drm/i915/skl+: change WM calc to fixed point 16.16
authorMahesh Kumar <mahesh1.kumar@intel.com>
Thu, 1 Dec 2016 15:49:37 +0000 (21:19 +0530)
committerPaulo Zanoni <paulo.r.zanoni@intel.com>
Wed, 7 Dec 2016 18:30:34 +0000 (16:30 -0200)
This patch changes Watermak calculation to fixed point calculation.
Problem with current calculation is during plane_blocks_per_line
calculation we divide intermediate blocks with min_scanlines and
takes floor of the result because of integer operation.
hence we end-up assigning less blocks than required. Which leads to
flickers.

Changes since V1:
 - Add fixed point data type as per Paulo's review
Changes since V2:
 - use fixed_point instead of fp_16_16
Changes since V3:
 - rebase
Changes since V4 (from Paulo):
 - My original renaming suggestion was misunderstood, so implement it
 - Simplify fixed_16_16_to_u32 implementation
 - Fix indentation

Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Mahesh Kumar <mahesh1.kumar@intel.com>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20161201154940.24446-6-mahesh1.kumar@intel.com
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/intel_pm.c

index 4eea9255dc5329fbb8f6be5457d0dcd15558be6d..b9138cd75a39cad7e92623b819cb5a8a361685e0 100644 (file)
@@ -119,6 +119,90 @@ bool __i915_inject_load_failure(const char *func, int line);
 #define i915_inject_load_failure() \
        __i915_inject_load_failure(__func__, __LINE__)
 
+typedef struct {
+       uint32_t val;
+} uint_fixed_16_16_t;
+
+#define FP_16_16_MAX ({ \
+       uint_fixed_16_16_t fp; \
+       fp.val = UINT_MAX; \
+       fp; \
+})
+
+static inline uint_fixed_16_16_t u32_to_fixed_16_16(uint32_t val)
+{
+       uint_fixed_16_16_t fp;
+
+       WARN_ON(val >> 16);
+
+       fp.val = val << 16;
+       return fp;
+}
+
+static inline uint32_t fixed_16_16_to_u32_round_up(uint_fixed_16_16_t fp)
+{
+       return DIV_ROUND_UP(fp.val, 1 << 16);
+}
+
+static inline uint32_t fixed_16_16_to_u32(uint_fixed_16_16_t fp)
+{
+       return fp.val >> 16;
+}
+
+static inline uint_fixed_16_16_t min_fixed_16_16(uint_fixed_16_16_t min1,
+                                                uint_fixed_16_16_t min2)
+{
+       uint_fixed_16_16_t min;
+
+       min.val = min(min1.val, min2.val);
+       return min;
+}
+
+static inline uint_fixed_16_16_t max_fixed_16_16(uint_fixed_16_16_t max1,
+                                                uint_fixed_16_16_t max2)
+{
+       uint_fixed_16_16_t max;
+
+       max.val = max(max1.val, max2.val);
+       return max;
+}
+
+static inline uint_fixed_16_16_t fixed_16_16_div_round_up(uint32_t val,
+                                                         uint32_t d)
+{
+       uint_fixed_16_16_t fp, res;
+
+       fp = u32_to_fixed_16_16(val);
+       res.val = DIV_ROUND_UP(fp.val, d);
+       return res;
+}
+
+static inline uint_fixed_16_16_t fixed_16_16_div_round_up_u64(uint32_t val,
+                                                             uint32_t d)
+{
+       uint_fixed_16_16_t res;
+       uint64_t interm_val;
+
+       interm_val = (uint64_t)val << 16;
+       interm_val = DIV_ROUND_UP_ULL(interm_val, d);
+       WARN_ON(interm_val >> 32);
+       res.val = (uint32_t) interm_val;
+
+       return res;
+}
+
+static inline uint_fixed_16_16_t mul_u32_fixed_16_16(uint32_t val,
+                                                    uint_fixed_16_16_t mul)
+{
+       uint64_t intermediate_val;
+       uint_fixed_16_16_t fp;
+
+       intermediate_val = (uint64_t) val * mul.val;
+       WARN_ON(intermediate_val >> 32);
+       fp.val = (uint32_t) intermediate_val;
+       return fp;
+}
+
 static inline const char *yesno(bool v)
 {
        return v ? "yes" : "no";
index 315a1b3392572e8eb2cdada36d4fd04199ed3388..9171431558a34146e7c97901ada5fecdbd33e06e 100644 (file)
@@ -3509,32 +3509,35 @@ skl_allocate_pipe_ddb(struct intel_crtc_state *cstate,
  * 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 cpp, uint32_t latency)
+static uint_fixed_16_16_t skl_wm_method1(uint32_t pixel_rate, uint8_t cpp,
+                                        uint32_t latency)
 {
-       uint32_t wm_intermediate_val, ret;
+       uint32_t wm_intermediate_val;
+       uint_fixed_16_16_t ret;
 
        if (latency == 0)
-               return UINT_MAX;
-
-       wm_intermediate_val = latency * pixel_rate * cpp / 512;
-       ret = DIV_ROUND_UP(wm_intermediate_val, 1000);
+               return FP_16_16_MAX;
 
