On Tue, Nov 1, 2016 at 2:57 PM, Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> wrote:
> @@ -140,6 +437,106 @@ int i915_oa_select_metric_set_hsw(On Fri, Oct 28, 2016 at 03:14:29AM +0100, Robert Bragg wrote:
> This adds 'compute', 'compute extended', 'memory reads', 'memory writes'
> and 'sampler balance' metric sets for Haswell.
>
> The code is auto generated from an XML description of metric sets,
> currently maintained in gputop, ref:
>
> https://github.com/rib/gputop
> > gputop-data/oa-*.xml
> > scripts/i915-perf-kernelgen.py
>
> $ make -C gputop-data -f Makefile.xml
>
> Signed-off-by: Robert Bragg <robert@xxxxxxxxxxxxx>
> Reviewed-by: Matthew Auld <matthew.auld@xxxxxxxxx>
> ---
> drivers/gpu/drm/i915/i915_oa_hsw.c | 559 ++++++++++++++++++++++++++++++ ++++++-
> 1 file changed, 558 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/i915/i915_oa_hsw.c b/drivers/gpu/drm/i915/i915_ oa_hsw.c
> index 6af25cf..4ddf756 100644
> --- a/drivers/gpu/drm/i915/i915_oa_hsw.c
> +++ b/drivers/gpu/drm/i915/i915_oa_hsw.c
> @@ -31,9 +31,14 @@
>
> enum metric_set_id {
> METRIC_SET_ID_RENDER_BASIC = 1,
> + METRIC_SET_ID_COMPUTE_BASIC,
> + METRIC_SET_ID_COMPUTE_EXTENDED,
> + METRIC_SET_ID_MEMORY_READS,
> + METRIC_SET_ID_MEMORY_WRITES,
> + METRIC_SET_ID_SAMPLER_BALANCE,
> };
>
> -int i915_oa_n_builtin_metric_sets_hsw = 1;
> +int i915_oa_n_builtin_metric_sets_hsw = 6;
>
> static const struct i915_oa_reg b_counter_config_render_basic[] = {
> { _MMIO(0x2724), 0x00800000 },
> @@ -112,6 +117,298 @@ get_render_basic_mux_config(struct drm_i915_private *dev_priv,
> return mux_config_render_basic;
> }
>
> +static const struct i915_oa_reg b_counter_config_compute_basic[] = {
> + { _MMIO(0x2710), 0x00000000 },
> + { _MMIO(0x2714), 0x00800000 },
> + { _MMIO(0x2718), 0xaaaaaaaa },
> + { _MMIO(0x271c), 0xaaaaaaaa },
> + { _MMIO(0x2720), 0x00000000 },
> + { _MMIO(0x2724), 0x00800000 },
> + { _MMIO(0x2728), 0xaaaaaaaa },
> + { _MMIO(0x272c), 0xaaaaaaaa },
> + { _MMIO(0x2740), 0x00000000 },
> + { _MMIO(0x2744), 0x00000000 },
> + { _MMIO(0x2748), 0x00000000 },
> + { _MMIO(0x274c), 0x00000000 },
> + { _MMIO(0x2750), 0x00000000 },
> + { _MMIO(0x2754), 0x00000000 },
> + { _MMIO(0x2758), 0x00000000 },
> + { _MMIO(0x275c), 0x00000000 },
> + { _MMIO(0x236c), 0x00000000 },
> +};
> +
> +static const struct i915_oa_reg mux_config_compute_basic[] = {
> + { _MMIO(0x253a4), 0x00000000 },
> + { _MMIO(0x2681c), 0x01f00800 },
> + { _MMIO(0x26820), 0x00001000 },
> + { _MMIO(0x2781c), 0x01f00800 },
> + { _MMIO(0x26520), 0x00000007 },
> + { _MMIO(0x265a0), 0x00000007 },
> + { _MMIO(0x25380), 0x00000010 },
> + { _MMIO(0x2538c), 0x00300000 },
> + { _MMIO(0x25384), 0xaa8aaaaa },
> + { _MMIO(0x25404), 0xffffffff },
> + { _MMIO(0x26800), 0x00004202 },
> + { _MMIO(0x26808), 0x00605817 },
> + { _MMIO(0x2680c), 0x10001005 },
> + { _MMIO(0x26804), 0x00000000 },
> + { _MMIO(0x27800), 0x00000102 },
> + { _MMIO(0x27808), 0x0c0701e0 },
> + { _MMIO(0x2780c), 0x000200a0 },
> + { _MMIO(0x27804), 0x00000000 },
> + { _MMIO(0x26484), 0x44000000 },
> + { _MMIO(0x26704), 0x44000000 },
> + { _MMIO(0x26500), 0x00000006 },
> + { _MMIO(0x26510), 0x00000001 },
> + { _MMIO(0x26504), 0x88000000 },
> + { _MMIO(0x26580), 0x00000006 },
> + { _MMIO(0x26590), 0x00000020 },
> + { _MMIO(0x26584), 0x00000000 },
> + { _MMIO(0x26104), 0x55822222 },
> + { _MMIO(0x26184), 0xaa866666 },
> + { _MMIO(0x25420), 0x08320c83 },
> + { _MMIO(0x25424), 0x06820c83 },
> + { _MMIO(0x2541c), 0x00000000 },
> + { _MMIO(0x25428), 0x00000c03 },
> +};
> +
> +static const struct i915_oa_reg *
