[Mesa-dev] [PATCH] nvc0: add support for performance monitoring metrics on Fermi
Ilia Mirkin
imirkin at alum.mit.edu
Fri Oct 16 14:22:39 PDT 2015
On Fri, Oct 16, 2015 at 5:29 PM, Samuel Pitoiset
<samuel.pitoiset at gmail.com> wrote:
> As explained in the CUDA toolkit documentation, "a metric is a
> characteristic of an application that is calculated from one or more
> event values."
>
> Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
> ---
> src/gallium/drivers/nouveau/Makefile.sources | 2 +
> src/gallium/drivers/nouveau/nvc0/nvc0_query_hw.c | 19 +-
> .../drivers/nouveau/nvc0/nvc0_query_hw_metric.c | 444 +++++++++++++++++++++
> .../drivers/nouveau/nvc0/nvc0_query_hw_metric.h | 42 ++
> 4 files changed, 504 insertions(+), 3 deletions(-)
> create mode 100644 src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_metric.c
> create mode 100644 src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_metric.h
>
> diff --git a/src/gallium/drivers/nouveau/Makefile.sources b/src/gallium/drivers/nouveau/Makefile.sources
> index edc6cf4..c18e9f5 100644
> --- a/src/gallium/drivers/nouveau/Makefile.sources
> +++ b/src/gallium/drivers/nouveau/Makefile.sources
> @@ -154,6 +154,8 @@ NVC0_C_SOURCES := \
> nvc0/nvc0_query.h \
> nvc0/nvc0_query_hw.c \
> nvc0/nvc0_query_hw.h \
> + nvc0/nvc0_query_hw_metric.c \
> + nvc0/nvc0_query_hw_metric.h \
> nvc0/nvc0_query_hw_sm.c \
> nvc0/nvc0_query_hw_sm.h \
> nvc0/nvc0_query_sw.c \
> diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw.c b/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw.c
> index 91254be..90ee82f 100644
> --- a/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw.c
> +++ b/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw.c
> @@ -25,6 +25,7 @@
>
> #include "nvc0/nvc0_context.h"
> #include "nvc0/nvc0_query_hw.h"
> +#include "nvc0/nvc0_query_hw_metric.h"
> #include "nvc0/nvc0_query_hw_sm.h"
>
> #define NVC0_HW_QUERY_STATE_READY 0
> @@ -371,6 +372,12 @@ nvc0_hw_create_query(struct nvc0_context *nvc0, unsigned type, unsigned index)
> return (struct nvc0_query *)hq;
> }
>
> + hq = nvc0_hw_metric_create_query(nvc0, type);
> + if (hq) {
> + hq->base.funcs = &hw_query_funcs;
> + return (struct nvc0_query *)hq;
> + }
> +
> hq = CALLOC_STRUCT(nvc0_hw_query);
> if (!hq)
> return NULL;
> @@ -435,14 +442,20 @@ int
> nvc0_hw_get_driver_query_info(struct nvc0_screen *screen, unsigned id,
> struct pipe_driver_query_info *info)
> {
> - int num_hw_sm_queries = 0;
> + int num_hw_sm_queries = 0, num_hw_metric_queries = 0;
>
> num_hw_sm_queries = nvc0_hw_sm_get_driver_query_info(screen, 0, NULL);
> + num_hw_metric_queries =
> + nvc0_hw_metric_get_driver_query_info(screen, 0, NULL);
>
> if (!info)
> - return num_hw_sm_queries;
> + return num_hw_sm_queries + num_hw_metric_queries;
> +
> + if (id < num_hw_sm_queries)
> + return nvc0_hw_sm_get_driver_query_info(screen, id, info);
>
> - return nvc0_hw_sm_get_driver_query_info(screen, id, info);
> + return nvc0_hw_metric_get_driver_query_info(screen,
> + id - num_hw_sm_queries, info);
> }
>
> void
> diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_metric.c b/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_metric.c
> new file mode 100644
