Re: [PATCH 3/4] media: platform: qcom/iris: add support for vpu33

[neil.armstrong@linaro.org]

On 04/03/2025 09:38, Dikshita Agarwal wrote:
> 
> 
> On 3/3/2025 10:33 PM, Neil Armstrong wrote:
>> On 28/02/2025 12:39, Dikshita Agarwal wrote:
>>>
>>>
>>> On 2/25/2025 2:35 PM, Neil Armstrong wrote:
>>>> The IRIS acceleration found in the SM8650 platforms uses the vpu33
>>>> hardware version, and requires a slighly different reset and power off
>>>> sequences in order to properly get out of runtime suspend.
>>>>
>>>> Signed-off-by: Neil Armstrong <neil.armstrong@linaro.org>
>>>> ---
>>>>    drivers/media/platform/qcom/iris/Makefile          |   1 +
>>>>    drivers/media/platform/qcom/iris/iris_vpu33.c      | 315
>>>> +++++++++++++++++++++
>>>>    drivers/media/platform/qcom/iris/iris_vpu_common.h |   1 +
>>>>    3 files changed, 317 insertions(+)
>>>>
>>>> diff --git a/drivers/media/platform/qcom/iris/Makefile
>>>> b/drivers/media/platform/qcom/iris/Makefile
>>>> index
>>>> 35390534534e93f4617c1036a05ca0921567ba1d..6b64c9988505afd9707c704449d60bb53209229f 100644
>>>> --- a/drivers/media/platform/qcom/iris/Makefile
>>>> +++ b/drivers/media/platform/qcom/iris/Makefile
>>>> @@ -21,6 +21,7 @@ qcom-iris-objs += \
>>>>                 iris_vdec.o \
>>>>                 iris_vpu2.o \
>>>>                 iris_vpu3.o \
>>>> +             iris_vpu33.o \
>>>>                 iris_vpu_buffer.o \
>>>>                 iris_vpu_common.o \
>>>>    diff --git a/drivers/media/platform/qcom/iris/iris_vpu33.c
>>>> b/drivers/media/platform/qcom/iris/iris_vpu33.c
>>>> new file mode 100644
>>>> index
>>>> 0000000000000000000000000000000000000000..128a050f206f99ec0d43b97ff995fa50d5684150
>>>> --- /dev/null
>>>> +++ b/drivers/media/platform/qcom/iris/iris_vpu33.c
>>>> @@ -0,0 +1,315 @@
>>>> +// SPDX-License-Identifier: GPL-2.0-only
>>>> +/*
>>>> + * Copyright (c) 2022-2024 Qualcomm Innovation Center, Inc. All rights
>>>> reserved.
>>>> + */
>>>> +
>>>> +#include <linux/iopoll.h>
>>>> +#include <linux/reset.h>
>>>> +
>>>> +#include "iris_instance.h"
>>>> +#include "iris_vpu_common.h"
>>>> +#include "iris_vpu_register_defines.h"
>>>> +
>>>> +#define WRAPPER_TZ_BASE_OFFS            0x000C0000
>>>> +#define AON_BASE_OFFS                0x000E0000
>>>> +#define AON_MVP_NOC_RESET            0x0001F000
>>>> +
>>>> +#define WRAPPER_DEBUG_BRIDGE_LPI_CONTROL    (WRAPPER_BASE_OFFS + 0x54)
>>>> +#define WRAPPER_DEBUG_BRIDGE_LPI_STATUS        (WRAPPER_BASE_OFFS + 0x58)
>>>> +#define WRAPPER_IRIS_CPU_NOC_LPI_CONTROL    (WRAPPER_BASE_OFFS + 0x5C)
>>>> +#define REQ_POWER_DOWN_PREP            BIT(0)
