Google Git
Sign in
llvm/llvm-project.git/refs/heads/main/./llvm/docs/NVPTXUsage.md
blob: 923a2b10f6be904061df5ca1b7a0f31e18077f2f [file] [view] [edit]
# User Guide for NVPTX Back-end
```{contents}
:depth: 3
:local: true
```
## Introduction
To support GPU programming, the NVPTX back-end supports a subset of LLVM IR
along with a defined set of conventions used to represent GPU programming
concepts. This document provides an overview of the general usage of the back-
end, including a description of the conventions used and the set of accepted
LLVM IR.
:::{note}
This document assumes a basic familiarity with CUDA and the PTX
assembly language. Information about the CUDA Driver API and the PTX assembly
language can be found in the [CUDA documentation](http://docs.nvidia.com/cuda/index.html).
:::
## Conventions
### Marking Functions as Kernels
In PTX, there are two types of functions: *device functions*, which are only
callable by device code, and *kernel functions*, which are callable by host
code. By default, the back-end will emit device functions. The `ptx_kernel`
calling convention is used to declare a function as a kernel function.
The following example shows a kernel function calling a device function in LLVM
IR. The function `@my_kernel` is callable from host code, but `@my_fmad` is
not.
```llvm
define float @my_fmad(float %x, float %y, float %z) {
%mul = fmul float %x, %y
%add = fadd float %mul, %z
ret float %add
}
define ptx_kernel void @my_kernel(ptr %ptr) {
%val = load float, ptr %ptr
%ret = call float @my_fmad(float %val, float %val, float %val)
store float %ret, ptr %ptr
ret void
}
```
When compiled, the PTX kernel functions are callable by host-side code.
### Parameter Attributes
`"nvvm.grid_constant"`
: This attribute may be attached to a `byval` parameter of a kernel function
to indicate that the parameter should be lowered as a direct reference to
the grid-constant memory of the parameter, as opposed to a copy of the
parameter in local memory. Writing to a grid-constant parameter is
undefined behavior. Unlike a normal `byval` parameter, the address of a
grid-constant parameter is not unique to a given function invocation but
instead is shared by all kernels in the grid.
(nvptx-fnattrs)=
### Function Attributes
`"nvvm.maxclusterrank"="<n>"`
: This attribute specifies the maximum number of blocks per cluster. Must be
non-zero. Only supported for Hopper+.
`"nvvm.minctasm"="<n>"`
: This indicates a hint/directive to the compiler/driver, asking it to put at
least these many CTAs on an SM.
`"nvvm.maxnreg"="<n>"`
: This attribute indicates the maximum number of registers to be used for the
kernel function.
`"nvvm.maxntid"="<x>[,<y>[,<z>]]"`
: This attribute declares the maximum number of threads in the thread block
(CTA). The maximum number of threads is the product of the maximum extent in
each dimension. Exceeding the maximum number of threads results in a runtime
error or kernel launch failure.
`"nvvm.reqntid"="<x>[,<y>[,<z>]]"`
: This attribute declares the exact number of threads in the thread block
(CTA). The number of threads is the product of the value in each dimension.
Specifying a different CTA dimension at launch will result in a runtime
error or kernel launch failure.
`"nvvm.cluster_dim"="<x>[,<y>[,<z>]]"`
: This attribute declares the number of thread blocks (CTAs) in the cluster.
The total number of CTAs is the product of the number of CTAs in each
dimension. Specifying a different cluster dimension at launch will result in
a runtime error or kernel launch failure. Only supported for Hopper+.
`"nvvm.blocksareclusters"`
: This attribute implies that the grid launch configuration for the
corresponding kernel function is specifying the number of clusters instead
of the number of thread blocks. This attribute is only allowed for kernel
functions and requires `nvvm.reqntid` and `nvvm.cluster_dim` attributes.
(address-spaces)=
### Address Spaces
The NVPTX back-end uses the following address space mapping:
| Address Space | Memory Space |
| ------------- | -------------- |
| 0 | Generic |
| 1 | Global |
| 2 | Internal Use |
| 3 | Shared |
| 4 | Constant |
| 5 | Local |
| 7 | Shared Cluster |
Every global variable and pointer type is assigned to one of these address
spaces, with 0 being the default address space. Intrinsics are provided which
can be used to convert pointers between the generic and non-generic address
spaces.
As an example, the following IR will define an array `@g` that resides in
global device memory.
```llvm
@g = internal addrspace(1) global [4 x i32] [ i32 0, i32 1, i32 2, i32 3 ]
```
LLVM IR functions can read and write to this array, and host-side code can
copy data to it by name with the CUDA Driver API.
Note that since address space 0 is the generic space, it is illegal to have
global variables in address space 0. Address space 0 is the default address
space in LLVM, so the `addrspace(N)` annotation is *required* for global
variables.
### Triples
The NVPTX target uses the module triple to select between 32/64-bit code
generation and the driver-compiler interface to use. The triple architecture
can be one of `nvptx` (32-bit PTX) or `nvptx64` (64-bit PTX). The
operating system should be one of `cuda` or `nvcl`, which determines the
interface used by the generated code to communicate with the driver. Most
users will want to use `cuda` as the operating system, which makes the
generated PTX compatible with the CUDA Driver API.
Example: 32-bit PTX for CUDA Driver API: `nvptx-nvidia-cuda`
Example: 64-bit PTX for CUDA Driver API: `nvptx64-nvidia-cuda`
(nvptx-arch-hierarchy)=
## NVPTX Architecture Hierarchy and Ordering
GPU architectures: sm_2Y/sm_3Y/sm_5Y/sm_6Y/sm_7Y/sm_8Y/sm_9Y/sm_10Y/sm_12Y
('Y' represents version within the architecture). The architectures have name of
the form `sm_XYz` where:
- `X` represent the generation number
- `Y` represents the version within the architecture, and
- `z` represents the optional feature suffix.
If `X1Y1 <= X2Y2`, then GPU capabilities of `sm_X1Y1` are included in
`sm_X2Y2`. For example, take `sm_90` (9 represents `X`, 0 represents
`Y`, and no feature suffix) and `sm_103` architectures (10 represents `X`,
3 represents `Y`, and no feature suffix). Since 90 <= 103, `sm_90` is
compatible with `sm_103`.
The family-specific variants have `f` feature suffix and they follow
following order:
`sm_X{Y2}f > sm_X{Y1}f` iff `Y2 > Y1`
`sm_XY{f} > sm_{XY}{}`
For example, take `sm_100f` (10 represents `X`, 0 represents `Y`, and
`f` represents `z`) and `sm_103f` (10 represents `X`, 3 represents
`Y`, and `f` represents `z`) architecture variants. Since `Y1 < Y2`,
`sm_100f` is compatible with `sm_103f`. Similarly based on the second rule,
`sm_90` is compatible with `sm_103f`.
Some counter examples, take `sm_100f` and `sm_120f` (12 represents `X`, 0
represents `Y`, and `f` represents `z`) architecture variants. Since both
belongs to different family i.e. `X1 != X2`, `sm_100f` is not compatible
with `sm_120f`.
The architecture-specific variants have `a` feature suffix and they follow
following order:
`sm_XY{a} > sm_XY{f} > sm_{XY}{}`
For example, take `sm_103a` (10 represents `X`, 3 represents `Y`, and
`a` represents `z`), `sm_103f`, and `sm_103` architecture variants. The
`sm_103` is compatible with `sm_103a` and `sm_103f`, and `sm_103f` is
compatible with `sm_103a`.
Encoding := Arch * 10 + 2 (for 'f') + 1 (for 'a')
Arch := X * 10 + Y
For example, `sm_103f` is encoded as 1032 (103 * 10 + 2) and `sm_103a` is
encoded as 1033 (103 * 10 + 2 + 1).
This encoding allows simple partial ordering of the architectures.
- Compare Family and Arch by dividing FullSMVersion by 100 and 10
respectively before the comparison.
- Compare within the family by comparing FullSMVersion, given both belongs to
the same family.
- Detect `a` variants by checking FullSMVersion & 1.
(nvptx-intrinsics)=
## NVPTX Intrinsics
### Reading PTX Special Registers
#### '`llvm.nvvm.read.ptx.sreg.*`'
##### Syntax:
```llvm
declare i32 @llvm.nvvm.read.ptx.sreg.tid.x()
declare i32 @llvm.nvvm.read.ptx.sreg.tid.y()
declare i32 @llvm.nvvm.read.ptx.sreg.tid.z()
declare i32 @llvm.nvvm.read.ptx.sreg.ntid.x()
declare i32 @llvm.nvvm.read.ptx.sreg.ntid.y()
declare i32 @llvm.nvvm.read.ptx.sreg.ntid.z()
declare i32 @llvm.nvvm.read.ptx.sreg.ctaid.x()
declare i32 @llvm.nvvm.read.ptx.sreg.ctaid.y()
declare i32 @llvm.nvvm.read.ptx.sreg.ctaid.z()
declare i32 @llvm.nvvm.read.ptx.sreg.nctaid.x()
declare i32 @llvm.nvvm.read.ptx.sreg.nctaid.y()
declare i32 @llvm.nvvm.read.ptx.sreg.nctaid.z()
declare i32 @llvm.nvvm.read.ptx.sreg.warpsize()
```
##### Overview:
The '`@llvm.nvvm.read.ptx.sreg.*`' intrinsics provide access to the PTX
special registers, in particular the kernel launch bounds. These registers
map in the following way to CUDA builtins:
| CUDA Builtin | PTX Special Register Intrinsic |
| ------------ | ----------------------------------- |
| `threadId` | `@llvm.nvvm.read.ptx.sreg.tid.*` |
| `blockIdx` | `@llvm.nvvm.read.ptx.sreg.ctaid.*` |
| `blockDim` | `@llvm.nvvm.read.ptx.sreg.ntid.*` |
| `gridDim` | `@llvm.nvvm.read.ptx.sreg.nctaid.*` |
#### '`llvm.nvvm.read.ptx.sreg.*_smem_size`'
##### Syntax:
```llvm
declare i32 @llvm.nvvm.read.ptx.sreg.total_smem_size()
declare i32 @llvm.nvvm.read.ptx.sreg.aggr_smem_size()
declare i32 @llvm.nvvm.read.ptx.sreg.dynamic_smem_size()
```
##### Overview:
The '`@llvm.nvvm.read.ptx.sreg.total_smem_size`' intrinsic reads the PTX
special register that holds the total amount of shared memory allocated per CTA
for the kernel at launch.
The reported value includes both statically allocated and dynamically requested
shared memory, but excludes any shared memory reserved for system use. The size
is expressed in units of the architecture-specific shared memory allocation
granularity. For targets sm_8x and newer, this granularity is 128 bytes.
The '`aggr_smem_size`' variant returns the aggregate shared memory size,
including the portion reserved for system software use.
The '`dynamic_smem_size`' variant returns the amount of dynamic shared
memory allocated per CTA for the kernel at launch time.
#### '`llvm.nvvm.read.ptx.sreg.reserved_smem_offset_*`'
##### Syntax:
```llvm
declare i32 @llvm.nvvm.read.ptx.sreg.reserved_smem_offset_begin()
declare i32 @llvm.nvvm.read.ptx.sreg.reserved_smem_offset_end()
declare i32 @llvm.nvvm.read.ptx.sreg.reserved_smem_offset_cap()
declare i32 @llvm.nvvm.read.ptx.sreg.reserved_smem_offset_0()
declare i32 @llvm.nvvm.read.ptx.sreg.reserved_smem_offset_1()
```
##### Overview:
The '`@llvm.nvvm.read.ptx.sreg.reserved_smem_offset_*`' intrinsics provide
access to PTX special registers that hold information about reserved shared
memory offsets.
The '`reserved_smem_offset_begin`' intrinsic reads the start offset of
the reserved shared memory region.
The '`reserved_smem_offset_end`' intrinsic reads the end offset of the
reserved shared memory region.
The '`reserved_smem_offset_cap`' intrinsic reads the capacity limit of
the reserved shared memory region.
The '`reserved_smem_offset_0`' and '`reserved_smem_offset_1`' intrinsics
read additional offsets in the reserved shared memory region.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#special-registers-reserved-smem).
### Barriers
#### '`llvm.nvvm.barrier.cta.*`'
##### Syntax:
```llvm
declare void @llvm.nvvm.barrier.cta.sync.count(i32 %id, i32 %n)
declare void @llvm.nvvm.barrier.cta.sync.all(i32 %id)
declare void @llvm.nvvm.barrier.cta.sync.aligned.count(i32 %id, i32 %n)
declare void @llvm.nvvm.barrier.cta.sync.aligned.all(i32 %id)
declare void @llvm.nvvm.barrier.cta.arrive.count(i32 %id, i32 %n)
declare void @llvm.nvvm.barrier.cta.arrive.aligned.count(i32 %id, i32 %n)
declare i32 @llvm.nvvm.barrier.cta.red.popc.count(i32 %id, i32 %n, i1 %pred)
declare i32 @llvm.nvvm.barrier.cta.red.popc.all(i32 %id, i1 %pred)
declare i32 @llvm.nvvm.barrier.cta.red.popc.aligned.count(i32 %id, i32 %n, i1 %pred)
declare i32 @llvm.nvvm.barrier.cta.red.popc.aligned.all(i32 %id, i1 %pred)
declare i32 @llvm.nvvm.barrier.cta.red.and.count(i32 %id, i32 %n, i1 %pred)
declare i32 @llvm.nvvm.barrier.cta.red.and.all(i32 %id, i1 %pred)
declare i32 @llvm.nvvm.barrier.cta.red.and.aligned.count(i32 %id, i32 %n, i1 %pred)
declare i32 @llvm.nvvm.barrier.cta.red.and.aligned.all(i32 %id, i1 %pred)
declare i32 @llvm.nvvm.barrier.cta.red.or.count(i32 %id, i32 %n, i1 %pred)
declare i32 @llvm.nvvm.barrier.cta.red.or.all(i32 %id, i1 %pred)
declare i32 @llvm.nvvm.barrier.cta.red.or.aligned.count(i32 %id, i32 %n, i1 %pred)
declare i32 @llvm.nvvm.barrier.cta.red.or.aligned.all(i32 %id, i1 %pred)
```
##### Overview:
The '`@llvm.nvvm.barrier.cta.*`' family of intrinsics perform barrier
synchronization and communication within a CTA. They can be used by the threads
within the CTA for synchronization and communication.
##### Semantics:
Operand %id specifies a logical barrier resource and must fall within the range
0 through 15. When present, operand %n specifies the number of threads
participating in the barrier. When specifying a thread count, the value must be
a multiple of the warp size. With the '`@llvm.nvvm.barrier.cta.sync.*`'
variants, the '`.all`' suffix indicates that all threads in the CTA should
participate in the barrier while the '`.count`' suffix indicates that only
the threads specified by the %n operand should participate in the barrier.
All forms of the '`@llvm.nvvm.barrier.cta.*`' intrinsic cause the executing
thread to wait for all non-exited threads from its warp and then marks the
warp's arrival at the barrier. In addition to signaling its arrival at the
barrier, the '`@llvm.nvvm.barrier.cta.red.*`' and
'`@llvm.nvvm.barrier.cta.sync.*`' intrinsics cause the executing thread to
wait for non-exited threads of all other warps participating in the barrier to
arrive. On the other hand, the '`@llvm.nvvm.barrier.cta.arrive.*`' intrinsic
does not cause the executing thread to wait for threads of other participating
warps.
When a barrier completes, the waiting threads are restarted without delay,
and the barrier is reinitialized so that it can be immediately reused.
The '`@llvm.nvvm.barrier.cta.red.*`' intrinsics perform a reduction operation
across threads. The %pred operands from all threads in the CTA are combined
using the specified reduction operator. Once the barrier count is reached, the
final value is returned in all threads waiting at the barrier.
The reduction operations for '`@llvm.nvvm.barrier.cta.red.*`' are
population-count ('`.popc`'), all-threads-true ('`.and`'),
and any-thread-true ('`.or`'). The result of '`.popc`' is the number of
threads with a true predicate, while '`.and`' and '`.or`' indicate if all
the threads had a true predicate or if any of the threads had a true predicate.
The '`@llvm.nvvm.barrier.cta.*`' intrinsic has an optional '`.aligned`'
modifier to indicate textual alignment of the barrier. When specified, it
indicates that all threads in the CTA will execute the same
'`@llvm.nvvm.barrier.cta.*`' instruction. In conditionally executed code, an
aligned '`@llvm.nvvm.barrier.cta.*`' instruction should only be used if it is
known that all threads in the CTA evaluate the condition identically, otherwise
behavior is undefined.
### MBarrier family of Intrinsics
#### Overview:
An `mbarrier` is a barrier created in shared memory that supports:
- Synchronizing any subset of threads within a CTA.
- One-way synchronization of threads across CTAs of a cluster.
Threads can perform only `arrive` operations but not `*_wait` on an
mbarrier located in shared::cluster space.
- Waiting for completion of asynchronous memory operations initiated by a
thread and making them visible to other threads.
Unlike `bar{.cta}/barrier{.cta}` instructions which can access a limited
number of barriers per CTA, `mbarrier` objects are user-defined and are
only limited by the total shared memory size available.
An mbarrier object is an opaque object in shared memory with an
alignment of 8-bytes. It keeps track of:
- Current phase of the mbarrier object
- Count of pending arrivals for the current phase of the mbarrier object
- Count of expected arrivals for the next phase of the mbarrier object
- Count of pending asynchronous memory operations (or transactions)
tracked by the current phase of the mbarrier object. This is also
referred to as `tx-count`. The unit of `tx-count` is specified
by the asynchronous memory operation (for example,
`llvm.nvvm.cp.async.bulk.tensor.g2s.*`).
The `phase` of an mbarrier object is the number of times the mbarrier
object has been used to synchronize threads/track async operations.
In each phase, threads perform:
- arrive/expect-tx/complete-tx operations to progress the current phase.
- test_wait/try_wait operations to check for completion of the current phase.
An mbarrier object completes the current phase when:
- The count of the pending arrivals has reached zero AND
- The tx-count has reached zero.
When an mbarrier object completes the current phase, below
actions are performed `atomically`:
- The mbarrier object transitions to the next phase.
- The pending arrival count is reinitialized to the expected arrival count.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#parallel-synchronization-and-communication-instructions-mbarrier).
#### '`llvm.nvvm.mbarrier.init`'
##### Syntax:
```llvm
declare void @llvm.nvvm.mbarrier.init(ptr %addr, i32 %count)
declare void @llvm.nvvm.mbarrier.init.shared(ptr addrspace(3) %addr, i32 %count)
```
##### Overview:
The '`@llvm.nvvm.mbarrier.init.*`' intrinsics are used to initialize
an mbarrier object located at `addr` with the value `count`.
`count` is a 32-bit unsigned integer value and must be within
the range [1...2^20-1]. During initialization:
- The tx-count and the current phase of the mbarrier object are set to 0.
- The expected and pending arrival counts are set to `count`.
##### Semantics:
The `.shared` variant explicitly uses shared memory address space for
the `addr` operand. If the `addr` does not fall within the
shared::cta space, then the behavior of this intrinsic is undefined.
Performing `mbarrier.init` on a valid mbarrier object is undefined;
use `mbarrier.inval` before reusing the memory for another mbarrier
or any other purpose.
#### '`llvm.nvvm.mbarrier.inval`'
##### Syntax:
```llvm
declare void @llvm.nvvm.mbarrier.inval(ptr %addr)
declare void @llvm.nvvm.mbarrier.inval.shared(ptr addrspace(3) %addr)
```
##### Overview:
The '`@llvm.nvvm.mbarrier.inval.*`' intrinsics invalidate the mbarrier
object at the address specified by `addr`.
##### Semantics:
The `.shared` variant explicitly uses shared memory address space for
the `addr` operand. If the `addr` does not fall within the
shared::cta space, then the behavior of this intrinsic is undefined.
It is expected that `addr` was previously initialized using
`mbarrier.init`; otherwise, the behavior is undefined.
#### '`llvm.nvvm.mbarrier.expect.tx`'
##### Syntax:
```llvm
declare void @llvm.nvvm.mbarrier.expect.tx.scope.cta.space.cta(ptr addrspace(3) %addr, i32 %tx_count)
declare void @llvm.nvvm.mbarrier.expect.tx.scope.cluster.space.cta(ptr addrspace(3) %addr, i32 %tx_count)
declare void @llvm.nvvm.mbarrier.expect.tx.scope.cta.space.cluster(ptr addrspace(7) %addr, i32 %tx_count)
declare void @llvm.nvvm.mbarrier.expect.tx.scope.cluster.space.cluster(ptr addrspace(7) %addr, i32 %tx_count)
```
##### Overview:
The '`@llvm.nvvm.mbarrier.expect.tx.*`' intrinsics increase the transaction
count of the mbarrier object at `%addr` by `%tx_count`. The `%tx_count`
is a 32-bit unsigned integer value.
##### Semantics:
The `.space.{cta/cluster}` indicates the address space where the mbarrier
object resides.
The `.scope.{cta/cluster}` denotes the set of threads that can directly
observe the synchronizing effect of the mbarrier operation. When scope is
"cta", all threads executing in the same CTA (as the current thread) can
directly observe the effect of the `expect.tx` operation. Similarly,
when scope is "cluster", all threads executing in the same Cluster
(as the current thread) can directly observe the effect of the operation.
If the `addr` does not fall within shared::cta or shared::cluster space,
then the behavior of this intrinsic is undefined. This intrinsic has
`relaxed` semantics and hence does not provide any memory ordering
or visibility guarantees.
#### '`llvm.nvvm.mbarrier.complete.tx`'
##### Syntax:
```llvm
declare void @llvm.nvvm.mbarrier.complete.tx.scope.cta.space.cta(ptr addrspace(3) %addr, i32 %tx_count)
declare void @llvm.nvvm.mbarrier.complete.tx.scope.cluster.space.cta(ptr addrspace(3) %addr, i32 %tx_count)
declare void @llvm.nvvm.mbarrier.complete.tx.scope.cta.space.cluster(ptr addrspace(7) %addr, i32 %tx_count)
declare void @llvm.nvvm.mbarrier.complete.tx.scope.cluster.space.cluster(ptr addrspace(7) %addr, i32 %tx_count)
```
##### Overview:
The '`@llvm.nvvm.mbarrier.complete.tx.*`' intrinsics decrease the transaction
count of the mbarrier object at `%addr` by `%tx_count`. The `%tx_count`
is a 32-bit unsigned integer value. As a result of this decrement,
the mbarrier can potentially complete its current phase and transition
to the next phase.
##### Semantics:
The semantics of these intrinsics are identical to those of the
`llvm.nvvm.mbarrier.expect.tx.*` intrinsics described above.
