Skip to contentSkip to conversation

drm/bridge: dw-hdmi-qp: compute audio CTS from N when not in TMDS table

PATCH In review Simon Wright 2 participants 3 messages May 20 – May 21

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 drm/bridge: dw-hdmi-qp: compute audio CTS from N when not in TMDS table

dw_hdmi_qp_find_cts() returns 0 for any TMDS character rate not present in common_tmds_cts_table[], which terminates at 148.5 MHz. In that case dw_hdmi_qp_set_sample_rate() calls dw_hdmi_qp_set_cts_n() with cts == 0, which leaves AUDPKT_ACR_CTS_OVR_EN clear and falls back to the controller's internal CTS auto-measurement.

On at least the RK3576 hdptx integration, the auto-measure path produces incorrect ACR timing on the wire at TMDS rates above the table's coverage. Strict HDMI sinks that cross-check the ACR CTS against the AVI/GCP and the actual TMDS clock then mute audio.

The issue is reproducible at any rate not present in the table, and is not specific to HDMI 2.x: 1920x1080@60 with 10-bit deep colour (185.625 MHz, HDMI 1.4) is affected, as is 3840x2160@60 8-bit (594 MHz, HDMI 2.0).

The driver already has the symmetric machinery for the N-table-miss
case: dw_hdmi_qp_compute_n() falls back to a dynamic search via dw_hdmi_qp_audio_math_diff() ((pixel_clk * n) / (128 * freq)) when no table entry matches. The CTS path lacks the equivalent fallback.

Compute CTS inline in dw_hdmi_qp_set_sample_rate() from N per the HDMI spec (CTS = TMDS * N / (128 * Fs)) when find_cts() returns 0. The standard override path then supplies the correct value on the wire instead of falling through to auto-measure.

The legacy DesignWare HDMI driver (drivers/gpu/drm/bridge/synopsys/ dw-hdmi.c, hdmi_set_clk_regenerator()) computes CTS from N via the same formula when AHB or GP audio is active, so the pattern is already established within the family.

Tested on R76S (RK3576) on Armbian-edge mainline 7.0.1 with Cristian Ciocaltea's hdptx-clk-fixes v1 series applied, against four sinks at four TMDS rates spanning HDMI 1.4 and HDMI 2.0:

  TMDS         Mode                In table?  G3     C4     TCL    Kogan
  148.5 MHz    1080p60  8-bit      yes        audio  audio  audio  audio
  185.625 MHz  1080p60 10-bit      no         audio  audio  audio  audio
  297 MHz      3840p30  8-bit      no         audio  audio  audio  audio
  594 MHz      3840p60  8-bit      no         audio  N/A    audio  audio

Without this fix, all four sinks mute audio at the rates marked "no" above. With the fix, audio plays cleanly. The 148.5 MHz row is a regression check confirming the in-table path is unchanged.

The LG C4 OLED's CTA-861 SVDs do not advertise 3840x2160@60 over TMDS (it is signalled FRL-only in this model's EDID), so that specific case is untestable on the C4 over the TMDS path. The same audio-path code is exercised on the C4 at 297 MHz and behaves identically to the G3.

More-permissive sinks (older HDMI 2.0 TVs tested informally) play audio at all rates with or without the fix because they do not strictly cross-check ACR CTS against the TMDS clock.

---
Changes in v2:
- Resend after v1 was mailer-mangled and unapplyable (reported by Luca Ceresoli).
- Body split into shorter paragraphs for readability per Luca's review.
- Original LG G3 report (linux-rockchip 070633) moved from inline prose to a Closes: trailer paired with Reported-by:, per Documentation/process/submitting-patches.rst.
- Add Assisted-by: trailer per Documentation/process/coding-assistants.rst.
- No code or test-result changes.

