C++ language extensions#
HIP provides a C++ syntax that is suitable for compiling most code that commonly appears in compute kernels (classes, namespaces, operator overloading, and templates). HIP also defines other language features that are designed to target accelerators, such as:
A kernel-launch syntax that uses standard C++ (this resembles a function call and is portable to all HIP targets)
Short-vector headers that can serve on a host or device
Math functions that resemble those in
math.h, which is included with standard C++ compilersBuilt-in functions for accessing specific GPU hardware capabilities
Note
This chapter describes the built-in variables and functions that are accessible from the HIP kernel. It’s intended for users who are familiar with CUDA kernel syntax and want to learn how HIP differs from CUDA.
Features are labeled with one of the following keywords:
Supported: HIP supports the feature with a CUDA-equivalent function
Not supported: HIP does not support the feature
Under development: The feature is under development and not yet available
Function-type qualifiers#
__device__#
Supported __device__ functions are:
Run on the device
Called from the device only
You can combine __device__ with the host keyword (__host__).
__global__#
Supported __global__ functions are:
Run on the device
Called (launched) from the host
HIP __global__ functions must have a void return type.
HIP doesn’t support dynamic-parallelism, which means that you can’t call __global__ functions from
the device.
__host__#
Supported __host__ functions are:
Run on the host
Called from the host
You can combine __host__ with __device__; in this case, the function compiles for the host and the
device. Note that these functions can’t use the HIP grid coordinate functions (e.g., threadIdx.x). If
you need to use HIP grid coordinate functions, you can pass the necessary coordinate information as
an argument.
You can’t combine __host__ with __global__.
HIP parses the __noinline__ and __forceinline__ keywords and converts them into the appropriate
Clang attributes.
Calling __global__ functions#
__global__ functions are often referred to as kernels. When you call a global function, you’re launching a kernel. When launching a kernel, you must specify an execution configuration that includes the grid and block dimensions. The execution configuration can also include other information for the launch, such as the amount of additional shared memory to allocate and the stream where you want to execute the kernel.
HIP introduces a standard C++ calling convention (hipLaunchKernelGGL) to pass the run
configuration to the kernel. However, you can also use the CUDA <<< >>> syntax.
When using hipLaunchKernelGGL, your first five parameters must be:
symbol kernelName: The name of the kernel you want to launch. To support template kernels that contain",", use theHIP_KERNEL_NAMEmacro (HIPIFY tools insert this automatically).
dim3 gridDim: 3D-grid dimensions that specify the number of blocks to launch.
dim3 blockDim: 3D-block dimensions that specify the number of threads in each block.
size_t dynamicShared: The amount of additional shared memory that you want to allocate when launching the kernel (see __shared__).
hipStream_t: The stream where you want to run the kernel. A value of0corresponds to the NULL stream (see Synchronization functions).
You can include your kernel arguments after these parameters.
// Example hipLaunchKernelGGL pseudocode:
__global__ void MyKernel(float *A, float *B, float *C, size_t N)
{
...
}
MyKernel<<<dim3(gridDim), dim3(groupDim), 0, 0>>> (a,b,c,n);
// Alternatively, you can launch the kernel using:
// hipLaunchKernelGGL(MyKernel, dim3(gridDim), dim3(groupDim), 0/*dynamicShared*/, 0/*stream), a, b, c, n);
You can use HIPIFY tools to convert CUDA launch syntax to hipLaunchKernelGGL. This includes the
conversion of optional <<< >>> arguments into the five required hipLaunchKernelGGL
parameters.
Note
HIP doesn’t support dimension sizes of \(gridDim * blockDim \ge 2^{32}\) when launching a kernel.