> index 0000000..dbe350a
> --- /dev/null
> +++ b/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_metric.c
> @@ -0,0 +1,444 @@
> +/*
> + * Copyright 2015 Samuel Pitoiset
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + */
> +
> +#include "nvc0/nvc0_context.h"
> +#include "nvc0/nvc0_query_hw_metric.h"
> +#include "nvc0/nvc0_query_hw_sm.h"
> +
> +/* === PERFORMANCE MONITORING METRICS for NVC0:NVE4 === */
> +static const char *nvc0_hw_metric_names[] =
> +{
> + "metric-achieved_occupancy",
> + "metric-branch_efficiency",
> + "metric-inst_issued",
> + "metric-inst_per_wrap",
> + "metric-inst_replay_overhead",
> + "metric-issued_ipc",
> + "metric-issue_slots",
> + "metric-issue_slot_utilization",
> + "metric-ipc",
> +};
> +
> +struct nvc0_hw_metric_query_cfg {
> + uint32_t queries[8];
> + uint32_t num_queries;
> +};
> +
> +#define _SM(n) NVC0_HW_SM_QUERY(NVC0_HW_SM_QUERY_ ##n)
> +#define _M(n, c) [NVC0_HW_METRIC_QUERY_##n] = c
> +
> +/* ==== Compute capability 2.0 (GF100/GF110) ==== */
> +static const struct nvc0_hw_metric_query_cfg
> +sm20_achieved_occupancy =
> +{
> + .queries[0] = _SM(ACTIVE_WARPS),
> + .queries[1] = _SM(ACTIVE_CYCLES),
> + .num_queries = 2,
> +};
> +
> +static const struct nvc0_hw_metric_query_cfg
> +sm20_branch_efficiency =
> +{
> + .queries[0] = _SM(BRANCH),
> + .queries[1] = _SM(DIVERGENT_BRANCH),
> + .num_queries = 2,
> +};
> +
> +static const struct nvc0_hw_metric_query_cfg
> +sm20_inst_per_wrap =
> +{
> + .queries[0] = _SM(INST_EXECUTED),
> + .queries[1] = _SM(WARPS_LAUNCHED),
> + .num_queries = 2,
> +};
> +
> +static const struct nvc0_hw_metric_query_cfg
> +sm20_inst_replay_overhead =
> +{
> + .queries[0] = _SM(INST_ISSUED),
> + .queries[1] = _SM(INST_EXECUTED),
> + .num_queries = 2,
> +};
> +
> +static const struct nvc0_hw_metric_query_cfg
> +sm20_issued_ipc =
> +{
> + .queries[0] = _SM(INST_ISSUED),
> + .queries[1] = _SM(ACTIVE_CYCLES),
> + .num_queries = 2,
> +};
> +
> +static const struct nvc0_hw_metric_query_cfg
> +sm20_ipc =
> +{
> + .queries[0] = _SM(INST_EXECUTED),
> + .queries[1] = _SM(ACTIVE_CYCLES),
> + .num_queries = 2,
> +};
> +
> +static const struct nvc0_hw_metric_query_cfg *sm20_hw_metric_queries[] =
> +{
> + _M(ACHIEVED_OCCUPANCY, &sm20_achieved_occupancy),
> + _M(BRANCH_EFFICIENCY, &sm20_branch_efficiency),
> + _M(INST_ISSUED, NULL),
> + _M(INST_PER_WRAP, &sm20_inst_per_wrap),
> + _M(INST_REPLAY_OVERHEAD, &sm20_inst_replay_overhead),
> + _M(ISSUED_IPC, &sm20_issued_ipc),
> + _M(ISSUE_SLOTS, NULL),
> + _M(ISSUE_SLOT_UTILIZATION, &sm20_issued_ipc),
> + _M(IPC, &sm20_ipc),
> +};
> +
> +/* ==== Compute capability 2.1 (GF108+ except GF110) ==== */
> +static const struct nvc0_hw_metric_query_cfg
> +sm21_inst_issued =
> +{
> + .queries[0] = _SM(INST_ISSUED1_0),
> + .queries[1] = _SM(INST_ISSUED1_1),
> + .queries[2] = _SM(INST_ISSUED2_0),
> + .queries[3] = _SM(INST_ISSUED2_1),
> + .num_queries = 4,
> +};
> +
> +static const struct nvc0_hw_metric_query_cfg
> +sm21_inst_replay_overhead =
> +{
> + .queries[0] = _SM(INST_ISSUED1_0),
> + .queries[1] = _SM(INST_ISSUED1_1),
> + .queries[2] = _SM(INST_ISSUED2_0),
> + .queries[3] = _SM(INST_ISSUED2_1),
> + .queries[4] = _SM(INST_EXECUTED),
> + .num_queries = 5,