>>>> +#define WRAPPER_IRIS_CPU_NOC_LPI_STATUS        (WRAPPER_BASE_OFFS + 0x60)
>>>> +#define WRAPPER_CORE_CLOCK_CONFIG        (WRAPPER_BASE_OFFS + 0x88)
>>>> +#define CORE_CLK_RUN                0x0
>>>> +
>>>> +#define WRAPPER_TZ_CTL_AXI_CLOCK_CONFIG        (WRAPPER_TZ_BASE_OFFS +
>>>> 0x14)
>>>> +#define CTL_AXI_CLK_HALT            BIT(0)
>>>> +#define CTL_CLK_HALT                BIT(1)
>>>> +
>>>> +#define WRAPPER_TZ_QNS4PDXFIFO_RESET        (WRAPPER_TZ_BASE_OFFS + 0x18)
>>>> +#define RESET_HIGH                BIT(0)
>>>> +
>>>> +#define CPU_CS_AHB_BRIDGE_SYNC_RESET        (CPU_CS_BASE_OFFS + 0x160)
>>>> +#define CORE_BRIDGE_SW_RESET            BIT(0)
>>>> +#define CORE_BRIDGE_HW_RESET_DISABLE        BIT(1)
>>>> +
>>>> +#define CPU_CS_X2RPMH                (CPU_CS_BASE_OFFS + 0x168)
>>>> +#define MSK_SIGNAL_FROM_TENSILICA        BIT(0)
>>>> +#define MSK_CORE_POWER_ON            BIT(1)
>>>> +
>>>> +#define AON_WRAPPER_MVP_NOC_RESET_REQ        (AON_MVP_NOC_RESET + 0x000)
>>>> +#define VIDEO_NOC_RESET_REQ            (BIT(0) | BIT(1))
>>>> +
>>>> +#define AON_WRAPPER_MVP_NOC_RESET_ACK        (AON_MVP_NOC_RESET + 0x004)
>>>> +
>>>> +#define VCODEC_SS_IDLE_STATUSN            (VCODEC_BASE_OFFS + 0x70)
>>>> +
>>>> +#define AON_WRAPPER_MVP_NOC_LPI_CONTROL        (AON_BASE_OFFS)
>>>> +#define AON_WRAPPER_MVP_NOC_LPI_STATUS        (AON_BASE_OFFS + 0x4)
>>>> +
>>>> +#define AON_WRAPPER_MVP_NOC_CORE_SW_RESET    (AON_BASE_OFFS + 0x18)
>>>> +#define SW_RESET                BIT(0)
>>>> +#define AON_WRAPPER_MVP_NOC_CORE_CLK_CONTROL    (AON_BASE_OFFS + 0x20)
>>>> +#define NOC_HALT                BIT(0)
>>>> +#define AON_WRAPPER_SPARE            (AON_BASE_OFFS + 0x28)
>>>> +
>>>> +#define VCODEC_DMA_SPARE_3 0x87B8
>>>> +
>>>> +static int reset_control_bulk_assert_id(int num_rstcs,
>>>> +                    struct reset_control_bulk_data *rstcs,
>>>> +                    char *id)
>>>> +{
>>>> +    int i;
>>>> +
>>>> +    for (i = 0; i < num_rstcs; ++i) {
>>>> +        if (!strcmp(rstcs[i].id, id))
>>>> +            return reset_control_assert(rstcs[i].rstc);
>>>> +    }
>>>> +
>>>> +    return -ENODEV;
>>>> +}
>>>> +
>>>> +static int reset_control_bulk_deassert_id(int num_rstcs,
>>>> +                      struct reset_control_bulk_data *rstcs,
>>>> +                      char *id)
>>>> +{
>>>> +    int i;
>>>> +
>>>> +    for (i = 0; i < num_rstcs; ++i) {
>>>> +        if (!strcmp(rstcs[i].id, id))
>>>> +            return reset_control_deassert(rstcs[i].rstc);
>>>> +    }
>>>> +
>>>> +    return -ENODEV;
>>>> +}
>>>> +
>>>> +static bool iris_vpu33_hw_power_collapsed(struct iris_core *core)
>>>> +{