v1: https://lore.kernel.org/linux-rockchip/[email protected]/

 drivers/gpu/drm/bridge/synopsys/dw-hdmi-qp.c | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)
drivers/gpu/drm/bridge/synopsys/dw-hdmi-qp.c +17 -0
index 0dbb1274360..b7203787057 100644
@@ -461,6 +461,23 @@ static void dw_hdmi_qp_set_sample_rate(struct dw_hdmi_qp *hdmi, unsigned long lo
 	n = dw_hdmi_qp_find_n(hdmi, tmds_char_rate, sample_rate);
 	cts = dw_hdmi_qp_find_cts(hdmi, tmds_char_rate, sample_rate);
 
+	/*
+	 * When no CTS table entry exists for the given TMDS rate, compute
+	 * CTS from N rather than letting the hardware auto-measure.  The
+	 * auto-CTS circuit produces incorrect audio timing at out-of-table
+	 * rates (e.g. 185.625 MHz, 297 MHz, 594 MHz), causing strict HDMI
+	 * sinks to mute audio.  Computed CTS = (TMDS * N) / (128 * Fs) per
+	 * HDMI spec; the standard override path then supplies it on the
+	 * wire.  Mirrors hdmi_set_clk_regenerator() in the legacy dw-hdmi
+	 * driver.
+	 */
+	if (!cts && n) {
+		u64 computed = (u64)tmds_char_rate * n;
+
+		do_div(computed, 128ULL * sample_rate);
+		cts = (unsigned int)computed;
+	}
+
 	dw_hdmi_qp_set_cts_n(hdmi, cts, n);
 }
 

base-commit: c1079aebb4de218caa86c44f9a53700d1a582683

Signature · 1 line
2.53.0
reviewer
#1 lore raw

Hi Simon,

Quoted 93 lines · Simon Wright

dw_hdmi_qp_find_cts() returns 0 for any TMDS character rate not present in common_tmds_cts_table[], which terminates at 148.5 MHz. In that case dw_hdmi_qp_set_sample_rate() calls dw_hdmi_qp_set_cts_n() with cts == 0, which leaves AUDPKT_ACR_CTS_OVR_EN clear and falls back to the controller's internal CTS auto-measurement.

On at least the RK3576 hdptx integration, the auto-measure path produces incorrect ACR timing on the wire at TMDS rates above the table's coverage. Strict HDMI sinks that cross-check the ACR CTS against the AVI/GCP and the actual TMDS clock then mute audio.

The issue is reproducible at any rate not present in the table, and is not specific to HDMI 2.x: 1920x1080@60 with 10-bit deep colour (185.625 MHz, HDMI 1.4) is affected, as is 3840x2160@60 8-bit (594 MHz, HDMI 2.0).

The driver already has the symmetric machinery for the N-table-miss
case: dw_hdmi_qp_compute_n() falls back to a dynamic search via dw_hdmi_qp_audio_math_diff() ((pixel_clk * n) / (128 * freq)) when no table entry matches. The CTS path lacks the equivalent fallback.

Compute CTS inline in dw_hdmi_qp_set_sample_rate() from N per the HDMI spec (CTS = TMDS * N / (128 * Fs)) when find_cts() returns 0. The standard override path then supplies the correct value on the wire instead of falling through to auto-measure.

The legacy DesignWare HDMI driver (drivers/gpu/drm/bridge/synopsys/ dw-hdmi.c, hdmi_set_clk_regenerator()) computes CTS from N via the same formula when AHB or GP audio is active, so the pattern is already established within the family.

Tested on R76S (RK3576) on Armbian-edge mainline 7.0.1 with Cristian Ciocaltea's hdptx-clk-fixes v1 series applied, against four sinks at four TMDS rates spanning HDMI 1.4 and HDMI 2.0:

  TMDS         Mode                In table?  G3     C4     TCL    Kogan
  148.5 MHz    1080p60  8-bit      yes        audio  audio  audio  audio
  185.625 MHz  1080p60 10-bit      no         audio  audio  audio  audio
  297 MHz      3840p30  8-bit      no         audio  audio  audio  audio
  594 MHz      3840p60  8-bit      no         audio  N/A    audio  audio

Without this fix, all four sinks mute audio at the rates marked "no" above. With the fix, audio plays cleanly. The 148.5 MHz row is a regression check confirming the in-table path is unchanged.