#### '`llvm.nvvm.mbarrier.arrive`'
##### Syntax:
```llvm
declare i64 @llvm.nvvm.mbarrier.arrive.scope.cta.space.cta(ptr addrspace(3) %addr, i32 %count)
declare i64 @llvm.nvvm.mbarrier.arrive.scope.cluster.space.cta(ptr addrspace(3) %addr, i32 %count)
declare void @llvm.nvvm.mbarrier.arrive.scope.cta.space.cluster(ptr addrspace(7) %addr, i32 %count)
declare void @llvm.nvvm.mbarrier.arrive.scope.cluster.space.cluster(ptr addrspace(7) %addr, i32 %count)
declare i64 @llvm.nvvm.mbarrier.arrive.relaxed.scope.cta.space.cta(ptr addrspace(3) %addr, i32 %count)
declare i64 @llvm.nvvm.mbarrier.arrive.relaxed.scope.cluster.space.cta(ptr addrspace(3) %addr, i32 %count)
declare void @llvm.nvvm.mbarrier.arrive.relaxed.scope.cta.space.cluster(ptr addrspace(7) %addr, i32 %count)
declare void @llvm.nvvm.mbarrier.arrive.relaxed.scope.cluster.space.cluster(ptr addrspace(7) %addr, i32 %count)
```
##### Overview:
The `@llvm.nvvm.mbarrier.arrive.*` intrinsics signal the arrival of the
executing thread or completion of an asynchronous instruction associated with
an arrive operation on the mbarrier object at `%addr`. This operation
decrements the pending arrival count by `%count`, a 32-bit unsigned integer,
potentially completing the current phase and triggering a transition to the
next phase.
##### Semantics:
The `.space.{cta/cluster}` indicates the address space where the mbarrier
object resides. When the mbarrier is in shared::cta space, the intrinsics
return an opaque 64-bit value capturing the phase of the mbarrier object
\_prior\_ to this arrive operation. This value can be used with a try_wait
or test_wait operation to check for the completion of the mbarrier.
The `.scope.{cta/cluster}` denotes the set of threads that can directly
observe the synchronizing effect of the mbarrier operation. When scope is
"cta", all threads executing in the same CTA (as the current thread) can
directly observe the effect of the `arrive` operation. Similarly,
when scope is "cluster", all threads executing in the same Cluster
(as the current thread) can directly observe the effect of the operation.
If the `addr` does not fall within shared::cta or shared::cluster space,
then the behavior of this intrinsic is undefined.
These intrinsics have `release` semantics by default. The release semantics
ensure ordering of operations that occur in program order \_before\_ this arrive
instruction, making their effects visible to subsequent operations in other
threads of the CTA (or cluster, depending on scope). Threads performing
corresponding acquire operations (such as mbarrier.test.wait) synchronize
with this release. The `relaxed` variants of these intrinsics do not
provide any memory ordering or visibility guarantees.
#### '`llvm.nvvm.mbarrier.arrive.expect.tx`'
##### Syntax:
```llvm
declare i64 @llvm.nvvm.mbarrier.arrive.expect.tx.scope.cta.space.cta(ptr addrspace(3) %addr, i32 %tx_count)
declare i64 @llvm.nvvm.mbarrier.arrive.expect.tx.scope.cluster.space.cta(ptr addrspace(3) %addr, i32 %tx_count)
declare void @llvm.nvvm.mbarrier.arrive.expect.tx.scope.cta.space.cluster(ptr addrspace(7) %addr, i32 %tx_count)
declare void @llvm.nvvm.mbarrier.arrive.expect.tx.scope.cluster.space.cluster(ptr addrspace(7) %addr, i32 %tx_count)
declare i64 @llvm.nvvm.mbarrier.arrive.expect.tx.relaxed.scope.cta.space.cta(ptr addrspace(3) %addr, i32 %tx_count)
declare i64 @llvm.nvvm.mbarrier.arrive.expect.tx.relaxed.scope.cluster.space.cta(ptr addrspace(3) %addr, i32 %tx_count)
declare void @llvm.nvvm.mbarrier.arrive.expect.tx.relaxed.scope.cta.space.cluster(ptr addrspace(7) %addr, i32 %tx_count)
declare void @llvm.nvvm.mbarrier.arrive.expect.tx.relaxed.scope.cluster.space.cluster(ptr addrspace(7) %addr, i32 %tx_count)
```
##### Overview:
The `@llvm.nvvm.mbarrier.arrive.expect.tx.*` intrinsics are similar to
the `@llvm.nvvm.mbarrier.arrive` intrinsics except that they also
perform an `expect-tx` operation \_prior\_ to the `arrive` operation.
The `%tx_count` specifies the transaction count for the `expect-tx`
operation and the count for the `arrive` operation is assumed to be 1.
##### Semantics:
The semantics of these intrinsics are identical to those of the
`llvm.nvvm.mbarrier.arrive.*` intrinsics described above.
#### '`llvm.nvvm.mbarrier.arrive.drop`'
##### Syntax:
```llvm
declare i64 @llvm.nvvm.mbarrier.arrive.drop.scope.cta.space.cta(ptr addrspace(3) %addr, i32 %count)
declare i64 @llvm.nvvm.mbarrier.arrive.drop.scope.cluster.space.cta(ptr addrspace(3) %addr, i32 %count)
declare void @llvm.nvvm.mbarrier.arrive.drop.scope.cta.space.cluster(ptr addrspace(7) %addr, i32 %count)
declare void @llvm.nvvm.mbarrier.arrive.drop.scope.cluster.space.cluster(ptr addrspace(7) %addr, i32 %count)
declare i64 @llvm.nvvm.mbarrier.arrive.drop.relaxed.scope.cta.space.cta(ptr addrspace(3) %addr, i32 %count)
declare i64 @llvm.nvvm.mbarrier.arrive.drop.relaxed.scope.cluster.space.cta(ptr addrspace(3) %addr, i32 %count)
declare void @llvm.nvvm.mbarrier.arrive.drop.relaxed.scope.cta.space.cluster(ptr addrspace(7) %addr, i32 %count)
declare void @llvm.nvvm.mbarrier.arrive.drop.relaxed.scope.cluster.space.cluster(ptr addrspace(7) %addr, i32 %count)
```
##### Overview:
The `@llvm.nvvm.mbarrier.arrive.drop.*` intrinsics decrement the
expected arrival count of the mbarrier object at `%addr` by
`%count` and then perform an `arrive` operation with `%count`.
The `%count` is a 32-bit integer.
##### Semantics:
The semantics of these intrinsics are identical to those of the
`llvm.nvvm.mbarrier.arrive.*` intrinsics described above.
#### '`llvm.nvvm.mbarrier.arrive.drop.expect.tx`'
##### Syntax:
```llvm
declare i64 @llvm.nvvm.mbarrier.arrive.drop.expect.tx.scope.cta.space.cta(ptr addrspace(3) %addr, i32 %tx_count)
declare i64 @llvm.nvvm.mbarrier.arrive.drop.expect.tx.scope.cluster.space.cta(ptr addrspace(3) %addr, i32 %tx_count)
declare void @llvm.nvvm.mbarrier.arrive.drop.expect.tx.scope.cta.space.cluster(ptr addrspace(7) %addr, i32 %tx_count)
declare void @llvm.nvvm.mbarrier.arrive.drop.expect.tx.scope.cluster.space.cluster(ptr addrspace(7) %addr, i32 %tx_count)
declare i64 @llvm.nvvm.mbarrier.arrive.drop.expect.tx.relaxed.scope.cta.space.cta(ptr addrspace(3) %addr, i32 %tx_count)
declare i64 @llvm.nvvm.mbarrier.arrive.drop.expect.tx.relaxed.scope.cluster.space.cta(ptr addrspace(3) %addr, i32 %tx_count)
declare void @llvm.nvvm.mbarrier.arrive.drop.expect.tx.relaxed.scope.cta.space.cluster(ptr addrspace(7) %addr, i32 %tx_count)
declare void @llvm.nvvm.mbarrier.arrive.drop.expect.tx.relaxed.scope.cluster.space.cluster(ptr addrspace(7) %addr, i32 %tx_count)
```
##### Overview:
The `@llvm.nvvm.mbarrier.arrive.drop.expect.tx.*` intrinsics perform
the below operations on the mbarrier located at `%addr`.
- Perform an `expect-tx` operation i.e. increase the transaction count
of the mbarrier by `%tx_count`, a 32-bit unsigned integer value.
- Decrement the expected arrival count of the mbarrier by 1.
- Perform an `arrive` operation on the mbarrier with a value of 1.
##### Semantics:
The semantics of these intrinsics are identical to those of the
`llvm.nvvm.mbarrier.arrive.*` intrinsics described above.
#### '`llvm.nvvm.mbarrier.test.wait`'
##### Syntax:
```llvm
declare i1 @llvm.nvvm.mbarrier.test.wait.scope.cta.space.cta(ptr addrspace(3) %addr, i64 %state)
declare i1 @llvm.nvvm.mbarrier.test.wait.scope.cluster.space.cta(ptr addrspace(3) %addr, i64 %state)
declare i1 @llvm.nvvm.mbarrier.test.wait.parity.scope.cta.space.cta(ptr addrspace(3) %addr, i32 %phase)
declare i1 @llvm.nvvm.mbarrier.test.wait.parity.scope.cluster.space.cta(ptr addrspace(3) %addr, i32 %phase)
declare i1 @llvm.nvvm.mbarrier.test.wait.relaxed.scope.cta.space.cta(ptr addrspace(3) %addr, i64 %state)
declare i1 @llvm.nvvm.mbarrier.test.wait.relaxed.scope.cluster.space.cta(ptr addrspace(3) %addr, i64 %state)
declare i1 @llvm.nvvm.mbarrier.test.wait.parity.relaxed.scope.cta.space.cta(ptr addrspace(3) %addr, i32 %phase)
declare i1 @llvm.nvvm.mbarrier.test.wait.parity.relaxed.scope.cluster.space.cta(ptr addrspace(3) %addr, i32 %phase)
```
##### Overview:
The `@llvm.nvvm.mbarrier.test.wait.*` intrinsics test for the completion
of the current or the immediately preceding phase of an mbarrier object at
`%addr`. The test for completion can be done with either the `state` or
the `phase-parity` of the mbarrier object.
- When done through the `i64 %state` operand, the state must be
returned by an `llvm.nvvm.mbarrier.arrive.*` on the \_same\_
mbarrier object.
- The `.parity` variant of these intrinsics test for completion
of the phase indicated by the operand `i32 %phase`, which is
the integer parity of either the current phase or the immediately
preceding phase of the mbarrier object. An even phase has integer
parity 0 and an odd phase has integer parity of 1. So the valid
values for phase-parity are 0 and 1.
##### Semantics:
The `.scope.{cta/cluster}` denotes the set of threads that the
test_wait operation can directly synchronize with.
If the `addr` does not fall within shared::cta space, then the
the behavior of this intrinsic is undefined.
These intrinsics have `acquire` semantics by default. This acquire
pattern establishes memory ordering for operations occurring in program
order after this `test_wait` instruction by making operations from
other threads in the CTA (or cluster, depending on scope) visible to
subsequent operations in the current thread. When this wait completes,
it synchronizes with the corresponding release pattern from the
`mbarrier.arrive` operation. The `relaxed` variants of these intrinsics
do not provide any memory ordering or visibility guarantees.
This `test.wait` intrinsic is non-blocking and immediately returns
the completion status without suspending the executing thread.
The boolean return value indicates:
- True: The immediately preceding phase has completed
- False: The current phase is still incomplete
When this wait returns true, the following ordering guarantees hold:
- All memory accesses (except async operations) requested prior to
`mbarrier.arrive` having release semantics by participating
threads of a CTA (or cluster, depending on scope) are visible to
the executing thread.
- All `cp.async` operations requested prior to `cp.async.mbarrier.arrive`
by participating threads of a CTA are visible to the executing thread.
- All `cp.async.bulk` operations using the same mbarrier object requested
prior to `mbarrier.arrive` having release semantics by participating CTA
threads are visible to the executing thread.
- Memory accesses requested after this wait are not visible to memory
accesses performed prior to `mbarrier.arrive` by other participating
threads.
- No ordering guarantee exists for memory accesses by the same thread
between an `mbarrier.arrive` and this wait.
#### '`llvm.nvvm.mbarrier.try.wait`'
##### Syntax:
```llvm
declare i1 @llvm.nvvm.mbarrier.try.wait{.relaxed}.scope.cta.space.cta(ptr addrspace(3) %addr, i64 %state)
declare i1 @llvm.nvvm.mbarrier.try.wait{.relaxed}.scope.cluster.space.cta(ptr addrspace(3) %addr, i64 %state)
declare i1 @llvm.nvvm.mbarrier.try.wait.parity{.relaxed}.scope.cta.space.cta(ptr addrspace(3) %addr, i32 %phase)
declare i1 @llvm.nvvm.mbarrier.try.wait.parity{.relaxed}.scope.cluster.space.cta(ptr addrspace(3) %addr, i32 %phase)
declare i1 @llvm.nvvm.mbarrier.try.wait.tl{.relaxed}.scope.cta.space.cta(ptr addrspace(3) %addr, i64 %state, i32 %timelimit)
declare i1 @llvm.nvvm.mbarrier.try.wait.tl{.relaxed}.scope.cluster.space.cta(ptr addrspace(3) %addr, i64 %state, i32 %timelimit)
declare i1 @llvm.nvvm.mbarrier.try.wait.parity.tl{.relaxed}.scope.cta.space.cta(ptr addrspace(3) %addr, i32 %phase, i32 %timelimit)
declare i1 @llvm.nvvm.mbarrier.try.wait.parity.tl{.relaxed}.scope.cluster.space.cta(ptr addrspace(3) %addr, i32 %phase, i32 %timelimit)
```
##### Overview:
The `@llvm.nvvm.mbarrier.try.wait.*` intrinsics test for the completion of
the current or immediately preceding phase of an mbarrier object at `%addr`.
Unlike the `test.wait` intrinsics, which perform a non-blocking test, these
intrinsics may block the executing thread until the specified phase completes
or a system-dependent time limit expires. Suspended threads resume execution
when the phase completes or the time limit elapses. This time limit is
configurable through the `.tl` variants of these intrinsics, where the
`%timelimit` operand (an unsigned integer) specifies the limit in
nanoseconds. Other semantics are identical to those of the `test.wait`
intrinsics described above.
### Electing a thread
#### '`llvm.nvvm.elect.sync`'
##### Syntax:
```llvm
declare {i32, i1} @llvm.nvvm.elect.sync(i32 %membermask)
```
##### Overview:
The '`@llvm.nvvm.elect.sync`' intrinsic generates the `elect.sync`
PTX instruction, which elects one predicated active leader thread from
a set of threads specified by `membermask`. The behavior is undefined
if the executing thread is not in `membermask`. The laneid of the
elected thread is captured in the i32 return value. The i1 return
value is set to `True` for the leader thread and `False` for all
the other threads. Election of a leader thread happens deterministically,
i.e. the same leader thread is elected for the same `membermask`
every time. For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#parallel-synchronization-and-communication-instructions-elect-sync).
### Membar/Fences
#### '`llvm.nvvm.fence.acquire/release.sync_restrict.*`'
##### Syntax:
```llvm
declare void @llvm.nvvm.fence.acquire.sync_restrict.space.cluster.scope.cluster()
declare void @llvm.nvvm.fence.release.sync_restrict.space.cta.scope.cluster()
```
##### Overview:
The `nvvm.fence.{semantics}.sync_restrict.*` restrict the class of memory
operations for which the fence instruction provides the memory ordering
guarantees. When `.sync_restrict` is restricted to `shared_cta`, then memory
semantics must be `release` and the effect of the fence operation only applies
to operations performed on objects in `shared_cta` space. Likewise, when
`sync_restrict` is restricted to `shared_cluster`, then memory semantics must be
`acquire` and the effect of the fence operation only applies to operations
performed on objects in `shared_cluster` memory space. The scope for both
operations is `cluster`. For more details, please refer the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#parallel-synchronization-and-communication-instructions-membar).
#### '`llvm.nvvm.fence.mbarrier_init.release.cluster`'
##### Syntax:
```llvm
declare void @llvm.nvvm.fence.mbarrier_init.release.cluster()
```
##### Overview:
`nvvm.fence.mbarrier_init.release.cluster` intrinsic restrict the class of
memory operations for which the fence instruction provides the memory ordering
guarantees. The `mbarrier_init` modifiers restricts the synchronizing effect to
the prior `mbarrier_init` operation executed by the same thread on mbarrier
objects in `shared_cta` memory space. For more details, please refer the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#parallel-synchronization-and-communication-instructions-membar)
#### '`llvm.nvvm.fence.proxy.async_generic.acquire/release.sync_restrict`'
##### Syntax:
```llvm
declare void @llvm.nvvm.fence.proxy.async_generic.acquire.sync_restrict.space.cluster.scope.cluster()
declare void @llvm.nvvm.fence.proxy.async_generic.release.sync_restrict.space.cta.scope.cluster()
```
##### Overview:
`nvvm.fence.proxy.async_generic.{semantics}.sync_restrict` are used to establish
ordering between a prior memory access performed via the [async proxy](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#proxies)
and a subsequent memory access performed via the generic proxy.
`nvvm.fence.proxy.async_generic.release.sync_restrict` can form a release
sequence that synchronizes with an acquire sequence that contains the
`nvvm.fence.proxy.async_generic.acquire.sync_restrict` proxy fence. When
`.sync_restrict` is restricted to `shared_cta`, then memory semantics must
be `release` and the effect of the fence operation only applies to operations
performed on objects in `shared_cta` space. Likewise, when `sync_restrict` is
restricted to `shared_cluster`, then memory semantics must be `acquire` and the
effect of the fence operation only applies to operations performed on objects in
`shared_cluster` memory space. The scope for both operations is `cluster`.
For more details, please refer the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#parallel-synchronization-and-communication-instructions-membar)
#### '`llvm.nvvm.fence.proxy.<proxykind>`'
##### Syntax:
```llvm
declare void @llvm.nvvm.fence.proxy.alias()
declare void @llvm.nvvm.fence.proxy.async()
declare void @llvm.nvvm.fence.proxy.async.global()
declare void @llvm.nvvm.fence.proxy.async.shared_cluster()
declare void @llvm.nvvm.fence.proxy.async.shared_cta()
```
##### Overview:
`nvvm.fence.proxy.{proxykind}` intrinsics represent a fence with bi-directional
proxy ordering that is established between the memory accesses done between the
[generic proxy](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#proxies)
and the proxy specified by `proxykind`. A `bi-directional proxy` ordering between
two proxykinds establishes two `uni-directional` proxy orderings: one from the
first proxykind to the second proxykind and the other from the second proxykind
to the first proxykind.
`alias` proxykind refers to memory accesses performed using virtually aliased
addresses to the same memory location
`async` proxykind specifies that the memory ordering is established between the
`async proxy` and the `generic proxy`. The memory ordering is limited only to
operations performed on objects in the state space specified (`generic`, `global`,
`shared_cluster`, `shared_cta`). If no state space is specified, then the memory
ordering applies on all state spaces. For more details, please refer the
[PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#parallel-synchronization-and-communication-instructions-membar)
#### '`llvm.nvvm.fence.proxy.tensormap_generic.*`'
##### Syntax:
```llvm
declare void @llvm.nvvm.fence.proxy.tensormap_generic.release.cta()
declare void @llvm.nvvm.fence.proxy.tensormap_generic.release.cluster()
declare void @llvm.nvvm.fence.proxy.tensormap_generic.release.gpu()
declare void @llvm.nvvm.fence.proxy.tensormap_generic.release.sys()
declare void @llvm.nvvm.fence.proxy.tensormap_generic.acquire.cta(ptr %addr, i32 %size)
declare void @llvm.nvvm.fence.proxy.tensormap_generic.acquire.cluster(ptr %addr, i32 %size)
declare void @llvm.nvvm.fence.proxy.tensormap_generic.acquire.gpu(ptr %addr, i32 %size)
declare void @llvm.nvvm.fence.proxy.tensormap_generic.acquire.sys(ptr %addr, i32 %size)
```
##### Overview:
The `@llvm.nvvm.fence.proxy.tensormap_generic.*` is a uni-directional fence
used to establish ordering between a prior memory access performed via the
generic [proxy](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#proxies)
and a subsequent memory access performed via the tensormap proxy.
`nvvm.fence.proxy.tensormap_generic.release` can form a release sequence that
synchronizes with an acquire sequence that contains the
`nvvm.fence.proxy.tensormap_generic.acquire` proxy fence. The following table
describes the mapping between LLVM Intrinsic and the PTX instruction:
| NVVM Intrinsic | PTX Instruction |
| ---------------------------------------------------- | ------------------------------------------------------- |
| `@llvm.nvvm.fence.proxy.tensormap_generic.release.*` | `fence.proxy.tensormap::generic.release.*` |
| `@llvm.nvvm.fence.proxy.tensormap_generic.acquire.*` | `fence.proxy.tensormap::generic.acquire.* [addr], size` |
The address operand `addr` and the operand `size` together specify the
memory range `[addr, addr+size)` on which the ordering guarantees on the
memory accesses across the proxies is to be provided. The only supported value
for the `size` operand is `128` and must be an immediate. Generic Addressing
is used unconditionally, and the address specified by the operand addr must fall
within the `.global` state space. Otherwise, the behavior is undefined. For
more information, see [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#parallel-synchronization-and-communication-instructions-membar).
### Address Space Intrinsics
#### '`llvm.nvvm.isspacep.*`' Intrinsics
##### Syntax:
```llvm
declare i1 @llvm.nvvm.isspacep.const(ptr %p)
declare i1 @llvm.nvvm.isspacep.global(ptr %p)
declare i1 @llvm.nvvm.isspacep.local(ptr %p)
declare i1 @llvm.nvvm.isspacep.shared(ptr %p)
declare i1 @llvm.nvvm.isspacep.shared.cluster(ptr %p)
```
##### Overview:
The '`llvm.nvvm.isspacep.*`' intrinsics determine whether the provided generic
pointer references memory which falls within a particular address space.
##### Semantics:
If the given pointer in the generic address space refers to memory which falls
within the state space of the intrinsic (and therefore could be safely address
space casted to this space), 1 is returned, otherwise 0 is returned.
#### '`llvm.nvvm.mapa.*`' Intrinsics
##### Syntax:
```llvm
declare ptr @llvm.nvvm.mapa(ptr %p, i32 %rank)
declare ptr addrspace(7) @llvm.nvvm.mapa.shared.cluster(ptr addrspace(3) %p, i32 %rank)
```
##### Overview:
The '`llvm.nvvm.mapa.*`' intrinsics map a shared memory pointer `p` of
another CTA with `%rank` to the current CTA. The `llvm.nvvm.mapa` form
expects a generic pointer to shared memory and returns a generic pointer to
shared cluster memory. The `llvm.nvvm.mapa.shared.cluster` form expects a
pointer to shared memory and returns a pointer to shared cluster memory. They
corresponds directly to the `mapa` and `mapa.shared.cluster` PTX
instructions.
##### Semantics:
If the given pointer in the generic address space refers to memory which falls
within the state space of the intrinsic (and therefore could be safely address
space casted to this space), 1 is returned, otherwise 0 is returned.
### Narrow Floating-Point Conversion intrinsics
These intrinsics perform conversions involving narrow floating-point formats.
The following table describes the rounding modes used across these intrinsics:
(narrow-fp-rounding-modes)=
```{list-table} Narrow Floating-Point Conversion Rounding Modes
:widths: 30 60
:header-rows: 1
* - Rounding Mode
- Description
* - `rn` (default)
- Round to nearest, with ties to even
* - `rz`
- Round towards zero
* - `rp`
- Round towards positive infinity
* - `rs`
- Stochastic rounding which is achieved through the use of the supplied
random bits (`%rnd_bits`). The result s rounded in the direction towards
zero or away from zero based on the carry out of the integer addition of
the of mantissa from the input.
```
(scale-factor)=
Some conversions involve a scale factor which is provided as a packed 16-bit
integer containing two scaling factors of type `ue8m0`, one for each input.
For down conversion, inputs are divided by `scale_factor` and then the
conversion is performed. For up-conversion, inputs are converted to destination
type and then multiplied by `scale_factor`.
