jellyfin/jellyfin-ffmpeg
1
0
Fork 0
mirror of https://github.com/jellyfin/jellyfin-ffmpeg.git synced 2026-09-03 05:09:58 +03:00
Code Issues Packages Projects Releases Wiki Activity

avfilter/tonemap_opencl: vectorize dovi reshaping

Browse Source 
avfilter/tonemap_opencl: vectorize Dolby Vision reshaping

Vectorize the Dolby Vision reshaping process to handle 4 pixels at a time, similar
to what is used in the tone mapping stage. This change allows us to read the dovi
buffer once and reuse parameters for 4 pixels which reduces memory load.
Most of the pipeline, including the common polynomial reshaping, benefits from
vectorization. The only exceptions are the coefficients lookup for intensity channel
and MMR reshaping: intensity lookup remains scalar because it processes 4 vec4
values that must be handled sequentially, and MMR reshaping cannot be vectorized
due to its inter-channel dot products.
With this optimization, 4K60 Dolby Vision tone mapping is now reliably achievable
on the RK3588.
This commit is contained in:
gnattu
2025-05-11 22:01:56 +08:00
parent e34b5cb73f
commit 01866d77af
1 changed files with 172 additions and 68 deletions
Download Patch File Download Diff File Expand all files Collapse all files
debian/patches/0007-add-bt2390-eetf-and-code-refactor-to-opencl-tonemap.patch
Vendored
+172 -68
View File
@@ -610,7 +610,7 @@ Index: FFmpeg/libavfilter/opencl/tonemap.cl
float j = tone_param;
float a, b;
@@ -71,202 +106,687 @@ float mobius(float s, float peak) {
@@ -71,202 +106,790 @@ float mobius(float s, float peak) {
return s;
a = -j * j * (peak - 1.0f) / (j * j - 2.0f * j + peak);
@@ -853,10 +853,8 @@ Index: FFmpeg/libavfilter/opencl/tonemap.cl
+ float3 c = yuv2lrgb(yuv);
+#endif
+ return c;
}
-float3 map_one_pixel_rgb(float3 rgb, float peak, float average) {
- float sig = max(max(rgb.x, max(rgb.y, rgb.z)), 1e-6f);
+}
+
+// Map from source space YUV to destination space RGB
+float3 map_to_dst_space_from_yuv(float3 yuv) {
+#ifdef DOVI_RESHAPE
@@ -869,6 +867,13 @@ Index: FFmpeg/libavfilter/opencl/tonemap.cl
+ c = lrgb2lrgb(c);
+#endif
+ return c;
}
-float3 map_one_pixel_rgb(float3 rgb, float peak, float average) {
- float sig = max(max(rgb.x, max(rgb.y, rgb.z)), 1e-6f);
+#ifdef DOVI_RESHAPE
+vec4 reshape_polyx4(vec4 s, vec4 coeffsx, vec4 coeffsy, vec4 coeffsz) {
+ return (coeffsz * s + coeffsy) * s + coeffsx;
+}
- // Rescale the variables in order to bring it into a representation where
@@ -877,13 +882,6 @@ Index: FFmpeg/libavfilter/opencl/tonemap.cl
- if (target_peak > 1.0f) {
- sig *= 1.0f / target_peak;
- peak *= 1.0f / target_peak;
- }
+#ifdef DOVI_RESHAPE
+vec reshape_poly(vec s, vec4 coeffs) {
+ return (coeffs.z * s + coeffs.y) * s + coeffs.x;
+}
- float sig_old = sig;
+vec reshape_mmr(vec3 sig,
+ vec4 coeffs,
+ __global const vec4 *dovi_mmr,
