Precalculate CSC params

6 files changed, 98 insertions, 89 deletions

diff --git a/src/host1x-gr2d.c b/src/host1x-gr2d.c
index aea1b31..a459bac 100644
--- a/src/host1x-gr2d.c
+++ b/src/host1x-gr2d.c
@@ -23,24 +23,9 @@
 
 #include "vdpau_tegra.h"
 
-#define FLOAT_TO_FIXED_6_12(fp) \
-    (((int32_t) (fp * 4096.0f + 0.5f)) & ((1 << 18) - 1))
-
-#define FLOAT_TO_FIXED_s2_7(fp) \
-    (((fp < 0.0f) << 9) | (((int32_t) (fabs(fp) * 128.0f)) & ((1 << 9) - 1)))
-
-#define FLOAT_TO_FIXED_s1_7(fp) \
-    (((fp < 0.0f) << 8) | (((int32_t) (fabs(fp) * 128.0f)) & ((1 << 8) - 1)))
-
-#define FLOAT_TO_FIXED_0_8(fp) \
-    (((int32_t) (fp * 256.0f + 0.5f)) & ((1 << 8) - 1))
-
-static float fclamp(float f, float min, float max)
-{
-    if (f < min) return min;
-    if (f > max) return max;
-    return f;
-}
+struct host1x_csc_params csc_rgb_default = {
+    .cvr = 0x80, .cub = 0x80, .cyx = 0x80,
+};
 
 int host1x_gr2d_clear(struct tegra_stream *stream,
                       struct host1x_pixelbuffer *pixbuf,
@@ -339,7 +324,7 @@ static uint32_t sb_offset(struct host1x_pixelbuffer *pixbuf,
 int host1x_gr2d_surface_blit(struct tegra_stream *stream,
                              struct host1x_pixelbuffer *src,
                              struct host1x_pixelbuffer *dst,
-                             VdpCSCMatrix *cscmat,
+                             struct host1x_csc_params *csc,
                              unsigned int sx, unsigned int sy,
                              unsigned int src_width, unsigned int src_height,
                              unsigned int dx, unsigned int dy,
@@ -487,30 +472,15 @@ coords_check:
     if (src->format != PIX_BUF_FMT_YV12) {
         tegra_stream_push(stream, HOST1X_OPCODE_MASK(0x15, 0x787));
 
-        tegra_stream_push(stream,
-                /* cvr */ FLOAT_TO_FIXED_s2_7(1.0f) << 12 |
-                /* cub */ FLOAT_TO_FIXED_s2_7(1.0f)); /* cscfirst */
-        tegra_stream_push(stream,
-                /* cyx */ FLOAT_TO_FIXED_s1_7(1.0f) << 24 |
-                /* cur */ FLOAT_TO_FIXED_s2_7(0.0f) << 12 |
-                /* cug */ FLOAT_TO_FIXED_s1_7(0.0f)); /* cscsecond */
-        tegra_stream_push(stream,
-                /* cvb */ FLOAT_TO_FIXED_s2_7(0.0f) << 16 |
-                /* cvg */ FLOAT_TO_FIXED_s1_7(0.0f)); /* cscthird */
+        tegra_stream_push(stream, csc->yos << 24 | csc->cvr << 12 | csc->cub); /* cscfirst */
+        tegra_stream_push(stream, csc->cyx << 24 | csc->cur << 12 | csc->cug); /* cscsecond */
+        tegra_stream_push(stream, csc->cvb << 16 | csc->cvg); /* cscthird */
     } else {
         tegra_stream_push(stream, HOST1X_OPCODE_MASK(0x15, 0x7E7));
 
-        tegra_stream_push(stream,
-                /* yos */ (-16) << 24 |
-                /* cvr */ FLOAT_TO_FIXED_s2_7( fclamp((*cscmat)[0][2], -3.98f, 3.98f) ) << 12 |
-                /* cub */ FLOAT_TO_FIXED_s2_7( fclamp((*cscmat)[2][1], -3.98f, 3.98f) )); /* cscfirst */
-        tegra_stream_push(stream,
-                /* cyx */ FLOAT_TO_FIXED_s1_7( fclamp((*cscmat)[0][0], -1.98f, 1.98f) ) << 24 |
-                /* cur */ FLOAT_TO_FIXED_s2_7( fclamp((*cscmat)[0][1], -3.98f, 3.98f) ) << 12 |
-                /* cug */ FLOAT_TO_FIXED_s1_7( fclamp((*cscmat)[1][1], -1.98f, 1.98f) )); /* cscsecond */
-        tegra_stream_push(stream,
-                /* cvb */ FLOAT_TO_FIXED_s2_7( fclamp((*cscmat)[2][2], -3.98f, 3.98f) ) << 16 |
-                /* cvg */ FLOAT_TO_FIXED_s1_7( fclamp((*cscmat)[1][2], -1.98f, 1.98f) )); /* cscthird */
+        tegra_stream_push(stream, csc->yos << 24 | csc->cvr << 12 | csc->cub); /* cscfirst */
+        tegra_stream_push(stream, csc->cyx << 24 | csc->cur << 12 | csc->cug); /* cscsecond */
+        tegra_stream_push(stream, csc->cvb << 16 | csc->cvg); /* cscthird */
 