#### `fp8` Conversion Intrinsics
##### Syntax:
```llvm
declare i16 @llvm.nvvm.ff.to{.e4m3x2, .e5m2x2}.rn{.relu}(float %a, float %b)
declare i16 @llvm.nvvm.ff.to.ue8m0x2{.rz, .rp}{.satfinite}(float %a, float %b)
declare i16 @llvm.f16x2.to{.e4m3x2, .e5m2x2}.rn{.relu}(<2 x half> %a)
declare i16 @llvm.bf16x2.to{.e4m3x2, .e5m2x2}.rn{.relu}.satfinite(<2 x bfloat> %a)
declare i16 @llvm.bf16x2.to.ue8m0x2{.rz, .rp}{.satfinite}(<2 x bfloat> %a)
declare <2 x half> @llvm.nvvm{.e4m3x2, .e5m2x2}.to.f16x2.rn{.relu}(i16 %a)
declare <2 x bfloat> @llvm.nvvm{.e4m3x2, .e5m2x2}.to.bf16x2.rn{.relu}{.satfinite}.scale.n2.ue8m0(i16 %a, i16 %scale_factor)
declare <2 x bfloat> @llvm.nvvm.ue8m0x2.to.bf16x2(i16 %a)
declare <4 x i8> @llvm.nvvm.f32x4.to{.e4m3x4, .e5m2x4}.rs{.relu}.satfinite(<4 x f32> %a, i32 %rnd_bits)
```
##### Overview:
These intrinsics perform conversions involving the `e4m3` and `e5m2` narrow
floating-point formats. In case of two inputs, the value converted from input
`%a` is stored in the upper 8-bits of the result, and the value converted
from input `%b` is stored in the lower 8-bits of the result.
For rounding modes, see {ref}`narrow-fp-rounding-modes`.
The `relu` modifier clamps negative results to 0.
When `satfinite` is specified, if the absolute value of input (ignoring sign)
is greater than `MAX_NORM` of the specified destination format, then the
result is sign-preserved `MAX_NORM` of the destination format and a positive
`MAX_NORM` in `.ue8m0x2` for which the destination sign is not supported.
Also, if the input value is `NaN`, then the result is `NaN` in the
specified destination format. The `satfinite` modifier is assumed to be
present for conversions involving `e4m3` and `e5m2` types as the
destination.
For scale factor, see {ref}`scale-factor <scale-factor>`.
For more information, see [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt).
#### `s2f6` Conversion Intrinsics
##### Syntax:
```llvm
declare i16 @llvm.nvvm.ff.to.s2f6x2.rn{.relu}.satfinite.scale.n2.ue8m0(float %a, float %b, i16 %scale_factor)
declare i16 @llvm.nvvm.bf16x2.to.s2f6x2.rn{.relu}.satfinite.scale.n2.ue8m0(<2 x bfloat> %a, i16 %scale_factor)
declare <2 x bfloat> @llvm.nvvm.s2f6x2.to.bf16x2.rn{.relu}{.satfinite}.scale.n2.ue8m0(i16 %a, i16 %scale_factor)
```
##### Overview:
These intrinsics perform conversions involving the `s2f6` narrow
floating-point format. In case of two inputs, the value converted from input
`%a` is stored in the upper 8-bits of the result, and the value converted
from input `%b` is stored in the lower 8-bits of the result.
For rounding modes, see {ref}`narrow-fp-rounding-modes`.
The `relu` modifier clamps negative results to 0.
When `satfinite` is specified, if the absolute value of input (ignoring sign)
is greater than `MAX_NORM` of the specified destination format, then the
result is sign-preserved `MAX_NORM` of the destination format. Also, if the
input is `NaN`, then the result is the positive `MAX_NORM` of the
destination format.
For scale factor, see {ref}`scale-factor <scale-factor>`.
For more information, see [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt).
#### `fp6` Conversion Intrinsics
##### Syntax:
```llvm
declare i16 @llvm.nvvm.ff.to{.e2m3x2, .e3m2x2}.rn{.relu}.satfinite(float %a, float %b)
declare i16 @llvm.nvvm.f16x2.to{.e2m3x2, .e3m2x2}.rn{.relu}.satfinite(<2 x half> %a)
declare i16 @llvm.nvvm.bf16x2.to{.e2m3x2, .e3m2x2}.rn{.relu}.satfinite(<2 x bfloat> %a)
declare <2 x half> @llvm.nvvm{.e2m3x2, .e3m2x2}.to.f16x2.rn{.relu}(i16 %a)
declare <2 x bfloat> @llvm.nvvm{.e2m3x2, .e3m2x2}.to.bf16x2.rn{.relu}{.satfinite}.scale.n2.ue8m0(i16 %a, i16 %scale_factor)
declare <4 x i8> @llvm.nvvm.f32x4.to{.e2m3x4, .e3m2x4}.rs{.relu}.satfinite(<4 x f32> %a, i32 %rnd_bits)
```
##### Overview:
These intrinsics perform conversions involving the `e2m3` and `e3m2` narrow
floating-point formats. In case of two inputs, the value converted from input
`%a` is stored in the upper 8-bits of the result, and the value converted
from input `%b` is stored in the lower 8-bits of the result with 2 MSBs
padded with 0s in both cases.
For rounding modes, see {ref}`narrow-fp-rounding-modes`.
The `relu` modifier clamps negative results to 0.
When `satfinite` is specified, if the absolute value of input (ignoring sign)
is greater than `MAX_NORM` of the specified destination format, then the
result is sign-preserved `MAX_NORM` of the destination format. Also, if the
input is `NaN`, then the result is the positive `MAX_NORM` of the
destination format.
For scale factor, see {ref}`scale-factor <scale-factor>`.
For more information, see [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt).
#### `fp4` Conversion Intrinsics
##### Syntax:
```llvm
declare i16 @llvm.nvvm.ff.to.e2m1x2.rn{.relu}.satfinite(float %a, float %b)
declare i16 @llvm.nvvm.f16x2.to.e2m1x2.rn{.relu}.satfinite(<2 x half> %a)
declare i16 @llvm.nvvm.bf16x2.to.e2m1x2.rn{.relu}.satfinite(<2 x bfloat> %a)
declare <2 x half> @llvm.nvvm.e2m1x2.to.f16x2.rn{.relu}(i16 %a)
declare <2 x bfloat> @llvm.nvvm.e2m1x2.to.bf16x2.rn{.relu}{.satfinite}.scale.n2.ue8m0(i16 %a, i16 %scale_factor)
declare i16 @llvm.nvvm.f32x4.to.e2m1x4.rs{.relu}.satfinite(<4 x f32> %a, i32 %rnd_bits)
```
##### Overview:
These intrinsics perform conversions involving the `e2m1` narrow
floating-point format. For conversions involving `e2m1x2`, the packed
`e2m1x2` value is stored in the lower byte of the `i16` argument or result.
In case of two inputs, the value converted from input
`%a` is stored in the upper 4-bits of the result, and the value converted
from input `%b` is stored in the lower 4-bits of the result.
For rounding modes, see {ref}`narrow-fp-rounding-modes`.
The `relu` modifier clamps negative results to 0.
When `satfinite` is specified, if the absolute value of input (ignoring sign)
is greater than `MAX_NORM` of the specified destination format, then the
result is sign-preserved `MAX_NORM` of the destination format. Also, if the
input is `NaN`, then the result is the positive `MAX_NORM` of the
destination format.
For scale factor, see {ref}`scale-factor <scale-factor>`.
For more information, see [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt).
### Arithmetic Intrinsics
#### '`llvm.nvvm.fabs.*`' Intrinsic
##### Syntax:
```llvm
declare float @llvm.nvvm.fabs.f32(float %a)
declare double @llvm.nvvm.fabs.f64(double %a)
declare half @llvm.nvvm.fabs.f16(half %a)
declare <2 x half> @llvm.nvvm.fabs.v2f16(<2 x half> %a)
declare bfloat @llvm.nvvm.fabs.bf16(bfloat %a)
declare <2 x bfloat> @llvm.nvvm.fabs.v2bf16(<2 x bfloat> %a)
```
##### Overview:
The '`llvm.nvvm.fabs.*`' intrinsics return the absolute value of the operand.
##### Semantics:
Unlike, '`llvm.fabs.*`', these intrinsics do not perfectly preserve NaN
values. Instead, a NaN input yields an unspecified NaN output.
#### '`llvm.nvvm.fabs.ftz.*`' Intrinsic
##### Syntax:
```llvm
declare float @llvm.nvvm.fabs.ftz.f32(float %a)
declare half @llvm.nvvm.fabs.ftz.f16(half %a)
declare <2 x half> @llvm.nvvm.fabs.ftz.v2f16(<2 x half> %a)
```
##### Overview:
The '`llvm.nvvm.fabs.ftz.*`' intrinsics return the absolute value of the
operand, flushing subnormals to sign preserving zero.
##### Semantics:
Before the absolute value is taken, the input is flushed to sign preserving
zero if it is a subnormal. In addition, unlike '`llvm.fabs.*`', a NaN input
yields an unspecified NaN output.
#### '`llvm.nvvm.idp2a.[us].[us]`' Intrinsics
##### Syntax:
```llvm
declare i32 @llvm.nvvm.idp2a.s.s(i32 %a, i32 %b, i1 immarg %is.hi, i32 %c)
declare i32 @llvm.nvvm.idp2a.s.u(i32 %a, i32 %b, i1 immarg %is.hi, i32 %c)
declare i32 @llvm.nvvm.idp2a.u.s(i32 %a, i32 %b, i1 immarg %is.hi, i32 %c)
declare i32 @llvm.nvvm.idp2a.u.u(i32 %a, i32 %b, i1 immarg %is.hi, i32 %c)
```
##### Overview:
The '`llvm.nvvm.idp2a.[us].[us]`' intrinsics performs a 2-element vector dot
product followed by addition. They corresponds directly to the `dp2a` PTX
instruction.
##### Semantics:
The 32-bit value in `%a` is broken into 2 16-bit values which are extended to
32 bits. For the '`llvm.nvvm.idp2a.u.[us]`' variants zero-extension is used,
while for the '`llvm.nvvm.idp2a.s.[us]`' sign-extension is used. Two bytes are
selected from `%b`, if `%is.hi` is true, the most significant bytes are
selected, otherwise the least significant bytes are selected. These bytes are
then extended to 32-bits. For the '`llvm.nvvm.idp2a.[us].u`' variants
zero-extension is used, while for the '`llvm.nvvm.idp2a.[us].s`'
sign-extension is used. The dot product of these 2-element vectors is added to
`%c` to produce the return.
#### '`llvm.nvvm.idp4a.[us].[us]`' Intrinsics
##### Syntax:
```llvm
declare i32 @llvm.nvvm.idp4a.s.s(i32 %a, i32 %b, i32 %c)
declare i32 @llvm.nvvm.idp4a.s.u(i32 %a, i32 %b, i32 %c)
declare i32 @llvm.nvvm.idp4a.u.s(i32 %a, i32 %b, i32 %c)
declare i32 @llvm.nvvm.idp4a.u.u(i32 %a, i32 %b, i32 %c)
```
##### Overview:
The '`llvm.nvvm.idp4a.[us].[us]`' intrinsics perform a 4-element vector dot
product followed by addition. They corresponds directly to the `dp4a` PTX
instruction.
##### Semantics:
Each of the 4 bytes in both `%a` and `%b` are extended to 32-bit integers
forming 2 `<4 x i32>`. For `%a`, zero-extension is used in the
'`llvm.nvvm.idp4a.u.[us]`' variants, while sign-extension is used with
'`llvm.nvvm.idp4a.s.[us]`' variants. Similarly, for `%b`, zero-extension is
used in the '`llvm.nvvm.idp4a.[us].u`' variants, while sign-extension is used
with '`llvm.nvvm.idp4a.[us].s`' variants. The dot product of these 4-element
vectors is added to `%c` to produce the return.
#### '`llvm.nvvm.add.*`' Half-precision Intrinsics
##### Syntax:
```llvm
declare half @llvm.nvvm.add.rn.sat.f16(half %a, half %b)
declare <2 x half> @llvm.nvvm.add.rn.sat.v2f16(<2 x half> %a, <2 x half> %b)
declare half @llvm.nvvm.add.rn.ftz.sat.f16(half %a, half %b)
declare <2 x half> @llvm.nvvm.add.rn.ftz.sat.v2f16(<2 x half> %a, <2 x half> %b)
```
##### Overview:
The '`llvm.nvvm.add.*`' intrinsics perform an addition operation with the
specified rounding mode and modifiers.
##### Semantics:
The '`.sat`' modifier performs a saturating addition where the result is
clamped to `[0.0, 1.0]` and `NaN` results are flushed to `+0.0f`.
The '`.ftz`' modifier flushes subnormal inputs and results to sign-preserving
zero.
#### '`llvm.nvvm.mul.*`' Half-precision Intrinsics
##### Syntax:
```llvm
declare half @llvm.nvvm.mul.rn.sat.f16(half %a, half %b)
declare <2 x half> @llvm.nvvm.mul.rn.sat.v2f16(<2 x half> %a, <2 x half> %b)
declare half @llvm.nvvm.mul.rn.ftz.sat.f16(half %a, half %b)
declare <2 x half> @llvm.nvvm.mul.rn.ftz.sat.v2f16(<2 x half> %a, <2 x half> %b)
```
##### Overview:
The '`llvm.nvvm.mul.*`' intrinsics perform a multiplication operation with
the specified rounding mode and modifiers.
##### Semantics:
The '`.sat`' modifier performs a saturating multiplication where the result is
clamped to `[0.0, 1.0]` and `NaN` results are flushed to `+0.0f`.
The '`.ftz`' modifier flushes subnormal inputs and results to sign-preserving
zero.
#### '`llvm.nvvm.fma.*`' Half-precision Intrinsics
##### Syntax:
```llvm
declare half @llvm.nvvm.fma.rn{.ftz}.f16(half %a, half %b, half %c)
declare <2 x half> @llvm.nvvm.fma.rn{.ftz}.f16x2(<2 x half> %a, <2 x half> %b, <2 x half> %c)
declare bfloat @llvm.nvvm.fma.rn.bf16(bfloat %a, bfloat %b, bfloat %c)
declare <2 x bfloat> @llvm.nvvm.fma.rn.bf16x2(<2 x bfloat> %a, <2 x bfloat> %b, <2 x bfloat> %c)
declare half @llvm.nvvm.fma.rn{.ftz}.sat.f16(half %a, half %b, half %c)
declare <2 x half> @llvm.nvvm.fma.rn{.ftz}.sat.f16x2(<2 x half> %a, <2 x half> %b, <2 x half> %c)
declare half @llvm.nvvm.fma.rn{.ftz}.relu.f16(half %a, half %b, half %c)
declare <2 x half> @llvm.nvvm.fma.rn{.ftz}.relu.f16x2(<2 x half> %a, <2 x half> %b, <2 x half> %c)
declare bfloat @llvm.nvvm.fma.rn.relu.bf16(bfloat %a, bfloat %b, bfloat %c)
declare <2 x bfloat> @llvm.nvvm.fma.rn.relu.bf16x2(<2 x bfloat> %a, <2 x bfloat> %b, <2 x bfloat> %c)
declare half @llvm.nvvm.fma.rn.oob{.relu}.f16(half %a, half %b, half %c)
declare <2 x half> @llvm.nvvm.fma.rn.oob{.relu}.v2f16(<2 x half> %a, <2 x half> %b, <2 x half> %c)
declare bfloat @llvm.nvvm.fma.rn.oob{.relu}.bf16(bfloat %a, bfloat %b, bfloat %c)
declare <2 x bfloat> @llvm.nvvm.fma.rn.oob{.relu}.v2bf16(<2 x bfloat> %a, <2 x bfloat> %b, <2 x bfloat> %c)
```
##### Overview:
The '`llvm.nvvm.fma.*`' intrinsics perform a fused multiply-add with no loss
of precision in the intermediate product and addition.
##### Semantics:
The '`.sat`' modifier performs a saturating operation where the result is
clamped to `[0.0, 1.0]` and `NaN` results are flushed to `+0.0f`.
The '`.ftz`' modifier flushes subnormal inputs and results to sign-preserving
zero.
The '`.relu`' modifier clamps the result to `0` if negative and `NaN`
results are flushed to canonical `NaN`.
The '`.oob`' modifier clamps the result to `0` if either of the operands is
an `OOB NaN` (defined under [Tensors](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensors)) value.
### Bit Manipulation Intrinsics
#### '`llvm.nvvm.fshl.clamp.*`' Intrinsic
##### Syntax:
```llvm
declare i32 @llvm.nvvm.fshl.clamp.i32(i32 %hi, i32 %lo, i32 %n)
```
##### Overview:
The '`llvm.nvvm.fshl.clamp`' family of intrinsics performs a clamped funnel
shift left. These intrinsics are very similar to '`llvm.fshl`', except the
shift amount is clamped at the integer width (instead of modulo it). Currently,
only `i32` is supported.
##### Semantics:
The '`llvm.nvvm.fshl.clamp`' family of intrinsic functions performs a clamped
funnel shift left: the first two values are concatenated as { %hi : %lo } (%hi
is the most significant bits of the wide value), the combined value is shifted
left, and the most significant bits are extracted to produce a result that is
the same size as the original arguments. The shift amount is the minimum of the
value of %n and the bit width of the integer type.
#### '`llvm.nvvm.fshr.clamp.*`' Intrinsic
##### Syntax:
```llvm
declare i32 @llvm.nvvm.fshr.clamp.i32(i32 %hi, i32 %lo, i32 %n)
```
##### Overview:
The '`llvm.nvvm.fshr.clamp`' family of intrinsics perform a clamped funnel
shift right. These intrinsics are very similar to '`llvm.fshr`', except the
shift amount is clamped at the integer width (instead of modulo it). Currently,
only `i32` is supported.
##### Semantics:
The '`llvm.nvvm.fshr.clamp`' family of intrinsic functions performs a clamped
funnel shift right: the first two values are concatenated as { %hi : %lo } (%hi
is the most significant bits of the wide value), the combined value is shifted
right, and the least significant bits are extracted to produce a result that is
the same size as the original arguments. The shift amount is the minimum of the
value of %n and the bit width of the integer type.
#### '`llvm.nvvm.flo.u.*`' Intrinsic
##### Syntax:
```llvm
declare i32 @llvm.nvvm.flo.u.i32(i32 %a, i1 %shiftamt)
declare i32 @llvm.nvvm.flo.u.i64(i64 %a, i1 %shiftamt)
```
##### Overview:
The '`llvm.nvvm.flo.u`' family of intrinsics identifies the bit position of
the leading one, returning either it's offset from the most or least significant
bit.
##### Semantics:
The '`llvm.nvvm.flo.u`' family of intrinsics returns the bit position of the
most significant 1. If %shiftamt is true, The result is the shift amount needed
to left-shift the found bit into the most-significant bit position, otherwise
the result is the shift amount needed to right-shift the found bit into the
least-significant bit position. 0xffffffff is returned if no 1 bit is found.
#### '`llvm.nvvm.flo.s.*`' Intrinsic
##### Syntax:
```llvm
declare i32 @llvm.nvvm.flo.s.i32(i32 %a, i1 %shiftamt)
declare i32 @llvm.nvvm.flo.s.i64(i64 %a, i1 %shiftamt)
```
##### Overview:
The '`llvm.nvvm.flo.s`' family of intrinsics identifies the bit position of
the leading non-sign bit, returning either it's offset from the most or least
significant bit.
##### Semantics:
The '`llvm.nvvm.flo.s`' family of intrinsics returns the bit position of the
most significant 0 for negative inputs and the most significant 1 for
non-negative inputs. If %shiftamt is true, The result is the shift amount needed
to left-shift the found bit into the most-significant bit position, otherwise
the result is the shift amount needed to right-shift the found bit into the
least-significant bit position. 0xffffffff is returned if no 1 bit is found.
#### '`llvm.nvvm.{zext,sext}.{wrap,clamp}`' Intrinsics
##### Syntax:
```llvm
declare i32 @llvm.nvvm.zext.wrap(i32 %a, i32 %b)
declare i32 @llvm.nvvm.zext.clamp(i32 %a, i32 %b)
declare i32 @llvm.nvvm.sext.wrap(i32 %a, i32 %b)
declare i32 @llvm.nvvm.sext.clamp(i32 %a, i32 %b)
```
##### Overview:
The '`llvm.nvvm.{zext,sext}.{wrap,clamp}`' family of intrinsics extracts the
low bits of the input value, and zero- or sign-extends them back to the original
width.
##### Semantics:
The '`llvm.nvvm.{zext,sext}.{wrap,clamp}`' family of intrinsics returns
extension of N lowest bits of operand %a. For the '`wrap`' variants, N is the
value of operand %b modulo 32. For the '`clamp`' variants, N is the value of
operand %b clamped to the range [0, 32]. The N lowest bits are then
zero-extended the case of the '`zext`' variants, or sign-extended the case of
the '`sext`' variants. If N is 0, the result is 0.
#### '`llvm.nvvm.bmsk.{wrap,clamp}`' Intrinsic
##### Syntax:
```llvm
declare i32 @llvm.nvvm.bmsk.wrap(i32 %a, i32 %b)
declare i32 @llvm.nvvm.bmsk.clamp(i32 %a, i32 %b)
```
##### Overview:
The '`llvm.nvvm.bmsk.{wrap,clamp}`' family of intrinsics creates a bit mask
given a starting bit position and a bit width.
##### Semantics:
The '`llvm.nvvm.bmsk.{wrap,clamp}`' family of intrinsics returns a value with
all bits set to 0 except for %b bits starting at bit position %a. For the
'`wrap`' variants, the values of %a and %b modulo 32 are used. For the
'`clamp`' variants, the values of %a and %b are clamped to the range [0, 32],
which in practice is equivalent to using them as is.
#### '`llvm.nvvm.prmt`' Intrinsic
##### Syntax:
```llvm
declare i32 @llvm.nvvm.prmt(i32 %lo, i32 %hi, i32 %selector)
```
##### Overview:
The '`llvm.nvvm.prmt`' constructs a permutation of the bytes of the first two
operands, selecting based on the third operand.
##### Semantics:
The bytes in the first two source operands are numbered from 0 to 7:
`{%hi, %lo} = {{b7, b6, b5, b4}, {b3, b2, b1, b0}}`. For each byte in the
target register, a 4-bit selection value is defined.
The 3 lsbs of the selection value specify which of the 8 source bytes should be
moved into the target position. The msb defines if the byte value should be
copied, or if the sign (msb of the byte) should be replicated over all 8 bits
of the target position (sign extend of the byte value); msb=0 means copy the
literal value; msb=1 means replicate the sign.
These 4-bit selection values are pulled from the lower 16-bits of the %selector
operand, with the least significant selection value corresponding to the least
significant byte of the destination.
#### '`llvm.nvvm.prmt.*`' Intrinsics
##### Syntax:
```llvm
declare i32 @llvm.nvvm.prmt.f4e(i32 %lo, i32 %hi, i32 %selector)
declare i32 @llvm.nvvm.prmt.b4e(i32 %lo, i32 %hi, i32 %selector)
declare i32 @llvm.nvvm.prmt.rc8(i32 %lo, i32 %selector)
declare i32 @llvm.nvvm.prmt.ecl(i32 %lo, i32 %selector)
declare i32 @llvm.nvvm.prmt.ecr(i32 %lo, i32 %selector)
declare i32 @llvm.nvvm.prmt.rc16(i32 %lo, i32 %selector)
```
##### Overview:
The '`llvm.nvvm.prmt.*`' family of intrinsics constructs a permutation of the
bytes of the first one or two operands, selecting based on the 2 least
significant bits of the final operand.
