From 2537fdc510ba2b5eca3e5c7a2b6d773498daa418 Mon Sep 17 00:00:00 2001 From: Hubert Mazur Date: Fri, 28 Oct 2022 11:34:39 +0000 Subject: [PATCH] sw_scale: Add specializations for hscale 16 to 19 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 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 (Note, the checkasm tests for these functions haven't been merged since they fail on x86.) Signed-off-by: Hubert Mazur Signed-off-by: Martin Storsjö --- 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 b571a5dc21..8d4dcb2541 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 diff --git a/libswscale/aarch64/swscale.c b/libswscale/aarch64/swscale.c index 5d0887cfe5..bbd9719a44 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)