[v1] drm/i915: Add Exec param to control data port coherency.

The patch adds a parameter to control the data port coherency functionality
on a per-exec call basis. When data port coherency flag value is different
than what it was in previous call for the context, a command to switch data
port coherency state is added before the buffer to be executed.

Rationale:

The OpenCL driver develpers requested a functionality to control cache
coherency at data port level. Keeping the coherency at that level is disabled
by default due to its performance costs. OpenCL driver is planning to
enable it for a small subset of submissions, when such functionality is
required. Below are answers to basic question explaining background
of the functionality and reasoning for the proposed implementation:

1. Why do we need a coherency enable/disable switch for memory that is shared
between CPU and GEN (GPU)?

Memory coherency between CPU and GEN, while being a great feature that enables
CL_MEM_SVM_FINE_GRAIN_BUFFER OCL capability on Intel GEN architecture, adds
overhead related to tracking (snooping) memory inside different cache units
(L1$, L2$, L3$, LLC$, etc.). At the same time, minority of modern OCL
applications actually use CL_MEM_SVM_FINE_GRAIN_BUFFER (and hence require
memory coherency between CPU and GPU). The goal of coherency enable/disable
switch is to remove overhead of memory coherency when memory coherency is not
needed.

2. Why do we need a global coherency switch?

In order to support I/O commands from within EUs (Execution Units), Intel GEN
ISA (GEN Instruction Set Assembly) contains dedicated "send" instructions.
These send instructions provide several addressing models. One of these
addressing models (named "stateless") provides most flexible I/O using plain
virtual addresses (as opposed to buffer_handle+offset models). This "stateless"
model is similar to regular memory load/store operations available on typical
CPUs. Since this model provides I/O using arbitrary virtual addresses, it
enables algorithmic designs that are based on pointer-to-pointer (e.g. buffer
of pointers) concepts. For instance, it allows creating tree-like data
structures such as:
                   ________________
                  |      NODE1     |
                  | uint64_t data  |
                  +----------------|
                  | NODE*  |  NODE*|
                  +--------+-------+
                    /              \
   ________________/                \________________
  |      NODE2     |                |      NODE3     |
  | uint64_t data  |                | uint64_t data  |
  +----------------|                +----------------|
  | NODE*  |  NODE*|                | NODE*  |  NODE*|
  +--------+-------+                +--------+-------+

Please note that pointers inside such structures can point to memory locations
in different OCL allocations  - e.g. NODE1 and NODE2 can reside in one OCL
allocation while NODE3 resides in a completely separate OCL allocation.
Additionally, such pointers can be shared with CPU (i.e. using SVM - Shared
Virtual Memory feature). Using pointers from different allocations doesn't
affect the stateless addressing model which even allows scattered reading from
different allocations at the same time (i.e. by utilizing SIMD-nature of send
instructions).

When it comes to coherency programming, send instructions in stateless model
can be encoded (at ISA level) to either use or disable coherency. However, for
generic OCL applications (such as example with tree-like data structure), OCL
compiler is not able to determine origin of memory pointed to by an arbitrary
pointer - i.e. is not able to track given pointer back to a specific
allocation. As such, it's not able to decide whether coherency is needed or not
for specific pointer (or for specific I/O instruction). As a result, compiler
encodes all stateless sends as coherent (doing otherwise would lead to
functional issues resulting from data corruption). Please note that it would be
possible to workaround this (e.g. based on allocations map and pointer bounds
checking prior to each I/O instruction) but the performance cost of such
workaround would be many times greater than the cost of keeping coherency
always enabled. As such, enabling/disabling memory coherency at GEN ISA level
is not feasible and alternative method is needed.

Such alternative solution is to have a global coherency switch that allows
disabling coherency for single (though entire) GPU submission. This is
beneficial because this way we:
* can enable (and pay for) coherency only in submissions that actually need
coherency (submissions that use CL_MEM_SVM_FINE_GRAIN_BUFFER resources)
* don't care about coherency at GEN ISA granularity (no performance impact)

3. Will coherency switch be used frequently?

There are scenarios that will require frequent toggling of the coherency
switch.
E.g. an application has two OCL compute kernels: kern_master and kern_worker.
kern_master uses, concurrently with CPU, some fine grain SVM resources
(CL_MEM_SVM_FINE_GRAIN_BUFFER). These resources contain descriptors of
computational work that needs to be executed. kern_master analyzes incoming
work descriptors and populates a plain OCL buffer (non-fine-grain) with payload
for kern_worker. Once kern_master is done, kern_worker kicks-in and processes
the payload that kern_master produced. These two kernels work in a loop, one
after another. Since only kern_master requires coherency, kern_worker should
not be forced to pay for it. This means that we need to have the ability to
toggle coherency switch on or off per each GPU submission:
(ENABLE COHERENCY) kern_master -> (DISABLE COHERENCY)kern_worker -> (ENABLE
COHERENCY) kern_master -> (DISABLE COHERENCY)kern_worker -> ...

4. Why the execlist flag approach was chosen?

There are two other ways of providing the functionality to UMDs, besides the
execlist flag:

a) Chicken bit register whitelisting.

This approach would allow adding the functionality without any change to
KMDs interface. Also, it has been determined that whitelisting is safe for
gen10 and gen11. The issue is with gen9, where hardware whitelisting cannot
be used, and OCL driver needs support for it. A workaround there would be to
use command parser, which verifies buffers before execution. But such parsing
comes at a considerable performance cost.

b) Providing the flag as context IOCTL setting.

The data port coherency switch could be implemented as a context parameter,
which would schedule submission of a buffer to switch the coherency flag.
That is an elegant solution with bounds the flag to context, which matches
the hardware placement of the feature. This solution was not accepted
because of OCL driver performance concerns. The OCL driver is constructed
with emphasis on creating small, but very frequent submissions. With such
architecture, adding IOCTL setparam call before submission has considerable
impact on the performance.

Bspec: 11419
Signed-off-by: Tomasz Lis <tomasz.lis@intel.com>
---
 drivers/gpu/drm/i915/i915_drv.c            |  3 ++
 drivers/gpu/drm/i915/i915_gem_context.h    |  1 +
 drivers/gpu/drm/i915/i915_gem_execbuffer.c | 17 ++++++++++
 drivers/gpu/drm/i915/intel_lrc.c           | 53 ++++++++++++++++++++++++++++++
 drivers/gpu/drm/i915/intel_lrc.h           |  3 ++
 include/uapi/drm/i915_drm.h                | 12 ++++++-
 6 files changed, 88 insertions(+), 1 deletion(-)

[v1] drm/i915: Add Exec param to control data port coherency.

Commit Message

Comments

Patch