drm/i915/skl: SKL Watermark Computation
authorPradeep Bhat <pradeep.bhat@intel.com>
Tue, 4 Nov 2014 17:06:42 +0000 (17:06 +0000)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Fri, 7 Nov 2014 17:42:04 +0000 (18:42 +0100)
This patch implements the watermark algorithm and its necessary
functions. Two function pointers skl_update_wm and
skl_update_sprite_wm are provided. The skl_update_wm will update
the watermarks for the crtc provided as an argument and then
checks for change in DDB allocation for other active pipes and
recomputes the watermarks for those Pipes and planes as well.
Finally it does the register programming for all dirty pipes.
The trigger of the Watermark double buffer registers will have
to be once the plane configurations are done by the caller.

v2: fixed the divide-by-0 error in the results computation func.
    Also reworked the PLANE_WM register values computation func to
    make it more compact. Incorporated all other review comments
    from Damien.

v3: Changed the skl_compute_plane_wm function to now return success
    or failure. Also the result blocks and lines are computed here
    instead of in skl_compute_wm_results function.

v4: Adjust skl_ddb_alloc_changed() to the new planes/cursor split
    (Damien)

v5: Reworked the affected functions to implement new plane/cursor
    split.

v6: Rework the logic that triggers the DDB allocation and WM computation
    of skl_update_other_pipe_wm() to not depend on non-computed DDB
    values.
    Always give a valid cursor_width (at boot it's 0) to keep the
    invariant that we consider the cursor plane always enabled.
    Otherwise we end up dividing by 0 in skl_compute_plane_wm()
    (Damien Lespiau)

v7: Spell out allocation
    skl_ddb_ functions should have the ddb as first argument
    Make the skl_ddb_alloc_changed() parameters const
    (Damien)

v8: Rebase on top of the crtc->primary changes

v9: Split the staging results structure to not exceed the 1Kb stack
    allocation in skl_update_wm()

v10: Make skl_pipe_pixel_rate() take a pointer to the pipe config
     Add a comment about overflow considerations for skl_wm_method1()
     Various additions of const
     Various use of sizeof(variable) instead of sizeof(type)
     Various move of variable definitons to a narrower scope
     Zero initialize some stack allocated structures to make sure we
     don't have garbage in case we don't write all the values
     (Ville)

v11: Remove non-necessary default number of blocks/lines when the plane
     is disabled (Ville)

Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Pradeep Bhat <pradeep.bhat@intel.com>
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/intel_pm.c

index aa8607a9e49c2dfa76a346221cb16673b1add3a7..fbe18aed254dacc3732a23baee640459a4d78441 100644 (file)
@@ -1722,8 +1722,18 @@ struct drm_i915_private {
                 */
                uint16_t skl_latency[8];
 
+               /*
+                * The skl_wm_values structure is a bit too big for stack
+                * allocation, so we keep the staging struct where we store
+                * intermediate results here instead.
+                */
+               struct skl_wm_values skl_results;
+
                /* current hardware state */
-               struct ilk_wm_values hw;
+               union {
+                       struct ilk_wm_values hw;
+                       struct skl_wm_values skl_hw;
+               };
        } wm;
 
        struct i915_runtime_pm pm;
index ca2080bf486f923c81060fa5ebbebfa4c2d79522..5500c5343e0efb46b8ea3463439f0999e6285e13 100644 (file)
@@ -2973,6 +2973,426 @@ static bool ilk_disable_lp_wm(struct drm_device *dev)
        return _ilk_disable_lp_wm(dev_priv, WM_DIRTY_LP_ALL);
 }
 