         tegra_stream_push_reloc(stream, src->bos[1], src->bo_offset[1]); /* uba */
         tegra_stream_push_reloc(stream, src->bos[2], src->bo_offset[2]); /* vba */
diff --git a/src/host1x.h b/src/host1x.h
index 06a20ca..15125f0 100644
--- a/src/host1x.h
+++ b/src/host1x.h
@@ -25,6 +25,18 @@
 #ifndef HOST1X_H
 #define HOST1X_H
 
+#define FLOAT_TO_FIXED_6_12(fp) \
+    (((int32_t) (fp * 4096.0f + 0.5f)) & ((1 << 18) - 1))
+
+#define FLOAT_TO_FIXED_s2_7(fp) \
+    (((fp < 0.0f) << 9) | (((int32_t) (fabs(fp) * 128.0f)) & ((1 << 9) - 1)))
+
+#define FLOAT_TO_FIXED_s1_7(fp) \
+    (((fp < 0.0f) << 8) | (((int32_t) (fabs(fp) * 128.0f)) & ((1 << 8) - 1)))
+
+#define FLOAT_TO_FIXED_0_8(fp) \
+    (((int32_t) (fp * 256.0f + 0.5f)) & ((1 << 8) - 1))
+
 #define HOST1X_OPCODE_SETCL(offset, classid, mask) \
     ((0x0 << 28) | (((offset) & 0xfff) << 16) | (((classid) & 0x3ff) << 6) | ((mask) & 0x3f))
 #define HOST1X_OPCODE_INCR(offset, count) \
@@ -78,6 +90,14 @@ enum layout_format {
     PIX_BUF_LAYOUT_TILED_16x16,
 };
 
+struct host1x_csc_params {
+    uint32_t yos, cvr, cub;
+    uint32_t cyx, cur, cug;
+    uint32_t cvb, cvg;
+};
+
+extern struct host1x_csc_params csc_rgb_default;
+
 struct host1x_pixelbuffer {
     union {
         struct drm_tegra_bo *bo;
@@ -149,7 +169,7 @@ int host1x_gr2d_blit(struct tegra_stream *stream,
 int host1x_gr2d_surface_blit(struct tegra_stream *stream,
                              struct host1x_pixelbuffer *src,
                              struct host1x_pixelbuffer *dst,
-                             VdpCSCMatrix *cscmat,
+                             struct host1x_csc_params *csc,
                              unsigned int sx, unsigned int sy,
                              unsigned int src_width, unsigned int src_height,
                              unsigned int dx, unsigned int dy,
diff --git a/src/surface_mixer.c b/src/surface_mixer.c
index eba10b9..c6b68e7 100644
--- a/src/surface_mixer.c
+++ b/src/surface_mixer.c
@@ -19,6 +19,44 @@
 
 #include "vdpau_tegra.h"
 