Kernel launch example#
// Example showing device function, __device__ __host__
// <- compile for both device and host
#include <hip/hip_runtime.h>
// Example showing device function, __device__ __host__
__host__ __device__ float PlusOne(float x) // <- compile for both device and host
{
return x + 1.0;
}
__global__ void MyKernel (const float *a, const float *b, float *c, unsigned N)
{
const int gid = threadIdx.x + blockIdx.x * blockDim.x; // <- coordinate index function
if (gid < N) {
c[gid] = a[gid] + PlusOne(b[gid]);
}
}
void callMyKernel()
{
float *a, *b, *c; // initialization not shown...
unsigned N = 1000000;
const unsigned blockSize = 256;
const int gridSize = (N + blockSize - 1)/blockSize;
MyKernel<<<dim3(gridSize), dim3(blockSize), 0, 0>>> (a,b,c,N);
// Alternatively, kernel can be launched by
// hipLaunchKernelGGL(MyKernel, dim3(gridSize), dim3(blockSize), 0, 0, a,b,c,N);
}
Variable type qualifiers#
__constant__#
The host writes constant memory before launching the kernel. This memory is read-only from the GPU while the kernel is running. The functions for accessing constant memory are:
hipGetSymbolAddress()hipGetSymbolSize()hipMemcpyToSymbol()hipMemcpyToSymbolAsync()hipMemcpyFromSymbol()hipMemcpyFromSymbolAsync()
__managed__#
Managed memory, including the __managed__ keyword, is supported in HIP combined host/device
compilation.
__restrict__#
__restrict__ tells the compiler that the associated memory pointer not to alias with any other pointer
in the kernel or function. This can help the compiler generate better code. In most use cases, every
pointer argument should use this keyword in order to achieve the benefit.
Built-in variables#
Coordinate built-ins#
The kernel uses coordinate built-ins (thread*, block*, grid*) to determine the coordinate index
and bounds for the active work item.
Built-ins are defined in amd_hip_runtime.h, rather than being implicitly defined by the compiler.
Coordinate variable definitions for built-ins are the same for HIP and CUDA. For example: threadIdx.x,
blockIdx.y, and gridDim.y. The products gridDim.x * blockDim.x, gridDim.y * blockDim.y, and
gridDim.z * blockDim.z are always less than 2^32.
Coordinate built-ins are implemented as structures for improved performance. When used with
printf, they must be explicitly cast to integer types.
warpSize#
The warpSize variable type is int. It contains the warp size (in threads) for the target device.
warpSize should only be used in device functions that develop portable wave-aware code.
Note
NVIDIA devices return 32 for this variable; AMD devices return 64 for gfx9 and 32 for gfx10 and above.
Vector types#
The following vector types are defined in hip_runtime.h. They are not automatically provided by the
compiler.
Short vector types#
Short vector types derive from basic integer and floating-point types. These structures are defined in
hip_vector_types.h. The first, second, third, and fourth components of the vector are defined by the
x, y, z, and w fields, respectively. All short vector types support a constructor function of the
form make_<type_name>(). For example, float4 make_float4(float x, float y, float z, float w) creates
a vector with type float4 and value (x,y,z,w).
HIP supports the following short vector formats:
Signed Integers:
char1,char2,char3,char4short1,short2,short3,short4int1,int2,int3,int4long1,long2,long3,long4longlong1,longlong2,longlong3,longlong4
Unsigned Integers:
uchar1,uchar2,uchar3,uchar4ushort1,ushort2,ushort3,ushort4uint1,uint2,uint3,uint4ulong1,ulong2,ulong3,ulong4ulonglong1,ulonglong2,ulonglong3,ulonglong4
Floating Points:
float1,float2,float3,float4double1,double2,double3,double4
dim3#
dim3 is a three-dimensional integer vector type that is commonly used to specify grid and group
dimensions.
The dim3 constructor accepts between zero and three arguments. By default, it initializes unspecified dimensions to 1.
typedef struct dim3 {
uint32_t x;
uint32_t y;
uint32_t z;
dim3(uint32_t _x=1, uint32_t _y=1, uint32_t _z=1) : x(_x), y(_y), z(_z) {};
};
Memory fence instructions#
HIP supports __threadfence() and __threadfence_block(). If you’re using threadfence_system() in the HIP-Clang path, you can use the following workaround:
Build HIP with the
HIP_COHERENT_HOST_ALLOCenvironment variable enabled.Modify kernels that use
__threadfence_system()as follows:
Ensure the kernel operates only on fine-grained system memory, which should be allocated with
hipHostMalloc().Remove
memcpyfor all allocated fine-grained system memory regions.
Synchronization functions#
Synchronization functions causes all threads in the group to wait at this synchronization point, and for all shared and global memory accesses by the threads to complete, before running synchronization. This guarantees the visibility of accessed data for all threads in the group.