> +get_compute_basic_mux_config(struct drm_i915_private *dev_priv,
> + int *len)
> +{
> + *len = ARRAY_SIZE(mux_config_compute_basic);
> + return mux_config_compute_basic;
> +}
struct drm_i915_private *dev_priv) > int
> ARRAY_SIZE(b_counter_config_render_basic);
>
> return 0;
> + case METRIC_SET_ID_COMPUTE_BASIC:
> + dev_priv->perf.oa.mux_regs =
> + get_compute_basic_mux_config(dev_priv,
> + &dev_priv->perf.oa.mux_regs_len);
> + if (!dev_priv->perf.oa.mux_regs) {
> + DRM_DEBUG_DRIVER("No suitable MUX config for \"COMPUTE_BASIC\" metric set");
> +
> + /* EINVAL because *_register_sysfs already checked this
> + * and so it wouldn't have been advertised so userspace and
> + * so shouldn't have been requested
> + */
> + return -EINVAL;
> + }
> +
> + dev_priv->perf.oa.b_counter_regs =
> + b_counter_config_compute_basic;
> + dev_priv->perf.oa.b_counter_regs_len =
> + ARRAY_SIZE(b_counter_config_compute_basic);
> +
> + return 0;
> i915_perf_register_sysfs_hsw(struct drm_i915_private *dev_priv) Why not use the derived state in dev_priv->perf.oa.mux_regs? Then we
> {
> @@ -178,9 +685,49 @@ i915_perf_register_sysfs_hsw(struct drm_i915_private *dev_priv)
> if (ret)
> goto error_render_basic;
> }
> + if (get_compute_basic_mux_config(dev_priv, &mux_len)) {
only expose what is initialised.
Although for Haswell none of our metric sets have conditional MUX configurations, the generated code should already be in shape to only advertising metric sets applicable to the system (which becomes an issue for gen8+). This was changed relatively recently in the gen8+ series after Mark Janes was hitting issues on Skylake in some of his tooling due to Mesa advertising one of the compute metric sets that wasn't really available on the system he had, which was only discoverable as a GL error when attempting to use it.
The perf.oa.mux_regs state only pertains to one current metric set that the OA unit has been configured with, after calling the generated i915_oa_select_metric_set_hsw() function in hsw_enable_metric_set(). Until an OA stream is opened and enabled perf.oa.mux_regs won't be initialised.
Notably the recent change for gen8+ mentioned above was to have the _select_metric_set_<gen>() code and the _register_sysfs_<gen>() code both work in terms of the get_<metric_set>_mux_config() functions since it's these functions that will check the fiddly sku specfic details on gen8+ to select the right MUX config or potentially fail if the metric set isn't available on the current system. So for gen8+ we can expect get_compute_basic_mux_config() will fail if the config isn't available and then won't be advertised via sysfs. On Haswell it looks a little redundant having these get_ functions unconditionally return a pointer to a corresponding array.
Hope that clarifies,Notably the recent change for gen8+ mentioned above was to have the _select_metric_set_<gen>() code and the _register_sysfs_<gen>() code both work in terms of the get_<metric_set>_mux_config() functions since it's these functions that will check the fiddly sku specfic details on gen8+ to select the right MUX config or potentially fail if the metric set isn't available on the current system. So for gen8+ we can expect get_compute_basic_mux_config() will fail if the config isn't available and then won't be advertised via sysfs. On Haswell it looks a little redundant having these get_ functions unconditionally return a pointer to a corresponding array.
- Robert
-Chris
--
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/intel-gfx
