[PATCH v2 0/4] drm/msm/adreno: Add Adreno 620

[PATCH v2 0/4] drm/msm/adreno: Add Adreno 620

[Ishu Brar via B4 Relay]

This adds Adreno 620 GPU found in Qualcomm SM7250 (Snapdragon
765/765G/768G) to msm driver. It is A650 derivative using 
a650 SQE/GMU firmware and a620_zap.mbn. On top of catalog entry it
needs two A620 specific GMU changes:

 - A620 has no PDC setup from AOP, so PDC sequencer is
   programmed from driver (xo/mx/cx addresses from cmd-db). Without
   it RSC never acks second power-up and device hangs.
 - A bandwidth table with BCM addresses of this SoC. The generic
   A630 one does not exist on SM7250, so first GX performance vote
   times out.

Tested on OnePlus Nord (SM7250) running mainline: the GPU, GMU and
Vulkan work. This needs the SM7250 SMMU series [1] as well, otherwise
the device is reset when the first per-process pagetable is created.
The SoC devicetree nodes are not part of this series; they will be sent
separately once the SoC dtsi is accepted.

[1] https://lore.kernel.org/all/20261023-smmu-sm7250-v2-0-8bb881ce19f5@brar.si/

Signed-off-by: Ishu Brar <patches@brar.si>
---
Changes in v2:
- Patch 3: also program the VRM command (vrm.soc + 4) in TCS1 and TCS3
  when cmd-db has it, as the downstream driver does (Konrad).
- Patch 4: GMU clock gating mode is 0x00020202 as downstream (was
  0x00020200); add the SH0/MC0/ACV BCM list for the generic bandwidth
  table. The bus widths are the ones the other A6xx/A7xx parts use.
- Patch 2 is kept for now: the generic table is only used when the OPP
  table has bandwidth levels, which needs the SM7250 interconnect
  driver and DT that are not upstream yet. Without it the A620 would
  fall back to the A630 table.
- Testing: v2 was re-tested on the OnePlus Nord (these changes on top
  of my 7.3-rc7 port tree, not on msm-next), with the interconnect
  path and OPP bandwidth levels in the GPU node. Vulkan workloads ran
  across GPU power cycles and all four GPU frequencies; each frequency
  selected its own bandwidth level, the generated BCM votes scale
  linearly with the OPP bandwidth and the GMU accepted the table. I
  could not read back the resulting DDR frequency.
- Protect map (patch 4 follow-up suggested by Konrad): not included
  yet, waiting for the downstream reference of the 0x1c800 range.
- Link to v1: https://patch.msgid.link/20261016-adreno-a620-v1-0-4a7ed6210d7d@brar.si

To: Rob Clark <robin.clark@oss.qualcomm.com>
To: Dmitry Baryshkov <lumag@kernel.org>
To: Abhinav Kumar <abhinav.kumar@linux.dev>
To: Jessica Zhang <jesszhan0024@gmail.com>
To: Sean Paul <sean@poorly.run>
To: Marijn Suijten <marijn.suijten@somainline.org>
To: David Airlie <airlied@gmail.com>
To: Simona Vetter <simona@ffwll.ch>
To: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
To: Maxime Ripard <mripard@kernel.org>
To: Thomas Zimmermann <tzimmermann@suse.de>
To: Rob Herring <robh@kernel.org>
To: Krzysztof Kozlowski <krzk+dt@kernel.org>
To: Conor Dooley <conor+dt@kernel.org>
To: Konrad Dybcio <konradybcio@kernel.org>
To: Akhil P Oommen <akhilpo@oss.qualcomm.com>
Cc: linux-arm-msm@vger.kernel.org
Cc: dri-devel@lists.freedesktop.org
Cc: freedreno@lists.freedesktop.org
Cc: devicetree@vger.kernel.org
Cc: linux-kernel@vger.kernel.org

---
Ishu Brar (4):
      dt-bindings: display: msm: Document Adreno 620 and its GMU
      drm/msm/a6xx: Add a bandwidth table for the Adreno 620 GMU
      drm/msm/a6xx: Program the PDC sequencer on Adreno 620
      drm/msm/adreno: Add Adreno 620