@@ -913,8 +911,10 @@ Index: FFmpeg/libavfilter/opencl/tonemap.cl
+ s += dot(dovi_mmr[mmr_idx + 4].xyz, sig2 * sig);
+ s += dot(dovi_mmr[mmr_idx + 5], sigX2 * sigX);
+ }
+ }
}
- float sig_old = sig;
-
- // Scale the signal to compensate for differences in the average brightness
- float slope = min(1.0f, sdr_avg / average);
- sig *= slope;
@@ -929,26 +929,63 @@ Index: FFmpeg/libavfilter/opencl/tonemap.cl
- coeff = native_powr(coeff, 10.0f / desat_param);
- rgb = mix(rgb, (float3)luma, (float3)coeff);
- sig = mix(sig, luma * slope, coeff);
+void reshape_dovi_yuv(float3 *yuv,
+ __global const vec *src_dovi_params,
+ __global const vec *src_dovi_pivots,
+ __global const vec4 *src_dovi_coeffs,
+ __global const vec4 *src_dovi_mmr)
+vec4 reshape_mmrx4(vec4 sig_i4,
+ vec4 sig_p4,
+ vec4 sig_t4,
+ vec4 coeffsx,
+ vec4 coeffsy,
+ vec4 coeffsz,
+ vec4 coeffsw,
+ __global const vec4 *dovi_mmr,
+ uchar dovi_mmr_single,
+ uchar dovi_min_order,
+ uchar dovi_max_order)
+{
+ bool t, has_mmr_poly;
+ #ifdef DOVI_PERF_FP16
+ vec3 sig = convert_half3(clamp((*yuv).xyz, 0.0f, 1.0f));
+ #else
+ vec3 sig = clamp((*yuv).xyz, 0.0f, 1.0f);
+ #endif
+ vec s;
+ vec4 coeffs;
+ vec4 coeffs = (vec4)(coeffsx.x, coeffsy.x, coeffsz.x, coeffsw.x);
+ vec4 result = (vec4)M_ZERO_VEC;
+ result.x = reshape_mmr((vec3)(sig_i4.x, sig_p4.x, sig_t4.x), coeffs, dovi_mmr,
+ dovi_mmr_single, dovi_min_order, dovi_max_order);
+ coeffs = (vec4)(coeffsx.y, coeffsy.y, coeffsz.y, coeffsw.y);
+ result.y = reshape_mmr((vec3)(sig_i4.y, sig_p4.y, sig_t4.y), coeffs, dovi_mmr,
+ dovi_mmr_single, dovi_min_order, dovi_max_order);
+ coeffs = (vec4)(coeffsx.z, coeffsy.z, coeffsz.z, coeffsw.z);
+ result.z = reshape_mmr((vec3)(sig_i4.z, sig_p4.z, sig_t4.z), coeffs, dovi_mmr,
+ dovi_mmr_single, dovi_min_order, dovi_max_order);
+ coeffs = (vec4)(coeffsx.w, coeffsy.w, coeffsz.w, coeffsw.w);
+ result.w = reshape_mmr((vec3)(sig_i4.w, sig_p4.w, sig_t4.w), coeffs, dovi_mmr,
+ dovi_mmr_single, dovi_min_order, dovi_max_order);
+ return result;
+}
+
+void reshape_dovi_iptx4(float3 *ipt0,
+ float3 *ipt1,
+ float3 *ipt2,
+ float3 *ipt3,
+ __global const vec *src_dovi_params,
+ __global const vec *src_dovi_pivots,
+ __global const vec4 *src_dovi_coeffs,
+ __global const vec4 *src_dovi_mmr)
+{
+ bool has_mmr_poly, t;
+ vec4 do_poly, coeffw_is_zero;
+ vec4 coeffs, coeffsx, coeffsy, coeffsz, coeffsw, sx4;
+ float4 result;
+ uchar dovi_num_pivots, dovi_has_mmr, dovi_has_poly;
+ uchar dovi_mmr_single, dovi_min_order, dovi_max_order;
+ vec dovi_lo, dovi_hi;
+ __global const vec *dovi_params;
+ __global const vec *dovi_pivots;
+ __global const vec4 *dovi_coeffs, *dovi_mmr;
+ vec dovi_lo, dovi_hi;
+
+ #ifdef DOVI_PERF_FP16
+ vec4 sig_i4 = convert_half4(clamp((vec4)((*ipt0).x,(*ipt1).x,(*ipt2).x,(*ipt3).x), 0.0f, 1.0f));
+ vec4 sig_p4 = convert_half4(clamp((vec4)((*ipt0).y,(*ipt1).y,(*ipt2).y,(*ipt3).y), 0.0f, 1.0f));