+       wm_intermediate_val = latency * pixel_rate * cpp;
+       ret = fixed_16_16_div_round_up_u64(wm_intermediate_val, 1000 * 512);
        return ret;
 }
 
-static uint32_t skl_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
-                              uint32_t latency, uint32_t plane_blocks_per_line)
+static uint_fixed_16_16_t skl_wm_method2(uint32_t pixel_rate,
+                       uint32_t pipe_htotal,
+                       uint32_t latency,
+                       uint_fixed_16_16_t plane_blocks_per_line)
 {
-       uint32_t ret;
        uint32_t wm_intermediate_val;
+       uint_fixed_16_16_t ret;
 
        if (latency == 0)
-               return UINT_MAX;
+               return FP_16_16_MAX;
 
        wm_intermediate_val = latency * pixel_rate;
-       ret = DIV_ROUND_UP(wm_intermediate_val, pipe_htotal * 1000) *
-                               plane_blocks_per_line;
-
+       wm_intermediate_val = DIV_ROUND_UP(wm_intermediate_val,
+                                          pipe_htotal * 1000);
+       ret = mul_u32_fixed_16_16(wm_intermediate_val, plane_blocks_per_line);
        return ret;
 }
 
@@ -3574,14 +3577,17 @@ static int skl_compute_plane_wm(const struct drm_i915_private *dev_priv,
        struct drm_plane_state *pstate = &intel_pstate->base;
        struct drm_framebuffer *fb = pstate->fb;
        uint32_t latency = dev_priv->wm.skl_latency[level];
-       uint32_t method1, method2;
-       uint32_t plane_bytes_per_line, plane_blocks_per_line;
+       uint_fixed_16_16_t method1, method2;
+       uint_fixed_16_16_t plane_blocks_per_line;
+       uint_fixed_16_16_t selected_result;
+       uint32_t interm_pbpl;
+       uint32_t plane_bytes_per_line;
        uint32_t res_blocks, res_lines;
-       uint32_t selected_result;
        uint8_t cpp;
        uint32_t width = 0, height = 0;
        uint32_t plane_pixel_rate;
-       uint32_t y_tile_minimum, y_min_scanlines;
+       uint_fixed_16_16_t y_tile_minimum;
+       uint32_t y_min_scanlines;
        struct intel_atomic_state *state =
                to_intel_atomic_state(cstate->base.state);
        bool apply_memory_bw_wa = skl_needs_memory_bw_wa(state);
@@ -3640,14 +3646,16 @@ static int skl_compute_plane_wm(const struct drm_i915_private *dev_priv,
 
        plane_bytes_per_line = width * cpp;
        if (y_tiled) {
+               interm_pbpl = DIV_ROUND_UP(plane_bytes_per_line *
+                                          y_min_scanlines, 512);
                plane_blocks_per_line =
-                     DIV_ROUND_UP(plane_bytes_per_line * y_min_scanlines, 512);
-               plane_blocks_per_line /= y_min_scanlines;
+                     fixed_16_16_div_round_up(interm_pbpl, y_min_scanlines);
        } else if (x_tiled) {
-               plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512);
+               interm_pbpl = DIV_ROUND_UP(plane_bytes_per_line, 512);
+               plane_blocks_per_line = u32_to_fixed_16_16(interm_pbpl);
        } else {
-               plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512)
-                                       + 1;
+               interm_pbpl = DIV_ROUND_UP(plane_bytes_per_line, 512) + 1;
+               plane_blocks_per_line = u32_to_fixed_16_16(interm_pbpl);
        }
 
        method1 = skl_wm_method1(plane_pixel_rate, cpp, latency);
@@ -3656,26 +3664,29 @@ static int skl_compute_plane_wm(const struct drm_i915_private *dev_priv,
                                 latency,
                                 plane_blocks_per_line);
 
-       y_tile_minimum = plane_blocks_per_line * y_min_scanlines;
+       y_tile_minimum = mul_u32_fixed_16_16(y_min_scanlines,
+                                            plane_blocks_per_line);
 
        if (y_tiled) {
-               selected_result = max(method2, y_tile_minimum);
+               selected_result = max_fixed_16_16(method2, y_tile_minimum);
        } else {
                if ((cpp * cstate->base.adjusted_mode.crtc_htotal / 512 < 1) &&
                    (plane_bytes_per_line / 512 < 1))
                        selected_result = method2;
-               else if ((ddb_allocation / plane_blocks_per_line) >= 1)
-                       selected_result = min(method1, method2);
+               else if ((ddb_allocation /
+                       fixed_16_16_to_u32_round_up(plane_blocks_per_line)) >= 1)
+                       selected_result = min_fixed_16_16(method1, method2);
                else
                        selected_result = method1;
        }
 
-       res_blocks = selected_result + 1;
-       res_lines = DIV_ROUND_UP(selected_result, plane_blocks_per_line);
+       res_blocks = fixed_16_16_to_u32_round_up(selected_result) + 1;
+       res_lines = DIV_ROUND_UP(selected_result.val,
+                                plane_blocks_per_line.val);
 
        if (level >= 1 && level <= 7) {
                if (y_tiled) {
-                       res_blocks += y_tile_minimum;
+                       res_blocks += fixed_16_16_to_u32_round_up(y_tile_minimum);
                        res_lines += y_min_scanlines;
                } else {
                        res_blocks++;