+static uint32_t skl_pipe_pixel_rate(const struct intel_crtc_config *config)
+{
+       /* TODO: Take into account the scalers once we support them */
+       return config->adjusted_mode.crtc_clock;
+}
+
+/*
+ * The max latency should be 257 (max the punit can code is 255 and we add 2us
+ * for the read latency) and bytes_per_pixel should always be <= 8, so that
+ * should allow pixel_rate up to ~2 GHz which seems sufficient since max
+ * 2xcdclk is 1350 MHz and the pixel rate should never exceed that.
+*/
+static uint32_t skl_wm_method1(uint32_t pixel_rate, uint8_t bytes_per_pixel,
+                              uint32_t latency)
+{
+       uint32_t wm_intermediate_val, ret;
+
+       if (latency == 0)
+               return UINT_MAX;
+
+       wm_intermediate_val = latency * pixel_rate * bytes_per_pixel;
+       ret = DIV_ROUND_UP(wm_intermediate_val, 1000);
+
+       return ret;
+}
+
+static uint32_t skl_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
+                              uint32_t horiz_pixels, uint8_t bytes_per_pixel,
+                              uint32_t latency)
+{
+       uint32_t ret, plane_bytes_per_line, wm_intermediate_val;
+
+       if (latency == 0)
+               return UINT_MAX;
+
+       plane_bytes_per_line = horiz_pixels * bytes_per_pixel;
+       wm_intermediate_val = latency * pixel_rate;
+       ret = DIV_ROUND_UP(wm_intermediate_val, pipe_htotal * 1000) *
+                               plane_bytes_per_line;
+
+       return ret;
+}
+
+static void skl_compute_transition_wm(struct drm_crtc *crtc,
+                                 struct skl_pipe_wm_parameters *params,
+                                 struct skl_pipe_wm *pipe_wm)
+{
+       /*
+        * For now it is suggested to use the LP0 wm val of corresponding
+        * plane as transition wm val. This is done while computing results.
+        */
+       if (!params->active)
+               return;
+}
+
+static uint32_t
+skl_compute_linetime_wm(struct drm_crtc *crtc, struct skl_pipe_wm_parameters *p)
+{
+       if (!intel_crtc_active(crtc))
+               return 0;
+
+       return DIV_ROUND_UP(8 * p->pipe_htotal * 1000, p->pixel_rate);
+
+}
+
+static bool skl_ddb_allocation_changed(const struct skl_ddb_allocation *new_ddb,
+                                      const struct intel_crtc *intel_crtc)
+{
+       struct drm_device *dev = intel_crtc->base.dev;
+       struct drm_i915_private *dev_priv = dev->dev_private;
+       const struct skl_ddb_allocation *cur_ddb = &dev_priv->wm.skl_hw.ddb;
+       enum pipe pipe = intel_crtc->pipe;
+
+       if (memcmp(new_ddb->plane[pipe], cur_ddb->plane[pipe],
+                  sizeof(new_ddb->plane[pipe])))
+               return true;
+
+       if (memcmp(&new_ddb->cursor[pipe], &cur_ddb->cursor[pipe],
+                   sizeof(new_ddb->cursor[pipe])))
+               return true;
+
+       return false;
+}
+
+static void skl_compute_wm_global_parameters(struct drm_device *dev,
+                                            struct intel_wm_config *config)
+{
+       struct drm_crtc *crtc;
+       struct drm_plane *plane;
+
+       list_for_each_entry(crtc, &dev->mode_config.crtc_list, head)
+               config->num_pipes_active += intel_crtc_active(crtc);
+
+       /* FIXME: I don't think we need those two global parameters on SKL */
+       list_for_each_entry(plane, &dev->mode_config.plane_list, head) {
+               struct intel_plane *intel_plane = to_intel_plane(plane);
+
+               config->sprites_enabled |= intel_plane->wm.enabled;
+               config->sprites_scaled |= intel_plane->wm.scaled;
+       }
+}
+
+static void skl_compute_wm_pipe_parameters(struct drm_crtc *crtc,
+                                          struct skl_pipe_wm_parameters *p)
+{
+       struct drm_device *dev = crtc->dev;
+       struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
+       enum pipe pipe = intel_crtc->pipe;
+       struct drm_plane *plane;
+       int i = 1; /* Index for sprite planes start */
+
+       p->active = intel_crtc_active(crtc);
+       if (p->active) {
+               p->pipe_htotal = intel_crtc->config.adjusted_mode.crtc_htotal;
+               p->pixel_rate = skl_pipe_pixel_rate(&intel_crtc->config);
+
+               /*
