        add                 x14, x13, x7                // filter + 2
+        add                 x9, x3, x9, lsl #1          // srcp + filterPos 1
+        add                 x15, x14, x7                // filter + 3
+        add                 x10, x3, x10, lsl #1        // srcp + filterPos 2
+        mov                 w0, w6                      // save the filterSize to temporary variable
+        add                 x11, x3, x11, lsl #1        // srcp + filterPos 3
+        add                 x5, x5, #16                 // advance filter position
+        mov                 x16, xzr                    // clear the register x16 used for offsetting the filter values
+
+2:
+        ldr                 q4, [x8], #16               // load src values for idx 0
+        ldr                 q5, [x9], #16               // load src values for idx 1
+        uxtl                v26.4s, v4.4h
+        uxtl2               v4.4s, v4.8h
+        ldr                 q31, [x12, x16]             // load filter values for idx 0
+        ldr                 q6, [x10], #16              // load src values for idx 2
+        sxtl                v22.4s, v31.4h
+        sxtl2               v31.4s, v31.8h
+        mla                 v16.4s, v26.4s, v22.4s      // multiplication of lower half for idx 0
+        uxtl                v25.4s, v5.4h
+        uxtl2               v5.4s, v5.8h
+        ldr                 q30, [x13, x16]             // load filter values for idx 1
+        ldr                 q7, [x11], #16              // load src values for idx 3
+        mla                 v16.4s, v4.4s, v31.4s       // multiplication of upper half for idx 0
+        uxtl                v24.4s, v6.4h
+        sxtl                v8.4s, v30.4h
+        sxtl2               v30.4s, v30.8h
+        mla                 v17.4s, v25.4s, v8.4s       // multiplication of lower half for idx 1
+        ldr                 q29, [x14, x16]             // load filter values for idx 2
+        uxtl2               v6.4s, v6.8h
+        sxtl                v9.4s, v29.4h
+        sxtl2               v29.4s, v29.8h
+        mla                 v17.4s, v5.4s, v30.4s       // multiplication of upper half for idx 1
+        ldr                 q28, [x15, x16]             // load filter values for idx 3
+        mla                 v18.4s, v24.4s, v9.4s       // multiplication of lower half for idx 2
+        uxtl                v23.4s, v7.4h
+        sxtl                v10.4s, v28.4h
+        mla                 v18.4s, v6.4s, v29.4s       // multiplication of upper half for idx 2
+        uxtl2               v7.4s, v7.8h
+        sxtl2               v28.4s, v28.8h
+        mla                 v19.4s, v23.4s, v10.4s      // multiplication of lower half for idx 3
+        sub                 w0, w0, #8
+        cmp                 w0, #8
+        mla                 v19.4s, v7.4s, v28.4s       // multiplication of upper half for idx 3
+
+        add                 x16, x16, #16               // advance filter values indexing
+
+        b.ge                2b
+
+        // 4 iterations left
+
+        sub                 x17, x7, #8                 // step back to wrap up the filter pos for last 4 elements
+
+        ldr                 d4, [x8]                    // load src values for idx 0
+        ldr                 d31, [x12, x17]             // load filter values for idx 0
+        uxtl                v4.4s, v4.4h
+        sxtl                v31.4s, v31.4h
+        ldr                 d5, [x9]                    // load src values for idx 1
+        mla                 v16.4s, v4.4s, v31.4s       // multiplication of upper half for idx 0
+        ldr                 d30, [x13, x17]             // load filter values for idx 1
+        uxtl                v5.4s, v5.4h
+        sxtl                v30.4s, v30.4h
+        ldr                 d6, [x10]                   // load src values for idx 2
+        mla                 v17.4s, v5.4s, v30.4s       // multiplication of upper half for idx 1
+        ldr                 d29, [x14, x17]             // load filter values for idx 2
+        uxtl                v6.4s, v6.4h
+        sxtl                v29.4s, v29.4h
+        ldr                 d7, [x11]                   // load src values for idx 3
+        ldr                 d28, [x15, x17]             // load filter values for idx 3
+        mla                 v18.4s, v6.4s, v29.4s       // multiplication of upper half for idx 2
+        uxtl                v7.4s, v7.4h
+        sxtl                v28.4s, v28.4h
+        addp                v16.4s, v16.4s, v17.4s
+        mla                 v19.4s, v7.4s, v28.4s       // multiplication of upper half for idx 3
+        subs                w2, w2, #4
+        addp                v18.4s, v18.4s, v19.4s
+        addp                v16.4s, v16.4s, v18.4s
+        sshl                v16.4s, v16.4s, v20.4s
+        smin                v16.4s, v16.4s, v21.4s
+
+        st1                 {v16.4s}, [x1], #16
+        add                 x4, x4, x7, lsl #2
+        b.gt                1b
+
+        ldp                 d8, d9, [sp]
+        ldp                 d10, d11, [sp, #0x10]
+
+        add                 sp, sp, #0x20
+
+        ret
+endfunc
\ No newline at end of file
diff --git a/libswscale/aarch64/swscale.c b/libswscale/aarch64/swscale.c
index 5d0887cfe51..bbd9719a44b 100644
--- a/libswscale/aarch64/swscale.c
+++ b/libswscale/aarch64/swscale.c
@@ -31,6 +31,15 @@ void ff_hscale16to15_X8_neon_asm(int shift, int16_t *_dst, int dstW,
 void ff_hscale16to15_X4_neon_asm(int shift, int16_t *_dst, int dstW,
                       const uint8_t *_src, const int16_t *filter,
                       const int32_t *filterPos, int filterSize);
+void ff_hscale16to19_4_neon_asm(int shift, int16_t *_dst, int dstW,
+                      const uint8_t *_src, const int16_t *filter,
+                      const int32_t *filterPos, int filterSize);
+void ff_hscale16to19_X8_neon_asm(int shift, int16_t *_dst, int dstW,
+                      const uint8_t *_src, const int16_t *filter,
+                      const int32_t *filterPos, int filterSize);
+void ff_hscale16to19_X4_neon_asm(int shift, int16_t *_dst, int dstW,
+                      const uint8_t *_src, const int16_t *filter,
+                      const int32_t *filterPos, int filterSize);
 
