Integrate film grain in the task threading system

1 files changed, 65 insertions, 44 deletions

diff --git a/src/fg_apply_tmpl.c b/src/fg_apply_tmpl.c
index c254a3d..1ffc6e0 100644
--- a/src/fg_apply_tmpl.c
+++ b/src/fg_apply_tmpl.c
@@ -89,14 +89,13 @@ static void generate_scaling(const int bitdepth,
 }
 
 #ifndef UNIT_TEST
-void bitfn(dav1d_apply_grain)(const Dav1dFilmGrainDSPContext *const dsp,
-                              Dav1dPicture *const out,
-                              const Dav1dPicture *const in)
+void bitfn(dav1d_prep_grain)(const Dav1dFilmGrainDSPContext *const dsp,
+                             Dav1dPicture *const out,
+                             const Dav1dPicture *const in,
+                             uint8_t scaling[3][SCALING_SIZE],
+                             entry grain_lut[3][GRAIN_HEIGHT+1][GRAIN_WIDTH])
 {
     const Dav1dFilmGrainData *const data = &out->frame_hdr->film_grain.data;
-
-    ALIGN_STK_16(entry, grain_lut, 3,[GRAIN_HEIGHT + 1][GRAIN_WIDTH]);
-    uint8_t scaling[3][SCALING_SIZE];
 #if BITDEPTH != 8
     const int bitdepth_max = (1 << out->p.bpc) - 1;
 #endif
@@ -150,60 +149,82 @@ void bitfn(dav1d_apply_grain)(const Dav1dFilmGrainDSPContext *const dsp,
                 memcpy(out->data[2], in->data[2], sz);
         }
     }
+}
 
+void bitfn(dav1d_apply_grain_row)(const Dav1dFilmGrainDSPContext *const dsp,
+                                  Dav1dPicture *const out,
+                                  const Dav1dPicture *const in,
+                                  const uint8_t scaling[3][SCALING_SIZE],
+                                  const entry grain_lut[3][GRAIN_HEIGHT+1][GRAIN_WIDTH],
+                                  const int row)
+{
     // Synthesize grain for the affected planes
-    const int rows = (out->p.h + 31) >> 5;
+    const Dav1dFilmGrainData *const data = &out->frame_hdr->film_grain.data;
     const int ss_y = in->p.layout == DAV1D_PIXEL_LAYOUT_I420;
     const int ss_x = in->p.layout != DAV1D_PIXEL_LAYOUT_I444;
     const int cpw = (out->p.w + ss_x) >> ss_x;
     const int is_id = out->seq_hdr->mtrx == DAV1D_MC_IDENTITY;
-    for (int row = 0; row < rows; row++) {
-        pixel *const luma_src =
-            ((pixel *) in->data[0]) + row * BLOCK_SIZE * PXSTRIDE(in->stride[0]);
-
-        if (data->num_y_points) {
-            const int bh = imin(out->p.h - row * BLOCK_SIZE, BLOCK_SIZE);
-            dsp->fgy_32x32xn(((pixel *) out->data[0]) + row * BLOCK_SIZE * PXSTRIDE(out->stride[0]),
-                             luma_src, out->stride[0], data,
-                             out->p.w, scaling[0], grain_lut[0], bh, row HIGHBD_TAIL_SUFFIX);
-        }
+    pixel *const luma_src =
+        ((pixel *) in->data[0]) + row * BLOCK_SIZE * PXSTRIDE(in->stride[0]);
+#if BITDEPTH != 8
+    const int bitdepth_max = (1 << out->p.bpc) - 1;
+#endif
 
-        if (!data->num_uv_points[0] && !data->num_uv_points[1] &&
-            !data->chroma_scaling_from_luma)
-        {
-            continue;
-        }
+    if (data->num_y_points) {
+        const int bh = imin(out->p.h - row * BLOCK_SIZE, BLOCK_SIZE);
+        dsp->fgy_32x32xn(((pixel *) out->data[0]) + row * BLOCK_SIZE * PXSTRIDE(out->stride[0]),
+                         luma_src, out->stride[0], data,
+                         out->p.w, scaling[0], grain_lut[0], bh, row HIGHBD_TAIL_SUFFIX);
+    }
 