 .../devicetree/bindings/display/msm/gmu.yaml       |  2 ++
 .../devicetree/bindings/display/msm/gpu.yaml       |  1 +
 drivers/gpu/drm/msm/adreno/a6xx_catalog.c          | 35 ++++++++++++++++++++++
 drivers/gpu/drm/msm/adreno/a6xx_gmu.c              | 35 ++++++++++++++++++----
 drivers/gpu/drm/msm/adreno/a6xx_hfi.c              | 34 +++++++++++++++++++++
 drivers/gpu/drm/msm/adreno/adreno_gpu.h            |  5 ++++
 6 files changed, 106 insertions(+), 6 deletions(-)
---
base-commit: d024361c631b3f020a26c07d683e5103404b0fbf
change-id: 20261016-adreno-a620-faee699b382f

Best regards,
--  
Ishu Brar <patches@brar.si>

[PATCH v2 1/4] dt-bindings: display: msm: Document Adreno 620 and its GMU

[Ishu Brar via B4 Relay]

From: Ishu Brar <patches@brar.si>

Add the Adreno 620 found in the Qualcomm SM7250 and its GMU. The GMU
has the same registers, clocks, interrupts and power domains as the
650.2 one. As on the other A6xx parts, the GPU clocks are defined in
the GMU node, so the GPU node has none.

Signed-off-by: Ishu Brar <patches@brar.si>
---
 Documentation/devicetree/bindings/display/msm/gmu.yaml | 2 ++
 Documentation/devicetree/bindings/display/msm/gpu.yaml | 1 +
 2 files changed, 3 insertions(+)

diff --git a/Documentation/devicetree/bindings/display/msm/gmu.yaml b/Documentation/devicetree/bindings/display/msm/gmu.yaml
index 07f0e547b..3478adebd 100644
--- a/Documentation/devicetree/bindings/display/msm/gmu.yaml
+++ b/Documentation/devicetree/bindings/display/msm/gmu.yaml
@@ -217,6 +217,7 @@ allOf:
         compatible:
           contains:
             enum:
+              - qcom,adreno-gmu-620.0
               - qcom,adreno-gmu-650.2
     then:
       properties:
@@ -238,6 +239,7 @@ allOf:
         compatible:
           contains:
             enum:
+              - qcom,adreno-gmu-620.0
               - qcom,adreno-gmu-640.1
               - qcom,adreno-gmu-650.2
     then:
diff --git a/Documentation/devicetree/bindings/display/msm/gpu.yaml b/Documentation/devicetree/bindings/display/msm/gpu.yaml
index c30de7feb..cf015fdee 100644
--- a/Documentation/devicetree/bindings/display/msm/gpu.yaml
+++ b/Documentation/devicetree/bindings/display/msm/gpu.yaml
@@ -435,6 +435,7 @@ allOf:
               - qcom,adreno-615.0
               - qcom,adreno-618.0
               - qcom,adreno-619.0
+              - qcom,adreno-620.0
               - qcom,adreno-621.0
               - qcom,adreno-623.0
               - qcom,adreno-630.2

-- 
2.56.0

[PATCH v2 2/4] drm/msm/a6xx: Add a bandwidth table for the Adreno 620 GMU

[Ishu Brar via B4 Relay]

From: Ishu Brar <patches@brar.si>

Without an A620 specific table the GMU falls back to the generic A630
one, whose BCM addresses do not exist on SM7250, and the first GX
performance vote times out.

Like the A650 one, send a single "off" entry just to get things running,
with the BCM addresses looked up in cmd-db. This is only used while the
GPU OPP table has no bandwidth levels; with them, the table is generated
from the BCMs listed in the catalog entry. Add adreno_is_a620() for the
selection.