##### Semantics:
As with the generic '`llvm.nvvm.prmt`' intrinsic, the bytes in the first one
or two source operands are numbered. The first source operand (%lo) is numbered
{b3, b2, b1, b0}, in the case of the '`f4e`' and '`b4e`' variants, the
second source operand (%hi) is numbered {b7, b6, b5, b4}.
Depending on the 2 least significant bits of the %selector operand, the result
of the permutation is defined as follows:
```{eval-rst}
+------------+----------------+--------------+
| Mode | %selector[1:0] | Output |
+------------+----------------+--------------+
| ``f4e`` | 0 | {3, 2, 1, 0} |
| +----------------+--------------+
| | 1 | {4, 3, 2, 1} |
| +----------------+--------------+
| | 2 | {5, 4, 3, 2} |
| +----------------+--------------+
| | 3 | {6, 5, 4, 3} |
+------------+----------------+--------------+
| ``b4e`` | 0 | {5, 6, 7, 0} |
| +----------------+--------------+
| | 1 | {6, 7, 0, 1} |
| +----------------+--------------+
| | 2 | {7, 0, 1, 2} |
| +----------------+--------------+
| | 3 | {0, 1, 2, 3} |
+------------+----------------+--------------+
| ``rc8`` | 0 | {0, 0, 0, 0} |
| +----------------+--------------+
| | 1 | {1, 1, 1, 1} |
| +----------------+--------------+
| | 2 | {2, 2, 2, 2} |
| +----------------+--------------+
| | 3 | {3, 3, 3, 3} |
+------------+----------------+--------------+
| ``ecl`` | 0 | {3, 2, 1, 0} |
| +----------------+--------------+
| | 1 | {3, 2, 1, 1} |
| +----------------+--------------+
| | 2 | {3, 2, 2, 2} |
| +----------------+--------------+
| | 3 | {3, 3, 3, 3} |
+------------+----------------+--------------+
| ``ecr`` | 0 | {0, 0, 0, 0} |
| +----------------+--------------+
| | 1 | {1, 1, 1, 0} |
| +----------------+--------------+
| | 2 | {2, 2, 1, 0} |
| +----------------+--------------+
| | 3 | {3, 2, 1, 0} |
+------------+----------------+--------------+
| ``rc16`` | 0 | {1, 0, 1, 0} |
| +----------------+--------------+
| | 1 | {3, 2, 3, 2} |
| +----------------+--------------+
| | 2 | {1, 0, 1, 0} |
| +----------------+--------------+
| | 3 | {3, 2, 3, 2} |
+------------+----------------+--------------+
```
### TMA family of Intrinsics
#### '`llvm.nvvm.cp.async.bulk.global.to.shared.cluster`'
##### Syntax:
```llvm
declare void @llvm.nvvm.cp.async.bulk.global.to.shared.cluster(ptr addrspace(7) %dst, ptr addrspace(3) %mbar, ptr addrspace(1) %src, i32 %size, i16 %mc, i64 %ch, i1 %flag_mc, i1 %flag_ch)
```
##### Overview:
The '`@llvm.nvvm.cp.async.bulk.global.to.shared.cluster`' intrinsic
corresponds to the `cp.async.bulk.shared::cluster.global.*` family of PTX
instructions. These instructions initiate an asynchronous copy of bulk data from
global memory to shared::cluster memory. The 32-bit operand `%size` specifies
the amount of memory to be copied and it must be a multiple of 16.
- The last two arguments to these intrinsics are boolean flags indicating
support for cache_hint and/or multicast modifiers. These flag arguments must
be compile-time constants. The backend looks through these flags and lowers
the intrinsics appropriately.
- The Nth argument (denoted by `i1 %flag_ch`) when set, indicates a valid
cache_hint (`i64 %ch`) and generates the `.L2::cache_hint` variant of the
PTX instruction.
- The [N-1]th argument (denoted by `i1 %flag_mc`) when set, indicates the
presence of a multicast mask (`i16 %mc`) and generates the PTX instruction
with the `.multicast::cluster` modifier.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#data-movement-and-conversion-instructions-cp-async-bulk).
#### '`llvm.nvvm.cp.async.bulk.global.to.shared.cta`'
##### Syntax:
```llvm
declare void @llvm.nvvm.cp.async.bulk.global.to.shared.cta(ptr addrspace(3) %dst, ptr addrspace(3) %mbar, ptr addrspace(1) %src, i32 %size, i64 %ch, i1 %flag_ch)
```
##### Overview:
The '`@llvm.nvvm.cp.async.bulk.global.to.shared.cta`' intrinsic corresponds to
the `cp.async.bulk.shared::cta.global.*` family of PTX instructions. These
instructions initiate an asynchronous copy of bulk data from global memory to
shared::cta memory. The 32-bit operand `%size` specifies the amount of memory
to be copied and it must be a multiple of 16. The last argument (denoted by
`i1 %flag_ch`) is a compile-time constant. When set, it indicates a valid
cache_hint (`i64 %ch`) and generates the `.L2::cache_hint` variant of the
PTX instruction.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#data-movement-and-conversion-instructions-cp-async-bulk).
#### '`llvm.nvvm.cp.async.bulk.shared.cta.to.global`'
##### Syntax:
```llvm
declare void @llvm.nvvm.cp.async.bulk.shared.cta.to.global(ptr addrspace(1) %dst, ptr addrspace(3) %src, i32 %size, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.shared.cta.to.global.bytemask(..., i32 %size, i64 %ch, i1 %flag_ch, i16 %mask)
```
##### Overview:
The '`@llvm.nvvm.cp.async.bulk.shared.cta.to.global`' intrinsic corresponds to
the `cp.async.bulk.global.shared::cta.*` set of PTX instructions. These
instructions initiate an asynchronous copy from shared::cta to global memory.
The 32-bit operand `%size` specifies the amount of memory to be copied
(in bytes) and it must be a multiple of 16. For the `.bytemask` variant, the
16-bit wide mask operand specifies whether the i-th byte of each 16-byte wide
chunk of source data is copied to the destination.
- The `i1 %flag_ch` argument to these intrinsics is a boolean flag indicating
support for cache_hint. This flag argument must be a compile-time constant.
When set, it indicates a valid cache_hint (`i64 %ch`) and generates the
`.L2::cache_hint` variant of the PTX instruction.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#data-movement-and-conversion-instructions-cp-async-bulk).
#### '`llvm.nvvm.cp.async.bulk.shared.cta.to.cluster`'
##### Syntax:
```llvm
declare void @llvm.nvvm.cp.async.bulk.shared.cta.to.cluster(ptr addrspace(7) %dst, ptr addrspace(3) %mbar, ptr addrspace(3) %src, i32 %size)
```
##### Overview:
The '`@llvm.nvvm.cp.async.bulk.shared.cta.to.cluster`' intrinsic corresponds
to the `cp.async.bulk.shared::cluster.shared::cta.*` PTX instruction. This
instruction initiates an asynchronous copy from shared::cta to shared::cluster
memory. The destination has to be in the shared memory of a different CTA within
the cluster. The 32-bit operand `%size` specifies the amount of memory to be
copied and it must be a multiple of 16.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#data-movement-and-conversion-instructions-cp-async-bulk).
#### '`llvm.nvvm.cp.async.bulk.prefetch.L2`'
##### Syntax:
```llvm
declare void @llvm.nvvm.cp.async.bulk.prefetch.L2(ptr addrspace(1) %src, i32 %size, i64 %ch, i1 %flag_ch)
```
##### Overview:
The '`@llvm.nvvm.cp.async.bulk.prefetch.L2`' intrinsic corresponds to the
`cp.async.bulk.prefetch.L2.*` family of PTX instructions. These instructions
initiate an asynchronous prefetch of bulk data from global memory to the L2
cache. The 32-bit operand `%size` specifies the amount of memory to be
prefetched in terms of bytes and it must be a multiple of 16.
- The last argument to these intrinsics is boolean flag indicating support for
cache_hint. These flag argument must be compile-time constant. When set, it
indicates a valid cache_hint (`i64 %ch`) and generates the
`.L2::cache_hint` variant of the PTX instruction.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-prefetch).
#### '`llvm.nvvm.prefetch.*`'
##### Syntax:
```llvm
declare void @llvm.nvvm.prefetch.global.L1(ptr addrspace(1) %global_ptr)
declare void @llvm.nvvm.prefetch.global.L2(ptr addrspace(1) %global_ptr)
declare void @llvm.nvvm.prefetch.local.L1(ptr addrspace(5) %local_ptr)
declare void @llvm.nvvm.prefetch.local.L2(ptr addrspace(5) %local_ptr)
declare void @llvm.nvvm.prefetch.L1(ptr %ptr)
declare void @llvm.nvvm.prefetch.L2(ptr %ptr)
declare void @llvm.nvvm.prefetch.tensormap.p0(ptr %ptr)
declare void @llvm.nvvm.prefetch.tensormap.p4(ptr addrspace(4) %const_ptr)
declare void @llvm.nvvm.prefetch.tensormap.p101(ptr addrspace(101) %param_ptr)
declare void @llvm.nvvm.prefetch.global.L2.evict.normal(ptr addrspace(1) %global_ptr)
declare void @llvm.nvvm.prefetch.global.L2.evict.last(ptr addrspace(1) %global_ptr)
declare void @llvm.nvvm.prefetchu.L1(ptr %ptr)
```
##### Overview:
The '`@llvm.nvvm.prefetch.*`' and '`@llvm.nvvm.prefetchu.*`' intrinsic
correspond to the '`prefetch.*`;' and '`prefetchu.*`' family of PTX
instructions. The '`prefetch.*`' instructions bring the cache line containing
the specified address in global or local memory address space into the specified
cache level (L1 or L2). If the '`.tensormap`' qualifier is specified then the
prefetch instruction brings the cache line containing the specified address in
the '`.const`' or '`.param memory`' state space for subsequent use by the
'`cp.async.bulk.tensor`' instruction. The '`prefetchu.*`' instruction brings
the cache line containing the specified generic address into the specified
uniform cache level. If no address space is specified, it is assumed to be
generic address. The intrinsic uses and eviction priority which can be accessed
by the '`.level::eviction_priority`' modifier.
- A prefetch to a shared memory location performs no operation.
- A prefetch into the uniform cache requires a generic address, and no operation
occurs if the address maps to a const, local, or shared memory location.
For more information, refer to the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-prefetch-prefetchu).
#### '`llvm.nvvm.applypriority.*`'
##### Syntax:
```llvm
declare void @llvm.nvvm.applypriority.global.L2.evict.normal(ptr addrspace(1) %global_ptr, i64 %size)
declare void @llvm.nvvm.applypriority.L2.evict.normal(ptr %ptr, i64 %size)
```
##### Overview:
The '`@llvm.nvvm.applypriority.*`' applies the cache eviction priority
specified by the .level::eviction_priority qualifier to the address range
\[a..a+size) in the specified cache level. If no state space is specified then
generic addressing is used. If the specified address does not fall within the
address window of .global state space then the behavior is undefined. The
operand size is an integer constant that specifies the amount of data, in bytes,
in the specified cache level on which the priority is to be applied. The only
supported value for the size operand is 128.
For more information, refer to the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-applypriority).
#### `llvm.nvvm.discard.*`'
##### Syntax:
```llvm
declare void @llvm.nvvm.discard.global.L2(ptr addrspace(1) %global_ptr, i64 immarg)
declare void @llvm.nvvm.discard.L2(ptr %ptr, i64 immarg)
```
##### Overview:
The *effects* of the `@llvm.nvvm.discard.L2*` intrinsics are those of a
non-atomic non-volatile `llvm.memset` that writes `undef` to the destination
address range `[%ptr, %ptr + immarg)`. The `%ptr` must be aligned by 128
bytes. Subsequent reads from the address range may read `undef` until the
memory is overwritten with a different value. These operations *hint* the
implementation that data in the L2 cache can be destructively discarded without
writing it back to memory. The operand `immarg` is an integer constant that
specifies the length in bytes of the address range `[%ptr, %ptr + immarg)` to
write `undef` into. The only supported value for the `immarg` operand is
`128`. If generic addressing is used and the specified address does not fall
within the address window of global memory (`addrspace(1)`) the behavior is
undefined.
```llvm
call void @llvm.nvvm.discard.L2(ptr %p, i64 128) ;; writes `undef` to [p, p+128)
%a = load i64, ptr %p. ;; loads 8 bytes containing undef
%b = load i64, ptr %p ;; loads 8 bytes containing undef
;; comparing %a and %b compares `undef` values!
%fa = freeze i64 %a ;; freezes undef to stable bit-pattern
%fb = freeze i64 %b ;; freezes undef to stable bit-pattern
;; %fa may compare different to %fb!
```
For more information, refer to the [CUDA C++ discard documentation](https://nvidia.github.io/cccl/libcudacxx/extended_api/memory_access_properties/discard_memory.html)
and to the [PTX ISA discard documentation](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-discard) .
#### '`llvm.nvvm.cp.async.bulk.tensor.g2s.tile.[1-5]d`'
##### Syntax:
```llvm
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.tile.1d(ptr addrspace(7) %dst, ptr addrspace(3) %bar, ptr %tensor_map, i32 %d0, i16 %mc, i64 %ch, i1 %flag_mc, i1 %flag_ch, i32 %flag_cta_group)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d(..., i32 %d0, i32 %d1, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.tile.3d(..., i32 %d0, i32 %d1, i32 %d2, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.tile.4d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.tile.5d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.tile.gather4.2d(ptr addrspace(7) %dst, ptr addrspace(3) %bar, ptr %tensor_map, i32 %x0, i32 %y0, i32 %y1, i32 %y2, i32 %y3, i16 %mc, i64 %ch, i1 %flag_mc, i1 %flag_ch, i32 %flag_cta_group)
```
##### Overview:
The '`@llvm.nvvm.cp.async.bulk.tensor.g2s.tile.[1-5]d`' intrinsics correspond
to the `cp.async.bulk.tensor.[1-5]d.*` set of PTX instructions. These
instructions initiate an asynchronous copy of tensor data from global memory to
shared::cluster memory (indicated by the `g2s` prefix) in `tile` mode. In
tile mode, the multi-dimensional layout of the source tensor is preserved at the
destination. The dimension of the tensor data ranges from 1d to 5d with the
coordinates specified by the `i32 %d0 ... i32 %d4` arguments. In
`tile.gather4` mode, four rows in a 2D tensor are combined to form a single 2D
destination tensor. The first coordinate `i32 %x0` denotes the column index
followed by four coordinates indicating the four row-indices. So, this mode
takes a total of 5 coordinates as input arguments. For more information on
`gather4` mode, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensor-tiled-scatter4-gather4-modes).
- The last three arguments to these intrinsics are flags indicating support for
multicast, cache_hint and cta_group::1/2 modifiers. These flag arguments must
be compile-time constants. The backend looks through these flags and lowers
the intrinsics appropriately.
- The argument denoted by `i1 %flag_ch` when set, indicates a valid cache_hint
(`i64 %ch`) and generates the `.L2::cache_hint` variant of the PTX
instruction.
- The argument denoted by `i1 %flag_mc` when set, indicates the presence of a
multicast mask (`i16 %mc`) and generates the PTX instruction with the
`.multicast::cluster` modifier.
- The argument denoted by `i32 %flag_cta_group` takes values within the range
\[0, 3) i.e. {0,1,2}. When the value of `%flag_cta_group` is not within the
range, it may raise an error from the Verifier. The default value is '0' with
no cta_group modifier in the instruction. The values of '1' and '2' lower to
`cta_group::1` and `cta_group::2` variants of the PTX instruction
respectively.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#data-movement-and-conversion-instructions-cp-async-bulk-tensor).
#### '`llvm.nvvm.cp.async.bulk.tensor.g2s.im2col.[3-5]d`'
##### Syntax:
```llvm
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.im2col.3d(ptr addrspace(7) %dst, ptr addrspace(3) %bar, ptr %tensor_map, i32 %d0, i32 %d1, i32 %d2, i16 %im2col0, i16 %mc, i64 %ch, i1 %flag_mc, i1 %flag_ch, i32 %flag_cta_group)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.im2col.4d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, i16 %im2col0, i16 %im2col1, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.im2col.5d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i16 %im2col0, i16 %im2col1, i16 %im2col2, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.im2col.w.3d(ptr addrspace(7) %dst, ptr addrspace(3) %bar, ptr %tensor_map, i32 %d0, i32 %d1, i32 %d2, i16 %wHalo, i16 %wOffset, i16 %mc, i64 %ch, i1 %flag_mc, i1 %flag_ch, i32 %flag_cta_group)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.im2col.w.4d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.im2col.w.5d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.im2col.w.128.3d(ptr addrspace(7) %dst, ptr addrspace(3) %bar, ptr %tensor_map, i32 %d0, i32 %d1, i32 %d2, i16 %wHalo, i16 %wOffset, i16 %mc, i64 %ch, i1 %flag_mc, i1 %flag_ch, i32 %flag_cta_group)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.im2col.w.128.4d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.im2col.w.128.5d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, ...)
```
##### Overview:
The '`@llvm.nvvm.cp.async.bulk.tensor.g2s.im2col.[3-5]d`' intrinsics
correspond to the `cp.async.bulk.tensor.[1-5]d.*` set of PTX instructions.
These instructions initiate an asynchronous copy of tensor data from global
memory to shared::cluster memory (indicated by the `g2s` prefix) in `im2col`
mode. In im2col mode, some dimensions of the source tensor are unrolled into a
single dimensional column at the destination. In this mode, the tensor has to be
at least three-dimensional. Along with the tensor coordinates, im2col offsets
are also specified (denoted by `i16 im2col0...i16 %im2col2`). For the
`im2col` mode, the number of offsets is two less than the number of dimensions
of the tensor operation. For the `im2col.w` and `im2col.w.128` mode, the
number of offsets is always 2, denoted by `i16 %wHalo` and `i16 %wOffset`
arguments. For more information on `im2col.w` and `im2col.w.128` modes,
refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensor-im2col-w-w128-modes).
The last three arguments to these intrinsics are flags, with the same
functionality as described in the `tile` mode intrinsics above.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#data-movement-and-conversion-instructions-cp-async-bulk-tensor).
#### '`llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.[1-5]d`'
##### Syntax:
```llvm
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.1d(ptr addrspace(3) %dst, ptr addrspace(3) %bar, ptr %tensor_map, i32 %d0, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.2d(..., i32 %d0, i32 %d1, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.3d(..., i32 %d0, i32 %d1, i32 %d2, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.4d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.5d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.gather4.2d(ptr addrspace(3) %dst, ptr addrspace(3) %bar, ptr %tensor_map, i32 %x0, i32 %y0, i32 %y1, i32 %y2, i32 %y3, i64 %ch, i1 %flag_ch)
```
##### Overview:
The '`@llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.[1-5]d`' intrinsics
correspond to the `cp.async.bulk.tensor.[1-5]d.shared::cta.global.*` set of
PTX instructions. These instructions initiate an asynchronous copy of tensor
data from global memory to shared::cta memory in `tile` mode. In tile mode,
the multi-dimensional layout of the source tensor is preserved at the
destination. The dimension of the tensor data ranges from 1d to 5d with the
coordinates specified by the `i32 %d0 ... i32 %d4` arguments. In
`tile.gather4` mode, four rows in a 2D tensor are combined to form a single 2D
destination tensor. The first coordinate `i32 %x0` denotes the column index
followed by four coordinates indicating the four row-indices. So, this mode
takes a total of 5 coordinates as input arguments. For more information on
`gather4` mode, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensor-tiled-scatter4-gather4-modes).
- The last argument to these intrinsics is a boolean flag indicating support for
cache_hint. This flag argument must be a compile-time constant. When set, it
indicates a valid cache_hint (`i64 %ch`) and generates the
`.L2::cache_hint` variant of the PTX instruction.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#data-movement-and-conversion-instructions-cp-async-bulk-tensor).
#### '`llvm.nvvm.cp.async.bulk.tensor.g2s.cta.im2col.[3-5]d`'
##### Syntax:
```llvm
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.im2col.3d(ptr addrspace(3) %dst, ptr addrspace(3) %bar, ptr %tensor_map, i32 %d0, i32 %d1, i32 %d2, i16 %im2col0, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.im2col.4d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, i16 %im2col0, i16 %im2col1, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.im2col.5d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i16 %im2col0, i16 %im2col1, i16 %im2col2, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.im2col.w.3d(ptr addrspace(3) %dst, ptr addrspace(3) %bar, ptr %tensor_map, i32 %d0, i32 %d1, i32 %d2, i16 %wHalo, i16 %wOffset, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.im2col.w.4d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.im2col.w.5d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.im2col.w.128.3d(ptr addrspace(3) %dst, ptr addrspace(3) %bar, ptr %tensor_map, i32 %d0, i32 %d1, i32 %d2, i16 %wHalo, i16 %wOffset, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.im2col.w.128.4d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.im2col.w.128.5d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, ...)
```
##### Overview:
The '`@llvm.nvvm.cp.async.bulk.tensor.g2s.cta.im2col.[3-5]d`' intrinsics
correspond to the `cp.async.bulk.tensor.[1-5]d.shared::cta.global.*` set of
PTX instructions. These instructions initiate an asynchronous copy of tensor
data from global memory to shared::cta memory in `im2col` mode. In im2col
mode, some dimensions of the source tensor are unrolled into a single
dimensional column at the destination. In this mode, the tensor has to be at
least three-dimensional. Along with the tensor coordinates, im2col offsets are
also specified (denoted by `i16 im2col0...i16 %im2col2`). For the `im2col`
mode, the number of offsets is two less than the number of dimensions of the
tensor operation. For the `im2col.w` and `im2col.w.128` mode, the number of
offsets is always 2, denoted by `i16 %wHalo` and `i16 %wOffset` arguments.
For more information on `im2col.w` and `im2col.w.128` modes, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensor-im2col-w-w128-modes).
- The last argument to these intrinsics is a boolean flag indicating support
for cache_hint. This flag argument must be a compile-time constant. When set,
it indicates a valid cache_hint (`i64 %ch`) and generates the
`.L2::cache_hint` variant of the PTX instruction.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#data-movement-and-conversion-instructions-cp-async-bulk-tensor).
#### '`llvm.nvvm.cp.async.bulk.tensor.s2g.tile.[1-5]d`'
##### Syntax:
```llvm
declare void @llvm.nvvm.cp.async.bulk.tensor.s2g.tile.1d(ptr addrspace(3) %src, ptr %tensor_map, i32 %d0, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.tensor.s2g.tile.2d(..., i32 %d0, i32 %d1, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.s2g.tile.3d(..., i32 %d0, i32 %d1, i32 %d2, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.s2g.tile.4d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.s2g.tile.5d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.s2g.tile.scatter4.2d(ptr addrspace(3) %src, ptr %tensor_map, i32 %x0, i32 %y0, i32 %y1, i32 %y2, i32 %y3, i64 %ch, i1 %flag_ch)
```
##### Overview:
The '`@llvm.nvvm.cp.async.bulk.tensor.s2g.tile.[1-5]d`' intrinsics correspond
to the `cp.async.bulk.tensor.[1-5]d.*` set of PTX instructions. These
instructions initiate an asynchronous copy of tensor data from shared::cta to
global memory (indicated by the `s2g` prefix) in `tile` mode. The dimension
of the tensor data ranges from 1d to 5d with the coordinates specified by the
`i32 %d0 ... i32 %d4` arguments. In `tile.scatter4` mode, a single 2D source
tensor is divided into four rows in the 2D destination tensor. The first
coordinate `i32 %x0` denotes the column index followed by four coordinates
indicating the four row-indices. So, this mode takes a total of 5 coordinates as
input arguments. For more information on `scatter4` mode, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensor-tiled-scatter4-gather4-modes).