+                * For now, assume primary and cursor planes are always enabled.
+                */
+               p->plane[0].enabled = true;
+               p->plane[0].bytes_per_pixel =
+                       crtc->primary->fb->bits_per_pixel / 8;
+               p->plane[0].horiz_pixels = intel_crtc->config.pipe_src_w;
+               p->plane[0].vert_pixels = intel_crtc->config.pipe_src_h;
+
+               p->cursor.enabled = true;
+               p->cursor.bytes_per_pixel = 4;
+               p->cursor.horiz_pixels = intel_crtc->cursor_width ?
+                                        intel_crtc->cursor_width : 64;
+       }
+
+       list_for_each_entry(plane, &dev->mode_config.plane_list, head) {
+               struct intel_plane *intel_plane = to_intel_plane(plane);
+
+               if (intel_plane->pipe == pipe)
+                       p->plane[i++] = intel_plane->wm;
+       }
+}
+
+static bool skl_compute_plane_wm(struct skl_pipe_wm_parameters *p,
+                                  struct intel_plane_wm_parameters *p_params,
+                                  uint16_t max_page_buff_alloc,
+                                  uint32_t mem_value,
+                                  uint16_t *res_blocks, /* out */
+                                  uint8_t *res_lines /* out */)
+{
+       uint32_t method1, method2, plane_bytes_per_line;
+       uint32_t result_bytes;
+
+       if (!p->active || !p_params->enabled)
+               return false;
+
+       method1 = skl_wm_method1(p->pixel_rate,
+                                p_params->bytes_per_pixel,
+                                mem_value);
+       method2 = skl_wm_method2(p->pixel_rate,
+                                p->pipe_htotal,
+                                p_params->horiz_pixels,
+                                p_params->bytes_per_pixel,
+                                mem_value);
+
+       plane_bytes_per_line = p_params->horiz_pixels *
+                                       p_params->bytes_per_pixel;
+
+       /* For now xtile and linear */
+       if (((max_page_buff_alloc * 512) / plane_bytes_per_line) >= 1)
+               result_bytes = min(method1, method2);
+       else
+               result_bytes = method1;
+
+       *res_blocks = DIV_ROUND_UP(result_bytes, 512) + 1;
+       *res_lines = DIV_ROUND_UP(result_bytes, plane_bytes_per_line);
+
+       return true;
+}
+
+static void skl_compute_wm_level(const struct drm_i915_private *dev_priv,
+                                struct skl_ddb_allocation *ddb,
+                                struct skl_pipe_wm_parameters *p,
+                                enum pipe pipe,
+                                int level,
+                                int num_planes,
+                                struct skl_wm_level *result)
+{
+       uint16_t latency = dev_priv->wm.skl_latency[level];
+       uint16_t ddb_blocks;
+       int i;
+
+       for (i = 0; i < num_planes; i++) {
+               ddb_blocks = skl_ddb_entry_size(&ddb->plane[pipe][i]);
+
+               result->plane_en[i] = skl_compute_plane_wm(p, &p->plane[i],
+                                               ddb_blocks,
+                                               latency,
+                                               &result->plane_res_b[i],
+                                               &result->plane_res_l[i]);
+       }
+
+       ddb_blocks = skl_ddb_entry_size(&ddb->cursor[pipe]);
+       result->cursor_en = skl_compute_plane_wm(p, &p->cursor, ddb_blocks,
+                                                latency, &result->cursor_res_b,
+                                                &result->cursor_res_l);
+}
+
+static void skl_compute_pipe_wm(struct drm_crtc *crtc,
+                               struct skl_ddb_allocation *ddb,
+                               struct skl_pipe_wm_parameters *params,
+                               struct skl_pipe_wm *pipe_wm)
+{
+       struct drm_device *dev = crtc->dev;
+       const struct drm_i915_private *dev_priv = dev->dev_private;
+       struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
+       int level, max_level = ilk_wm_max_level(dev);
+
+       for (level = 0; level <= max_level; level++) {
+               skl_compute_wm_level(dev_priv, ddb, params, intel_crtc->pipe,