>>>> +    u32 value, pwr_status;
>>>> +
>>>> +    value = readl(core->reg_base + WRAPPER_CORE_POWER_STATUS);
>>>> +    pwr_status = value & BIT(1);
>>>> +
>>>> +    return pwr_status ? false : true;
>>>> +}
>>>> +
>>>> +static void iris_vpu33_power_off_hardware(struct iris_core *core)
>>>> +{
>>>> +    u32 reg_val = 0, value, i;
>>>> +    int ret;
>>>> +    int count = 0;
>>>> +
>>>> +    if (iris_vpu33_hw_power_collapsed(core))
>>>> +        goto disable_power;
>>>> +
>>>> +    value = readl(core->reg_base + WRAPPER_CORE_CLOCK_CONFIG);
>>>> +    if (value)
>>>> +        writel(CORE_CLK_RUN, core->reg_base + WRAPPER_CORE_CLOCK_CONFIG);
>>>> +
>>>> +    value = readl(core->reg_base + VCODEC_DMA_SPARE_3);
>>>> +    value |= BIT(0);
>>>> +    writel(value, core->reg_base + VCODEC_DMA_SPARE_3)> +
>>>> +    for (i = 0; i < core->iris_platform_data->num_vpp_pipe; i++) {
>>>> +        ret = readl_poll_timeout(core->reg_base +
>>>> VCODEC_SS_IDLE_STATUSN + 4 * i,
>>>> +                     reg_val, reg_val & 0x400000, 2000, 20000);
>>>> +        if (ret)
>>>> +            goto disable_power;
>>>> +    }
>>>> +
>>>> +    /* set MNoC to low power, set PD_NOC_QREQ (bit 0) */
>>>> +    value = readl(core->reg_base + AON_WRAPPER_MVP_NOC_LPI_CONTROL);
>>>> +    value |= BIT(0);
>>>> +    writel(value, core->reg_base + AON_WRAPPER_MVP_NOC_LPI_CONTROL);
>>>> +
>>>> +    value = readl(core->reg_base + AON_WRAPPER_MVP_NOC_LPI_STATUS);
>>>> +
>>>> +    while ((!(value & BIT(0))) && (value & BIT(1) || value & BIT(2))) {
>>>> +        value = readl(core->reg_base + AON_WRAPPER_MVP_NOC_LPI_CONTROL);
>>>> +        value &= ~BIT(0);
>>>> +        writel(value, core->reg_base + AON_WRAPPER_MVP_NOC_LPI_CONTROL);
>>>> +
>>>> +        usleep_range(10, 20);
>>>> +
>>>> +        value = readl(core->reg_base + AON_WRAPPER_MVP_NOC_LPI_CONTROL);
>>>> +        value |= BIT(0);
>>>> +        writel(value, core->reg_base + AON_WRAPPER_MVP_NOC_LPI_CONTROL);
>>>> +
>>>> +        usleep_range(10, 20);
>>>> +
>>>> +        value = readl(core->reg_base + AON_WRAPPER_MVP_NOC_LPI_STATUS);
>>>> +
>>>> +        ++count;
>>>> +        if (count >= 1000)
>>>> +            break;
>>>> +    }
>>>> +
>>>> +    if (count < 1000) {
>>>> +        value = readl(core->reg_base + AON_WRAPPER_MVP_NOC_LPI_CONTROL);
>>>> +        value &= ~BIT(0);
>>>> +        writel(value, core->reg_base + AON_WRAPPER_MVP_NOC_LPI_CONTROL);
>>>> +    }
>>>> +
>>>> +    writel(VIDEO_NOC_RESET_REQ, core->reg_base +
>>>> AON_WRAPPER_MVP_NOC_RESET_REQ);
>>>> +
>>>> +    ret = readl_poll_timeout(core->reg_base +
>>>> AON_WRAPPER_MVP_NOC_RESET_ACK,