The __syncthreads() built-in function is supported in HIP. The __syncthreads_count(int),
__syncthreads_and(int), and __syncthreads_or(int) functions are under development.
The Cooperative Groups API offer options to do synchronization on a developer defined set of thread groups. For further information, check Cooperative Groups API or Cooperative Groups how to.
Math functions#
HIP-Clang supports a set of math operations that are callable from the device. HIP supports most of the device functions supported by CUDA. These are described on Math API page.
Texture functions#
The supported texture functions are listed in texture_fetch_functions.h and
texture_indirect_functions.h header files in the
HIP-AMD backend repository.
Texture functions are not supported on some devices. To determine if texture functions are supported
on your device, use Macro __HIP_NO_IMAGE_SUPPORT == 1. You can query the attribute
hipDeviceAttributeImageSupport to check if texture functions are supported in the host runtime
code.
Surface functions#
The supported surface functions are located on Surface object reference page.
Timer functions#
To read a high-resolution timer from the device, HIP provides the following built-in functions:
Returning the incremental counter value for every clock cycle on a device:
clock_t clock() long long int clock64()
The difference between the values that are returned represents the cycles used.
Returning the wall clock count at a constant frequency on the device:
long long int wall_clock64()
This can be queried using the HIP API with the
hipDeviceAttributeWallClockRateattribute of the device in HIP application code. For example:int wallClkRate = 0; //in kilohertz HIPCHECK(hipDeviceGetAttribute(&wallClkRate, hipDeviceAttributeWallClockRate, deviceId));
Where
hipDeviceAttributeWallClockRateis a device attribute. Note that wall clock frequency is a per-device attribute.Note that
clock()andclock64()do not work properly on AMD RDNA3 (GFX11) graphic processors.
Atomic functions#
Atomic functions are run as read-modify-write (RMW) operations that reside in global or shared memory. No other device or thread can observe or modify the memory location during an atomic operation. If multiple instructions from different devices or threads target the same memory location, the instructions are serialized in an undefined order.
To support system scope atomic operations, you can use the HIP APIs that contain the _system suffix.
For example:
atomicAnd: This function is atomic and coherent within the GPU device running the functionatomicAnd_system: This function extends the atomic operation from the GPU device to other CPUs and GPU devices in the system.
HIP supports the following atomic operations.
Function |
Supported in HIP |
Supported in CUDA |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✗ |
|
✓ |
✗ |
|
✓ |
✗ |
|
✓ |
✗ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✗ |
✓ |
|
✗ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
Unsafe floating-point atomic RMW operations#
Some HIP devices support fast atomic RMW operations on floating-point values. For example,
atomicAdd on single- or double-precision floating-point values may generate a hardware RMW
instruction that is faster than emulating the atomic operation using an atomic compare-and-swap
(CAS) loop.
On some devices, fast atomic RMW instructions can produce results that differ from the same functions implemented with atomic CAS loops. For example, some devices will use different rounding or denormal modes, and some devices produce incorrect answers if fast floating-point atomic RMW instructions target fine-grained memory allocations.
The HIP-Clang compiler offers a compile-time option, so you can choose fast–but potentially
unsafe–atomic instructions for your code. On devices that support these instructions, you can include
the -munsafe-fp-atomics option. This flag indicates to the compiler that all floating-point atomic
function calls are allowed to use an unsafe version, if one exists. For example, on some devices, this
flag indicates to the compiler that no floating-point atomicAdd function can target fine-grained
memory.
If you want to avoid using unsafe use a floating-point atomic RMW operations, you can use the
-mno-unsafe-fp-atomics option. Note that the compiler default is to not produce unsafe
floating-point atomic RMW instructions, so the -mno-unsafe-fp-atomics option is not necessarily
required. However, passing this option to the compiler is good practice.
When you pass -munsafe-fp-atomics or -mno-unsafe-fp-atomics to the compiler’s command line,
the option is applied globally for the entire compilation. Note that if some of the atomic RMW function
calls cannot safely use the faster floating-point atomic RMW instructions, you must use
-mno-unsafe-fp-atomics in order to ensure that your atomic RMW function calls produce correct
results.