+                                    level, intel_num_planes(intel_crtc),
+                                    &pipe_wm->wm[level]);
+       }
+       pipe_wm->linetime = skl_compute_linetime_wm(crtc, params);
+
+       skl_compute_transition_wm(crtc, params, pipe_wm);
+}
+
+static void skl_compute_wm_results(struct drm_device *dev,
+                                  struct skl_pipe_wm_parameters *p,
+                                  struct skl_pipe_wm *p_wm,
+                                  struct skl_wm_values *r,
+                                  struct intel_crtc *intel_crtc)
+{
+       int level, max_level = ilk_wm_max_level(dev);
+       enum pipe pipe = intel_crtc->pipe;
+
+       for (level = 0; level <= max_level; level++) {
+               uint16_t ddb_blocks;
+               uint32_t temp;
+               int i;
+
+               for (i = 0; i < intel_num_planes(intel_crtc); i++) {
+                       temp = 0;
+                       ddb_blocks = skl_ddb_entry_size(&r->ddb.plane[pipe][i]);
+
+                       if ((p_wm->wm[level].plane_res_b[i] > ddb_blocks) ||
+                               (p_wm->wm[level].plane_res_l[i] > 31))
+                               p_wm->wm[level].plane_en[i] = false;
+
+                       temp |= p_wm->wm[level].plane_res_l[i] <<
+                                       PLANE_WM_LINES_SHIFT;
+                       temp |= p_wm->wm[level].plane_res_b[i];
+                       if (p_wm->wm[level].plane_en[i])
+                               temp |= PLANE_WM_EN;
+
+                       r->plane[pipe][i][level] = temp;
+                       /* Use the LP0 WM value for transition WM for now. */
+                       if (level == 0)
+                               r->plane_trans[pipe][i] = temp;
+               }
+
+               temp = 0;
+               ddb_blocks = skl_ddb_entry_size(&r->ddb.cursor[pipe]);
+
+               if ((p_wm->wm[level].cursor_res_b > ddb_blocks) ||
+                       (p_wm->wm[level].cursor_res_l > 31))
+                       p_wm->wm[level].cursor_en = false;
+
+               temp |= p_wm->wm[level].cursor_res_l << PLANE_WM_LINES_SHIFT;
+               temp |= p_wm->wm[level].cursor_res_b;
+
+               if (p_wm->wm[level].cursor_en)
+                       temp |= PLANE_WM_EN;
+
+               r->cursor[pipe][level] = temp;
+               /* Use the LP0 WM value for transition WM for now. */
+               if (level == 0)
+                       r->cursor_trans[pipe] = temp;
+
+       }
+
+       r->wm_linetime[pipe] = p_wm->linetime;
+}
+
+static void skl_write_wm_values(struct drm_i915_private *dev_priv,
+                               const struct skl_wm_values *new)
+{
+       struct drm_device *dev = dev_priv->dev;
+       struct intel_crtc *crtc;
+
+       list_for_each_entry(crtc, &dev->mode_config.crtc_list, base.head) {
+               int i, level, max_level = ilk_wm_max_level(dev);
+               enum pipe pipe = crtc->pipe;
+
+               if (new->dirty[pipe]) {
+                       I915_WRITE(PIPE_WM_LINETIME(pipe),
+                                       new->wm_linetime[pipe]);
+
+                       for (level = 0; level <= max_level; level++) {
+                               for (i = 0; i < intel_num_planes(crtc); i++)
+                                       I915_WRITE(PLANE_WM(pipe, i, level),
+                                               new->plane[pipe][i][level]);
+                               I915_WRITE(CUR_WM(pipe, level),
+                                       new->cursor[pipe][level]);
+                       }
+                       for (i = 0; i < intel_num_planes(crtc); i++)
+                               I915_WRITE(PLANE_WM_TRANS(pipe, i),
+                                               new->plane_trans[pipe][i]);
+                       I915_WRITE(CUR_WM_TRANS(pipe), new->cursor_trans[pipe]);
+               }
+       }
+
+       dev_priv->wm.skl_hw = *new;
+}
+
+static bool skl_update_pipe_wm(struct drm_crtc *crtc,
+                              struct skl_pipe_wm_parameters *params,
+                              struct intel_wm_config *config,
+                              struct skl_ddb_allocation *ddb, /* out */
+                              struct skl_pipe_wm *pipe_wm /* out */)
+{