>>>> +                 reg_val, reg_val & 0x3, 200, 2000);
>>>> +    if (ret)
>>>> +        goto disable_power;
>>>> +> +    writel(0x0, core->reg_base + AON_WRAPPER_MVP_NOC_RESET_REQ);
>>>> +
>>>> +    ret = readl_poll_timeout(core->reg_base +
>>>> AON_WRAPPER_MVP_NOC_RESET_ACK,
>>>> +                 reg_val, !(reg_val & 0x3), 200, 2000);
>>>> +    if (ret)
>>>> +        goto disable_power;
>>>> +> +    writel(CORE_BRIDGE_SW_RESET | CORE_BRIDGE_HW_RESET_DISABLE,
>>>> +           core->reg_base + CPU_CS_AHB_BRIDGE_SYNC_RESET);
>>>> +    writel(CORE_BRIDGE_HW_RESET_DISABLE, core->reg_base +
>>>> CPU_CS_AHB_BRIDGE_SYNC_RESET);
>>>> +    writel(0x0, core->reg_base + CPU_CS_AHB_BRIDGE_SYNC_RESET);
>>>> +
>>>> +disable_power:
>>>> +    iris_vpu_power_off_hw(core);
>>>> +}
>>>> +
>>>> +static int iris_vpu33_power_off_controller(struct iris_core *core)
>>>> +{
>>>> +    u32 rst_tbl_size = core->iris_platform_data->clk_rst_tbl_size;
>>>> +    u32 val = 0;
>>>> +    int ret;
>>>> +
>>>> +    writel(MSK_SIGNAL_FROM_TENSILICA | MSK_CORE_POWER_ON,
>>>> core->reg_base + CPU_CS_X2RPMH);
>>>> +
>>>> +    writel(REQ_POWER_DOWN_PREP, core->reg_base +
>>>> WRAPPER_IRIS_CPU_NOC_LPI_CONTROL);
>>>> +
> Here, WRAPPER_IRIS_CPU_NOC_LPI_CONTROL is set in vpu33 as well, I didn't
> get, what's the concern here?

It's AON_WRAPPER_MVP_NOC_LPI_CONTROL which is set in power_off_hardware for vpu33,
if we leave in power_off_controller the HW doesn't properly wake up on resume.

Neil

> 
> Thanks,
> Dikshita
>>>> +    ret = readl_poll_timeout(core->reg_base +
>>>> WRAPPER_IRIS_CPU_NOC_LPI_STATUS,
>>>> +                 val, val & BIT(0), 200, 2000);
>>>> +    if (ret)
>>>> +        goto disable_power;
>>>> +
>>>> +    writel(0x0, core->reg_base + WRAPPER_DEBUG_BRIDGE_LPI_CONTROL);
>>>> +
>>>> +    ret = readl_poll_timeout(core->reg_base +
>>>> WRAPPER_DEBUG_BRIDGE_LPI_STATUS,
>>>> +                 val, val == 0, 200, 2000);
>>>> +    if (ret)
>>>> +        goto disable_power;
>>>> +
>>>> +    writel(CTL_AXI_CLK_HALT | CTL_CLK_HALT,
>>>> +           core->reg_base + WRAPPER_TZ_CTL_AXI_CLOCK_CONFIG);
>>>> +    writel(RESET_HIGH, core->reg_base + WRAPPER_TZ_QNS4PDXFIFO_RESET);
>>>> +    writel(0x0, core->reg_base + WRAPPER_TZ_QNS4PDXFIFO_RESET);
>>>> +    writel(0x0, core->reg_base + WRAPPER_TZ_CTL_AXI_CLOCK_CONFIG);
>>>> +
>>> The code till here in this API is common with
>>> iris_vpu_power_off_controller(), please check the possibility of reusing it.
>>
>> Not exactly, the common code sets WRAPPER_IRIS_CPU_NOC_LPI_CONTROL, which
>> is not in the vpu33 sequence, not sure how I'll be able to reuse it.