+ vec4 sig_t4 = convert_half4(clamp((vec4)((*ipt0).z,(*ipt1).z,(*ipt2).z,(*ipt3).z), 0.0f, 1.0f));
+ #else
+ vec4 sig_i4 = clamp((vec4)((*ipt0).x,(*ipt1).x,(*ipt2).x,(*ipt3).x), 0.0f, 1.0f);
+ vec4 sig_p4 = clamp((vec4)((*ipt0).y,(*ipt1).y,(*ipt2).y,(*ipt3).y), 0.0f, 1.0f);
+ vec4 sig_t4 = clamp((vec4)((*ipt0).z,(*ipt1).z,(*ipt2).z,(*ipt3).z), 0.0f, 1.0f);
+ #endif
+
+ // Reshape I
+ dovi_params = src_dovi_params;
@@ -964,8 +1001,11 @@ Index: FFmpeg/libavfilter/opencl/tonemap.cl
+ dovi_lo = dovi_params[6];
+ dovi_hi = dovi_params[7];
+
+ s = sig.x;
+ coeffs = dovi_coeffs[0];
+ coeffsx = (vec4)coeffs.x;
+ coeffsy = (vec4)coeffs.y;
+ coeffsz = (vec4)coeffs.z;
+ coeffsw = (vec4)coeffs.w;
+
+ if (dovi_num_pivots > 2) {
+ #ifdef DOVI_PERF_FP16
@@ -983,19 +1023,57 @@ Index: FFmpeg/libavfilter/opencl/tonemap.cl
+ const vec8 coeffs2 = (vec8)(dovi_coeffs[4], dovi_coeffs[5]);
+ const vec8 coeffs3 = (vec8)(dovi_coeffs[6], dovi_coeffs[7]);
+ #endif
+ coeffs = mix(mix(mix(coeffs0.lo, coeffs0.hi, (vec4)(s >= pivots[0])),
+ mix(coeffs1.lo, coeffs1.hi, (vec4)(s >= pivots[2])),
+ (vec4)(s >= pivots[1])),
+ mix(mix(coeffs2.lo, coeffs2.hi, (vec4)(s >= pivots[4])),
+ mix(coeffs3.lo, coeffs3.hi, (vec4)(s >= pivots[6])),
+ (vec4)(s >= pivots[5])),
+ (vec4)(s >= pivots[3]));
+ vec4 coeffs_temp = mix(mix(mix(coeffs0.lo, coeffs0.hi, (vec4)(sig_i4.x >= pivots[0])),
+ mix(coeffs1.lo, coeffs1.hi, (vec4)(sig_i4.x >= pivots[2])),
+ (vec4)(sig_i4.x >= pivots[1])),
+ mix(mix(coeffs2.lo, coeffs2.hi, (vec4)(sig_i4.x >= pivots[4])),
+ mix(coeffs3.lo, coeffs3.hi, (vec4)(sig_i4.x >= pivots[6])),
+ (vec4)(sig_i4.x >= pivots[5])),
+ (vec4)(sig_i4.x >= pivots[3]));
+ coeffsx.x = coeffs_temp.x;
+ coeffsy.x = coeffs_temp.y;
+ coeffsz.x = coeffs_temp.z;
+ coeffsw.x = coeffs_temp.w;
+
+ coeffs_temp = mix(mix(mix(coeffs0.lo, coeffs0.hi, (vec4)(sig_i4.y >= pivots[0])),
+ mix(coeffs1.lo, coeffs1.hi, (vec4)(sig_i4.y >= pivots[2])),
+ (vec4)(sig_i4.y >= pivots[1])),
+ mix(mix(coeffs2.lo, coeffs2.hi, (vec4)(sig_i4.y >= pivots[4])),
+ mix(coeffs3.lo, coeffs3.hi, (vec4)(sig_i4.y >= pivots[6])),
+ (vec4)(sig_i4.y >= pivots[5])),
+ (vec4)(sig_i4.y >= pivots[3]));
+ coeffsx.y = coeffs_temp.x;
+ coeffsy.y = coeffs_temp.y;
+ coeffsz.y = coeffs_temp.z;
+ coeffsw.y = coeffs_temp.w;
+
+ coeffs_temp = mix(mix(mix(coeffs0.lo, coeffs0.hi, (vec4)(sig_i4.z >= pivots[0])),
+ mix(coeffs1.lo, coeffs1.hi, (vec4)(sig_i4.z >= pivots[2])),
+ (vec4)(sig_i4.z >= pivots[1])),
+ mix(mix(coeffs2.lo, coeffs2.hi, (vec4)(sig_i4.z >= pivots[4])),
+ mix(coeffs3.lo, coeffs3.hi, (vec4)(sig_i4.z >= pivots[6])),
+ (vec4)(sig_i4.z >= pivots[5])),
+ (vec4)(sig_i4.z >= pivots[3]));
+ coeffsx.z = coeffs_temp.x;
+ coeffsy.z = coeffs_temp.y;
+ coeffsz.z = coeffs_temp.z;
+ coeffsw.z = coeffs_temp.w;
+