+static bool custom_csc(VdpCSCMatrix const csc_matrix)
+{
+    int i, k;
+
+    if (memcmp(csc_matrix, CSC_BT_601, sizeof(VdpCSCMatrix)) == 0 ||
+            memcmp(csc_matrix, CSC_BT_709, sizeof(VdpCSCMatrix)) == 0)
+                return false;
+
+    for (i = 0; i < 3; i++)
+        for (k = 0; k < 3; k++)
+            if (fabs(csc_matrix[i][k] - CSC_BT_601[i][k]) > 0.01f)
+                goto check_709;
+
+    return false;
+
+    /* XXX: Tegra's CSC is hardcoded to BT601 in the kernel driver */
+check_709:
+    for (i = 0; i < 3; i++)
+        for (k = 0; k < 3; k++)
+            if (fabs(csc_matrix[i][k] - CSC_BT_709[i][k]) > 0.01f)
+                return true;
+
+    return false;
+}
+
+static void mixer_apply_vdp_csc(tegra_mixer *mix, VdpCSCMatrix const cscmat)
+{
+    mix->csc.yos = -16;
+    mix->csc.cvr = FLOAT_TO_FIXED_s2_7( CLAMP(cscmat[0][2], -3.98f, 3.98f) );
+    mix->csc.cub = FLOAT_TO_FIXED_s2_7( CLAMP(cscmat[2][1], -3.98f, 3.98f) );
+    mix->csc.cyx = FLOAT_TO_FIXED_s1_7( CLAMP(cscmat[0][0], -1.98f, 1.98f) );
+    mix->csc.cur = FLOAT_TO_FIXED_s2_7( CLAMP(cscmat[0][1], -3.98f, 3.98f) );
+    mix->csc.cug = FLOAT_TO_FIXED_s1_7( CLAMP(cscmat[1][1], -1.98f, 1.98f) );
+    mix->csc.cvb = FLOAT_TO_FIXED_s2_7( CLAMP(cscmat[2][2], -3.98f, 3.98f) );
+    mix->csc.cvg = FLOAT_TO_FIXED_s1_7( CLAMP(cscmat[1][2], -1.98f, 1.98f) );
+    mix->custom_csc = custom_csc(cscmat);
+}
+
 VdpStatus vdp_video_mixer_query_feature_support(VdpDevice device,
                                                 VdpVideoMixerFeature feature,
                                                 VdpBool *is_supported)
@@ -207,6 +245,8 @@ VdpStatus vdp_video_mixer_create(VdpDevice device,
     ref_device(dev);
     mix->dev = dev;
 
+    mixer_apply_vdp_csc(mix, CSC_BT_709);
+
     *mixer = i;
 
     return VDP_STATUS_OK;
@@ -243,8 +283,7 @@ VdpStatus vdp_video_mixer_set_attribute_values(
     while (count--) {
         switch (attributes[count]) {
         case VDP_VIDEO_MIXER_ATTRIBUTE_CSC_MATRIX:
-            memcpy(&mix->csc_matrix, attribute_values[count],
-                   sizeof(VdpCSCMatrix));
+            mixer_apply_vdp_csc(mix, attribute_values[count]);
             break;
 
         case VDP_VIDEO_MIXER_ATTRIBUTE_BACKGROUND_COLOR:
@@ -465,7 +504,7 @@ VdpStatus vdp_video_mixer_render(
         host1x_gr2d_surface_blit(mix->dev->stream,
                                  bg_surf->pixbuf,
                                  dest_surf->pixbuf,
-                                 &mix->csc_matrix,
+                                 &csc_rgb_default,
                                  bg_x0, bg_y0,
                                  bg_width, bg_height,
                                  0,
@@ -475,17 +514,18 @@ VdpStatus vdp_video_mixer_render(
     }
 
     if (!draw_background) {
-        shared = create_shared_surface(dest_surf,
-                                       video_surf,
-                                       &mix->csc_matrix,
-                                       src_vid_x0,
-                                       src_vid_y0,
-                                       src_vid_width,
-                                       src_vid_height,
-                                       dst_vid_x0,
-                                       dst_vid_y0,
-                                       dst_vid_width,
-                                       dst_vid_height);
+        if (!mix->custom_csc)
+            shared = create_shared_surface(dest_surf,
+                                           video_surf,
+                                           &mix->csc,
+                                           src_vid_x0,
+                                           src_vid_y0,
+                                           src_vid_width,
+                                           src_vid_height,
+                                           dst_vid_x0,
+                                           dst_vid_y0,
+                                           dst_vid_width,
+                                           dst_vid_height);
         if (!shared) {
             ret = dynamic_alloc_surface_data(dest_surf);
             if (ret) {
@@ -509,7 +549,7 @@ VdpStatus vdp_video_mixer_render(
         host1x_gr2d_surface_blit(mix->dev->stream,
                                  video_surf->pixbuf,
                                  dest_surf->pixbuf,
-                                 &mix->csc_matrix,
+                                 &mix->csc,
                                  src_vid_x0,
                                  src_vid_y0,
                                  src_vid_width,
diff --git a/src/surface_output.c b/src/surface_output.c
index f5ebdd9..adb55c9 100644
--- a/src/surface_output.c
+++ b/src/surface_output.c
@@ -315,7 +315,7 @@ VdpStatus vdp_output_surface_render_bitmap_surface(
         host1x_gr2d_surface_blit(dst_surf->dev->stream,
                                  src_surf->pixbuf,
                                  dst_surf->pixbuf,
-                                 NULL,
+                                 &csc_rgb_default,
                                  src_x0, src_y0,
                                  src_width, src_height,
                                  dst_x0, dst_y0,
diff --git a/src/surface_shared.c b/src/surface_shared.c
index 0958933..8a0d6ae 100644
--- a/src/surface_shared.c
+++ b/src/surface_shared.c
@@ -61,34 +61,9 @@ static XvImage * create_video_xv(tegra_surface *video)
     return xv_img;
 }
 