Assisted-by: Claude
Signed-off-by: Ishu Brar <patches@brar.si>
---
 drivers/gpu/drm/msm/adreno/a6xx_hfi.c   | 34 +++++++++++++++++++++++++++++++++
 drivers/gpu/drm/msm/adreno/adreno_gpu.h |  5 +++++
 2 files changed, 39 insertions(+)

diff --git a/drivers/gpu/drm/msm/adreno/a6xx_hfi.c b/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
index 6b34e704a..b59853678 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
+++ b/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
@@ -553,6 +553,38 @@ static void a650_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
 	msg->cnoc_cmds_data[1][0] =  0x60000001;
 }
 
+static void a620_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
+{
+	/*
+	 * Used when the OPP table has no bandwidth levels, otherwise the
+	 * table is generated from the BCMs in the catalog. Send a single
+	 * "off" entry just to get things running.
+	 */
+	msg->bw_level_num = 1;
+
+	msg->ddr_cmds_num = 3;
+	msg->ddr_wait_bitmask = 0x01;
+
+	msg->ddr_cmds_addrs[0] = cmd_db_read_addr("MC0");
+	msg->ddr_cmds_addrs[1] = cmd_db_read_addr("SH0");
+	msg->ddr_cmds_addrs[2] = cmd_db_read_addr("ACV");
+
+	msg->ddr_cmds_data[0][0] =  0x40000000;
+	msg->ddr_cmds_data[0][1] =  0x40000000;
+	msg->ddr_cmds_data[0][2] =  0x40000000;
+
+	/*
+	 * These are the CX (CNOC) votes - these are used by the GMU but the
+	 * votes are known and fixed for the target
+	 */
+	msg->cnoc_cmds_num = 1;
+	msg->cnoc_wait_bitmask = 0x01;
+
+	msg->cnoc_cmds_addrs[0] = cmd_db_read_addr("CN0");
+	msg->cnoc_cmds_data[0][0] =  0x40000000;
+	msg->cnoc_cmds_data[1][0] =  0x60000001;
+}
+
 static void a690_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
 {
 	/*
@@ -818,6 +850,8 @@ static int a6xx_hfi_send_bw_table(struct a6xx_gmu *gmu)
 		a619_build_bw_table(msg);
 	else if (adreno_is_a640_family(adreno_gpu))
 		a640_build_bw_table(msg);
+	else if (adreno_is_a620(adreno_gpu))
+		a620_build_bw_table(msg);
 	else if (adreno_is_a650(adreno_gpu))
 		a650_build_bw_table(msg);
 	else if (adreno_is_7c3(adreno_gpu))
diff --git a/drivers/gpu/drm/msm/adreno/adreno_gpu.h b/drivers/gpu/drm/msm/adreno/adreno_gpu.h
index 4b2156825..b77502446 100644
--- a/drivers/gpu/drm/msm/adreno/adreno_gpu.h
+++ b/drivers/gpu/drm/msm/adreno/adreno_gpu.h
@@ -443,6 +443,11 @@ static inline int adreno_is_a619_holi(const struct adreno_gpu *gpu)
 	return adreno_is_a619(gpu) && adreno_has_gmu_wrapper(gpu);
 }
 
+static inline int adreno_is_a620(const struct adreno_gpu *gpu)
+{
+	return gpu->info->chip_ids[0] == 0x06020000;
+}
+
 static inline int adreno_is_a621(const struct adreno_gpu *gpu)
 {
 	return gpu->info->chip_ids[0] == 0x06020100;

-- 
2.56.0

[PATCH v2 3/4] drm/msm/a6xx: Program the PDC sequencer on Adreno 620

[Ishu Brar via B4 Relay]

From: Ishu Brar <patches@brar.si>

Unlike the rest of the A650 family, the A620 does not get its PDC set up
by AOP (the downstream devicetree has no pdc_in_aop for it). Without the
sequence, the RSC never acks the second power-up ("Unable to power on
the GPU RSC") and the device hangs.

Program the sequence from the driver, taking the xo, mx and cx resource
addresses from cmd-db instead of the hardcoded ones.