-        const int bh = (imin(out->p.h - row * BLOCK_SIZE, BLOCK_SIZE) + ss_y) >> ss_y;
+    if (!data->num_uv_points[0] && !data->num_uv_points[1] &&
+        !data->chroma_scaling_from_luma)
+    {
+        return;
+    }
 
-        // extend padding pixels
-        if (out->p.w & ss_x) {
-            pixel *ptr = luma_src;
-            for (int y = 0; y < bh; y++) {
-                ptr[out->p.w] = ptr[out->p.w - 1];
-                ptr += PXSTRIDE(in->stride[0]) << ss_y;
-            }
+    const int bh = (imin(out->p.h - row * BLOCK_SIZE, BLOCK_SIZE) + ss_y) >> ss_y;
+
+    // extend padding pixels
+    if (out->p.w & ss_x) {
+        pixel *ptr = luma_src;
+        for (int y = 0; y < bh; y++) {
+            ptr[out->p.w] = ptr[out->p.w - 1];
+            ptr += PXSTRIDE(in->stride[0]) << ss_y;
         }
+    }
 
-        const ptrdiff_t uv_off = row * BLOCK_SIZE * PXSTRIDE(out->stride[1]) >> ss_y;
-        if (data->chroma_scaling_from_luma) {
-            for (int pl = 0; pl < 2; pl++)
+    const ptrdiff_t uv_off = row * BLOCK_SIZE * PXSTRIDE(out->stride[1]) >> ss_y;
+    if (data->chroma_scaling_from_luma) {
+        for (int pl = 0; pl < 2; pl++)
+            dsp->fguv_32x32xn[in->p.layout - 1](((pixel *) out->data[1 + pl]) + uv_off,
+                                                ((const pixel *) in->data[1 + pl]) + uv_off,
+                                                in->stride[1], data, cpw,
+                                                scaling[0], grain_lut[1 + pl],
+                                                bh, row, luma_src, in->stride[0],
+                                                pl, is_id HIGHBD_TAIL_SUFFIX);
+    } else {
+        for (int pl = 0; pl < 2; pl++)
+            if (data->num_uv_points[pl])
                 dsp->fguv_32x32xn[in->p.layout - 1](((pixel *) out->data[1 + pl]) + uv_off,
                                                     ((const pixel *) in->data[1 + pl]) + uv_off,
                                                     in->stride[1], data, cpw,
-                                                    scaling[0], grain_lut[1 + pl],
+                                                    scaling[1 + pl], grain_lut[1 + pl],
                                                     bh, row, luma_src, in->stride[0],
                                                     pl, is_id HIGHBD_TAIL_SUFFIX);
-        } else {
-            for (int pl = 0; pl < 2; pl++)
-                if (data->num_uv_points[pl])
-                    dsp->fguv_32x32xn[in->p.layout - 1](((pixel *) out->data[1 + pl]) + uv_off,
-                                                        ((const pixel *) in->data[1 + pl]) + uv_off,
-                                                        in->stride[1], data, cpw,
-                                                        scaling[1 + pl], grain_lut[1 + pl],
-                                                        bh, row, luma_src, in->stride[0],
-                                                        pl, is_id HIGHBD_TAIL_SUFFIX);
-        }
     }
 }
+
+void bitfn(dav1d_apply_grain)(const Dav1dFilmGrainDSPContext *const dsp,
+                              Dav1dPicture *const out,
+                              const Dav1dPicture *const in)
+{
+    ALIGN_STK_16(entry, grain_lut, 3,[GRAIN_HEIGHT + 1][GRAIN_WIDTH]);
+    uint8_t scaling[3][SCALING_SIZE];
+    const int rows = (out->p.h + 31) >> 5;
+
+    bitfn(dav1d_prep_grain)(dsp, out, in, scaling, grain_lut);
+    for (int row = 0; row < rows; row++)
+        bitfn(dav1d_apply_grain_row)(dsp, out, in, scaling, grain_lut, row);
+}
 #endif