Qualcomm Technologies, Inc. GPU Qualcomm Technologies, Inc. Adreno GPU Required properties: - compatible: Must be "qcom,kgsl-3d0". May also includes "qcom,adreno-gpu-*" for few targets. Must include "qcom,adreno-gpu-a619-holi" for Holi target. Must include "qcom,adreno-gpu-a621" for Neo target. Must include "qcom,adreno-gpu-a660-shima" for Shima target. Must include "qcom,adreno-gpu-gen7-0-0" for Waipio target. Must include "qcom,adreno-gpu-gen7-0-1" for Waipio V2 target. Must include "qcom,adreno-gpu-gen7-2-0" for Kalama target. Must include "qcom,adreno-gpu-gen7-2-1" for Kalama V2 target. Must include "qcom,adreno-gpu-gen7-4-0" for Cape target. Must include "qcom,adreno-gpu-gen7-3-0" for Parrot target. Must include "qcom,adreno-gpu-gen7-9-0" for Pineapple target. Must include "qcom,adreno-gpu-gen7-9-1" for Pineapple V2 target. Must include "qcom,adreno-gpu-gen7-17-0" for Kera target. Must include "qcom,adreno-gpu-gen8-6-0" for Tuna target. - reg: Specifies the list of register regions for the device. - reg-names: Resource names used for the register regions specified in reg. - interrupts: Interrupt mapping for GPU nterrupts. - interrupt-names: String property to describe the names of the interrupts. - qcom,gpu-bimc-interface-clk-freq: GPU-BIMC interface clock needs to set to this value for targets where B/W requirements does not meet GPU Turbo use cases. - clocks: List of phandle and clock specifier pairs, one pair for each clock input to the device. - clock-names: List of clock input name strings sorted in the same order as the clocks property. - qcom,base-leakage-coefficient: Dynamic leakage coefficient. - qcom,lm-limit: Current limit for GPU limit management. - qcom,isense-clk-on-level: below or equal this power level isense clock is at XO rate, above this powerlevel isense clock is at working frequency. Bus Scaling Data: - qcom,gpu-bus-table: Defines a bus voting table with the below properties. Multiple sets of bus voting tables can be defined for given platform based on the type of ddr system. Properties: - compatible: Must be "qcom,gpu-bus-table". Additionally, "qcom,gpu-bus-table-ddr" must also be provided, with the ddr type value(integer) appended to the string. - qcom,msm-bus,name: String property to describe the name of the 3D graphics processor. - qcom,msm-bus,num-cases: This is the the number of Bus Scaling use cases defined in the vectors property. - qcom,msm-bus,active-only: A boolean flag indicating if it is active only. - qcom,msm-bus,num-paths: This represents the number of paths in each Bus Scaling Usecase. - qcom,msm-bus,vectors-KBps: A series of 4 cell properties, format of which is: , , // For Bus Scaling Usecase 1 , , // For Bus Scaling Usecase 2 <.. .. .. ..>, <.. .. .. ..>; // For Bus Scaling Usecase n This property is a series of all vectors for all Bus Scaling Usecases. Each set of vectors for each usecase describes bandwidth votes for a combination of src/dst ports. The driver will set the desired use case based on the selected power level and the desired bandwidth vote will be registered for the port pairs. Current values of src are: 0 = MSM_BUS_MASTER_GRAPHICS_3D 1 = MSM_BUS_MASTER_GRAPHICS_3D_PORT1 2 = MSM_BUS_MASTER_V_OCMEM_GFX3D Current values of dst are: 0 = MSM_BUS_SLAVE_EBI_CH0 1 = MSM_BUS_SLAVE_OCMEM ab: Represents aggregated bandwidth. This value is 0 for Graphics. ib: Represents instantaneous bandwidth. This value has a range <0 8000 MB/s> - qcom,ocmem-bus-client: Container for another set of bus scaling properties qcom,msm-bus,name