- The last argument to these intrinsics is a boolean flag indicating support for
cache_hint. This flag argument must be a compile-time constant. When set, it
indicates a valid cache_hint (`i64 %ch`) and generates the
`.L2::cache_hint` variant of the PTX instruction.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#data-movement-and-conversion-instructions-cp-async-bulk-tensor).
#### '`llvm.nvvm.cp.async.bulk.tensor.s2g.im2col.[3-5]d`'
##### Syntax:
```llvm
declare void @llvm.nvvm.cp.async.bulk.tensor.s2g.im2col.3d(ptr addrspace(3) %src, ptr %tensor_map, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.tensor.s2g.im2col.4d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.s2g.im2col.5d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, ...)
```
##### Overview:
The '`@llvm.nvvm.cp.async.bulk.tensor.s2g.im2col.[1-5]d`' intrinsics
correspond to the `cp.async.bulk.tensor.[1-5]d.*` set of PTX instructions.
These instructions initiate an asynchronous copy of tensor data from
shared::cta to global memory (indicated by the `s2g` prefix) in `im2col`
mode. In this mode, the tensor has to be at least three-dimensional. Unlike the
`g2s` variants, there are no im2col_offsets for these intrinsics. The last
argument to these intrinsics is a boolean flag, with the same functionality as
described in the `s2g.tile` mode intrinsics above.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#data-movement-and-conversion-instructions-cp-async-bulk-tensor).
#### '`llvm.nvvm.cp.async.bulk.tensor.prefetch.tile.[1-5]d`'
##### Syntax:
```llvm
declare void @llvm.nvvm.cp.async.bulk.tensor.prefetch.tile.1d(ptr %tensor_map, i32 %d0, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.tensor.prefetch.tile.2d(..., i32 %d0, i32 %d1, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.prefetch.tile.3d(..., i32 %d0, i32 %d1, i32 %d2, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.prefetch.tile.4d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.prefetch.tile.5d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.prefetch.tile.gather4.2d(ptr %tensor_map, i32 %x0, i32 %y0, i32 %y1, i32 %y2, i32 %y3, i64 %ch, i1 %flag_ch)
```
##### Overview:
The '`@llvm.nvvm.cp.async.bulk.tensor.prefetch.tile.[1-5]d`' intrinsics
correspond to the `cp.async.bulk.prefetch.tensor.[1-5]d.L2.global*` set of
PTX instructions. These instructions initiate an asynchronous prefetch of tensor
data from global memory to the L2 cache. In tile mode, the multi-dimensional
layout of the source tensor is preserved at the destination. The dimension of
the tensor data ranges from 1d to 5d with the coordinates specified by the
`i32 %d0 ... i32 %d4` arguments.
In `tile.gather4` mode, four rows in the 2-dimnesional source tensor are
fetched to the L2 cache. The first coordinate `i32 %x0` denotes the column
index followed by four coordinates indicating the four row-indices. So, this
mode takes a total of 5 coordinates as input arguments. For more information
on `gather4` mode, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensor-tiled-scatter4-gather4-modes).
- The last argument to these intrinsics is a boolean flag indicating support
for cache_hint. This flag argument must be a compile-time constant. When set,
it indicates a valid cache_hint (`i64 %ch`) and generates the
`.L2::cache_hint` variant of the PTX instruction.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-prefetch-tensor).
#### '`llvm.nvvm.cp.async.bulk.tensor.prefetch.im2col.[3-5]d`'
##### Syntax:
```llvm
declare void @llvm.nvvm.cp.async.bulk.tensor.prefetch.im2col.3d(ptr %tensor_map, i32 %d0, i32 %d1, i32 %d2, i16 %im2col0, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.tensor.prefetch.im2col.4d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, i16 %im2col0, i16 %im2col1, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.prefetch.im2col.5d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i16 %im2col0, i16 %im2col1, i16 %im2col2, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.prefetch.im2col.w.3d(ptr %tensor_map, i32 %d0, i32 %d1, i32 %d2, i16 %wHalo, i16 %wOffset, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.tensor.prefetch.im2col.w.4d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.prefetch.im2col.w.5d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.prefetch.im2col.w.128.3d(ptr %tensor_map, i32 %d0, i32 %d1, i32 %d2, i16 %wHalo, i16 %wOffset, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.tensor.prefetch.im2col.w.128.4d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.prefetch.im2col.w.128.5d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, ...)
```
##### Overview:
The '`@llvm.nvvm.cp.async.bulk.tensor.prefetch.im2col.[3-5]d`' intrinsics
correspond to the `cp.async.bulk.prefetch.tensor.[1-5]d.L2.global*` set of PTX
instructions. These instructions initiate an asynchronous prefetch of tensor
data from global memory to the L2 cache. In im2col mode, some dimensions of the
source tensor are unrolled into a single dimensional column at the destination.
In this mode, the tensor has to be at least three-dimensional. Along with the
tensor coordinates, im2col offsets are also specified (denoted by
`i16 im2col0...i16 %im2col2`). For `im2col` mode, the number of offsets is
two less than the number of dimensions of the tensor operation. For the
`im2col.w` and `im2col.w.128` modes, the number of offsets is always 2,
denoted by `i16 %wHalo` and `i16 %wOffset` arguments. For more information
on `im2col.w` and `im2col.w.128` modes, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensor-im2col-w-w128-modes).
The last argument to these intrinsics is a boolean flag, with the same
functionality as described in the `tile` mode intrinsics above.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-prefetch-tensor).
#### '`llvm.nvvm.cp.async.bulk.tensor.reduce.[red_op].tile.[1-5]d`'
##### Syntax:
```llvm
declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.tile.1d(ptr addrspace(3) %src, ptr %tensor_map, i32 %d0, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.tile.1d(ptr addrspace(3) %src, ptr %tensor_map, i32 %d0, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.tile.1d(ptr addrspace(3) %src, ptr %tensor_map, i32 %d0, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.tile.1d(ptr addrspace(3) %src, ptr %tensor_map, i32 %d0, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.tile.1d(ptr addrspace(3) %src, ptr %tensor_map, i32 %d0, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.tile.1d(ptr addrspace(3) %src, ptr %tensor_map, i32 %d0, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.tile.1d(ptr addrspace(3) %src, ptr %tensor_map, i32 %d0, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.tile.1d(ptr addrspace(3) %src, ptr %tensor_map, i32 %d0, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.<red_op>.tile.2d(..., i32 %d0, i32 %d1, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.<red_op>.tile.3d(..., i32 %d0, i32 %d1, i32 %d2, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.<red_op>.tile.4d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.<red_op>.tile.5d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, ...)
```
##### Overview:
The '`@llvm.nvvm.cp.async.bulk.tensor.reduce.<red_op>.tile.[1-5]d`' intrinsics
correspond to the `cp.reduce.async.bulk.tensor.[1-5]d.*` set of PTX
instructions. These instructions initiate an asynchronous reduction operation of
tensor data in global memory with the tensor data in shared\{::cta} memory, using
`tile` mode. The dimension of the tensor data ranges from 1d to 5d with the
coordinates specified by the `i32 %d0 ... i32 %d4` arguments. The supported
reduction operations are {add, min, max, inc, dec, and, or, xor} as described in
the `tile.1d` intrinsics.
- The last argument to these intrinsics is a boolean flag indicating support for
cache_hint. This flag argument must be a compile-time constant. When set, it
indicates a valid cache_hint (`i64 %ch`) and generates the
`.L2::cache_hint` variant of the PTX instruction.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#data-movement-and-conversion-instructions-cp-reduce-async-bulk-tensor).
#### '`llvm.nvvm.cp.async.bulk.tensor.reduce.[red_op].im2col.[3-5]d`'
##### Syntax:
```llvm
declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.<red_op>.im2col.3d(ptr addrspace(3) %src, ptr %tensor_map, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 %flag_ch)
declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.<red_op>.im2col.4d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, ...)
declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.<red_op>.im2col.5d(..., i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, ...)
```
##### Overview:
The '`@llvm.nvvm.cp.async.bulk.tensor.reduce.<red_op>.im2col.[3-5]d`'
intrinsics correspond to the `cp.reduce.async.bulk.tensor.[3-5]d.*` set of PTX
instructions. These instructions initiate an asynchronous reduction operation of
tensor data in global memory with the tensor data in shared\{::cta} memory, using
`im2col` mode. In this mode, the tensor has to be at least three-dimensional.
The supported reduction operations supported are the same as the ones in the
tile mode. The last argument to these intrinsics is a boolean flag, with the
same functionality as described in the `tile` mode intrinsics above.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#data-movement-and-conversion-instructions-cp-reduce-async-bulk-tensor).
### Warp Group Intrinsics
#### '`llvm.nvvm.wgmma.fence.sync.aligned`'
##### Syntax:
```llvm
declare void @llvm.nvvm.wgmma.fence.sync.aligned()
```
##### Overview:
The '`@llvm.nvvm.wgmma.fence.sync.aligned`' intrinsic generates the
`wgmma.fence.sync.aligned` PTX instruction, which establishes an ordering
between prior accesses to any warpgroup registers and subsequent accesses to
the same registers by a `wgmma.mma_async` instruction.
The `wgmma.fence` instruction must be issued by all warps of the warpgroup in
the following locations:
- Before the first `wgmma.mma_async` operation in a warpgroup.
- Between a register access by a thread in the warpgroup and any
`wgmma.mma_async` instruction that accesses the same registers, except when
these are accumulator register accesses across multiple `wgmma.mma_async`
instructions of the same shape in which case an ordering guarantee is
provided by default.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#asynchronous-warpgroup-level-matrix-instructions-wgmma-fence).
#### '`llvm.nvvm.wgmma.commit_group.sync.aligned`'
##### Syntax:
```llvm
declare void @llvm.nvvm.wgmma.commit_group.sync.aligned()
```
##### Overview:
The '`@llvm.nvvm.wgmma.commit_group.sync.aligned`' intrinsic generates the
`wgmma.commit_group.sync.aligned` PTX instruction, which creates a new
wgmma-group per warpgroup and batches all prior `wgmma.mma_async`
instructions initiated by the executing warp but not committed to any
wgmma-group into the new wgmma-group. If there are no uncommitted `wgmma
mma_async` instructions then, `wgmma.commit_group` results in an empty
wgmma-group.
An executing thread can wait for the completion of all `wgmma.mma_async`
operations in a wgmma-group by using `wgmma.wait_group`.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#asynchronous-warpgroup-level-matrix-instructions-wgmma-commit-group).
#### '`llvm.nvvm.wgmma.wait_group.sync.aligned`'
##### Syntax:
```llvm
declare void @llvm.nvvm.wgmma.wait_group.sync.aligned(i64 immarg N)
```
##### Overview:
The '`@llvm.nvvm.wgmma.wait_group.sync.aligned`' intrinsic generates the
`wgmma.commit_group.sync.aligned N` PTX instruction, which will cause the
executing thread to wait until only `N` or fewer of the most recent
wgmma-groups are pending and all the prior wgmma-groups committed by the
executing threads are complete. For example, when `N` is 0, the executing
thread waits on all the prior wgmma-groups to complete. Operand `N` is an
integer constant.
Accessing the accumulator register or the input register containing the
fragments of matrix A of a `wgmma.mma_async` instruction without first
performing a `wgmma.wait_group` instruction that waits on a wgmma-group
including that `wgmma.mma_async` instruction is undefined behavior.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#asynchronous-warpgroup-level-matrix-instructions-wgmma-wait-group).
#### '`llvm.nvvm.griddepcontrol.*`'
##### Syntax:
```llvm
declare void @llvm.nvvm.griddepcontrol.launch_dependents()
declare void @llvm.nvvm.griddepcontrol.wait()
```
##### Overview:
The `griddepcontrol` intrinsics allows the dependent grids and prerequisite
grids as defined by the runtime, to control execution in the following way:
`griddepcontrol.launch_dependents` intrinsic signals that the dependents can
be scheduled, before the current grid completes. The intrinsic can be invoked by
multiple threads in the current CTA and repeated invocations of the intrinsic
will have no additional side effects past that of the first invocation.
`griddepcontrol.wait` intrinsic causes the executing thread to wait until all
prerequisite grids in flight have completed and all the memory operations from
the prerequisite grids are performed and made visible to the current grid.
For more information, refer
[PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#parallel-synchronization-and-communication-instructions-griddepcontrol).
### Tensormap Replace Intrinsics
These intrinsics modify the fields of the tensor-map object at `%addr` in
`tile` mode.
For more information, refer to the
[PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-tensormap-replace).
#### '`llvm.nvvm.tensormap.replace.global.address`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tensormap.replace.global.address.p1(ptr addrspace(1) %addr, i64 %new_value)
declare void @llvm.nvvm.tensormap.replace.global.address.p3(ptr addrspace(3) %addr, i64 %new_value)
```
##### Overview:
The '`@llvm.nvvm.tensormap.replace.global.address.*`' intrinsics replace the
`global_address` field of the tensor-map object with `%new_value`.
#### '`llvm.nvvm.tensormap.replace.rank`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tensormap.replace.rank.p1(ptr addrspace(1) %addr, i32 %new_value)
declare void @llvm.nvvm.tensormap.replace.rank.p3(ptr addrspace(3) %addr, i32 %new_value)
```
##### Overview:
The '`@llvm.nvvm.tensormap.replace.rank.*`' intrinsics replace the `rank`
field of the tensor-map object with `%new_value` which must be one less than
the desired tensor rank as this field uses zero-based numbering.
#### '`llvm.nvvm.tensormap.replace.global.stride`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tensormap.replace.global.stride.p1(ptr addrspace(1) %addr, i32 immarg %ord, i64 %new_value)
declare void @llvm.nvvm.tensormap.replace.global.stride.p3(ptr addrspace(3) %addr, i32 immarg %ord, i64 %new_value)
```
##### Overview:
The '`@llvm.nvvm.tensormap.replace.global.stride.*`' intrinsics replace the
`%ord`-th element of the `global_stride` field of the tensor-map object
with `%new_value`. `%ord` must be in the range [0, 4].
#### '`llvm.nvvm.tensormap.replace.element.stride`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tensormap.replace.element.stride.p1(ptr addrspace(1) %addr, i32 immarg %ord, i32 %new_value)
declare void @llvm.nvvm.tensormap.replace.element.stride.p3(ptr addrspace(3) %addr, i32 immarg %ord, i32 %new_value)
```
##### Overview:
The '`@llvm.nvvm.tensormap.replace.element.stride.*`' intrinsics replace the
`%ord`-th element of the `element_stride` field of the tensor-map object
with `%new_value`. `%ord` must be in the range [0, 4].
#### '`llvm.nvvm.tensormap.replace.global.dim`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tensormap.replace.global.dim.p1(ptr addrspace(1) %addr, i32 immarg %ord, i32 %new_value)
declare void @llvm.nvvm.tensormap.replace.global.dim.p3(ptr addrspace(3) %addr, i32 immarg %ord, i32 %new_value)
```
##### Overview:
The '`@llvm.nvvm.tensormap.replace.global.dim.*`' intrinsics replace the
`%ord`-th element of the `global_dim` field of the tensor-map object
with `%new_value`. `%ord` must be in the range [0, 4].
#### '`llvm.nvvm.tensormap.replace.box.dim`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tensormap.replace.box.dim.p1(ptr addrspace(1) %addr, i32 immarg %ord, i32 %new_value)
declare void @llvm.nvvm.tensormap.replace.box.dim.p3(ptr addrspace(3) %addr, i32 immarg %ord, i32 %new_value)
```
##### Overview:
The '`@llvm.nvvm.tensormap.replace.box.dim.*`' intrinsics replace the
`%ord`-th element of the `box_dim` field of the tensor-map object with
`%new_value`. `%ord` must be in the range [0, 4].
#### '`llvm.nvvm.tensormap.replace.elemtype`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tensormap.replace.elemtype.p1(ptr addrspace(1) %addr, i32 immarg %new_value)
declare void @llvm.nvvm.tensormap.replace.elemtype.p3(ptr addrspace(3) %addr, i32 immarg %new_value)
```
##### Overview:
The '`@llvm.nvvm.tensormap.replace.elemtype.*`' intrinsics replace the
`elemtype` field of the tensor-map object with the type specified by
`%new_value`.
##### Semantics:
The following table shows the mapping of `%new_value` to the actual element
type:
| Element Type | Value |
| ------------------------ | ----- |
| `u8` | 0 |
| `u16` | 1 |
| `u32` | 2 |
| `s32` | 3 |
| `u64` | 4 |
| `s64` | 5 |
| `f16` | 6 |
| `f32` | 7 |
| `f32.ftz` | 8 |
| `f64` | 9 |
| `bf16` | 10 |
| `tf32` | 11 |
| `tf32.ftz` | 12 |
| `b4x16` | 13 |
| `b4x16_p64` | 14 |
| `b6x16_p32` or `b6p2x16` | 15 |
#### '`llvm.nvvm.tensormap.replace.interleave.layout`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tensormap.replace.interleave.layout.p1(ptr addrspace(1) %addr, i32 immarg %new_value)
declare void @llvm.nvvm.tensormap.replace.interleave.layout.p3(ptr addrspace(3) %addr, i32 immarg %new_value)
```
##### Overview:
The '`@llvm.nvvm.tensormap.replace.interleave.layout.*`' intrinsics replace
the `interleave_layout` field of the tensor-map object with the layout
specified by `%new_value`.
##### Semantics:
The following table shows the mapping of `%new_value` to the actual layout:
| Interleave Layout | Value |
| ----------------- | ----- |
| `No interleave` | 0 |
| `16B interleave` | 1 |
| `32B interleave` | 2 |
#### '`llvm.nvvm.tensormap.replace.swizzle_mode`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tensormap.replace.swizzle.mode.p1(ptr addrspace(1) %addr, i32 immarg %new_value)
declare void @llvm.nvvm.tensormap.replace.swizzle.mode.p3(ptr addrspace(3) %addr, i32 immarg %new_value)
```
##### Overview:
The '`@llvm.nvvm.tensormap.replace.swizzle.mode.*`' intrinsics replace the
`swizzle_mode` field of the tensor-map object with the swizzle mode specified
by `%new_value`.
##### Semantics:
The following table shows the mapping of `%new_value` to the actual swizzle
mode:
| Swizzle Mode | Value |
| -------------- | ----- |
| `No swizzle` | 0 |
| `32B swizzle` | 1 |
| `64B swizzle` | 2 |
| `128B swizzle` | 3 |
| `96B swizzle` | 4 |
#### '`llvm.nvvm.tensormap.replace.swizzle_atomicity`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tensormap.replace.swizzle.atomicity.p1(ptr addrspace(1) %addr, i32 immarg %new_value)
declare void @llvm.nvvm.tensormap.replace.swizzle.atomicity.p3(ptr addrspace(3) %addr, i32 immarg %new_value)
```
##### Overview:
The '`@llvm.nvvm.tensormap.replace.swizzle.atomicity.*`' intrinsics replace
the `swizzle_atomicity` field of the tensor-map object with the swizzle
atomicity specified by `%new_value`.
##### Semantics:
The following table shows the mapping of `%new_value` to the actual swizzle
atomicity:
| Swizzle Atomicity | Value |
| ----------------- | ----- |
| `16B` | 0 |
| `32B` | 1 |
| `32B + 8B flip` | 2 |
| `64B` | 3 |
#### '`llvm.nvvm.tensormap.replace.fill_mode`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tensormap.replace.fill.mode.p1(ptr addrspace(1) %addr, i32 immarg %new_value)
declare void @llvm.nvvm.tensormap.replace.fill.mode.p3(ptr addrspace(3) %addr, i32 immarg %new_value)
```
##### Overview:
The '`@llvm.nvvm.tensormap.replace.fill.mode.*`' intrinsics replace the
`fill_mode` field of the tensor-map object with the fill mode specified by
`%new_value`.
##### Semantics:
The following table shows the mapping of `%new_value` to the actual fill mode:
| Fill Mode | Value |
| -------------- | ----- |
| `Zero fill` | 0 |
| `OOB-NaN fill` | 1 |
### TCGEN05 family of Intrinsics
The llvm.nvvm.tcgen05.\* intrinsics model the TCGEN05 family of instructions
exposed by PTX. These intrinsics use 'Tensor Memory' (henceforth `tmem`).
NVPTX represents this memory using `addrspace(6)` and is always 32-bits.
For more information, refer to the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensor-memory).
The tensor-memory pointers may only be used with the tcgen05 intrinsics.
There are specialized load/store instructions provided (tcgen05.ld/st) to
work with tensor-memory.
For more information on tensor-memory load/store instructions, refer to [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensor-memory-and-register-load-store-instructions).
#### '`llvm.nvvm.tcgen05.alloc`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tcgen05.alloc.cg1(ptr %dst, i32 %ncols)
declare void @llvm.nvvm.tcgen05.alloc.cg2(ptr %dst, i32 %ncols)
declare void @llvm.nvvm.tcgen05.alloc.shared.cg1(ptr addrspace(3) %dst, i32 %ncols)
declare void @llvm.nvvm.tcgen05.alloc.shared.cg2(ptr addrspace(3) %dst, i32 %ncols)
```
##### Overview:
The '`@llvm.nvvm.tcgen05.alloc.*`' intrinsics correspond to the
`tcgen05.alloc.cta_group*.sync.aligned.b32` family of PTX instructions.
The `tcgen05.alloc` is a potentially blocking instruction which dynamically
allocates the specified number of columns in the Tensor Memory and writes the
address of the allocated Tensor Memory into shared memory at the location
specified by `%dst`. The 32-bit operand `%ncols` specifies the number of
columns to be allocated and it must be a power-of-two. The `.shared` variant
explicitly uses shared memory address space for the `%dst` operand. The
`.cg1` and `.cg2` variants generate `cta_group::1` and `cta_group::2`
variants of the instruction respectively.
For more information, refer to the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensor-memory-allocation-and-management-instructions).
#### '`llvm.nvvm.tcgen05.dealloc`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tcgen05.dealloc.cg1(ptr addrspace(6) %tmem_addr, i32 %ncols)
declare void @llvm.nvvm.tcgen05.dealloc.cg2(ptr addrspace(6) %tmem_addr, i32 %ncols)
```
##### Overview:
The '`@llvm.nvvm.tcgen05.dealloc.*`' intrinsics correspond to the
`tcgen05.dealloc.*` set of PTX instructions. The `tcgen05.dealloc`
instructions deallocates the Tensor Memory specified by the Tensor Memory
address `%tmem_addr`. The operand `%tmem_addr` must point to a previous
Tensor Memory allocation. The 32-bit operand `%ncols` specifies the number
of columns to be de-allocated. The `.cg1` and `.cg2` variants generate
`cta_group::1` and `cta_group::2` variants of the instruction respectively.
For more information, refer to the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensor-memory-allocation-and-management-instructions).