 static void ff_hscale16to15_4_neon(SwsContext *c, int16_t *_dst, int dstW,
                       const uint8_t *_src, const int16_t *filter,
@@ -79,6 +88,60 @@ static void ff_hscale16to15_X4_neon(SwsContext *c, int16_t *_dst, int dstW,
     ff_hscale16to15_X4_neon_asm(sh, _dst, dstW, _src, filter, filterPos, filterSize);
 }
 
+static void ff_hscale16to19_4_neon(SwsContext *c, int16_t *_dst, int dstW,
+                           const uint8_t *_src, const int16_t *filter,
+                           const int32_t *filterPos, int filterSize)
+{
+    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(c->srcFormat);
+    int bits            = desc->comp[0].depth - 1;
+    int sh              = bits - 4;
+
+    if ((isAnyRGB(c->srcFormat) || c->srcFormat==AV_PIX_FMT_PAL8) && desc->comp[0].depth<16) {
+        sh = 9;
+    } else if (desc->flags & AV_PIX_FMT_FLAG_FLOAT) { /* float input are process like uint 16bpc */
+        sh = 16 - 1 - 4;
+    }
+
+    ff_hscale16to19_4_neon_asm(sh, _dst, dstW, _src, filter, filterPos, filterSize);
+
+}
+
+static void ff_hscale16to19_X8_neon(SwsContext *c, int16_t *_dst, int dstW,
+                           const uint8_t *_src, const int16_t *filter,
+                           const int32_t *filterPos, int filterSize)
+{
+    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(c->srcFormat);
+    int bits            = desc->comp[0].depth - 1;
+    int sh              = bits - 4;
+
+    if ((isAnyRGB(c->srcFormat) || c->srcFormat==AV_PIX_FMT_PAL8) && desc->comp[0].depth<16) {
+        sh = 9;
+    } else if (desc->flags & AV_PIX_FMT_FLAG_FLOAT) { /* float input are process like uint 16bpc */
+        sh = 16 - 1 - 4;
+    }
+
+    ff_hscale16to19_X8_neon_asm(sh, _dst, dstW, _src, filter, filterPos, filterSize);
+
+}
+
+static void ff_hscale16to19_X4_neon(SwsContext *c, int16_t *_dst, int dstW,
+                           const uint8_t *_src, const int16_t *filter,
+                           const int32_t *filterPos, int filterSize)
+{
+    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(c->srcFormat);
+    int bits            = desc->comp[0].depth - 1;
+    int sh              = bits - 4;
+
+    if ((isAnyRGB(c->srcFormat) || c->srcFormat==AV_PIX_FMT_PAL8) && desc->comp[0].depth<16) {
+        sh = 9;
+    } else if (desc->flags & AV_PIX_FMT_FLAG_FLOAT) { /* float input are process like uint 16bpc */
+        sh = 16 - 1 - 4;
+    }
+
+    ff_hscale16to19_X4_neon_asm(sh, _dst, dstW, _src, filter, filterPos, filterSize);
+
+}
+
 #define SCALE_FUNC(filter_n, from_bpc, to_bpc, opt) \
 void ff_hscale ## from_bpc ## to ## to_bpc ## _ ## filter_n ## _ ## opt( \
                                                 SwsContext *c, int16_t *data, \
@@ -117,6 +180,9 @@ void ff_yuv2plane1_8_neon(
         if (c->dstBpc <= 14)                                            \
             hscalefn =                                                  \
                 ff_hscale16to15_ ## filtersize ## _ ## opt;             \
+        else                                                            \
+            hscalefn =                                                  \
+                ff_hscale16to19_ ## filtersize ## _ ## opt;             \
     }                                                                   \
 } while (0)
 