qcom,msm-bus,num-cases qcom,msm-bus,num-paths qcom,msm-bus,vectors-KBps to be used by ocmem msm bus scaling client. GDSC Oxili Regulators: - regulator-names: List of regulator name strings sorted in power-on order - vddcx-supply: Phandle for vddcx regulator device node. - vdd-supply: Phandle for vdd regulator device node. Power Domains: - power-domains: List of PM domain specifiers that reference each power-domain used by the GPU - power-domain-names: List of names that represent each of the specifiers in the 'power-domains' property. Includes 'cx', 'gx' and 'gmu_cx' which represent the power-domains for CX GDSC, GX GDSC and GMU CX GDSC respectively. IOMMU Data: - iommu: Phandle for the KGSL IOMMU device node GPU Power levels: - qcom,gpu-pwrlevel-bins: Container for sets of GPU power levels (see adreno-pwrlevels.txt) DCVS Core info - qcom,dcvs-core-info Container for the DCVS core info (see dcvs-core-info.txt) Optional Properties: - qcom,initial-powerlevel: This value indicates which qcom,gpu-pwrlevel should be used at start time and when coming back out of resume - qcom,throttle-pwrlevel: This value indicates which qcom,gpu-pwrlevel LM throttling may start to occur - qcom,bus-control: Boolean. Enables an independent bus vote from the gpu frequency - qcom,bus-width: Bus width in number of bytes. This enables dynamic AB bus voting based on bus width and actual bus transactions. - qcom,bus-accesses: Parameter for tuning bus dcvs. - qcom,bus-accesses-ddrX: Parameter for tuning bus dcvs for each DDR configuration where X will be the return value from of_fdt_get_ddrtype(). - qcom,gpubw-dev: a phandle to a device representing bus bandwidth requirements (see devdw.txt) - qcom,idle-timeout: This property represents the time in milliseconds for idle timeout. - qcom,no-nap: If it exists software clockgating will be disabled at boot time. - qcom,chipid: If it exists this property is used to replace the chip identification read from the GPU hardware. This is used to override faulty hardware readings. - qcom,gpu-model: If it exists this property is used for GPU model name. - qcom,vk-device-id: If it exists this property is used to specify vulkan device ID. - qcom,disable-wake-on-touch: Boolean. Disables the GPU power up on a touch input event. - qcom,disable-busy-time-burst: Boolean. Disables the busy time burst to avoid switching of power level for large frames based on the busy time limit. - qcom,pm-qos-active-latency: Right after GPU wakes up from sleep, driver votes for acceptable maximum latency to the pm-qos driver. This voting demands that the system can not go into any power save state *if* the latency to bring system back into active state is more than this value. Value is in microseconds. - qcom,pm-qos-wakeup-latency: Similar to the above. Driver votes against deep low power modes right before GPU wakes up from sleep. - qcom,l2pc-cpu-mask-latency: The CPU mask latency in microseconds to avoid L2PC on masked CPUs. - qcom,gpu-cx-ipeak: CX Ipeak is a mitigation scheme which throttles cDSP frequency if all the clients are running at their respective threshold frequencies to limit CX peak current. phandle - phandle of CX Ipeak device node bit - Every bit corresponds to a client of CX Ipeak driver in the relevant register. - qcom, gpu-cx-ipeak-freq: GPU frequency threshold for CX Ipeak voting. GPU votes to CX Ipeak driver when GPU clock crosses this threshold. CX Ipeak can limit peak current based on voting from other clients. - qcom,force-32bit: Force the GPU to use 32 