>>
>> Neil
>>
>>>> +    reset_control_bulk_assert_id(rst_tbl_size, core->resets, "bus");
>>>> +    reset_control_bulk_assert_id(rst_tbl_size, core->resets, "core");
>>>> +    usleep_range(1000, 1100);
>>>> +    reset_control_bulk_deassert_id(rst_tbl_size, core->resets, "bus");
>>>> +    reset_control_bulk_deassert_id(rst_tbl_size, core->resets, "core");
>>>> +
>>>> +    /* Disable MVP NoC clock */
>>>> +    val = readl(core->reg_base + AON_WRAPPER_MVP_NOC_CORE_CLK_CONTROL);
>>>> +    val |= NOC_HALT;
>>>> +    writel(val, core->reg_base + AON_WRAPPER_MVP_NOC_CORE_CLK_CONTROL);
>>>> +
>>>> +    /* enable MVP NoC reset */
>>>> +    val = readl(core->reg_base + AON_WRAPPER_MVP_NOC_CORE_SW_RESET);
>>>> +    val |= SW_RESET;
>>>> +    writel(val, core->reg_base + AON_WRAPPER_MVP_NOC_CORE_SW_RESET);
>>>> +
>>>> +    /* poll AON spare register bit0 to become zero with 50ms timeout */
>>>> +    ret = readl_poll_timeout(core->reg_base + AON_WRAPPER_SPARE,
>>>> +                 val, (val & BIT(0)) == 0, 1000, 50000);
>>>> +    if (ret)
>>>> +        goto disable_power;
>>>> +
>>>> +    /* enable bit(1) to avoid cvp noc xo reset */
>>>> +    val = readl(core->reg_base + AON_WRAPPER_SPARE);
>>>> +    val |= BIT(1);
>>>> +    writel(val, core->reg_base + AON_WRAPPER_SPARE);
>>>> +
>>>> +    reset_control_bulk_assert_id(rst_tbl_size, core->resets, "xo");
>>>> +
>>>> +    /* De-assert MVP NoC reset */
>>>> +    val = readl(core->reg_base + AON_WRAPPER_MVP_NOC_CORE_SW_RESET);
>>>> +    val &= ~SW_RESET;
>>>> +    writel(val, core->reg_base + AON_WRAPPER_MVP_NOC_CORE_SW_RESET);
>>>> +
>>>> +    usleep_range(80, 100);
>>>> +    reset_control_bulk_deassert_id(rst_tbl_size, core->resets, "xo");
>>>> +
>>>> +    /* reset AON spare register */
>>>> +    writel(0, core->reg_base + AON_WRAPPER_SPARE);
>>>> +
>>>> +    /* Enable MVP NoC clock */
>>>> +    val = readl(core->reg_base + AON_WRAPPER_MVP_NOC_CORE_CLK_CONTROL);
>>>> +    val &= ~NOC_HALT;
>>>> +    writel(val, core->reg_base + AON_WRAPPER_MVP_NOC_CORE_CLK_CONTROL);
>>>> +
>>>> +    iris_disable_unprepare_clock(core, IRIS_CTRL_CLK);
>>>> +
>>>> +disable_power:
>>>> +    iris_disable_power_domains(core,
>>>> core->pmdomain_tbl->pd_devs[IRIS_CTRL_POWER_DOMAIN]);
>>>> +    iris_disable_unprepare_clock(core, IRIS_AXI_CLK);
>>>> +
>>>> +    return 0;
>>>> +}
>>>> +
>>>> +static u64 iris_vpu33_calculate_frequency(struct iris_inst *inst,
>>>> size_t data_size)
>>>> +{
>>>> +    struct platform_inst_caps *caps =
>>>> inst->core->iris_platform_data->inst_caps;
>>>> +    struct v4l2_format *inp_f = inst->fmt_src;
>>>> +    u32 height, width, mbs_per_second, mbpf;
>>>> +    u64 fw_cycles, fw_vpp_cycles;
>>>> +    u64 vsp_cycles, vpp_cycles;
>>>> +    u32 fps = DEFAULT_FPS;
>>>> +
>>>> +    width = max(inp_f->fmt.pix_mp.width, inst->crop.width);
>>>> +    height = max(inp_f->fmt.pix_mp.height, inst->crop.height);