#### '`llvm.nvvm.tcgen05.relinq.alloc.permit`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tcgen05.relinq.alloc.permit.cg1()
declare void @llvm.nvvm.tcgen05.relinq.alloc.permit.cg2()
```
##### Overview:
The '`@llvm.nvvm.tcgen05.relinq.alloc.permit.*`' intrinsics correspond to the
`tcgen05.relinquish_alloc_permit.*` set of PTX instructions. This instruction
specifies that the CTA of the executing thread is relinquishing the right to
allocate Tensor Memory. So, it is illegal for a CTA to perform `tcgen05.alloc`
after any of its constituent threads execute
`tcgen05.relinquish_alloc_permit`. The `.cg1` and `.cg2` variants generate
`cta_group::1` and `cta_group::2` flavors of the instruction respectively.
For more information, refer to the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensor-memory-allocation-and-management-instructions).
#### '`llvm.nvvm.tcgen05.commit`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tcgen05.commit.{cg1,cg2}(ptr %mbar)
declare void @llvm.nvvm.tcgen05.commit.shared.{cg1,cg2}(ptr addrspace(3) %mbar)
declare void @llvm.nvvm.tcgen05.commit.mc.{cg1,cg2}(ptr %mbar, i16 %mc)
declare void @llvm.nvvm.tcgen05.commit.mc.shared.{cg1,cg2}(ptr addrspace(3) %mbar, i16 %mc)
```
##### Overview:
The '`@llvm.nvvm.tcgen05.commit.*`' intrinsics correspond to the
`tcgen05.commit.{cg1/cg2}.mbarrier::arrive::one.*` set of PTX instructions.
The `tcgen05.commit` is an asynchronous instruction which makes the mbarrier
object (`%mbar`) track the completion of all prior asynchronous tcgen05
operations. The `.mc` variants allow signaling on the mbarrier objects of
multiple CTAs (specified by `%mc`) in the cluster. The `.cg1` and `.cg2`
variants generate `cta_group::1` and `cta_group::2` flavors of the
instruction respectively.
For more information, refer to the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen-async-sync-operations-commit).
#### '`llvm.nvvm.tcgen05.wait`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tcgen05.wait.ld()
declare void @llvm.nvvm.tcgen05.wait.st()
```
##### Overview:
The '`@llvm.nvvm.tcgen05.wait.ld/st`' intrinsics correspond to the
`tcgen05.wait::{ld/st}.sync.aligned` pair of PTX instructions. The
`tcgen05.wait::ld` causes the executing thread to block until all prior
`tcgen05.ld` operations issued by the executing thread have completed. The
`tcgen05.wait::st` causes the executing thread to block until all prior
`tcgen05.st` operations issued by the executing thread have completed.
For more information, refer to the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-instructions-tcgen05-wait).
#### '`llvm.nvvm.tcgen05.fence`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tcgen05.fence.before.thread.sync()
declare void @llvm.nvvm.tcgen05.fence.after.thread.sync()
```
##### Overview:
The '`@llvm.nvvm.tcgen05.fence.*`' intrinsics correspond to the
`tcgen05.fence::{before/after}_thread_sync` pair of PTX instructions. These
instructions act as code motion fences for asynchronous tcgen05 operations.
For more information, refer to the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensorcore-5th-generation-instructions-tcgen05-fence).
#### '`llvm.nvvm.tcgen05.shift`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tcgen05.shift.down.cg1(ptr addrspace(6) %tmem_addr)
declare void @llvm.nvvm.tcgen05.shift.down.cg2(ptr addrspace(6) %tmem_addr)
```
##### Overview:
The '`@llvm.nvvm.tcgen05.shift.{cg1/cg2}`' intrinsics correspond to the
`tcgen05.shift.{cg1/cg2}` PTX instructions. The `tcgen05.shift` is an
asynchronous instruction which initiates the shifting of 32-byte elements
downwards across all the rows, except the last, by one row. The address operand
`%tmem_addr` specifies the base address of the matrix in the Tensor Memory
whose rows must be down shifted.
For more information, refer to the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-instructions-tcgen05-shift).
#### '`llvm.nvvm.tcgen05.cp`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tcgen05.cp.4x256b.{cg1,cg2}(ptr addrspace(6) %tmem_addr, i64 %sdesc)
declare void @llvm.nvvm.tcgen05.cp.128x256b.{cg1,cg2}(ptr addrspace(6) %tmem_addr, i64 %sdesc)
declare void @llvm.nvvm.tcgen05.cp.128x128b.{cg1,cg2}(ptr addrspace(6) %tmem_addr, i64 %sdesc)
declare void @llvm.nvvm.tcgen05.cp.32x128b_warpx4.{cg1,cg2}(ptr addrspace(6) %tmem_addr, i64 %sdesc)
declare void @llvm.nvvm.tcgen05.cp.64x128b_warpx2_02_13.{cg1,cg2}(ptr addrspace(6) %tmem_addr, i64 %sdesc)
declare void @llvm.nvvm.tcgen05.cp.64x128b_warpx2_01_23.{cg1,cg2}(ptr addrspace(6) %tmem_addr, i64 %sdesc)
declare void @llvm.nvvm.tcgen05.cp.4x256b.b6x16_p32.{cg1,cg2}(ptr addrspace(6) %tmem_addr, i64 %sdesc)
declare void @llvm.nvvm.tcgen05.cp.128x256b.b6x16_p32.{cg1,cg2}(ptr addrspace(6) %tmem_addr, i64 %sdesc)
declare void @llvm.nvvm.tcgen05.cp.128x128b.b6x16_p32.{cg1,cg2}(ptr addrspace(6) %tmem_addr, i64 %sdesc)
declare void @llvm.nvvm.tcgen05.cp.32x128b_warpx4.b6x16_p32.{cg1,cg2}(ptr addrspace(6) %tmem_addr, i64 %sdesc)
declare void @llvm.nvvm.tcgen05.cp.64x128b_warpx2_02_13.b6x16_p32.{cg1,cg2}(ptr addrspace(6) %tmem_addr, i64 %sdesc)
declare void @llvm.nvvm.tcgen05.cp.64x128b_warpx2_01_23.b6x16_p32.{cg1,cg2}(ptr addrspace(6) %tmem_addr, i64 %sdesc)
declare void @llvm.nvvm.tcgen05.cp.4x256b.b4x16_p64.{cg1,cg2}(ptr addrspace(6) %tmem_addr, i64 %sdesc)
declare void @llvm.nvvm.tcgen05.cp.128x256b.b4x16_p64.{cg1,cg2}(ptr addrspace(6) %tmem_addr, i64 %sdesc)
declare void @llvm.nvvm.tcgen05.cp.128x128b.b4x16_p64.{cg1,cg2}(ptr addrspace(6) %tmem_addr, i64 %sdesc)
declare void @llvm.nvvm.tcgen05.cp.32x128b_warpx4.b4x16_p64.{cg1,cg2}(ptr addrspace(6) %tmem_addr, i64 %sdesc)
declare void @llvm.nvvm.tcgen05.cp.64x128b_warpx2_02_13.b4x16_p64.{cg1,cg2}(ptr addrspace(6) %tmem_addr, i64 %sdesc)
declare void @llvm.nvvm.tcgen05.cp.64x128b_warpx2_01_23.b4x16_p64.{cg1,cg2}(ptr addrspace(6) %tmem_addr, i64 %sdesc)
```
##### Overview:
The '`@llvm.nvvm.tcgen05.cp.{shape}.{src_fmt}.{cg1/cg2}`' intrinsics
correspond to the `tcgen05.cp.*` family of PTX instructions. The
`tcgen05.cp` instruction initiates an asynchronous copy operation from shared
memory to the location specified by `%tmem_addr` in Tensor Memory. The 64-bit
register operand `%sdesc` is the matrix descriptor representing the source
matrix in shared memory that needs to be copied.
The valid shapes for the copy operation are:
{128x256b, 4x256b, 128x128b, 64x128b_warpx2_02_13, 64x128b_warpx2_01_23,
32x128b_warpx4}.
Shapes `64x128b` and `32x128b` require dedicated multicast qualifiers,
which are appended to the corresponding intrinsic names.
Optionally, the data can be decompressed from the source format in the shared
memory to the destination format in Tensor Memory during the copy operation.
Currently, only `.b8x16` is supported as destination format. The valid source
formats are `.b6x16_p32` and `.b4x16_p64`.
When the source format is `.b6x16_p32`, a contiguous set of 16 elements of
6-bits each followed by four bytes of padding (`_p32`) in shared memory is
decompressed into 16 elements of 8-bits (`.b8x16`) each in the Tensor Memory.
When the source format is `.b4x16_p64`, a contiguous set of 16 elements of
4-bits each followed by eight bytes of padding (`_p64`) in shared memory is
decompressed into 16 elements of 8-bits (`.b8x16`) each in the Tensor Memory.
For more information on the decompression schemes, refer to the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#optional-decompression).
For more information on the tcgen05.cp instruction, refer to the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-instructions-tcgen05-cp).
#### '`llvm.nvvm.tcgen05.ld.*`'
##### Syntax:
```llvm
declare <n x i32> @llvm.nvvm.tcgen05.ld.<shape>.<num>(ptr addrspace(6) %tmem_addr, i1 %pack)
declare <n x i32> @llvm.nvvm.tcgen05.ld.red.32x32b.<num>.i32(ptr addrspace(6) %tmem_addr, i32 %redOp)
declare <n x i32> @llvm.nvvm.tcgen05.ld.red.32x32b.<num>.f32(ptr addrspace(6) %tmem_addr, i32 %redOp, i1 %abs, i1 %nan)
declare <n x i32> @llvm.nvvm.tcgen05.ld.16x32bx2.<num>(ptr addrspace(6) %tmem_addr, i64 %offset, i1 %pack)
declare <n x i32> @llvm.nvvm.tcgen05.ld.red.16x32bx2.<num>.i32(ptr addrspace(6) %tmem_addr, i64 %offset, i32 %redOp)
declare <n x i32> @llvm.nvvm.tcgen05.ld.red.16x32bx2.<num>.f32(ptr addrspace(6) %tmem_addr, i64 %offset, i32 %redOp, i1 %abs, i1 %nan)
```
##### Overview:
This group of intrinsics asynchronously load data from the Tensor Memory at the
ocation specified by the 32-bit address operand `tmem_addr` into the destination
registers, collectively across all threads of the warps.
All the threads in the warp must specify the same value of `tmem_addr`, which
must be the base address of the collective load operation. Otherwise, the
behavior is undefined.
The `shape` qualifier and the `num` qualifier together determines the total
dimension of the data ('n') which is loaded from the Tensor Memory. The `shape`
qualifier indicates the base dimension of data. The `num` qualifier indicates
the repeat factor on the base dimension resulting in the total dimension of the
data that is accessed.
Allowed values for the `num` are `x1, x2, x4, x8, x16, x32, x64, x128` except
for `tcgen05.ld.red` which does not support `x1`
Allowed values for the 'shape' in the first intrinsic are
`16x64b, 16x128b, 16x256b, 32x32b`.
Allowed value for the 'shape' in the second intrinsic is `16x32bx2`.
The result of the intrinsic is a vector consisting of one or more 32-bit
registers derived from `shape` and `num` as shown below.
| num/shape | 16x32bx2/16x64b/32x32b | 16x128b | 16x256b |
| --------- | ---------------------- | ------- | ------- |
| x1 | 1 | 2 | 4 |
| x2 | 2 | 4 | 8 |
| x4 | 4 | 8 | 16 |
| x8 | 8 | 16 | 32 |
| x16 | 16 | 32 | 64 |
| x32 | 32 | 64 | 128 |
| x64 | 64 | 128 | NA |
| x128 | 128 | NA | NA |
The last argument `i1 %pack` is a compile-time constant which when set,
indicates that the adjacent columns are packed into a single 32-bit element
during the load
`tcgen05.ld.red` contains `%redOp` flag to specify the load reduction operation
and the f32 variant supports `%abs` and `%nan` bit flags for abs and nan
respectively
`%redOp` flag:
| value | operation |
| ----- | --------- |
| 0 | min |
| 1 | max |
For more information, refer to the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-instructions-tcgen05-ld).
#### '`llvm.nvvm.tcgen05.st.*`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tcgen05.st.<shape>.<num>(ptr addrspace(6) %tmem_addr, <n x i32> %args, i1 %unpack)
declare void @llvm.nvvm.tcgen05.st.16x32bx2.<num>(ptr addrspace(6) %tmem_addr, <n x i32> %args, i64 %offset, i1 %unpack)
```
##### Overview:
This group of intrinsics asynchronously store data from the source vector into
the Tensor Memory at the location specified by the 32-bit address operand
`tmem_addr` collectively across all threads of the warps.
All the threads in the warp must specify the same value of `tmem_addr`, which
must be the base address of the collective load operation. Otherwise, the
behavior is undefined.
The `shape` qualifier and the `num` qualifier together determines the total
dimension of the data ('n') which is loaded from the Tensor Memory. The `shape`
qualifier indicates the base dimension of data. The `num` qualifier indicates
the repeat factor on the base dimension resulting in the total dimension of the
data that is accessed.
Allowed values for the 'num' are `x1, x2, x4, x8, x16, x32, x64, x128`.
Allowed values for the 'shape' in the first intrinsic are
`16x64b, 16x128b, 16x256b, 32x32b`.
Allowed value for the 'shape' in the second intrinsic is `16x32bx2`.
`args` argument is a vector consisting of one or more 32-bit registers derived
from `shape` and `num` as listed in the table listed in the `tcgen05.ld`
section.
Each shape support an `unpack` mode to allow a 32-bit element in the register to
be unpacked into two 16-bit elements and store them in adjacent columns.
`unpack` mode can be enabled by setting the `%unpack` operand to 1 and can be
disabled by setting it to 0.
The last argument `i1 %unpack` is a compile-time constant which when set,
indicates that a 32-bit element in the register to be unpacked into two 16-bit
elements and store them in adjacent columns.
For more information, refer to the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-instructions-tcgen05-st).
### tcgen05.mma Intrinsics
#### Overview
`tcgen05.mma` operation of shape `M x N x K` perform matrix multiplication and
accumulation of the form: `D = A * B + D` where:
> - the `A` matrix has shape `M x K`, in either `Tensor Memory` or `Shared Memory`
> - the `B` matrix has shape `K x N`, in `Shared Memory` of the current CTA and, optionally in peer CTA
> - the `D` matrix is of the shape `M x N`, in `Tensor Memory`
Optionally an input predicate can be used to disable the input (`%enable_inp_d`)
from the accumulator matrix and the following operation can be performed as
`D = A * B`
The matrix multiplication and accumulation operations are categorized into
various kinds based on input types and the throughput of the multiplication
operation. The following table shows the different kinds of MMA operations
that are supported:
| .kind | Supported Input Types |
| -------- | ---------------------------------------------------------------------- |
| f16 | F16 and BF16 |
| tf32 | TF32 |
| f8f6f4 | All combinations of F8, F6, and F4 |
| i8 | Signed and Unsigned 8-bit Integers |
| mxf8f6f4 | MX-floating point formats |
| mxf4 | MX-floating point formats (FP4) |
| mxf4nvf4 | MXF4 + custom NVIDIA 4-bit floating point (with common scaling factor) |
`tcgen05.mma.sp` supports sparse variant of `A` with shape `M x K` stored in
packed form as `M X (K / 2)` in memory. The `%spmetadata` specifies the mapping
of the `K / 2` non-zero elements to the `K` elements before performing the MMA
operation.
`tcgen05.mma.block_scale` perform matrix multiplication with block scaling
`D = (A * scale_A) * (B * scale_B) + D` where scaling of input matrices from
memory to form the matrix `A` and matrix `B` before performing the MMA
operation. Scale factors for `A` and `B` matrices need to be duplicated to all
32 lane partitions of tensor memory. The shape of `%scale_a` and `%scale_b`
matrices depend on the `.scale_vectorsize` described in
[here](https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-mma-scale-valid-comb)
The sparsity metadata (`%spmetadata`) as well as the block-scale inputs for
`A / B` matrices (`%scale_a` and `%scale_b`) reside in Tensor Memory.
To facilitate opportunistic re-use of `A / B` matrix data across a sequence of
MMA operations, the `A/B` matrices are loaded into a collector buffer
(`%collector_usage_a_op_flag`, `%collector_usage_b_buffer_flag`, and
`%collector_usage_b_op_flag`). The flag value of the collector_usage flag in the
intrinsic specifies the nature of the re-use
There are three kinds of matrix descriptors used by the tcgen05 family of
instructions:
```{list-table}
:widths: 28 60 12
:header-rows: 1
* - Descriptor
- Description
- Size (bits)
* - Shared Memory Descriptor
- Describes properties of multiplicand matrix in shared memory, including
its location within the CTA's shared memory.
[PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-shared-memory-descriptor)
- 64
* - Instruction Descriptor
- Describes shapes, types, and details of all matrices and the MMA
operation.
[PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-zero-column-mask-descriptor)
- 32
* - Zero-Column Mask Descriptor
- Generates a mask specifying which columns of B matrix are zeroed in the
MMA operation, regardless of values in shared memory. Total mask size =
N bits
[PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-instruction-descriptor)
- 64
```
`tcgen05.mma` can be used for general matrix multiplication or for convolution
operations. In case of convolutions, the `activations` can be stored in either
matrix `A` or matrix `B` while the `weights` will be stored in the other matrix
`tcgen05.mma` has an optional collector qualifier to specify when an `A` or `B`
matrix is new to the sequence and should be loaded, unchanged within the
sequence and, should be reused, or the last use in the sequence and should be
discarded. The collector qualifier is used to give the TensorCore permission to
reuse a previously loaded `A` or `B` matrix; however reuse is opportunistic in
that the TensorCore may reload a matrix even when it has permission to reuse
that matrix. Thus, the source memory of an A or B matrix must not be modified
while the MMA instruction using those matrices has not completed - regardless of
collector qualifier permissions.
The `cta_group::1` specifies that the operation is performed on the Tensor
Memory of the executing thread’s CTA only. The `cta_group::2` specifies that the
MMA operation is performed on the Tensor Memory of the executing thread’s CTA
and its peer CTA.
The vector operand `%disable_output_lane` specifies the lane(s) in the Tensor
Memory that should be not be updated with the resultant matrix D. Elements of
the vector operand disable-output-lane forms a mask where each bit corresponds
to a lane of the Tensor Memory, with least significant bit of the first element
of the vector (leftmost in syntax) corresponding to the lane 0 of the Tensor
Memory. If a bit in the mask is 1, then the corresponding lane in the Tensor
Memory for the resultant matrix D will not be updated
#### Intrinsic Design:
Given the broad feature set of `tcgen05.mma` instruction modeling these
through intrinsics is highly complex, and the following table outlines the large
number of intrinsics required to fully support the `tcgen05.mma` instruction
set.
| variant | Configuration | Total Variants |
| ---------------------------------- | ------------------------------------------------------------------------------------------------- | -------------- |
| tcgen05.mma.shared | 2 (space) x 2 (sp) x 4 (kind) x 2 (cta_group) x 4 (collector_usage) | 128 |
| tcgen05.mma.tensor.ashift | 2 (sp) x 4 (kind) x 2 (cta_group) x 2 (collector_usage) | 32 |
| tcgen05.mma.scale_d | 2 (space) x 2 (sp) x 2 (kind) x 2 (cta_group) x 4 (collector_usage) | 128 |
| tcgen05.mma.scale_d.tensor.ashift | 2 (sp) x 2 (kind) x 2 (cta_group) x 2 (collector_usage) | 16 |
| tcgen05.mma.disable_output_lane | 2 (space) x 2 (sp) x 4 (kind) x 2 (cta_group) x 4 (collector_usage) | 128 |
| tcgen05.mma.disable_output_lane... | 2 (sp) x 4 (kind) x 2 (cta_group) x 2 (collector_usage) | 32 |
| tcgen05.mma.block_scale | 2 (space) x 1 (mxf4nvf4) x 2 (cta_group) x 2 (scale_vec_size) x 4 (collector_usage) | 32 |
| tcgen05.mma.block_scale | 2 (space) x 1 (mxf4) x 2 (cta_group) x 2 (scale_vec_size) x 4 (collector_usage) | 32 |
| tcgen05.mma.block_scale | 2 (space) x 1 (mxf8f6f4) x 2 (cta_group) x 2 (scale_vec_size) x 4 (collector_usage) | 32 |
| tcgen05.mma.ws | 2 (space) x 2 (sp) x 4 (kind) x 2 (zero_col_mask) x 4 (collector_usage_op) x 4 (collector_buffer) | 256 |
| Total | | 816 |
To reduce the number of possible intrinsic variations, we've modeled the
`tcgen05.mma` instructions using flag operands. We've added range checks to
these flags to prevent invalid values. We also expanded some flags back into
intrinsic modifiers to avoid supporting invalid combinations of features.
#### '`llvm.nvvm.tcgen05.mma.*`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tcgen05.mma.shared(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, i32 %kind_flag, i32 %cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.tensor<.ashift>(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, i32 %kind_flag, i32 %cta_group_flag, i32 %collector_usage_a_op_flag)
; .sp variants
declare void @llvm.nvvm.tcgen05.mma.sp.shared(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, ptr addrspace(6) %spmetadata, i1 %enable_inp_d, i32 %kind_flag, i32 %cta_group_flag, i32 %collector_usage_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.tensor<.ashift>(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, ptr addrspace(6) %spmetadata, i1 %enable_inp_d, i32 %kind_flag, i32 %cta_group_flag, i32 %collector_usage_a_op_flag)
; .scale_d variants
declare void @llvm.nvvm.tcgen05.mma.shared.scale_d(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, i64 %scale_d_imm, i32 %cta_group_flag, i32 %kind_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.tensor.scale_d<.ashift>(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, i64 %scale_d_imm, i32 %cta_group_flag, i32 %kind_flag, i32 %collector_usage_a_op_flag)
; sp.scale_d variants
declare void @llvm.nvvm.tcgen05.mma.sp.shared.scale_d(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, ptr addrspace(6) %spmetadata, i1 %enable_inp_d, i64 %scale_d_imm, i32 %cta_group_flag, i32 %collector_usage_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.tensor.scale_d<.ashift>(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, ptr addrspace(6) %spmetadata, i1 %enable_inp_d, i64 %scale_d_imm, i32 %cta_group, i32 %collector_usage_a_op_flag)
```
##### Overview:
`nvvm.tcgen05.mma` is an asynchronous intrinsic which initiates an `M x N x K`
matrix multiply and accumulate operation, `D = A * B + D` where the `A` matrix
is `M x K`, the `B` matrix is `K x N`, and the `D` matrix is `M x N`. The
operation of the form `D = A*B` is issued when the input predicate argument
`%enable_inp_d` is false. The optional immediate argument `%scale_d_imm` can be
specified to scale the input matrix `D` as follows:
`D = A * B + D * (2 ^ - %scale_d_imm)`. The valid range of values for argument
`%scale_d_imm` is `[0, 15]`. The 32-bit register operand idesc is the
instruction descriptor as described in
[Instruction descriptor](https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-instruction-descriptor)
`nvvm.tcgen05.mma` has single thread semantics, unlike the collective
instructions `nvvm.mma.sync` or the PTX `wgmma.mma_async` instruction. So, a
single thread issuing the `nvvm.tcgen05.mma` will result in the initiation of
the whole matrix and accumulate operation
When `.sp` is specifed, the dimension of A matrix is `M x (K/2)` and requires
specifying an additional `%spmetadata` argument
`.ashift` shifts the rows of the A matrix down by one row, except for the last
row in the Tensor Memory. `.ashift` is only allowed with M = 128 or M = 256.