Assisted-by: Claude
Signed-off-by: Ishu Brar <patches@brar.si>
---
 drivers/gpu/drm/msm/adreno/a6xx_gmu.c | 35 +++++++++++++++++++++++++++++------
 1 file changed, 29 insertions(+), 6 deletions(-)

diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
index 662f09bec..8d426f1f1 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
@@ -686,6 +686,7 @@ static void a6xx_gmu_rpmh_init(struct a6xx_gmu *gmu)
 	u32 seqmem0_drv0_reg = REG_A6XX_RSCC_SEQ_MEM_0_DRV0;
 	void __iomem *seqptr = NULL;
 	uint32_t pdc_address_offset;
+	u32 mx_addr = 0x30010, cx_addr = 0x30000, vrm_addr = 0;
 	void __iomem *pdcptr;
 	bool pdc_in_aop = false;
 
@@ -697,8 +698,18 @@ static void a6xx_gmu_rpmh_init(struct a6xx_gmu *gmu)
 	if (IS_ERR(pdcptr))
 		return;
 
-	if (adreno_is_a650_family(adreno_gpu) ||
-	    adreno_is_a7xx(adreno_gpu))
+	/* Unlike the rest of the A650 family, A620 has no PDC setup from AOP */
+	if (adreno_is_a620(adreno_gpu)) {
+		pdc_address_offset = cmd_db_read_addr("xo.lvl");
+		mx_addr = cmd_db_read_addr("mx.lvl");
+		cx_addr = cmd_db_read_addr("cx.lvl");
+		vrm_addr = cmd_db_read_addr("vrm.soc");
+		if (!pdc_address_offset || !mx_addr || !cx_addr) {
+			DRM_DEV_ERROR(gmu->dev, "No PDC resource addresses in cmd-db\n");
+			return;
+		}
+	} else if (adreno_is_a650_family(adreno_gpu) ||
+		 adreno_is_a7xx(adreno_gpu))
 		pdc_in_aop = true;
 	else if (adreno_is_a618(adreno_gpu) || adreno_is_a640_family(adreno_gpu))
 		pdc_address_offset = 0x30090;
@@ -766,25 +777,31 @@ static void a6xx_gmu_rpmh_init(struct a6xx_gmu *gmu)
 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_WAIT_FOR_CMPL_BANK, 0);
 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CONTROL, 0);
 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID, 0x10108);
-	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR, 0x30010);
+	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR, mx_addr);
 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA, 1);
 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 4, 0x10108);
-	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 4, 0x30000);
+	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 4, cx_addr);
 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 4, 0x0);
 
 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 8, 0x10108);
 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 8, pdc_address_offset);
 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 8, 0x0);
 
+	if (vrm_addr) {
+		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 12, 0x10108);
+		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 12, vrm_addr + 0x4);
+		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 12, 0x0);
+	}
+
 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_ENABLE_BANK, 7);
 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_WAIT_FOR_CMPL_BANK, 0);
 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CONTROL, 0);
 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID, 0x10108);
-	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR, 0x30010);
+	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR, mx_addr);
 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA, 2);
 
 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 4, 0x10108);
-	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 4, 0x30000);
+	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 4, cx_addr);
 	if (adreno_is_a618(adreno_gpu) || adreno_is_a619(adreno_gpu) ||
 			adreno_is_a650_family(adreno_gpu))
 		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x2);
@@ -794,6 +811,12 @@ static void a6xx_gmu_rpmh_init(struct a6xx_gmu *gmu)
 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 8, pdc_address_offset);
 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 8, 0x3);
 
+	if (vrm_addr) {
+		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 12, 0x10108);
+		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 12, vrm_addr + 0x4);
+		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 12, 0x1);
+	}
+
 	/* Setup GPU PDC */
 setup_pdc:
 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_SEQ_START_ADDR, 0);

-- 
2.56.0

[PATCH v2 4/4] drm/msm/adreno: Add Adreno 620

[Ishu Brar via B4 Relay]

From: Ishu Brar <patches@brar.si>

Add the catalog entry for the Adreno 620 (chip id 0x06020000) found in
the Snapdragon 765/765G/768G (SM7250). It is an A650 derivative and
uses its SQE and GMU firmware and protect table, together with the A620
zap shader.