bit data sizes even if it is capable of doing 64 bit. - qcom,gpu-speed-bin: GPU speed bin information in the format offset - offset of the efuse register from the base. mask - mask for the relevant bits in the efuse register. shift - number of bits to right shift to get the speed bin value. - qcom,gpu-disable-fuse: GPU disable fuse offset - offset of the efuse register from the base. mask - mask for the relevant bits in the efuse register. shift - number of bits to right shift to get the disable_gpu fuse bit value. - qcom,soc-hw-rev-efuse: SOC hardware revision fuse information in the format offset - offset of the efuse register from the base. bit_position - hardware revision starting bit in the efuse register. mask - mask for the relevant bits in the efuse register. - qcom,highest-bank-bit: Specify the bit of the highest DDR bank. This is programmed into protected registers and also passed to the user as a property. - qcom,min-access-length: Specify the minimum access length for the chip. Either 32 or 64 bytes. Based on the above options, program the appropriate bit into certain protected registers and also pass to the user as a property. - qcom,ubwc-mode: Specify the ubwc mode for this chip. 1: UBWC 1.0 2: UBWC 2.0 3: UBWC 3.0 4: UBWC 4.0 5: UBWC 5.0 Based on the ubwc mode, program the appropriate bit into certain protected registers and also pass to the user as a property. - qcom,l2pc-cpu-mask: Disables L2PC on masked CPUto the string.rendering thread is running on masked CPUs. Bit 0 is for CPU-0, bit 1 is for CPU-1... - qcom,l2pc-update-queue: Disables L2PC on masked CPUs at queue time when it's true. - qcom,snapshot-size: Specify the size of snapshot in bytes. This will override snapshot size defined in the driver code. - qcom,enable-ca-jump: Boolean. Enables use of context aware DCVS - qcom,ca-busy-penalty: This property represents the time in microseconds required to initiate context aware power level jump. - qcom,ca-target-pwrlevel: This value indicates which qcom,gpu-pwrlevel to jump on in case of context aware power level jump. - qcom,gpu-qdss-stm: baseAddr - base address of the gpu channels in the qdss stm memory region size - size of the gpu stm region - qcom,gpu-timer: baseAddr - base address of the qtimer memory region size - size of the qtimer region - qcom,tzone-names: Specify the names of GPU thermal zones. These will be used to get gpu temperature from the thermal driver API. nvmem-cells: A phandle to the configuration data such as gpu speed bin, gpu gaming mode, gpu model name provided by a nvmem device. If unspecified default values shall be used. nvmem-cell-names: Should be "speed_bin", "gaming_bin", "gpu_model" GPU Quirks: - qcom,gpu-quirk-two-pass-use-wfi: Signal the GPU to set Set TWOPASSUSEWFI bit in PC_DBG_ECO_CNTL (5XX and 6XX only) - qcom,gpu-quirk-critical-packets: Submit a set of critical PM4 packets when the GPU wakes up - qcom,gpu-quirk-fault-detect-mask: Mask out RB1-3 activity signals from HW hang detection logic - qcom,gpu-quirk-dp2clockgating-disable: Disable RB sampler data path clock gating optimization - qcom,gpu-quirk-lmloadkill-disable: Use register setting to disable local memory(LM) feature to avoid corner case error - qcom,gpu-quirk-hfi-use-reg: Use registers to replace DCVS HFI message to avoid GMU failure to access system memory during IFPC - qcom,gpu-quirk-limit-uche-gbif-rw: Limit number of read and write transactions from UCHE block to GBIF to avoid possible deadlock between GBIF, SMMU and MEMNOC. - qcom,gpu-quirk-mmu-secure-cb-alt: Select alternate secure context bank to generate SID1 for secure playback. KGSL Memory Pools: - qcom,gpu-mempools: Container for sets of GPU mempools.Multiple sets (pools) can be defined within qcom,gpu-mempools. Each mempool defines a pool order, reserved pages, allocation allowed. Properties: - compatible: Must be qcom,gpu-mempools. - qcom,mempool-max-pages: Max pages for all mempools, If not defined there is no limit. - qcom,gpu-mempool: Defines a set of mempools. Properties: - reg: Index of the pool (0 = lowest pool order). - qcom,mempool-page-size: Size of page. - qcom,mempool-reserved: Number of pages reserved at init time for a pool. - qcom,mempool-allocate: Allocate memory from the system memory when the reserved pool exhausted. - qcom,mempool-max-pages: Limit on max pages this pool can hold, If not defined there is no limit. GPU model configuration: - qcom,gpu-models: Container of sets of GPU model names specified by qcom,gpu-models. Properties: - compatible: Must be qcom,gpu-models. - qcom,gpu-model: Defines a GPU model name for specific GPU model ID. Properties: - compatible: May also include "qcom,adreno-gpu-*" for few targets. - qcom,gpu-model-id: Identifier for the specific GPU hardware configuration - must match the value read from the hardware. - qcom,gpu-model: GPU model name for a specific GPU hardware. - qcom,vk-device-id: Vulkan device id unique for specific GPU hardware model. SOC Hardware revisions: - qcom,soc-hw-revisions: Container of sets of SOC hardware revisions specified by qcom,soc-hw-revision. Properties: - compatible: Must be qcom,soc-hw-revisions. - qcom,soc-hw-revision: Defines a SOC hardware revision. Properties: - qcom,soc-hw-revision: Identifier for the hardware revision - must match the value read from the hardware. - qcom,chipid: GPU Chip ID to be used for this hardware revision. - qcom,gpu-quirk-*: GPU quirks applicable for this hardware revision. GPU LLC slice info: - cache-slice-names: List of LLC cache slices for GPU transactions and pagetable walk. - cache-slices: phandle to the system LLC driver, cache slice index. L3 Power levels: - qcom,l3-pwrlevels: Container for sets of L3 power levels, the L3 frequency is adjusted according to the performance hint received from userspace. Properties: - compatible: Must be qcom,l3-pwrlevels - qcom,l3-pwrlevel: A single L3 powerlevel Properties: - reg: Index of the L3 powerlevel 0 = powerlevel for no L3 vote 1 = powerlevel for medium L3 vote 2 = powerlevel for maximum L3 vote - qcom,l3-freq: The L3 frequency for the powerlevel (in Hz) GPU coresight info: The following properties are optional as collecting data via coresight might not be supported for every chipset. The documentation for coresight properties can be found in: Documentation/devicetree/bindings/coresight/coresight.txt - qcom,gpu-coresights: Container for sets of GPU coresight sources. - coresight-id: Unique integer identifier for the bus. - coresight-name: Unique descriptive name of the bus. - coresight-nr-inports: Number of input ports on the bus. - coresight-outports: List of output port numbers on the bus. - coresight-child-list: List of phandles pointing to the children of this component. - coresight-child-ports: List of input port numbers of the children. - coresight-atid: The unique ATID value of the coresight device Example of A330 GPU in MSM8916: &soc { msm_gpu: qcom,kgsl-3d0@1c00000 { label = "kgsl-3d0"; compatible = "qcom,kgsl-3d0", "qcom,kgsl-3d"; reg = <0x1c00000 0x10000 0x1c20000 0x20000>; reg-names = "kgsl_3d0_reg_memory" , "kgsl_3d0_shader_memory"; interrupts = <0 33 0>; interrupt-names = "kgsl_3d0_irq"; qcom,id = <0>; qcom,chipid = <0x03000600>; qcom,initial-pwrlevel = <1>; /* Idle Timeout = HZ/12 */ qcom,idle-timeout = <8>; qcom,strtstp-sleepwake; clocks = <&clock_gcc clk_gcc_oxili_gfx3d_clk>, <&clock_gcc clk_gcc_oxili_ahb_clk>, <&clock_gcc clk_gcc_oxili_gmem_clk>, <&clock_gcc clk_gcc_bimc_gfx_clk>, <&clock_gcc clk_gcc_bimc_gpu_clk>; clock-names = "core_clk", "iface_clk", "mem_clk", "mem_iface_clk", "alt_mem_iface_clk"; /* Bus Scale Settings */ qcom, gpu-bus-table { compatible="qcom,gpu-bus-table","qcom,gpu-bus-table-ddr7"; qcom,msm-bus,name = "grp3d"; qcom,msm-bus,num-cases = <4>; qcom,msm-bus,num-paths = <1>; qcom,msm-bus,vectors-KBps = <26 512 0 0>, <26 512 0 1600000>, <26 512 0 3200000>, <26 512 0 4264000>; }; /* GDSC oxili regulators */ vdd-supply = <&gdsc_oxili_gx>; nvmem-cells = <&gpu_speed_bin>, <&gpu_gaming_bin>, <&gpu_model_bin>; nvmem-cell-names = "speed_bin", "gaming_bin","gpu_model"; /* IOMMU Data */ iommu = <&gfx_iommu>; /* Trace bus */ coresight-id = <67>; coresight-name = "coresight-gfx"; coresight-nr-inports = <0>; coresight-outports = <0>; coresight-child-list = <&funnel_in0>; coresight-child-ports = <5>; /* Enable context aware freq. scaling */ qcom,enable-ca-jump; /* Context aware jump busy penalty in us */ qcom,ca-busy-penalty = <12000>; /* Context aware jump target power level */ qcom,ca-target-pwrlevel = <1>; qcom,soc-hw-revisions { #address-cells = <1>; #size-cells = <0>; compatible="qcom,soc-hw-revisions"; qcom,soc-hw-revision@0 { reg = <0>; qcom,chipid = <0x06010500>; qcom,gpu-quirk-hfi-use-reg; qcom,gpu-quirk-limit-uche-gbif-rw; }; qcom,soc-hw-revision@1 { reg = <1>; qcom,chipid = <0x06010501>; qcom,gpu-quirk-hfi-use-reg; }; }; qcom,gpu-models { #address-cells = <1>; #size-cells = <0>; compatible="qcom,gpu-models"; qcom,gpu-model@0 { compatible="qcom,adreno-gpu-a642l"; qcom,gpu-model-id = <0>; qcom,gpu-model = "Adreno642Lv1"; qcom,vk-device-id= <0x06030500>; }; qcom,gpu-model@1 { compatible="qcom,adreno-gpu-a645"; qcom,gpu-model-id = <190>; qcom,gpu-model = "Adreno645"; qcom,vk-device-id= <0x06030500>; }; } /* GPU Mempools */ qcom,gpu-mempools { #address-cells= <1>; #size-cells = <0>; compatible = "qcom,gpu-mempools"; /* 4K Page Pool configuration */ qcom,gpu-mempool@0 { reg = <0>; qcom,mempool-page-size = <4096>; qcom,mempool-reserved = <2048>; qcom,mempool-allocate; }; /* 8K Page Pool configuration */ qcom,gpu-mempool@1 { reg = <1>; qcom,mempool-page-size = <8192>; qcom,mempool-reserved = <1024>; qcom,mempool-allocate; }; /* 64K Page Pool configuration */ qcom,gpu-mempool@2 { reg = <2>; qcom,mempool-page-size = <65536>; qcom,mempool-reserved = <256>; }; /* 1M Page Pool configuration */ qcom,gpu-mempool@3 { reg = <3>; qcom,mempool-page-size = <1048576>; qcom,mempool-reserved = <32>; }; }; /* Power levels */ qcom,gpu-pwrlevels-bins { #address-cells = <1>; #size-cells = <0>; qcom,gpu-pwrlevels-0 { #address-cells = <1>; #size-cells = <0>; qcom,speed-bin = <0>; qcom,ca-target-pwrlevel = <1>; qcom,gpu-pwrlevel@0 { reg = <0>; qcom,gpu-freq = <400000000>; qcom,bus-freq = <3>; qcom,io-fraction = <33>; }; qcom,gpu-pwrlevel@1 { reg = <1>; qcom,gpu-freq = <310000000>; qcom,bus-freq = <2>; qcom,io-fraction = <66>; }; qcom,gpu-pwrlevel@2 { reg = <2>; qcom,gpu-freq = <200000000>; qcom,bus-freq = <1>; qcom,io-fraction = <100>; }; qcom,gpu-pwrlevel@3 { reg = <3>; qcom,gpu-freq = <27000000>; qcom,bus-freq = <0>; qcom,io-fraction = <0>; }; }; }; }; };