The `%collector_usage_a_op_flag` flag specifies the usage of collector buffer
for matrix `A`. It is illegal to specify either of `USE` or `FILL` for
`%collector_usage_a_op_flag` along with `.ashift`
For more information, refer to the
[PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-mma-instructions-mma)
The following tables describe the possible values of the flag arguments
`%kind_flag` flag:
| `kind_flag` | value |
| ----------- | ----- |
| F16 | 0 |
| TF32 | 1 |
| F8F6F4 | 2 |
| I8 | 3 |
`%cta_group_flag` flag:
| `cta_group_flag` | value |
| ---------------- | ----- |
| CG1 | 1 |
| CG2 | 2 |
`%collector_usage_a_op_flag` flag:
| `collector_usage_a_op_flag` | value |
| --------------------------- | ----- |
| DISCARD | 0 |
| LASTUSE | 1 |
| USE | 2 |
| FILL | 3 |
#### '`llvm.nvvm.tcgen05.mma.block_scale*`'
##### Syntax:
```llvm
; mxf8f6f4
declare void @llvm.nvvm.tcgen05.mma.shared.mxf8f6f4.block_scale(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.tensor.mxf8f6f4.block_scale(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.shared.mxf8f6f4.block_scale.block32(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.tensor.mxf8f6f4.block_scale.block32(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.shared.mxf8f6f4.block_scale(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, ptr addrspace(6) %spmetadata, i1 %enable_inp_d, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.tensor.mxf8f6f4.block_scale(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, ptr addrspace(6) %spmetadata, i1 %enable_inp_d, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.shared.mxf8f6f4.block_scale.block32(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, ptr addrspace(6) %spmetadata, i1 %enable_inp_d, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.tensor.mxf8f6f4.block_scale.block32(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, ptr addrspace(6) %spmetadata, i1 %enable_inp_d, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
; mxf4
declare void @llvm.nvvm.tcgen05.mma.shared.mxf4.block_scale(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.tensor.mxf4.block_scale(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.shared.mxf4.block_scale.block32(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.tensor.mxf4.block_scale.block32(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.shared.mxf4.block_scale(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmetadata, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.tensor.mxf4.block_scale(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmetadata, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.shared.mxf4.block_scale.block32(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmetadata, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.tensor.mxf4.block_scale.block32(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmetadata, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
; mxf4nvf4
declare void @llvm.nvvm.tcgen05.mma.shared.mxf4nvf4.block_scale.block16(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.tensor.mxf4nvf4.block_scale.block16(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.shared.mxf4nvf4.block_scale.block32(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.tensor.mxf4nvf4.block_scale.block32(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.shared.mxf4nvf4.block_scale.block16(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmetadata, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.tensor.mxf4nvf4.block_scale.block16(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmetadata, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.shared.mxf4nvf4.block_scale.block32(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmetadata, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.tensor.mxf4nvf4.block_scale.block32(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmetadata, ptr addrspace(6) %scale_a, ptr addrspace(6) %scale_b, i32 cta_group_flag, i32 %collector_usage_a_op_flag)
```
##### Overview:
`nvvm.tcgen05.mma.block_scale` is an asynchronous intrinsic which initiates
an `M x N x K` matrix multiply and accumulate operation
`D = (A * scale_a) * (B * scale_b) + D` where the `A` matrix is `M x K`, the
`B` matrix is `K x N`, and the `D` matrix is `M x N`. The matrices `A` and `B`
are scaled with `%scale_A` and `%scale_B` matrices respectively before
performing the matrix multiply and accumulate operation. The operation of the
form `D = A*B` is issued when the input predicate argument `%enable_inp_d` is
false. The 32-bit register operand idesc is the instruction descriptor as
described in [Instruction descriptor](https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-instruction-descriptor)
`nvvm.tcgen05.mma.block_scale` has single thread semantics, unlike the
collective instructions `nvvm.mma.sync` or the PTX `wgmma.mma_async`
instruction. So, a single thread issuing the `nvvm.tcgen05.mma.block_scale` will
result in the initiation of the whole matrix multiply and accumulate operation.
When `.sp` is specified, the dimension of A matrix is `M x (K / 2)` and requires
specifying an additional `%spmetadata` argument
The `%collector_usage_a_op_flag` flag specifies the usage of collector buffer
for matrix `A`
For more information, refer to the
[PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-mma-instructions-mma)
The following tables describe the possible values of the flag arguments
`%cta_group`:
| `cta_group` | value |
| ----------- | ----- |
| CG1 | 1 |
| CG2 | 2 |
`%collector_usage_a_op_flag`:
| `collector_usage_a_op_flag` | value |
| --------------------------- | ----- |
| DISCARD | 0 |
| LASTUSE | 1 |
| USE | 2 |
| FILL | 3 |
#### '`llvm.nvvm.tcgen05.mma.disable_output_lane*`'
##### Syntax:
```llvm
declare void @llvm.nvvm.tcgen05.mma.shared.disable_output_lane.cg1(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, <4 x i32> %disable_output_lane_v4, i32 %kind_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.shared.disable_output_lane.cg2(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, <8 x i32> %disable_output_lane_v8, i32 %kind_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1<.ashift>(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, <4 x i32> %disable_output_lane_v4, i32 %kind_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2<.ashift>(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, <8 x i32> %disable_output_lane_v8, i32 %kind_flag, i32 %collector_usage_a_op_flag)
; .sp variants
declare void @llvm.nvvm.tcgen05.mma.sp.shared.disable_output_lane.cg1(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, ptr addrspace(6) %spmetadata, i1 %enable_inp_d, <4 x i32> %disable_output_lane_v4, i32 %kind_flag, i32 %collector_usage_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.shared.disable_output_lane.cg2(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, ptr addrspace(6) %spmetadata, i1 %enable_inp_d, <8 x i32> %disable_output_lane_v8, i32 %kind_flag, i32 %collector_usage_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.tensor.disable_output_lane.cg1<.ashift>(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, ptr addrspace(6) %spmetadata, i1 %enable_inp_d, <4 x i32> %disable_output_lane_v4, i32 %kind_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.tensor.disable_output_lane.cg2<.ashift>(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, ptr addrspace(6) %spmetadata, i1 %enable_inp_d, <8 x i32> %disable_output_lane_v8, i32 %kind_flag, i32 %collector_usage_a_op_flag)
; .scale_d variants
declare void @llvm.nvvm.tcgen05.mma.shared.scale_d.disable_output_lane.cg1(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, i64 %scale_d_imm, <4 x i32> %disable_output_lane_v4, i32 %kind_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.shared.scale_d.disable_output_lane.cg2(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, i64 %scale_d_imm, <8 x i32> %disable_output_lane_v8, i32 %kind_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg1<.ashift>(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, i64 %scale_d_imm, <4 x i32> %disable_output_lane_v4, i32 %kind_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg2<.ashift>(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, i64 %scale_d_imm, <8 x i32> %disable_output_lane_v8, i32 %kind_flag, i32 %collector_usage_a_op_flag)
; .sp.scale_d variants
declare void @llvm.nvvm.tcgen05.mma.sp.shared.scale_d.disable_output_lane.cg1(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, ptr addrspace(6) %spmetadata, i1 %enable_inp_d, i64 %scale_d_imm, <4 x i32> %disable_output_lane_v4, i32 %kind_flag, i32 %collector_usage_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.shared.scale_d.disable_output_lane.cg2(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, ptr addrspace(6) %spmetadata, i1 %enable_inp_d, i64 %scale_d_imm, <8 x i32> %disable_output_lane_v8, i32 %kind_flag, i32 %collector_usage_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.tensor.scale_d.disable_output_lane.cg1<.ashift>(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, ptr addrspace(6) %spmetadata, i1 %enable_inp_d, i64 %scale_d_imm, <4 x i32> %disable_output_lane_v4, i32 %kind_flag, i32 %collector_usage_a_op_flag)
declare void @llvm.nvvm.tcgen05.mma.sp.tensor.scale_d.disable_output_lane.cg2<.ashift>(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, ptr addrspace(6) %spmetadata, i1 %enable_inp_d, i64 %scale_d_imm, <8 x i32> %disable_output_lane_v8, i32 %kind_flag, i32 %collector_usage_a_op_flag)
```
##### Overview:
`nvvm.tcgen05.mma.disable_output_lane` is an asynchronous intrinsic which
initiates an `M x N x K` matrix multiply and accumulate operation
`D = A * B + D` where the `A` matrix is `M x K`, the `B` matrix is `K x N`, and
the `D` matrix is `M x N`. The operation of the form `D = A*B` is issued when
the input predicate argument `%enable_inp_d` is false. The optional immediate
argument `%scale_d_imm` can be specified to scale the input matrix `D` as
follows: `D = A*B+D * (2 ^ - %scale_d_imm)`. The valid range of values for
argument `%scale_d_imm` is `[0, 15]`. The 32-bit register operand idesc is the
instruction descriptor as described in
[Instruction descriptor](https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-instruction-descriptor)
The vector operand `%disable_output_lane` specifies the lane(s) in the Tensor
Memory that should be not be updated with the resultant matrix `D`. Elements of
the vector operand `%disable_output_lane` forms a mask where each bit
corresponds to a lane of the Tensor Memory, with least significant bit of the
first element of the vector corresponding to the `lane 0` of the Tensor Memory.
If a bit in the mask is 1, then the corresponding lane in the Tensor Memory for
the resultant matrix `D` will not be updated
`nvvm.tcgen05.mma.disable_output_lane` has single thread semantics, unlike the
collective instructions `nvvm.mma.sync` or the PTX `wgmma.mma_async`
instruction. So, a single thread issuing the
`nvvm.tcgen05.mma.disable_output_lane` will result in the initiation of the
whole matrix multiply and accumulate operation
When `.sp` is specifed, the dimension of A matrix is `M x (K / 2)` and requires
specifiying an additional `%spmetadata` argument.
> `.ashift` shifts the rows of the A matrix down by one row, except for the last
> row in the Tensor Memory. `.ashift` is only allowed with M = 128 or M = 256.
The `%collector_usage_a_op_flag` flag specifies the usage of collector buffer
for matrix `A`. It is illegal to specify either of `USE` or `FILL` for
`%collector_usage_a_op_flag` along with `.ashift`
For more information, refer to the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-mma-instructions-mma)
The following tables describes the possible values of the flag arguments
`%kind_flag`:
| `kind_flag` | value |
| ----------- | ----- |
| F16 | 0 |
| TF32 | 1 |
| F8F6F4 | 2 |
| I8 | 3 |
`%cta_group_flag`:
| `cta_group_flag` | value |
| ---------------- | ----- |
| CG1 | 1 |
| CG2 | 2 |
`%collector_usage_a_op_flag`:
| `collector_usage_a_op_flag` | value |
| --------------------------- | ----- |
| DISCARD | 0 |
| LASTUSE | 1 |
| USE | 2 |
| FILL | 3 |
#### '`llvm.nvvm.tcgen05.mma.ws*`'
##### Syntax:
```llvm
// tcgen05.mma.ws
declare void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, i32 %kind_flag, i32 %collector_usage_b_buffer_flag, i32 %collector_usage_b_op_flag)
declare void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, i32 %kind_flag, i32 %collector_usage_b_buffer_flag, i32 %collector_usage_b_op_flag)
declare void @llvm.nvvm.tcgen05.mma.ws.shared.zero_col_mask(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 %kind_flag, i32 %collector_usage_b_buffer_flag, i32 %collector_usage_b_op_flag)
declare void @llvm.nvvm.tcgen05.mma.ws.shared.zero_col_mask(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 %kind_flag, i32 %collector_usage_b_buffer_flag, i32 %collector_usage_b_op_flag)
declare void @llvm.nvvm.tcgen05.mma.ws.tensor.zero_col_mask(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 %kind_flag, i32 %collector_usage_b_buffer_flag, i32 %collector_usage_b_op_flag)
; .sp variants
declare void @llvm.nvvm.tcgen05.mma.ws.sp.shared(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmetadata, i32 %kind_flag, i32 %collector_usage_b_buffer_flag, i32 %collector_usage_b_op_flag)
declare void @llvm.nvvm.tcgen05.mma.ws.sp.tensor(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmetadata, i32 %kind_flag, i32 %collector_usage_b_buffer_flag, i32 %collector_usage_b_op_flag)
declare void @llvm.nvvm.tcgen05.mma.ws.sp.shared.zero_col_mask(ptr addrspace(6) %d, i64 %adesc, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmetadata, i64 %zero_col_mask, i32 %kind_flag, i32 %collector_usage_b_buffer_flag, i32 %collector_usage_b_op_flag)
declare void @llvm.nvvm.tcgen05.mma.ws.sp.tensor.zero_col_mask(ptr addrspace(6) %d, ptr addrspace(6) %atensor, i64 %bdesc, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmetadata, i64 %zero_col_mask, i32 %kind_flag, i32 %collector_usage_b_buffer_flag, i32 %collector_usage_b_op_flag)
```
##### Overview:
`nvvm.tcgen05.mma.ws` is an asynchronous intrinsic which initiates an
`M x N x K` weight stationary convolution matrix multiply and accumulate
operation, `D = A * B + D` where the `A` matrix is `M x K`, the `B` matrix is
`K x N`, and the `D` matrix is `M x N`. The operation of the form `D = A*B` is
issued when the input predicate argument `%enable_inp_d` is false. The optional
immediate argument `%scale_d_imm` can be specified to scale the input matrix `D`
as follows: `D = A*B+D * (2 ^ - %scale_d_imm)`. The valid range of values for
argument `%scale_d_imm` is `[0, 15]`. The 32-bit register operand idesc is the
instruction descriptor as described in
[Instruction descriptor](https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-instruction-descriptor)
`nvvm.tcgen05.mma` has single thread semantics, unlike the collective
instructions `nvvm.mma.sync` or the PTX `wgmma.mma_async` instruction. So, a
single thread issuing the `nvvm.tcgen05.mma` will result in the initiation of
the whole matrix multiply and accumulate operation
When `.sp` is specifed, the dimension of A matrix is `M x (K / 2)` and requires
specifiying an additional `%spmetadata` argument
The operand `%zero_col_mask` is a 64-bit register which specifies the
[Zero-Column Mask Descriptor](https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-zero-column-mask-descriptor).
The zero-column mask descriptor is used to generate a mask that specifies which
columns of `B` matrix will have zero value for the matrix multiply and
accumulate operation regardless of the values present in the shared memory.
The `%collector_usage_b_buffer_flag` and `%collector_usage_b_op_flag` together
flag specifies the usage of collector buffer for Matrix `B`.
For more information, refer to the
[PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-mma-instructions-mma-ws)
The following tables describes the possible values of the flag arguments
`%kind_flag`:
| `kind_flag` | value |
| ----------- | ----- |
| F16 | 0 |
| TF32 | 1 |
| F8F6F4 | 2 |
| I8 | 3 |
`%collector_usage_b_buffer_flag`:
| `collector_usage_b_buffer_flag` | value |
| ------------------------------- | ----- |
| B0 | 0 |
| B1 | 1 |
| B2 | 2 |
| B3 | 3 |
`%collector_usage_b_op_flag`:
| `collector_usage_b_op_flag` | value |
| --------------------------- | ----- |
| DISCARD | 0 |
| LASTUSE | 1 |
| USE | 2 |
| FILL | 3 |
### Store Intrinsics
#### '`llvm.nvvm.st.bulk.*`'
##### Syntax:
```llvm
declare void @llvm.nvvm.st.bulk(ptr addrspace(1) %dst, i64 %size, i64 immarg %initval)
declare void @llvm.nvvm.st.bulk.shared.cta(ptr addrspace(3) %dst, i64 %size, i64 immarg %initval)
```
##### Overview:
The '`@llvm.nvvm.st.bulk.*`' intrinsics initialize a region of shared memory
starting from the location specified by the destination address operand `%dst`.
The integer operand `%size` specifies the amount of memory to be initialized in
terms of number of bytes and must be a multiple of 8. Otherwise, the behavior
is undefined.
The integer immediate operand `%initval` specifies the initialization value for
the memory locations. The only numeric value allowed is 0.
The `@llvm.nvvm.st.bulk.shared.cta` and `@llvm.nvvm.st.bulk` intrinsics are
similar but the latter uses generic addressing (see [Generic Addressing](https://docs.nvidia.com/cuda/parallel-thread-execution/#generic-addressing)).
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-st-bulk).
#### '`llvm.nvvm.st.async`'
##### Syntax:
```llvm
declare void @llvm.nvvm.st.async.i32(ptr addrspace(7) %dest_addr, i32 %value, ptr addrspace(7) %mbarrier_addr)
declare void @llvm.nvvm.st.async.i64(ptr addrspace(7) %dest_addr, i64 %value, ptr addrspace(7) %mbarrier_addr)
declare void @llvm.nvvm.st.async.i128(ptr addrspace(7) %dest_addr, i128 %value, ptr addrspace(7) %mbarrier_addr)
```
##### Overview:
The '`llvm.nvvm.st.async`' intrinsic initiates a weak
asynchronous store operation to shared memory that stores the value specified
by the `%value` operand to the destination address specified by the
`%dest_addr` operand. The `%value` operand must be an `i32`, `i64` or
`i128`, lowering to the `.b32`, `.b64` and `.b128` variants
of the `st.async` PTX instruction respectively.
The store operation is treated as a weak memory operation. The effects of this
operation become visible to other threads only when synchronization is
established by other means.
The operation is performed asynchronously and the completion is signalled using
the mbarrier object specified by the `%mbarrier_addr` operand. Upon completion,
a [complete-tx](https://docs.nvidia.com/cuda/parallel-thread-execution/#parallel-synchronization-and-communication-instructions-mbarrier-complete-tx-operation) operation is performed on the mbarrier object, with the
`completeCount` argument equal to the amount of data stored in bytes.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-st-async).
#### '`llvm.nvvm.st.async.{sys,gpu}`'
##### Syntax:
```llvm
; sys scope
declare void @llvm.nvvm.st.async.sys.i8(ptr addrspace(1) %dest_addr, i8 %value, i1 immarg %is_multimem)
declare void @llvm.nvvm.st.async.sys.i16(ptr addrspace(1) %dest_addr, i16 %value, i1 immarg %is_multimem)
declare void @llvm.nvvm.st.async.sys.i32(ptr addrspace(1) %dest_addr, i32 %value, i1 immarg %is_multimem)
declare void @llvm.nvvm.st.async.sys.i64(ptr addrspace(1) %dest_addr, i64 %value, i1 immarg %is_multimem)
; gpu scope
declare void @llvm.nvvm.st.async.gpu.i8(ptr addrspace(1) %dest_addr, i8 %value, i1 immarg %is_multimem)
declare void @llvm.nvvm.st.async.gpu.i16(ptr addrspace(1) %dest_addr, i16 %value, i1 immarg %is_multimem)
declare void @llvm.nvvm.st.async.gpu.i32(ptr addrspace(1) %dest_addr, i32 %value, i1 immarg %is_multimem)
declare void @llvm.nvvm.st.async.gpu.i64(ptr addrspace(1) %dest_addr, i64 %value, i1 immarg %is_multimem)
```
##### Overview:
The '`llvm.nvvm.st.async.sys`' and
'`llvm.nvvm.st.async.gpu`' intrinsics initiate an
asynchronous release store to global memory that stores the value specified by
the `%value` operand to the destination address specified by the `%dest_addr`
operand.
The `%is_multimem` immediate argument selects the variant of the store. When it
is `0`, a regular `st.async` is emitted. When it is `1`, a
`multimem.st.async` is emitted and the `%dest_addr` operand must be a
multimem address.
The store carries `.release` semantics — prior stores from the current thread
are made visible to other threads in the scope.
The scope of this operation can be `sys` or `gpu`.
These intrinsics lower to the `.b8`, `.b16`, `.b32`, and `.b64` variants of the
`st.async` instruction.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-st-async).
#### '`llvm.nvvm.st.async.mmio.sys`'
##### Syntax:
```llvm
declare void @llvm.nvvm.st.async.mmio.sys.i8(ptr addrspace(1) %dest_addr, i8 %value)
declare void @llvm.nvvm.st.async.mmio.sys.i16(ptr addrspace(1) %dest_addr, i16 %value)
declare void @llvm.nvvm.st.async.mmio.sys.i32(ptr addrspace(1) %dest_addr, i32 %value)
declare void @llvm.nvvm.st.async.mmio.sys.i64(ptr addrspace(1) %dest_addr, i64 %value)
```
##### Overview:
The '`llvm.nvvm.st.async.mmio.sys`' intrinsic performs an [MMIO](https://docs.nvidia.com/cuda/parallel-thread-execution/#mmio-operation)
store to global memory with `.release` semantics at the `sys` scope.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-st-async).
### clusterlaunchcontrol Intrinsics
#### '`llvm.nvvm.clusterlaunchcontrol.try_cancel*`' Intrinsics
##### Syntax:
```llvm
declare void @llvm.nvvm.clusterlaunchcontrol.try_cancel.async.shared(ptr addrspace(3) %addr, ptr addrspace(3) %mbar)
declare void @llvm.nvvm.clusterlaunchcontrol.try_cancel.async.multicast.shared(ptr addrspace(3) %addr, ptr addrspace(3) %mbar)
```
##### Overview:
The `clusterlaunchcontrol.try_cancel` intrinsics requests atomically cancelling
the launch of a cluster that has not started running yet. It asynchronously
non-atomically writes a 16-byte opaque response to shared memory, pointed to by
16-byte-aligned `addr` indicating whether the operation succeeded or failed.
`addr` and 8-byte-aligned `mbar` must refer to `shared::cta` otherwise the
behavior is undefined. The completion of the asynchronous operation is tracked
using the mbarrier completion mechanism at `.cluster` scope referenced by the
shared memory pointer, `mbar`. On success, the opaque response contains the
CTA id of the first CTA of the canceled cluster; no other successful response
from other `clusterlaunchcontrol.try_cancel` operations from the same grid
will contain that id.
The `multicast` variant specifies that the response is asynchronously
non-atomically written to the corresponding shared memory location of each CTA
in the requesting cluster. The completion of the write of each local response is
tracked by independent mbarriers at the corresponding shared memory location of
each CTA in theccluster.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/?a#parallel-synchronization-and-communication-instructions-clusterlaunchcontrol-try-cancel).
#### '`llvm.nvvm.clusterlaunchcontrol.query_cancel.is_canceled`' Intrinsic
##### Syntax:
```llvm
declare i1 @llvm.nvvm.clusterlaunchcontrol.query_cancel.is_canceled(i128 %try_cancel_response)
```
##### Overview:
The `llvm.nvvm.clusterlaunchcontrol.query_cancel.is_canceled` intrinsic
decodes the opaque response written by the
`llvm.nvvm.clusterlaunchcontrol.try_cancel` operation.
The intrinsic returns `0` (false) if the request failed. If the request
succeeded, it returns `1` (true). A true result indicates that:
- the thread block cluster whose first CTA id matches that of the response
handle will not run, and
- no other successful response of another `try_cancel` request in the grid
will contain the first CTA id of that cluster
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/?a#parallel-synchronization-and-communication-instructions-clusterlaunchcontrol-query-cancel).
