PR #24255 opened by zuxy URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/24255 Patch URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/24255.patch
Also fixed a bug in handling negative weights in 10-bit asm implementation. Signed-off-by: Zuxy Meng <[email protected]> >From 56cba6021b87fdf778290649549bb540530c34c1 Mon Sep 17 00:00:00 2001 From: Zuxy Meng <[email protected]> Date: Fri, 21 Aug 2026 22:04:42 -0700 Subject: [PATCH] checkasm/h264: add weight tests Also fixed a bug in handling negative weights in 10-bit asm implementation. Signed-off-by: Zuxy Meng <[email protected]> --- libavcodec/x86/h264_weight_10bit.asm | 9 ++- tests/checkasm/h264dsp.c | 103 +++++++++++++++++++++++++++ 2 files changed, 109 insertions(+), 3 deletions(-) diff --git a/libavcodec/x86/h264_weight_10bit.asm b/libavcodec/x86/h264_weight_10bit.asm index 356871bc62..7b7e367c9d 100644 --- a/libavcodec/x86/h264_weight_10bit.asm +++ b/libavcodec/x86/h264_weight_10bit.asm @@ -54,8 +54,10 @@ SECTION .text movd m2, r3m pslld m0, m2 ; 1<<log2_denom SPLATW m0, m0 + add r4w, r4w + movzx r4d, r4w shl r5, 19 ; *8, move to upper half of dword - lea r5, [r5+r4*2+0x10000] + lea r5, [r5+r4+0x10000] movd m3, r5d ; weight<<1 | 1+(offset<<(3)) pshufd m3, m3, 0 mova m4, [pw_pixel_max] @@ -82,7 +84,7 @@ SECTION .text psrad m6, m2 %if cpuflag(sse4) packusdw m5, m6 - pminsw m5, m4 + pminuw m5, m4 %else packssdw m5, m6 CLIPW m5, m7, m4 @@ -176,6 +178,7 @@ DECLARE_REG_TMP 7 %macro BIWEIGHT_SETUP 0 lea t0, [t0*4+1] ; (offset<<2)+1 or t0, 1 + and r5d, 0xFFFF shl r6, 16 or r5, r6 movd m4, r5d ; weightd | weights @@ -214,7 +217,7 @@ DECLARE_REG_TMP 7 psrad m2, m6 %if cpuflag(sse4) packusdw m0, m2 - pminsw m0, m3 + pminuw m0, m3 %else packssdw m0, m2 CLIPW m0, m7, m3 diff --git a/tests/checkasm/h264dsp.c b/tests/checkasm/h264dsp.c index 680fd91576..6d0649ebb4 100644 --- a/tests/checkasm/h264dsp.c +++ b/tests/checkasm/h264dsp.c @@ -500,6 +500,103 @@ static void check_loop_filter_intra(void) } } +static void check_weight(void) +{ + LOCAL_ALIGNED_16(uint8_t, dst, [32 * 32 * 2]); + LOCAL_ALIGNED_16(uint8_t, dst0, [32 * 32 * 2]); + LOCAL_ALIGNED_16(uint8_t, dst1, [32 * 32 * 2]); + H264DSPContext h; + declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *dst, ptrdiff_t stride, + int height, int log2_denom, int weight, int offset); + + for (int bit_depth = 8; bit_depth <= 10; bit_depth += 2) { + ff_h264dsp_init(&h, bit_depth, 1); + uint32_t mask = pixel_mask[bit_depth - 8]; + for (int w = 16; w >= 2; w >>= 1) { + int idx = 4 - av_log2(w); + + if (check_func(h.weight_pixels_tab[idx], "weight_%dx%d_%d", + w, 16, bit_depth)) { + for (int hgt = 16; hgt >= 2; hgt >>= 1) { + for (int i = 0; i < 32; i++) { + int stride = 32 * SIZEOF_PIXEL; + int log2_denom = rnd() % 8; + int weight = (rnd() % 256) - 128; + int offset = (rnd() % (1 << bit_depth)) - + (1 << (bit_depth - 1)); + for (int y = 0; y < hgt; y++) { + for (int x = 0; x < w * SIZEOF_PIXEL; x += 4) { + AV_WN32A(dst + y * stride + x, rnd() & mask); + } + } + memcpy(dst0, dst, 32 * 32 * 2); + memcpy(dst1, dst, 32 * 32 * 2); + call_ref(dst0, stride, hgt, log2_denom, weight, offset); + call_new(dst1, stride, hgt, log2_denom, weight, offset); + if (memcmp(dst0, dst1, 32 * 32 * 2)) + fail(); + bench_new(dst, stride, hgt, log2_denom, weight, offset); + } + } + } + } + } +} + +static void check_biweight(void) +{ + LOCAL_ALIGNED_16(uint8_t, dst, [32 * 32 * 2]); + LOCAL_ALIGNED_16(uint8_t, dst0, [32 * 32 * 2]); + LOCAL_ALIGNED_16(uint8_t, dst1, [32 * 32 * 2]); + LOCAL_ALIGNED_16(uint8_t, src, [32 * 32 * 2]); + LOCAL_ALIGNED_16(uint8_t, src0, [32 * 32 * 2]); + LOCAL_ALIGNED_16(uint8_t, src1, [32 * 32 * 2]); + H264DSPContext h; + declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *dst, uint8_t *src, + ptrdiff_t stride, int height, int log2_denom, + int weightd, int weights, int offset); + + for (int bit_depth = 8; bit_depth <= 10; bit_depth += 2) { + uint32_t mask = pixel_mask[bit_depth - 8]; + ff_h264dsp_init(&h, bit_depth, 1); + for (int w = 16; w >= 2; w >>= 1) { + int idx = 4 - av_log2(w); + + if (check_func(h.biweight_pixels_tab[idx], "biweight_%dx%d_%d", + w, 16, bit_depth)) { + for (int hgt = 16; hgt >= 4; hgt >>= 1) { + for (int i = 0; i < 32; i++) { + int stride = 32 * SIZEOF_PIXEL; + int log2_denom = rnd() % 4; // 8-bit fails if > 3 + int weightd = rnd() % 256 - 128; + int weights = rnd() % 256 - 128; + int offset = (rnd() % (1 << bit_depth)) - + (1 << (bit_depth - 1)); + for (int y = 0; y < hgt; y++) { + for (int x = 0; x < w * SIZEOF_PIXEL; x += 4) { + AV_WN32A(dst + y * stride + x, rnd() & mask); + AV_WN32A(src + y * stride + x, rnd() & mask); + } + } + memcpy(dst0, dst, 32 * 32 * 2); + memcpy(dst1, dst, 32 * 32 * 2); + memcpy(src0, src, 32 * 32 * 2); + memcpy(src1, src, 32 * 32 * 2); + call_ref(dst0, src0, stride, hgt, log2_denom, weightd, + weights, offset); + call_new(dst1, src1, stride, hgt, log2_denom, weightd, + weights, offset); + if (memcmp(dst0, dst1, 32 * 32 * 2)) + fail(); + bench_new(dst, src, stride, hgt, log2_denom, weightd, + weights, offset); + } + } + } + } + } +} + void checkasm_check_h264dsp(void) { check_idct(); @@ -512,4 +609,10 @@ void checkasm_check_h264dsp(void) check_loop_filter_intra(); report("loop_filter_intra"); + + check_weight(); + report("weight"); + + check_biweight(); + report("biweight"); } -- 2.52.0 _______________________________________________ ffmpeg-devel mailing list -- [email protected] To unsubscribe send an email to [email protected]