-static bool custom_csc(VdpCSCMatrix const *csc_matrix)
-{
-    int i, k;
-
-    if (memcmp(*csc_matrix, CSC_BT_601, sizeof(VdpCSCMatrix)) == 0 ||
-            memcmp(*csc_matrix, CSC_BT_709, sizeof(VdpCSCMatrix)) == 0)
-                return false;
-
-    for (i = 0; i < 3; i++)
-        for (k = 0; k < 3; k++)
-            if (fabs((*csc_matrix)[i][k] - CSC_BT_601[i][k]) > 0.01f)
-                goto check_709;
-
-    return false;
-
-    /* XXX: Tegra's CSC is hardcoded to BT601 in the kernel driver */
-check_709:
-    for (i = 0; i < 3; i++)
-        for (k = 0; k < 3; k++)
-            if (fabs((*csc_matrix)[i][k] - CSC_BT_709[i][k]) > 0.01f)
-                return true;
-
-    return false;
-}
-
 tegra_shared_surface *create_shared_surface(tegra_surface *disp,
                                             tegra_surface *video,
-                                            VdpCSCMatrix const *csc_matrix,
+                                            struct host1x_csc_params *csc,
                                             uint32_t src_x0,
                                             uint32_t src_y0,
                                             uint32_t src_width,
@@ -101,7 +76,7 @@ tegra_shared_surface *create_shared_surface(tegra_surface *disp,
     tegra_shared_surface *shared;
     int ret;
 
-    if (disp->data_dirty || custom_csc(csc_matrix)) {
+    if (disp->data_dirty) {
         return NULL;
     }
 
@@ -113,7 +88,7 @@ tegra_shared_surface *create_shared_surface(tegra_surface *disp,
     assert(disp->shared == NULL);
 
     atomic_set(&shared->refcnt, 1);
-    memcpy(&shared->csc_matrix, csc_matrix, sizeof(VdpCSCMatrix));
+    memcpy(&shared->csc, csc, sizeof(*csc));
 
     shared->xv_img = create_video_xv(video);
     shared->video = video;
@@ -238,7 +213,7 @@ int shared_surface_transfer_video(tegra_surface *disp)
     host1x_gr2d_surface_blit(video->dev->stream,
                              video->pixbuf,
                              disp->pixbuf,
-                             &shared->csc_matrix,
+                             &shared->csc,
                              shared->src_x0,
                              shared->src_y0,
                              shared->src_width,
diff --git a/src/vdpau_tegra.h b/src/vdpau_tegra.h
index 994e7c8..00bd1ea 100644
--- a/src/vdpau_tegra.h
+++ b/src/vdpau_tegra.h
@@ -102,6 +102,9 @@
     fprintf(stderr, "%s:%d/%s(): " fmt, \
             __FILE__, __LINE__, __func__, ##args)
 
+#define CLAMP(_v, _vmin, _vmax) \
+    (((_v) < (_vmin) ? (_vmin) : (((_v) > (_vmax)) ? (_vmax) : (_v))))
+
 #define UNIFIED_BUFFER  0
 
 extern VdpCSCMatrix CSC_BT_601;
@@ -131,7 +134,7 @@ typedef struct tegra_shared_surface {
     atomic_t refcnt;
     struct tegra_surface *video;
     struct tegra_surface *disp;
-    VdpCSCMatrix csc_matrix;
+    struct host1x_csc_params csc;
     uint32_t src_x0, src_y0, src_width, src_height;
     uint32_t dst_x0, dst_y0, dst_width, dst_height;
     XvImage *xv_img;
@@ -194,9 +197,10 @@ typedef struct tegra_decoder {
 } tegra_decoder;
 
 typedef struct tegra_mixer {
-    VdpCSCMatrix csc_matrix;
+    struct host1x_csc_params csc;
     VdpColor bg_color;
     tegra_device *dev;
+    bool custom_csc;
 } tegra_mixer;
 
 typedef struct tegra_pqt {
@@ -294,7 +298,7 @@ int sync_video_frame_dmabufs(tegra_surface *surf, enum frame_sync type);
 
 tegra_shared_surface *create_shared_surface(tegra_surface *disp,
                                             tegra_surface *video,
-                                            VdpCSCMatrix const *csc_matrix,
+                                            struct host1x_csc_params *csc,
                                             uint32_t src_x0,
                                             uint32_t src_y0,
                                             uint32_t src_width,