+ coeffs_temp = mix(mix(mix(coeffs0.lo, coeffs0.hi, (vec4)(sig_i4.w >= pivots[0])),
+ mix(coeffs1.lo, coeffs1.hi, (vec4)(sig_i4.w >= pivots[2])),
+ (vec4)(sig_i4.w >= pivots[1])),
+ mix(mix(coeffs2.lo, coeffs2.hi, (vec4)(sig_i4.w >= pivots[4])),
+ mix(coeffs3.lo, coeffs3.hi, (vec4)(sig_i4.w >= pivots[6])),
+ (vec4)(sig_i4.w >= pivots[5])),
+ (vec4)(sig_i4.w >= pivots[3]));
+ coeffsx.w = coeffs_temp.x;
+ coeffsy.w = coeffs_temp.y;
+ coeffsz.w = coeffs_temp.z;
+ coeffsw.w = coeffs_temp.w;
}
- sig = TONE_FUNC(sig, peak);
+ has_mmr_poly = dovi_has_mmr && dovi_has_poly;
+ t = (has_mmr_poly && coeffs.w == M_ZERO_VEC) || (!has_mmr_poly && dovi_has_poly);
-
- sig = min(sig, 1.0f);
- rgb *= (sig/sig_old);
- return rgb;
@@ -1007,14 +1085,30 @@ Index: FFmpeg/libavfilter/opencl/tonemap.cl
- c = lrgb2lrgb(c);
- return c;
-}
+ s = t ? reshape_poly(s, coeffs)
+ : reshape_mmr(sig, coeffs, dovi_mmr,
+ dovi_mmr_single, dovi_min_order, dovi_max_order);
+ has_mmr_poly = dovi_has_mmr && dovi_has_poly;
+ #ifdef DOVI_PERF_FP16
+ coeffw_is_zero = convert_half4(coeffsw == (vec4)M_ZERO_VEC);
+ #else
+ coeffw_is_zero = coeffsw == (vec4)M_ZERO_VEC;
+ #endif
+ do_poly = has_mmr_poly
+ ? coeffw_is_zero
+ : (vec4)(dovi_has_poly != M_ZERO_VEC);
+
+ sx4 = mix(reshape_mmrx4(sig_i4, sig_p4, sig_t4,
+ coeffsx, coeffsy, coeffsz, coeffsw, dovi_mmr,
+ dovi_mmr_single, dovi_min_order, dovi_max_order),
+ reshape_polyx4(sig_i4, coeffsx, coeffsy, coeffsz),
+ do_poly);
+ #ifdef DOVI_PERF_FP16
+ (*yuv).x = convert_float(clamp(s, dovi_lo, dovi_hi));
+ result = convert_float4(clamp(sx4, dovi_lo, dovi_hi));
+ #else
+ (*yuv).x = clamp(s, dovi_lo, dovi_hi);
+ result = clamp(sx4, dovi_lo, dovi_hi);
+ #endif
+ (*ipt0).x = result.x;
+ (*ipt1).x = result.y;
+ (*ipt2).x = result.z;
+ (*ipt3).x = result.w;
+
+ // Reshape P
+ dovi_params = src_dovi_params + 1*8;
@@ -1030,20 +1124,28 @@ Index: FFmpeg/libavfilter/opencl/tonemap.cl
+ dovi_lo = dovi_params[6];
+ dovi_hi = dovi_params[7];
+
+ s = sig.y;
+ coeffs = dovi_coeffs[0];
+ coeffsx = (vec4)coeffs.x;
+ coeffsy = (vec4)coeffs.y;
+ coeffsz = (vec4)coeffs.z;
+ coeffsw = (vec4)coeffs.w;
+
+ has_mmr_poly = dovi_has_mmr && dovi_has_poly;
+ t = (has_mmr_poly && coeffs.w == M_ZERO_VEC) || (!has_mmr_poly && dovi_has_poly);
+ t = has_mmr_poly ? coeffs.w == M_ZERO_VEC : dovi_has_poly != M_ZERO_VEC;
+
+ s = t ? reshape_poly(s, coeffs)
+ : reshape_mmr(sig, coeffs, dovi_mmr,
+ dovi_mmr_single, dovi_min_order, dovi_max_order);
+ #ifdef DOVI_PERF_FP16
+ (*yuv).y = convert_float(clamp(s, dovi_lo, dovi_hi));
+ #else
+ (*yuv).y = clamp(s, dovi_lo, dovi_hi);
+ #endif
+ sx4 = t ? reshape_polyx4(sig_p4, coeffsx, coeffsy, coeffsz)
+ : reshape_mmrx4(sig_i4, sig_p4, sig_t4,
+ coeffsx, coeffsy, coeffsz, coeffsw, dovi_mmr,
+ dovi_mmr_single, dovi_min_order, dovi_max_order);
+#ifdef DOVI_PERF_FP16