The speed-bin fuse values map to the OPP bins as in the downstream
"lito" devicetree: fuse 0 is the default bin and fuse 132 and 115 are
the two alternate ones.

The GMU clock gating mode is 0x00020202, which is what the downstream
driver programs on every A6xx except the A612 and A615 family. List the
SH0/MC0/ACV bus clock managers like the other A6xx and A7xx parts do, so
the bandwidth table is generated from the OPP table once it carries
bandwidth votes; until then the table added earlier in this series is
used. The bus widths are the ones the other parts use.

Signed-off-by: Ishu Brar <patches@brar.si>
---
 drivers/gpu/drm/msm/adreno/a6xx_catalog.c | 35 +++++++++++++++++++++++++++++++
 1 file changed, 35 insertions(+)

diff --git a/drivers/gpu/drm/msm/adreno/a6xx_catalog.c b/drivers/gpu/drm/msm/adreno/a6xx_catalog.c
index 4ef2bd6e7..bea11ba6c 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_catalog.c
+++ b/drivers/gpu/drm/msm/adreno/a6xx_catalog.c
@@ -943,6 +943,41 @@ static const struct adreno_info a6xx_gpus[] = {
 			{ 169, 2 },
 			{ 180, 1 },
 		),
+	}, {
+		.chip_ids = ADRENO_CHIP_IDS(0x06020000),
+		.family = ADRENO_6XX_GEN3,
+		.fw = {
+			[ADRENO_FW_SQE] = "a650_sqe.fw",
+			[ADRENO_FW_GMU] = "a650_gmu.bin",
+		},
+		.gmem = SZ_512K,
+		.inactive_period = DRM_MSM_INACTIVE_PERIOD,
+		.quirks = ADRENO_QUIRK_HAS_CACHED_COHERENT |
+			  ADRENO_QUIRK_HAS_HW_APRIV,
+		.funcs = &a6xx_gpu_funcs,
+		.zapfw = "a620_zap.mbn",
+		.a6xx = &(const struct a6xx_info) {
+			.hwcg = a620_hwcg,
+			.protect = &a650_protect,
+			.gbif_cx = a640_gbif,
+			.gmu_cgc_mode = 0x00020202,
+			.prim_fifo_threshold = 0x00010000,
+			.bcms = (const struct a6xx_bcm[]) {
+				{ .name = "SH0", .buswidth = 16 },
+				{ .name = "MC0", .buswidth = 4 },
+				{
+					.name = "ACV",
+					.fixed = true,
+					.perfmode = BIT(3),
+				},
+				{ /* sentinel */ },
+			},
+		},
+		.speedbins = ADRENO_SPEEDBINS(
+			{ 0,   0 },
+			{ 132, 1 },
+			{ 115, 3 },
+		),
 	}, {
 		.chip_ids = ADRENO_CHIP_IDS(0x06020100),
 		.family = ADRENO_6XX_GEN3,

-- 
2.56.0

Re: [PATCH v2 2/4] drm/msm/a6xx: Add a bandwidth table for the Adreno 620 GMU

[Konrad Dybcio]

On 10/23/26 10:49 AM, Ishu Brar via B4 Relay wrote:
> From: Ishu Brar <patches@brar.si>
> 
> Without an A620 specific table the GMU falls back to the generic A630
> one, whose BCM addresses do not exist on SM7250, and the first GX
> performance vote times out.
> 
> Like the A650 one, send a single "off" entry just to get things running,
> with the BCM addresses looked up in cmd-db. This is only used while the
> GPU OPP table has no bandwidth levels; with them, the table is generated
> from the BCMs listed in the catalog entry. Add adreno_is_a620() for the
> selection.

Please just add the BCMs for your GPU:

.bcms = (const struct a6xx_bcm[]) {
	{ .name = "SH0", .buswidth = 16 },
	{ .name = "MC0", .buswidth = 4 },
	{
		.name = "ACV",
		.fixed = true,
		.perfmode = BIT(3),
	},
	{ /* sentinel */ }
};

You can factor it out since it's common across many a6xx GPUs
and &refer to it from the entry for a620

Konrad