This is an automated email from the git hooks/post-receive script. Git pushed a commit to branch master in repository ffmpeg.
commit 7626d21bf9a2dd8fac9d9de67b6070b80641e60a Author: Kacper Michajłow <[email protected]> AuthorDate: Thu Oct 29 10:44:22 2026 +0100 Commit: Kacper Michajłow <[email protected]> CommitDate: Mon Nov 2 18:19:15 2026 +0000 swscale/uops: round every float operation of the reference backend With x87 math the compiler keeps the intermediate results of an expression with excess precision. Round them, so that they produce the same results. Signed-off-by: Kacper Michajłow <[email protected]> --- libswscale/uops.h | 2 +- libswscale/uops_tmpl.c | 52 ++++++++++++++++++++++++++++++++------------------ 2 files changed, 34 insertions(+), 20 deletions(-) diff --git a/libswscale/uops.h b/libswscale/uops.h index 506b3ff622..4d25107927 100644 --- a/libswscale/uops.h +++ b/libswscale/uops.h @@ -94,7 +94,7 @@ static inline SwsPixel ff_sws_pixel_from_q64(SwsPixelType type, AVRational64 val case SWS_PIXEL_U8: return (SwsPixel) { .u8 = val.num / val.den }; case SWS_PIXEL_U16: return (SwsPixel) { .u16 = val.num / val.den }; case SWS_PIXEL_U32: return (SwsPixel) { .u32 = val.num / val.den }; - case SWS_PIXEL_F32: return (SwsPixel) { .f32 = (float) val.num / val.den }; + case SWS_PIXEL_F32: return (SwsPixel) { .f32 = (float) val.num / (float) val.den }; case SWS_PIXEL_NONE: case SWS_PIXEL_TYPE_NB: break; } diff --git a/libswscale/uops_tmpl.c b/libswscale/uops_tmpl.c index e94bfca417..56e28749b1 100644 --- a/libswscale/uops_tmpl.c +++ b/libswscale/uops_tmpl.c @@ -61,6 +61,9 @@ # error Invalid BIT_DEPTH #endif +/* Every intermediate float result is cast to pixel_t, which rounds away the + * excess precision of x87 math for the same results as everywhere else */ + /********************************* * Generic read/write operations * *********************************/ @@ -285,10 +288,10 @@ DECL_READ(read_planar_fv, const SwsCompMask mask, const SwsPixelType type) SWS_LOOP for (int i = 0; i < SWS_BLOCK_SIZE; i++) { - if (X) xs.f32[i] += weight * in0[i]; - if (Y) ys.f32[i] += weight * in1[i]; - if (Z) zs.f32[i] += weight * in2[i]; - if (W) ws.f32[i] += weight * in3[i]; + if (X) xs.f32[i] += (float) (weight * in0[i]); + if (Y) ys.f32[i] += (float) (weight * in1[i]); + if (Z) zs.f32[i] += (float) (weight * in2[i]); + if (W) ws.f32[i] += (float) (weight * in3[i]); } if (X) in0 = bump_ptr(in0, exec->in_stride[0]); @@ -355,10 +358,10 @@ DECL_READ(read_planar_fh, const SwsCompMask mask, const SwsPixelType type) inter_t sx = 0, sy = 0, sz = 0, sw = 0; for (int j = 0; j < filter_size; j++) { const int weight = weights[j]; - if (X) sx += weight * start0[j]; - if (Y) sy += weight * start1[j]; - if (Z) sz += weight * start2[j]; - if (W) sw += weight * start3[j]; + if (X) sx += (inter_t) (weight * start0[j]); + if (Y) sy += (inter_t) (weight * start1[j]); + if (Z) sz += (inter_t) (weight * start2[j]); + if (W) sw += (inter_t) (weight * start3[j]); } if (X) xs.f32[i] = (float) sx * scale; @@ -838,7 +841,7 @@ DECL_FUNC(linear, const SwsCompMask mask, const uint32_t one, const uint32_t zer const pixel_t ww = w[i]; #define LIN_VAL(I, J, val) \ - ((one & SWS_MASK(I, J)) ? (val) : (uinter_t) c.m[I][J] * (val)) + ((one & SWS_MASK(I, J)) ? (val) : (pixel_t) ((uinter_t) c.m[I][J] * (val))) #define LIN_ROW(I, var) do { \ pixel_t tmp = (zero & SWS_MASK(I, 4)) ? 0 : c.k[I]; \ @@ -879,7 +882,18 @@ av_always_inline static vec3_t fn(vec3)(v3u16_t v) return (vec3_t) { v.x, v.y, v.z }; } -#define lerp(a, b, w) ((a) + (w) * ((pixel_t) (b) - (a))) +#define lerp(a, b, w) ((pixel_t) ((a) + (pixel_t) ((w) * (pixel_t) ((pixel_t) (b) - (a))))) + +av_always_inline static +pixel_t fn(sum4)(pixel_t w0, pixel_t a, pixel_t w1, pixel_t b, + pixel_t w2, pixel_t c, pixel_t w3, pixel_t d) +{ + pixel_t sum = (pixel_t) (w0 * a); + sum += (pixel_t) (w1 * b); + sum += (pixel_t) (w2 * c); + sum += (pixel_t) (w3 * d); + return sum; +} av_always_inline static vec3_t fn(lerp3)(vec3_t a, vec3_t b, pixel_t w) @@ -932,16 +946,16 @@ vec3_t fn(lut3d_static)(const SwsLut3D *restrict lut3d, vec3_t rgb) const vec3_t v3 = fn(vec3)(base[off0 + off1 + off2]); return (vec3_t) { - w0 * v0.x + w1 * v1.x + w2 * v2.x + w3 * v3.x, - w0 * v0.y + w1 * v1.y + w2 * v2.y + w3 * v3.y, - w0 * v0.z + w1 * v1.z + w2 * v2.z + w3 * v3.z, + fn(sum4)(w0, v0.x, w1, v1.x, w2, v2.x, w3, v3.x), + fn(sum4)(w0, v0.y, w1, v1.y, w2, v2.y, w3, v3.y), + fn(sum4)(w0, v0.z, w1, v1.z, w2, v2.z, w3, v3.z), }; } av_always_inline static vec3_t fn(lut3d_dynamic)(const SwsLut3D *restrict lut3d, vec3_t rgb) { - rgb.x *= (TONE_LUT_SIZE - 1) / (pixel_t) UINT16_MAX; + rgb.x *= (pixel_t) ((TONE_LUT_SIZE - 1) / (pixel_t) UINT16_MAX); /* Linear interpolation */ const int Ix = (int) rgb.x; @@ -955,13 +969,13 @@ vec3_t fn(lut3d_dynamic)(const SwsLut3D *restrict lut3d, vec3_t rgb) const pixel_t scale = k / (pixel_t) (1 << 15); rgb.x = lerp(a.x, b.x, If); - rgb.y = bias + scale * rgb.y; - rgb.z = bias + scale * rgb.z; + rgb.y = bias + (pixel_t) (scale * rgb.y); + rgb.z = bias + (pixel_t) (scale * rgb.z); /* Re-scale to output LUT size */ - rgb.x *= (OUTPUT_LUT_SIZE_I - 1) / (pixel_t) UINT16_MAX; - rgb.y *= (OUTPUT_LUT_SIZE_PT - 1) / (pixel_t) UINT16_MAX; - rgb.z *= (OUTPUT_LUT_SIZE_PT - 1) / (pixel_t) UINT16_MAX; + rgb.x *= (pixel_t) ((OUTPUT_LUT_SIZE_I - 1) / (pixel_t) UINT16_MAX); + rgb.y *= (pixel_t) ((OUTPUT_LUT_SIZE_PT - 1) / (pixel_t) UINT16_MAX); + rgb.z *= (pixel_t) ((OUTPUT_LUT_SIZE_PT - 1) / (pixel_t) UINT16_MAX); /* Trilinear interpolation */ const int lo0 = (int) rgb.x; -- To stop receiving notification emails like this one, please contact [email protected]. _______________________________________________ ffmpeg-cvslog mailing list -- [email protected] To unsubscribe send an email to [email protected]