>>>> +
>>>> +    mbpf = NUM_MBS_PER_FRAME(height, width);
>>>> +    mbs_per_second = mbpf * fps;
>>>> +
>>>> +    fw_cycles = fps * caps->mb_cycles_fw;
>>>> +    fw_vpp_cycles = fps * caps->mb_cycles_fw_vpp;
>>>> +
>>>> +    vpp_cycles = mult_frac(mbs_per_second, caps->mb_cycles_vpp,
>>>> (u32)inst->fw_caps[PIPE].value);
>>>> +    /* 21 / 20 is minimum overhead factor */
>>>> +    vpp_cycles += max(div_u64(vpp_cycles, 20), fw_vpp_cycles);
>>>> +
>>>> +    /* 1.059 is multi-pipe overhead */
>>>> +    if (inst->fw_caps[PIPE].value > 1)
>>>> +        vpp_cycles += div_u64(vpp_cycles * 59, 1000);
>>>> +
>>>> +    vsp_cycles = fps * data_size * 8;
>>>> +    vsp_cycles = div_u64(vsp_cycles, 2);
>>>> +    /* VSP FW overhead 1.05 */
>>>> +    vsp_cycles = div_u64(vsp_cycles * 21, 20);
>>>> +
>>>> +    if (inst->fw_caps[STAGE].value == STAGE_1)
>>>> +        vsp_cycles = vsp_cycles * 3;
>>>> +
>>>> +    return max3(vpp_cycles, vsp_cycles, fw_cycles);
>>>> +}
>>>> +
>>> This is exactly same as vpu3 calculation, pls reuse.
>>
>> Ack
>>
>>>> +static int iris_vpu33_reset_controller(struct iris_core *core)
>>>> +{
>>>> +    u32 rst_tbl_size = core->iris_platform_data->clk_rst_tbl_size;
>>>> +
>>>> +    reset_control_bulk_assert_id(rst_tbl_size, core->resets, "bus");
>>>> +    reset_control_bulk_assert_id(rst_tbl_size, core->resets, "core");
>>>> +
>>>> +    usleep_range(1000, 1100);
>>>> +
>>>> +    reset_control_bulk_deassert_id(rst_tbl_size, core->resets, "bus");
>>>> +    reset_control_bulk_deassert_id(rst_tbl_size, core->resets, "core");
>>>> +
>>>> +    return 0;
>>>> +}
>>>> +
>>> should be replacable with reset_control_bulk_reset. pls revisit.
>>
>>>> +const struct vpu_ops iris_vpu33_ops = {
>>>> +    .reset_controller = iris_vpu33_reset_controller,
>>>> +    .power_off_hw = iris_vpu33_power_off_hardware,
>>>> +    .power_off_controller = iris_vpu33_power_off_controller,
>>>> +    .calc_freq = iris_vpu33_calculate_frequency,
>>>> +};
>>> you can rename vpu3.c to vpu3x.c and move these ops to same file, this way
>>> common API can be reused and no need of redinfing the macros as well.
>>
>> I'll see how it turns out.
>>
>> Thanks,
>> Neil
>>
>>>
>>> Thanks,
>>> Dikshita
>>>> diff --git a/drivers/media/platform/qcom/iris/iris_vpu_common.h
>>>> b/drivers/media/platform/qcom/iris/iris_vpu_common.h
>>>> index
>>>> c948d8b5aee87ccf1fd53c5518a27294232d8fb8..c4d02a3b884881eb033dc0342f948848adae2819 100644
>>>> --- a/drivers/media/platform/qcom/iris/iris_vpu_common.h
>>>> +++ b/drivers/media/platform/qcom/iris/iris_vpu_common.h
>>>> @@ -10,6 +10,7 @@ struct iris_core;
>>>>      extern const struct vpu_ops iris_vpu2_ops;
>>>>    extern const struct vpu_ops iris_vpu3_ops;
>>>> +extern const struct vpu_ops iris_vpu33_ops;
>>>>      struct vpu_ops {
>>>>        int (*reset_controller)(struct iris_core *core);
>>>>
>>