+ result = convert_float4(clamp(sx4, dovi_lo, dovi_hi));
+#else
+ result = clamp(sx4, dovi_lo, dovi_hi);
+#endif
+ (*ipt0).y = result.x;
+ (*ipt1).y = result.y;
+ (*ipt2).y = result.z;
+ (*ipt3).y = result.w;
+
+ // Reshape T
+ dovi_params = src_dovi_params + 2*8;
@@ -1059,20 +1161,28 @@ Index: FFmpeg/libavfilter/opencl/tonemap.cl
+ dovi_lo = dovi_params[6];
+ dovi_hi = dovi_params[7];
+
+ s = sig.z;
+ coeffs = dovi_coeffs[0];
+ coeffsx = (vec4)coeffs.x;
+ coeffsy = (vec4)coeffs.y;
+ coeffsz = (vec4)coeffs.z;
+ coeffsw = (vec4)coeffs.w;
+
+ has_mmr_poly = dovi_has_mmr && dovi_has_poly;
+ t = (has_mmr_poly && coeffs.w == M_ZERO_VEC) || (!has_mmr_poly && dovi_has_poly);
+ t = has_mmr_poly ? coeffs.w == M_ZERO_VEC : dovi_has_poly != M_ZERO_VEC;
+
+ s = t ? reshape_poly(s, coeffs)
+ : reshape_mmr(sig, coeffs, dovi_mmr,
+ dovi_mmr_single, dovi_min_order, dovi_max_order);
+ sx4 = t ? reshape_polyx4(sig_t4, coeffsx, coeffsy, coeffsz)
+ : reshape_mmrx4(sig_i4, sig_p4, sig_t4,
+ coeffsx, coeffsy, coeffsz, coeffsw, dovi_mmr,
+ dovi_mmr_single, dovi_min_order, dovi_max_order);
+ #ifdef DOVI_PERF_FP16
+ (*yuv).z = convert_float(clamp(s, dovi_lo, dovi_hi));
+ result = convert_float4(clamp(sx4, dovi_lo, dovi_hi));
+ #else
+ (*yuv).z = clamp(s, dovi_lo, dovi_hi);
+ result = clamp(sx4, dovi_lo, dovi_hi);
+ #endif
+ (*ipt0).z = result.x;
+ (*ipt1).z = result.y;
+ (*ipt2).z = result.z;
+ (*ipt3).z = result.w;
+}
+#endif
+
@@ -1161,10 +1271,7 @@ Index: FFmpeg/libavfilter/opencl/tonemap.cl
+ __global const vec *dovi_pivots = dovi_buf + 24;
+ __global const vec4 *dovi_coeffs = (__global const vec4 *)(dovi_buf + 48);
+ __global const vec4 *dovi_mmr = (__global const vec4 *)(dovi_buf + 144);
+ reshape_dovi_yuv(&yuv0, dovi_params, dovi_pivots, dovi_coeffs, dovi_mmr);
+ reshape_dovi_yuv(&yuv1, dovi_params, dovi_pivots, dovi_coeffs, dovi_mmr);
+ reshape_dovi_yuv(&yuv2, dovi_params, dovi_pivots, dovi_coeffs, dovi_mmr);
+ reshape_dovi_yuv(&yuv3, dovi_params, dovi_pivots, dovi_coeffs, dovi_mmr);
+ reshape_dovi_iptx4(&yuv0, &yuv1, &yuv2, &yuv3, dovi_params, dovi_pivots, dovi_coeffs, dovi_mmr);
+#endif
+
+ float3 c0, c1, c2, c3;
@@ -1400,10 +1507,7 @@ Index: FFmpeg/libavfilter/opencl/tonemap.cl
+ __global const vec *dovi_pivots = dovi_buf + 24;
+ __global const vec4 *dovi_coeffs = (__global const vec4 *)(dovi_buf + 48);
+ __global const vec4 *dovi_mmr = (__global const vec4 *)(dovi_buf + 144);
+ reshape_dovi_yuv(&yuv0, dovi_params, dovi_pivots, dovi_coeffs, dovi_mmr);
+ reshape_dovi_yuv(&yuv1, dovi_params, dovi_pivots, dovi_coeffs, dovi_mmr);
+ reshape_dovi_yuv(&yuv2, dovi_params, dovi_pivots, dovi_coeffs, dovi_mmr);
+ reshape_dovi_yuv(&yuv3, dovi_params, dovi_pivots, dovi_coeffs, dovi_mmr);
+ reshape_dovi_iptx4(&yuv0, &yuv1, &yuv2, &yuv3, dovi_params, dovi_pivots, dovi_coeffs, dovi_mmr);
+#endif
+
+ float3 c0, c1, c2, c3;