#### '`llvm.nvvm.clusterlaunchcontrol.query_cancel.get_first_ctaid.*`' Intrinsics
##### Syntax:
```llvm
declare i32 @llvm.nvvm.clusterlaunchcontrol.query_cancel.get_first_ctaid.x(i128 %try_cancel_response)
declare i32 @llvm.nvvm.clusterlaunchcontrol.query_cancel.get_first_ctaid.y(i128 %try_cancel_response)
declare i32 @llvm.nvvm.clusterlaunchcontrol.query_cancel.get_first_ctaid.z(i128 %try_cancel_response)
```
##### Overview:
The `clusterlaunchcontrol.query_cancel.get_first_ctaid.*` intrinsic can be
used to decode the successful opaque response written by the
`llvm.nvvm.clusterlaunchcontrol.try_cancel` operation.
If the request succeeded:
- `llvm.nvvm.clusterlaunchcontrol.query_cancel.get_first_ctaid.{x,y,z}`
returns the coordinate of the first CTA in the canceled cluster, either x, y,
or z.
If the request failed, the behavior of these intrinsics is undefined.
For more information, refer [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/?a#parallel-synchronization-and-communication-instructions-clusterlaunchcontrol-query-cancel).
### Perf Monitor Event Intrinsics
#### '`llvm.nvvm.pm.event.mask`' Intrinsic
##### Syntax:
```llvm
declare void @llvm.nvvm.pm.event.mask(i16 immarg %mask_val)
```
##### Overview:
The '`llvm.nvvm.pm.event.mask`' intrinsic triggers one or more performance
monitor events. Each bit in the 16-bit immediate operand `%mask_val` controls
an event.
For more information on the pmevent instructions, refer to the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#miscellaneous-instructions-pmevent).
### Warp-level Matrix Transpose Intrinsics
#### '`llvm.nvvm.movmatrix.sync.aligned.m8n8.trans.b16`'
##### Syntax:
```llvm
declare i32 @llvm.nvvm.movmatrix.sync.aligned.m8n8.trans.b16(i32 %src)
```
##### Overview:
The '`@llvm.nvvm.movmatrix.sync.aligned.m8n8.trans.b16`' intrinsic
transposes an 8x8 matrix of 16-bit elements distributed across all 32
threads of a warp. Each thread provides a 32-bit register containing two
packed `.b16` elements, and receives back two packed `.b16` elements
from the transposed matrix in the same format.
The mandatory `.sync` qualifier indicates that `movmatrix` causes the
executing thread to wait until all threads in the warp execute the same
`movmatrix` intrinsic before resuming execution.
The mandatory `.aligned` qualifier indicates that all threads in the warp
must execute the same `movmatrix` intrinsic. In conditionally executed
code, a `movmatrix` intrinsic should only be used if it is known that
all threads in the warp evaluate the condition identically, otherwise the
behavior is undefined.
For more information, refer to the [PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#warp-level-matrix-instructions-movmatrix).
### Other Intrinsics
For the full set of NVPTX intrinsics, please see the
`include/llvm/IR/IntrinsicsNVVM.td` file in the LLVM source tree.
(libdevice)=
## Linking with Libdevice
The CUDA Toolkit comes with an LLVM bitcode library called `libdevice` that
implements many common mathematical functions. This library can be used as a
high-performance math library for any compilers using the LLVM NVPTX target.
The library can be found under `nvvm/libdevice/` in the CUDA Toolkit and
there is a separate version for each compute architecture.
For a list of all math functions implemented in libdevice, see
[libdevice Users Guide](http://docs.nvidia.com/cuda/libdevice-users-guide/index.html).
To accommodate various math-related compiler flags that can affect code
generation of libdevice code, the library code depends on a special LLVM IR
pass (`NVVMReflect`) to handle conditional compilation within LLVM IR. This
pass looks for calls to the `@__nvvm_reflect` function and replaces them
with constants based on the defined reflection parameters. Such conditional
code often follows a pattern:
```c++
float my_function(float a) {
if (__nvvm_reflect("FASTMATH"))
return my_function_fast(a);
else
return my_function_precise(a);
}
```
The default value for all unspecified reflection parameters is zero.
The `NVVMReflect` pass should be executed early in the optimization
pipeline, immediately after the link stage. The `internalize` pass is also
recommended to remove unused math functions from the resulting PTX. For an
input IR module `module.bc`, the following compilation flow is recommended:
The `NVVMReflect` pass will attempt to remove dead code even without
optimizations. This allows potentially incompatible instructions to be avoided
at all optimizations levels by using the `__CUDA_ARCH` argument.
1. Save list of external functions in `module.bc`
2. Link `module.bc` with `libdevice.compute_XX.YY.bc`
3. Internalize all functions not in list from (1)
4. Eliminate all unused internal functions
5. Run `NVVMReflect` pass
6. Run standard optimization pipeline
:::{note}
`linkonce` and `linkonce_odr` linkage types are not suitable for the
libdevice functions. It is possible to link two IR modules that have been
linked against libdevice using different reflection variables.
:::
Since the `NVVMReflect` pass replaces conditionals with constants, it will
often leave behind dead code of the form:
```llvm
entry:
..
br i1 true, label %foo, label %bar
foo:
..
bar:
; Dead code
..
```
Therefore, it is recommended that `NVVMReflect` is executed early in the
optimization pipeline before dead-code elimination.
The NVPTX TargetMachine knows how to schedule `NVVMReflect` at the beginning
of your pass manager; just use the following code when setting up your pass
manager and the PassBuilder will use `registerPassBuilderCallbacks` to let
NVPTXTargetMachine::registerPassBuilderCallbacks add the pass to the
pass manager:
```c++
std::unique_ptr<TargetMachine> TM = ...;
PassBuilder PB(TM);
ModulePassManager MPM;
PB.parsePassPipeline(MPM, ...);
```
### Reflection Parameters
The libdevice library currently uses the following reflection parameters to
control code generation:
| Flag | Description |
| ------------------ | ------------------------------------------------------ |
| `__CUDA_FTZ=[0,1]` | Use optimized code paths that flush subnormals to zero |
The value of this flag is determined by the "nvvm-reflect-ftz" module flag.
The following sets the ftz flag to 1.
```llvm
!llvm.module.flags = !{!0}
!0 = !{i32 4, !"nvvm-reflect-ftz", i32 1}
```
(`i32 4` indicates that the value set here overrides the value in another
module we link with. See the [LangRef](project:LangRef.md#module-flags-metadata)
for details.)
## Executing PTX
The most common way to execute PTX assembly on a GPU device is to use the CUDA
Driver API. This API is a low-level interface to the GPU driver and allows for
JIT compilation of PTX code to native GPU machine code.
Initializing the Driver API:
```c++
CUdevice device;
CUcontext context;
// Initialize the driver API
cuInit(0);
// Get a handle to the first compute device
cuDeviceGet(&device, 0);
// Create a compute device context
cuCtxCreate(&context, 0, device);
```
JIT compiling a PTX string to a device binary:
```c++
CUmodule module;
CUfunction function;
// JIT compile a null-terminated PTX string
cuModuleLoadData(&module, (void*)PTXString);
// Get a handle to the "myfunction" kernel function
cuModuleGetFunction(&function, module, "myfunction");
```
For full examples of executing PTX assembly, please see the [CUDA Samples](https://developer.nvidia.com/cuda-downloads) distribution.
## Common Issues
### ptxas complains of undefined function: \_\_nvvm_reflect
When linking with libdevice, the `NVVMReflect` pass must be used. See
{ref}`libdevice` for more information.
## Tutorial: A Simple Compute Kernel
To start, let us take a look at a simple compute kernel written directly in
LLVM IR. The kernel implements vector addition, where each thread computes one
element of the output vector C from the input vectors A and B. To make this
easier, we also assume that only a single CTA (thread block) will be launched,
and that it will be one dimensional.
### The Kernel
```llvm
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"
target triple = "nvptx64-nvidia-cuda"
; Intrinsic to read X component of thread ID
declare i32 @llvm.nvvm.read.ptx.sreg.tid.x() readnone nounwind
define ptx_kernel void @kernel(ptr addrspace(1) %A,
ptr addrspace(1) %B,
ptr addrspace(1) %C) {
entry:
; What is my ID?
%id = tail call i32 @llvm.nvvm.read.ptx.sreg.tid.x() readnone nounwind
; Compute pointers into A, B, and C
%ptrA = getelementptr float, ptr addrspace(1) %A, i32 %id
%ptrB = getelementptr float, ptr addrspace(1) %B, i32 %id
%ptrC = getelementptr float, ptr addrspace(1) %C, i32 %id
; Read A, B
%valA = load float, ptr addrspace(1) %ptrA, align 4
%valB = load float, ptr addrspace(1) %ptrB, align 4
; Compute C = A + B
%valC = fadd float %valA, %valB
; Store back to C
store float %valC, ptr addrspace(1) %ptrC, align 4
ret void
}
```
We can use the LLVM `llc` tool to directly run the NVPTX code generator:
```text
# llc -mcpu=sm_20 kernel.ll -o kernel.ptx
```
:::{note}
If you want to generate 32-bit code, change `p:64:64:64` to `p:32:32:32`
in the module data layout string and use `nvptx-nvidia-cuda` as the
target triple.
:::
The output we get from `llc` (as of LLVM 3.4):
```text
//
// Generated by LLVM NVPTX Back-End
//
.version 3.1
.target sm_20
.address_size 64
// .globl kernel
// @kernel
.visible .entry kernel(
.param .u64 kernel_param_0,
.param .u64 kernel_param_1,
.param .u64 kernel_param_2
)
{
.reg .f32 %f<4>;
.reg .s32 %r<2>;
.reg .s64 %rl<8>;
// %bb.0: // %entry
ld.param.u64 %rl1, [kernel_param_0];
mov.u32 %r1, %tid.x;
mul.wide.s32 %rl2, %r1, 4;
add.s64 %rl3, %rl1, %rl2;
ld.param.u64 %rl4, [kernel_param_1];
add.s64 %rl5, %rl4, %rl2;
ld.param.u64 %rl6, [kernel_param_2];
add.s64 %rl7, %rl6, %rl2;
ld.global.f32 %f1, [%rl3];
ld.global.f32 %f2, [%rl5];
add.f32 %f3, %f1, %f2;
st.global.f32 [%rl7], %f3;
ret;
}
```
### Dissecting the Kernel
Now let us dissect the LLVM IR that makes up this kernel.
#### Data Layout
The data layout string determines the size in bits of common data types, their
ABI alignment, and their storage size. For NVPTX, you should use one of the
following:
32-bit PTX:
```llvm
target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"
```
64-bit PTX:
```llvm
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"
```
#### Target Intrinsics
In this example, we use the `@llvm.nvvm.read.ptx.sreg.tid.x` intrinsic to
read the X component of the current thread's ID, which corresponds to a read
of register `%tid.x` in PTX. The NVPTX back-end supports a large set of
intrinsics. A short list is shown below; please see
`include/llvm/IR/IntrinsicsNVVM.td` for the full list.
| Intrinsic | CUDA Equivalent |
| --------------------------------------------- | ----------------- |
| `i32 @llvm.nvvm.read.ptx.sreg.tid.{x,y,z}` | threadIdx.{x,y,z} |
| `i32 @llvm.nvvm.read.ptx.sreg.ctaid.{x,y,z}` | blockIdx.{x,y,z} |
| `i32 @llvm.nvvm.read.ptx.sreg.ntid.{x,y,z}` | blockDim.{x,y,z} |
| `i32 @llvm.nvvm.read.ptx.sreg.nctaid.{x,y,z}` | gridDim.{x,y,z} |
| `void @llvm.nvvm.barrier0()` | \_\_syncthreads() |
#### Address Spaces
You may have noticed that all of the pointer types in the LLVM IR example had
an explicit address space specifier. What is address space 1? NVIDIA GPU
devices (generally) have four types of memory:
- Global: Large, off-chip memory
- Shared: Small, on-chip memory shared among all threads in a CTA
- Local: Per-thread, private memory
- Constant: Read-only memory shared across all threads
These different types of memory are represented in LLVM IR as address spaces.
There is also a fifth address space used by the NVPTX code generator that
corresponds to the "generic" address space. This address space can represent
addresses in any other address space (with a few exceptions). This allows
users to write IR functions that can load/store memory using the same
instructions. Intrinsics are provided to convert pointers between the generic
and non-generic address spaces.
See {ref}`address-spaces` and {ref}`nvptx-intrinsics` for more information.
### Running the Kernel
Generating PTX from LLVM IR is all well and good, but how do we execute it on
a real GPU device? The CUDA Driver API provides a convenient mechanism for
loading and JIT compiling PTX to a native GPU device, and launching a kernel.
The API is similar to OpenCL. A simple example showing how to load and
execute our vector addition code is shown below. Note that for brevity this
code does not perform much error checking!
:::{note}
You can also use the `ptxas` tool provided by the CUDA Toolkit to offline
compile PTX to machine code (SASS) for a specific GPU architecture. Such
binaries can be loaded by the CUDA Driver API in the same way as PTX. This
can be useful for reducing startup time by precompiling the PTX kernels.
:::
```c++
#include <iostream>
#include <fstream>
#include <cassert>
#include "cuda.h"
void checkCudaErrors(CUresult err) {
assert(err == CUDA_SUCCESS);
}
/// main - Program entry point
int main(int argc, char **argv) {
CUdevice device;
CUmodule cudaModule;
CUcontext context;
CUfunction function;
CUlinkState linker;
int devCount;
// CUDA initialization
checkCudaErrors(cuInit(0));
checkCudaErrors(cuDeviceGetCount(&devCount));
checkCudaErrors(cuDeviceGet(&device, 0));
char name[128];
checkCudaErrors(cuDeviceGetName(name, 128, device));
std::cout << "Using CUDA Device [0]: " << name << "\n";
int devMajor, devMinor;
checkCudaErrors(cuDeviceComputeCapability(&devMajor, &devMinor, device));
std::cout << "Device Compute Capability: "
<< devMajor << "." << devMinor << "\n";
if (devMajor < 2) {
std::cerr << "ERROR: Device 0 is not SM 2.0 or greater\n";
return 1;
}
std::ifstream t("kernel.ptx");
if (!t.is_open()) {
std::cerr << "kernel.ptx not found\n";
return 1;
}
std::string str((std::istreambuf_iterator<char>(t)),
std::istreambuf_iterator<char>());
// Create driver context
checkCudaErrors(cuCtxCreate(&context, 0, device));
// Create module for object
checkCudaErrors(cuModuleLoadDataEx(&cudaModule, str.c_str(), 0, 0, 0));
// Get kernel function
checkCudaErrors(cuModuleGetFunction(&function, cudaModule, "kernel"));
// Device data
CUdeviceptr devBufferA;
CUdeviceptr devBufferB;
CUdeviceptr devBufferC;
checkCudaErrors(cuMemAlloc(&devBufferA, sizeof(float)*16));
checkCudaErrors(cuMemAlloc(&devBufferB, sizeof(float)*16));
checkCudaErrors(cuMemAlloc(&devBufferC, sizeof(float)*16));
float* hostA = new float[16];
float* hostB = new float[16];
float* hostC = new float[16];
// Populate input
for (unsigned i = 0; i != 16; ++i) {
hostA[i] = (float)i;
hostB[i] = (float)(2*i);
hostC[i] = 0.0f;
}
checkCudaErrors(cuMemcpyHtoD(devBufferA, &hostA[0], sizeof(float)*16));
checkCudaErrors(cuMemcpyHtoD(devBufferB, &hostB[0], sizeof(float)*16));
unsigned blockSizeX = 16;
unsigned blockSizeY = 1;
unsigned blockSizeZ = 1;
unsigned gridSizeX = 1;
unsigned gridSizeY = 1;
unsigned gridSizeZ = 1;
// Kernel parameters
void *KernelParams[] = { &devBufferA, &devBufferB, &devBufferC };
std::cout << "Launching kernel\n";
// Kernel launch
checkCudaErrors(cuLaunchKernel(function, gridSizeX, gridSizeY, gridSizeZ,
blockSizeX, blockSizeY, blockSizeZ,
0, NULL, KernelParams, NULL));
// Retrieve device data
checkCudaErrors(cuMemcpyDtoH(&hostC[0], devBufferC, sizeof(float)*16));
std::cout << "Results:\n";
for (unsigned i = 0; i != 16; ++i) {
std::cout << hostA[i] << " + " << hostB[i] << " = " << hostC[i] << "\n";
}
// Clean up after ourselves
delete [] hostA;
delete [] hostB;
delete [] hostC;
// Clean-up
checkCudaErrors(cuMemFree(devBufferA));
checkCudaErrors(cuMemFree(devBufferB));
checkCudaErrors(cuMemFree(devBufferC));
checkCudaErrors(cuModuleUnload(cudaModule));
checkCudaErrors(cuCtxDestroy(context));
return 0;
}
```
You will need to link with the CUDA driver and specify the path to cuda.h.
```text
# clang++ sample.cpp -o sample -O2 -g -I/usr/local/cuda-5.5/include -lcuda
```
We don't need to specify a path to `libcuda.so` since this is installed in a
system location by the driver, not the CUDA toolkit.
If everything goes as planned, you should see the following output when
running the compiled program:
```text
Using CUDA Device [0]: GeForce GTX 680
Device Compute Capability: 3.0
Launching kernel
Results:
0 + 0 = 0
1 + 2 = 3
2 + 4 = 6
3 + 6 = 9
4 + 8 = 12
5 + 10 = 15
6 + 12 = 18
7 + 14 = 21
8 + 16 = 24
9 + 18 = 27
10 + 20 = 30
11 + 22 = 33
12 + 24 = 36
13 + 26 = 39
14 + 28 = 42
15 + 30 = 45
```
:::{note}
You will likely see a different device identifier based on your hardware
:::
## Tutorial: Linking with Libdevice
In this tutorial, we show a simple example of linking LLVM IR with the
libdevice library. We will use the same kernel as the previous tutorial,
except that we will compute `C = pow(A, B)` instead of `C = A + B`.
Libdevice provides an `__nv_powf` function that we will use.
```llvm
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"
target triple = "nvptx64-nvidia-cuda"
; Intrinsic to read X component of thread ID
declare i32 @llvm.nvvm.read.ptx.sreg.tid.x() readnone nounwind
; libdevice function
declare float @__nv_powf(float, float)
define ptx_kernel void @kernel(ptr addrspace(1) %A,
ptr addrspace(1) %B,
ptr addrspace(1) %C) {
entry:
; What is my ID?
%id = tail call i32 @llvm.nvvm.read.ptx.sreg.tid.x() readnone nounwind
; Compute pointers into A, B, and C
%ptrA = getelementptr float, ptr addrspace(1) %A, i32 %id
%ptrB = getelementptr float, ptr addrspace(1) %B, i32 %id
%ptrC = getelementptr float, ptr addrspace(1) %C, i32 %id
; Read A, B
%valA = load float, ptr addrspace(1) %ptrA, align 4
%valB = load float, ptr addrspace(1) %ptrB, align 4
; Compute C = pow(A, B)
%valC = call float @__nv_powf(float %valA, float %valB)
; Store back to C
store float %valC, ptr addrspace(1) %ptrC, align 4
ret void
}
```
To compile this kernel, we perform the following steps:
1. Link with libdevice
2. Internalize all but the public kernel function
3. Run `NVVMReflect` and set `__CUDA_FTZ` to 0
4. Optimize the linked module
5. Codegen the module
These steps can be performed by the LLVM `llvm-link`, `opt`, and `llc`
tools. In a complete compiler, these steps can also be performed entirely
programmatically by setting up an appropriate pass configuration (see
{ref}`libdevice`).
```text
# llvm-link t2.bc libdevice.compute_20.10.bc -o t2.linked.bc
# opt -internalize -internalize-public-api-list=kernel -nvvm-reflect-list=__CUDA_FTZ=0 -nvvm-reflect -O3 t2.linked.bc -o t2.opt.bc
# llc -mcpu=sm_20 t2.opt.bc -o t2.ptx
```
:::{note}
The `-nvvm-reflect-list=_CUDA_FTZ=0` is not strictly required, as any
undefined variables will default to zero. It is shown here for evaluation
purposes.
:::
This gives us the following PTX (excerpt):
```text
//
// Generated by LLVM NVPTX Back-End
//
.version 3.1
.target sm_20
.address_size 64
// .globl kernel
// @kernel
.visible .entry kernel(
.param .u64 kernel_param_0,
.param .u64 kernel_param_1,
.param .u64 kernel_param_2
)
{
.reg .pred %p<30>;
.reg .f32 %f<111>;
.reg .s32 %r<21>;
.reg .s64 %rl<8>;
// %bb.0: // %entry
ld.param.u64 %rl2, [kernel_param_0];
mov.u32 %r3, %tid.x;
ld.param.u64 %rl3, [kernel_param_1];
mul.wide.s32 %rl4, %r3, 4;
add.s64 %rl5, %rl2, %rl4;
ld.param.u64 %rl6, [kernel_param_2];
add.s64 %rl7, %rl3, %rl4;
add.s64 %rl1, %rl6, %rl4;
ld.global.f32 %f1, [%rl5];
ld.global.f32 %f2, [%rl7];
setp.eq.f32 %p1, %f1, 0f3F800000;
setp.eq.f32 %p2, %f2, 0f00000000;
or.pred %p3, %p1, %p2;
@%p3 bra BB0_1;
bra.uni BB0_2;
BB0_1:
mov.f32 %f110, 0f3F800000;
st.global.f32 [%rl1], %f110;
ret;
BB0_2: // %__nv_isnanf.exit.i
abs.f32 %f4, %f1;
setp.gtu.f32 %p4, %f4, 0f7F800000;
@%p4 bra BB0_4;
// %bb.3: // %__nv_isnanf.exit5.i
abs.f32 %f5, %f2;
setp.le.f32 %p5, %f5, 0f7F800000;
@%p5 bra BB0_5;
BB0_4: // %.critedge1.i
add.f32 %f110, %f1, %f2;
st.global.f32 [%rl1], %f110;
ret;
BB0_5: // %__nv_isinff.exit.i
...
BB0_26: // %__nv_truncf.exit.i.i.i.i.i
mul.f32 %f90, %f107, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f91, %f90;
mov.f32 %f92, 0fBF317200;
fma.rn.f32 %f93, %f91, %f92, %f107;
mov.f32 %f94, 0fB5BFBE8E;
fma.rn.f32 %f95, %f91, %f94, %f93;
mul.f32 %f89, %f95, 0f3FB8AA3B;
// inline asm
ex2.approx.ftz.f32 %f88,%f89;
// inline asm
add.f32 %f96, %f91, 0f00000000;
ex2.approx.f32 %f97, %f96;
mul.f32 %f98, %f88, %f97;
setp.lt.f32 %p15, %f107, 0fC2D20000;
selp.f32 %f99, 0f00000000, %f98, %p15;
setp.gt.f32 %p16, %f107, 0f42D20000;
selp.f32 %f110, 0f7F800000, %f99, %p16;
setp.eq.f32 %p17, %f110, 0f7F800000;
@%p17 bra BB0_28;
// %bb.27:
fma.rn.f32 %f110, %f110, %f108, %f110;
BB0_28: // %__internal_accurate_powf.exit.i
setp.lt.f32 %p18, %f1, 0f00000000;
setp.eq.f32 %p19, %f3, 0f3F800000;
and.pred %p20, %p18, %p19;
@!%p20 bra BB0_30;
bra.uni BB0_29;
BB0_29:
mov.b32 %r9, %f110;
xor.b32 %r10, %r9, -2147483648;
mov.b32 %f110, %r10;
BB0_30: // %__nv_powf.exit
st.global.f32 [%rl1], %f110;
ret;
}
```