+       struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
+
+       skl_compute_wm_pipe_parameters(crtc, params);
+       skl_compute_pipe_wm(crtc, ddb, params, pipe_wm);
+
+       if (!memcmp(&intel_crtc->wm.skl_active, pipe_wm, sizeof(*pipe_wm)))
+               return false;
+
+       intel_crtc->wm.skl_active = *pipe_wm;
+       return true;
+}
+
+static void skl_update_other_pipe_wm(struct drm_device *dev,
+                                    struct drm_crtc *crtc,
+                                    struct intel_wm_config *config,
+                                    struct skl_wm_values *r)
+{
+       struct intel_crtc *intel_crtc;
+       struct intel_crtc *this_crtc = to_intel_crtc(crtc);
+
+       /*
+        * If the WM update hasn't changed the allocation for this_crtc (the
+        * crtc we are currently computing the new WM values for), other
+        * enabled crtcs will keep the same allocation and we don't need to
+        * recompute anything for them.
+        */
+       if (!skl_ddb_allocation_changed(&r->ddb, this_crtc))
+               return;
+
+       /*
+        * Otherwise, because of this_crtc being freshly enabled/disabled, the
+        * other active pipes need new DDB allocation and WM values.
+        */
+       list_for_each_entry(intel_crtc, &dev->mode_config.crtc_list,
+                               base.head) {
+               struct skl_pipe_wm_parameters params = {};
+               struct skl_pipe_wm pipe_wm = {};
+               bool wm_changed;
+
+               if (this_crtc->pipe == intel_crtc->pipe)
+                       continue;
+
+               if (!intel_crtc->active)
+                       continue;
+
+               wm_changed = skl_update_pipe_wm(&intel_crtc->base,
+                                               &params, config,
+                                               &r->ddb, &pipe_wm);
+
+               /*
+                * If we end up re-computing the other pipe WM values, it's
+                * because it was really needed, so we expect the WM values to
+                * be different.
+                */
+               WARN_ON(!wm_changed);
+
+               skl_compute_wm_results(dev, &params, &pipe_wm, r, intel_crtc);
+               r->dirty[intel_crtc->pipe] = true;
+       }
+}
+
+static void skl_update_wm(struct drm_crtc *crtc)
+{
+       struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
+       struct drm_device *dev = crtc->dev;
+       struct drm_i915_private *dev_priv = dev->dev_private;
+       struct skl_pipe_wm_parameters params = {};
+       struct skl_wm_values *results = &dev_priv->wm.skl_results;
+       struct skl_pipe_wm pipe_wm = {};
+       struct intel_wm_config config = {};
+
+       memset(results, 0, sizeof(*results));
+
+       skl_compute_wm_global_parameters(dev, &config);
+
+       if (!skl_update_pipe_wm(crtc, &params, &config,
+                               &results->ddb, &pipe_wm))
+               return;
+
+       skl_compute_wm_results(dev, &params, &pipe_wm, results, intel_crtc);
+       results->dirty[intel_crtc->pipe] = true;
+
+       skl_update_other_pipe_wm(dev, crtc, &config, results);
+       skl_write_wm_values(dev_priv, results);
+}
+
+static void
+skl_update_sprite_wm(struct drm_plane *plane, struct drm_crtc *crtc,
+                    uint32_t sprite_width, uint32_t sprite_height,
+                    int pixel_size, bool enabled, bool scaled)
+{
+       struct intel_plane *intel_plane = to_intel_plane(plane);
+
+       intel_plane->wm.enabled = enabled;
+       intel_plane->wm.scaled = scaled;
+       intel_plane->wm.horiz_pixels = sprite_width;
+       intel_plane->wm.vert_pixels = sprite_height;
+       intel_plane->wm.bytes_per_pixel = pixel_size;
+
+       skl_update_wm(crtc);
+}
+
 static void ilk_update_wm(struct drm_crtc *crtc)
 {
        struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
@@ -6198,6 +6618,8 @@ void intel_init_pm(struct drm_device *dev)
                skl_setup_wm_latency(dev);
 
                dev_priv->display.init_clock_gating = gen9_init_clock_gating;
+               dev_priv->display.update_wm = skl_update_wm;
+               dev_priv->display.update_sprite_wm = skl_update_sprite_wm;
        } else if (HAS_PCH_SPLIT(dev)) {
                ilk_setup_wm_latency(dev);