drm/imx: lock scanout transfers for consecutive bursts

Because of its shallow queues and limited reordering ability, the i.MX6Q
memory controller likes AXI bursts of consecutive addresses a lot.
To optimize memory access performance, lock the IPU scanout channels for
a number of burst accesses each, before switching to the next channel.
The burst size and length of a locked burst chain are chosen not to
overshoot the stride.

Enabling the 8-burst channel lock on a single 1920x1080@60Hz RGBx
scanout (474 MiB/s of 64-byte IPU memory read accesses) reduces the
reported memory controller busy cycles from 46% to below 28% on an
otherwise idle i.MX6Q.
Tested-by: NLucas Stach <l.stach@pengutronix.de>
Signed-off-by: NPhilipp Zabel <p.zabel@pengutronix.de>

drivers/gpu/drm/imx/ipuv3-plane.c

@@ -496,6 +496,27 @@ static int ipu_chan_assign_axi_id(int ipu_chan)
 	}
 }
 
+static void ipu_calculate_bursts(u32 width, u32 cpp, u32 stride,
+				 u8 *burstsize, u8 *num_bursts)
+{
+	const unsigned int width_bytes = width * cpp;
+	unsigned int npb, bursts;
+
+	/* Maximum number of pixels per burst without overshooting stride */
+	for (npb = 64 / cpp; npb > 0; --npb) {
+		if (round_up(width_bytes, npb * cpp) <= stride)
+			break;
+	}
+	*burstsize = npb;
+
+	/* Maximum number of consecutive bursts without overshooting stride */
+	for (bursts = 8; bursts > 1; bursts /= 2) {
+		if (round_up(width_bytes, npb * cpp * bursts) <= stride)
+			break;
+	}
+	*num_bursts = bursts;
+}
+
 static void ipu_plane_atomic_update(struct drm_plane *plane,
 				    struct drm_plane_state *old_state)
 {
@@ -509,6 +530,9 @@ static void ipu_plane_atomic_update(struct drm_plane *plane,
 	unsigned long alpha_eba = 0;
 	enum ipu_color_space ics;
 	unsigned int axi_id = 0;
+	const struct drm_format_info *info;
+	u8 burstsize, num_bursts;
+	u32 width, height;
 	int active;
 
 	if (ipu_plane->dp_flow == IPU_DP_FLOW_SYNC_FG)
@@ -583,15 +607,21 @@ static void ipu_plane_atomic_update(struct drm_plane *plane,
 
 	ipu_dmfc_config_wait4eot(ipu_plane->dmfc, drm_rect_width(dst));
 
+	width = drm_rect_width(&state->src) >> 16;
+	height = drm_rect_height(&state->src) >> 16;
+	info = drm_format_info(fb->format->format);
+	ipu_calculate_bursts(width, info->cpp[0], fb->pitches[0],
+			     &burstsize, &num_bursts);
+
 	ipu_cpmem_zero(ipu_plane->ipu_ch);
-	ipu_cpmem_set_resolution(ipu_plane->ipu_ch,
-				 drm_rect_width(&state->src) >> 16,
-				 drm_rect_height(&state->src) >> 16);
+	ipu_cpmem_set_resolution(ipu_plane->ipu_ch, width, height);
 	ipu_cpmem_set_fmt(ipu_plane->ipu_ch, fb->format->format);
+	ipu_cpmem_set_burstsize(ipu_plane->ipu_ch, burstsize);
 	ipu_cpmem_set_high_priority(ipu_plane->ipu_ch);
 	ipu_idmac_set_double_buffer(ipu_plane->ipu_ch, 1);
 	ipu_cpmem_set_stride(ipu_plane->ipu_ch, fb->pitches[0]);
 	ipu_cpmem_set_axi_id(ipu_plane->ipu_ch, axi_id);
+
 	switch (fb->format->format) {
 	case DRM_FORMAT_YUV420:
 	case DRM_FORMAT_YVU420:
@@ -631,6 +661,7 @@ static void ipu_plane_atomic_update(struct drm_plane *plane,
 	case DRM_FORMAT_RGBX8888_A8:
 	case DRM_FORMAT_BGRX8888_A8:
 		alpha_eba = drm_plane_state_to_eba(state, 1);
+		num_bursts = 0;
 
 		dev_dbg(ipu_plane->base.dev->dev, "phys = %lu %lu, x = %d, y = %d",
 			eba, alpha_eba, state->src.x1 >> 16, state->src.y1 >> 16);
@@ -656,6 +687,7 @@ static void ipu_plane_atomic_update(struct drm_plane *plane,
 	}
 	ipu_cpmem_set_buffer(ipu_plane->ipu_ch, 0, eba);
 	ipu_cpmem_set_buffer(ipu_plane->ipu_ch, 1, eba);
+	ipu_idmac_lock_enable(ipu_plane->ipu_ch, num_bursts);
 	ipu_plane_enable(ipu_plane);
 }