> +};
> +
> +static const struct nvc0_hw_metric_query_cfg
> +sm21_issued_ipc =
> +{
> + .queries[0] = _SM(INST_ISSUED1_0),
> + .queries[1] = _SM(INST_ISSUED1_1),
> + .queries[2] = _SM(INST_ISSUED2_0),
> + .queries[3] = _SM(INST_ISSUED2_1),
> + .queries[4] = _SM(ACTIVE_CYCLES),
> + .num_queries = 5,
> +};
> +
> +static const struct nvc0_hw_metric_query_cfg *sm21_hw_metric_queries[] =
> +{
> + _M(ACHIEVED_OCCUPANCY, &sm20_achieved_occupancy),
> + _M(BRANCH_EFFICIENCY, &sm20_branch_efficiency),
> + _M(INST_ISSUED, &sm21_inst_issued),
> + _M(INST_PER_WRAP, &sm20_inst_per_wrap),
> + _M(INST_REPLAY_OVERHEAD, &sm21_inst_replay_overhead),
> + _M(ISSUED_IPC, &sm21_issued_ipc),
> + _M(ISSUE_SLOTS, &sm21_inst_issued),
> + _M(ISSUE_SLOT_UTILIZATION, &sm21_issued_ipc),
> + _M(IPC, &sm20_ipc),
> +};
> +
> +#undef _SM
> +#undef _M
> +
> +static inline const struct nvc0_hw_metric_query_cfg **
> +nvc0_hw_metric_get_queries(struct nvc0_screen *screen)
> +{
> + struct nouveau_device *dev = screen->base.device;
> +
> + if (dev->chipset == 0xc0 || dev->chipset == 0xc8)
> + return sm20_hw_metric_queries;
> + return sm21_hw_metric_queries;
> +}
> +
> +static const struct nvc0_hw_metric_query_cfg *
> +nvc0_hw_metric_query_get_cfg(struct nvc0_context *nvc0,
> + struct nvc0_hw_query *hq)
> +{
> + const struct nvc0_hw_metric_query_cfg **queries;
> + struct nvc0_screen *screen = nvc0->screen;
> + struct nvc0_query *q = &hq->base;
> +
> + queries = nvc0_hw_metric_get_queries(screen);
> + return queries[q->type - NVC0_HW_METRIC_QUERY(0)];
> +}
> +
> +static void
> +nvc0_hw_metric_destroy_query(struct nvc0_context *nvc0,
> + struct nvc0_hw_query *hq)
> +{
> + struct nvc0_hw_metric_query *hmq = nvc0_hw_metric_query(hq);
> + unsigned i;
> +
> + for (i = 0; i < hmq->num_queries; i++)
> + hmq->queries[i]->funcs->destroy_query(nvc0, hmq->queries[i]);
> + FREE(hmq);
> +}
> +
> +static boolean
> +nvc0_hw_metric_begin_query(struct nvc0_context *nvc0, struct nvc0_hw_query *hq)
> +{
> + struct nvc0_hw_metric_query *hmq = nvc0_hw_metric_query(hq);
> + boolean ret = false;
> + unsigned i;
> +
> + for (i = 0; i < hmq->num_queries; i++) {
> + ret = hmq->queries[i]->funcs->begin_query(nvc0, hmq->queries[i]);
> + if (!ret)
> + return ret;
> + }
> + return ret;
> +}
> +
> +static void
> +nvc0_hw_metric_end_query(struct nvc0_context *nvc0, struct nvc0_hw_query *hq)
> +{
> + struct nvc0_hw_metric_query *hmq = nvc0_hw_metric_query(hq);
> + unsigned i;
> +
> + for (i = 0; i < hmq->num_queries; i++)
> + hmq->queries[i]->funcs->end_query(nvc0, hmq->queries[i]);
> +}
> +
> +static uint64_t
> +sm20_hw_metric_calc_result(struct nvc0_hw_query *hq, uint64_t res64[8])
> +{
> + uint64_t value = 0;
> +
> + switch (hq->base.type - NVC0_HW_METRIC_QUERY(0)) {
> + case NVC0_HW_METRIC_QUERY_ACHIEVED_OCCUPANCY:
> + /* (active_warps / active_cycles) / max. number of warps on a MP */
> + if (res64[1])
> + value = (res64[0] / (float)res64[1]) / 48;
Why isn't this all just "return ...." and then a return 0 at the end?
i.e. why do you have the value variable?
Also I don't know how big these values get, but you might want to use
doubles instead of floats.
More information about the mesa-dev
mailing list