The LG C4 OLED's CTA-861 SVDs do not advertise 3840x2160@60 over TMDS (it is signalled FRL-only in this model's EDID), so that specific case is untestable on the C4 over the TMDS path. The same audio-path code is exercised on the C4 at 297 MHz and behaves identically to the G3.

More-permissive sinks (older HDMI 2.0 TVs tested informally) play audio at all rates with or without the fix because they do not strictly cross-check ACR CTS against the TMDS clock.

---
Changes in v2:
- Resend after v1 was mailer-mangled and unapplyable (reported by Luca Ceresoli).
- Body split into shorter paragraphs for readability per Luca's review.
- Original LG G3 report (linux-rockchip 070633) moved from inline prose to a Closes: trailer paired with Reported-by:, per Documentation/process/submitting-patches.rst.
- Add Assisted-by: trailer per Documentation/process/coding-assistants.rst.
- No code or test-result changes.

v1: https://lore.kernel.org/linux-rockchip/[email protected]/

 drivers/gpu/drm/bridge/synopsys/dw-hdmi-qp.c | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)
drivers/gpu/drm/bridge/synopsys/dw-hdmi-qp.c +6 -0
index 0dbb1274360..b7203787057 100644
@@ -461,6 +461,23 @@ static void dw_hdmi_qp_set_sample_rate(struct dw_hdmi_qp *hdmi, unsigned long lo
 	n = dw_hdmi_qp_find_n(hdmi, tmds_char_rate, sample_rate);
 	cts = dw_hdmi_qp_find_cts(hdmi, tmds_char_rate, sample_rate);
 
+	/*
+	 * When no CTS table entry exists for the given TMDS rate, compute
+	 * CTS from N rather than letting the hardware auto-measure.  The
+	 * auto-CTS circuit produces incorrect audio timing at out-of-table
+	 * rates (e.g. 185.625 MHz, 297 MHz, 594 MHz), causing strict HDMI
+	 * sinks to mute audio.  Computed CTS = (TMDS * N) / (128 * Fs) per

drivers/gpu/drm/bridge/synopsys/dw-hdmi-qp.c static void dw_hdmi_qp_set_sample_rate(struct dw_hdmi_qp *hdmi, unsigned long lo · line 479

Quoted 10 lines
drivers/gpu/drm/bridge/synopsys/dw-hdmi-qp.c +10 -0
+	 * HDMI spec; the standard override path then supplies it on the
+	 * wire.  Mirrors hdmi_set_clk_regenerator() in the legacy dw-hdmi
+	 * driver.
+	 */
+	if (!cts && n) {
+		u64 computed = (u64)tmds_char_rate * n;
+
+		do_div(computed, 128ULL * sample_rate);
+		cts = (unsigned int)computed;
+	}

Could this function instead be converted to use the helper drm_hdmi_acr_get_n_cts() to lookup/calculate the n and cts value?

At first glance it looks to produce same/similar values as this driver tries to do open-coded.

Regards,
Jonas

Quoted 6 lines
drivers/gpu/drm/bridge/synopsys/dw-hdmi-qp.c +1 -0
+
 	dw_hdmi_qp_set_cts_n(hdmi, cts, n);
 }
 

base-commit: c1079aebb4de218caa86c44f9a53700d1a582683

#2 lore raw
Quoted 2 lines · Jonas Karlman

Could this function instead be converted to use the helper drm_hdmi_acr_get_n_cts() to lookup/calculate the n and cts value?

Done in v3:

https://lore.kernel.org/linux-rockchip/[email protected]/

Re-tested on the LG G3 OLED (the original reporting sink) at 148.5 / 185.625 / 297 / 594 MHz; audio plays at all four. v3 also drops ~200 lines of open-coded N/CTS tables and search functions that the helper makes redundant.

Simon