-- 
2.37.1

_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".

^ permalink raw reply	[flat|nested] 5+ messages in thread

* Re: [FFmpeg-devel] [PATCH 0/3] sw_scale: Provide neon implementation for hscale
  2022-10-28 11:34 [FFmpeg-devel] [PATCH 0/3] sw_scale: Provide neon implementation for hscale Hubert Mazur
                   ` (2 preceding siblings ...)
  2022-10-28 11:34 ` [FFmpeg-devel] [PATCH 3/3] sw_scale: Add specializations for hscale 16 to 19 Hubert Mazur
@ 2022-11-01 13:26 ` Martin Storsjö
  3 siblings, 0 replies; 5+ messages in thread
From: Martin Storsjö @ 2022-11-01 13:26 UTC (permalink / raw)
  To: Hubert Mazur; +Cc: gjb, upstream, jswinney, ffmpeg-devel, mw, spop

On Fri, 28 Oct 2022, Hubert Mazur wrote:

> This patchset contains arm64 neon implementation of hscale functions.
> Fixed minor style issues and declared C function wrappers as static.
> This patchset do not contain the patch for checkasm tool, as the
> previous one did. The reason behind it was failing tests on x86 arch
> but not on aarch64 or loongarch. Probably the hscale functions on x86
> have some bugs. Currently the checkasm tool does not check
> the validity of hscale functions on x86 at all. Implementation of hscale
> for x86 should be fixed anyway. As it comes to aarch64 the tests were
> passing. Attaching a link to fate result in patchwork. After fixing x86
> the patch for checkasm could be merged.

It's too bad that we don't have that test mergeable at this point, but I 
guess we can't hold back the new aarch64 assembly due to that...

The patches seem fine as noted before; I found some more cases of missing 
trailing newlines at the end of the file, and some other minor stray 
changes in the patches, which I fixed. I also amended the commit messages 
saying that some of the benchmarks are from the non-merged checkasm test.

With that, I pushed these patches.

// Martin

_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".

^ permalink raw reply	[flat|nested] 5+ messages in thread

end of thread, other threads:[~2022-11-01 13:26 UTC | newest]

Thread overview: 5+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2022-10-28 11:34 [FFmpeg-devel] [PATCH 0/3] sw_scale: Provide neon implementation for hscale Hubert Mazur
2022-10-28 11:34 ` [FFmpeg-devel] [PATCH 1/3] sw_scale: Add specializations for hscale 8 to 19 Hubert Mazur
2022-10-28 11:34 ` [FFmpeg-devel] [PATCH 2/3] sw_scale: Add specializations for hscale 16 to 15 Hubert Mazur
2022-10-28 11:34 ` [FFmpeg-devel] [PATCH 3/3] sw_scale: Add specializations for hscale 16 to 19 Hubert Mazur
2022-11-01 13:26 ` [FFmpeg-devel] [PATCH 0/3] sw_scale: Provide neon implementation for hscale Martin Storsjö

Git Inbox Mirror of the ffmpeg-devel mailing list - see https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

This inbox may be cloned and mirrored by anyone:

	git clone --mirror https://master.gitmailbox.com/ffmpegdev/0 ffmpegdev/git/0.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 ffmpegdev ffmpegdev/ https://master.gitmailbox.com/ffmpegdev \
		ffmpegdev@gitmailbox.com
	public-inbox-index ffmpegdev

Example config snippet for mirrors.


AGPL code for this site: git clone https://public-inbox.org/public-inbox.git