HIP has four extra functions that you can use to more precisely control which floating-point atomic RMW functions produce unsafe atomic RMW instructions:
float unsafeAtomicAdd(float* address, float val)double unsafeAtomicAdd(double* address, double val)(Always produces fast atomic RMW instructions on devices that have them, even when-mno-unsafe-fp-atomicsis used)float safeAtomicAdd(float* address, float val)
double safeAtomicAdd(double* address, double val)(Always produces safe atomic RMW operations, even when-munsafe-fp-atomicsis used)
Warp cross-lane functions#
Threads in a warp are referred to as lanes and are numbered from 0 to warpSize - 1.
Warp cross-lane functions operate across all lanes in a warp. The hardware guarantees that all warp
lanes will execute in lockstep, so additional synchronization is unnecessary, and the instructions
use no shared memory.
Note that NVIDIA and AMD devices have different warp sizes. You can use warpSize built-ins in you
portable code to query the warp size.
Tip
Be sure to review HIP code generated from the CUDA path to ensure that it doesn’t assume a
waveSize of 32. “Wave-aware” code that assumes a waveSize of 32 can run on a wave-64
machine, but it only utilizes half of the machine’s resources.
To get the default warp size of a GPU device, use hipGetDeviceProperties in you host functions.
cudaDeviceProp props;
cudaGetDeviceProperties(&props, deviceID);
int w = props.warpSize;
// implement portable algorithm based on w (rather than assume 32 or 64)
Only use warpSize built-ins in device functions, and don’t assume warpSize to be a compile-time
constant.
Note that assembly kernels may be built for a warp size that is different from the default. All mask values either returned or accepted by these builtins are 64-bit unsigned integer values, even when compiled for a wave-32 device, where all the higher bits are unused. CUDA code ported to HIP requires changes to ensure that the correct type is used.
Note that the __sync variants are made available in ROCm 6.2, but disabled by
default to help with the transition to 64-bit masks. They can be enabled by
setting the preprocessor macro HIP_ENABLE_WARP_SYNC_BUILTINS. These builtins
will be enabled unconditionally in the next ROCm release. Wherever possible, the
implementation includes a static assert to check that the program source uses
the correct type for the mask.
Warp vote and ballot functions#
int __all(int predicate)
int __any(int predicate)
unsigned long long __ballot(int predicate)
unsigned long long __activemask()
int __all_sync(unsigned long long mask, int predicate)
int __any_sync(unsigned long long mask, int predicate)
unsigned long long __ballot_sync(unsigned long long mask, int predicate)
You can use __any and __all to get a summary view of the predicates evaluated by the
participating lanes.
__any(): Returns 1 if the predicate is non-zero for any participating lane, otherwise it returns 0.__all(): Returns 1 if the predicate is non-zero for all participating lanes, otherwise it returns 0.
To determine if the target platform supports the any/all instruction, you can use the hasWarpVote
device property or the HIP_ARCH_HAS_WARP_VOTE compiler definition.
__ballot returns a bit mask containing the 1-bit predicate value from each
lane. The nth bit of the result contains the 1 bit contributed by the nth warp
lane.
__activemask() returns a bit mask of currently active warp lanes. The nth bit
of the result is 1 if the nth warp lane is active.
Note that the __ballot and __activemask builtins in HIP have a 64-bit return
value (unlike the 32-bit value returned by the CUDA builtins). Code ported from
CUDA should be adapted to support the larger warp sizes that the HIP version
requires.
Applications can test whether the target platform supports the __ballot or
__activemask instructions using the hasWarpBallot device property in host
code or the HIP_ARCH_HAS_WARP_BALLOT macro defined by the compiler for device
code.
The _sync variants require a 64-bit unsigned integer mask argument that
specifies the lanes in the warp that will participate in cross-lane
communication with the calling lane. Each participating thread must have its own
bit set in its mask argument, and all active threads specified in any mask
argument must execute the same call with the same mask, otherwise the result is
undefined.
Warp match functions#
unsigned long long __match_any(T value)
unsigned long long __match_all(T value, int *pred)
unsigned long long __match_any_sync(unsigned long long mask, T value)
unsigned long long __match_all_sync(unsigned long long mask, T value, int *pred)
T can be a 32-bit integer type, 64-bit integer type or a single precision or
double precision floating point type.
__match_any returns a bit mask containing a 1-bit for every participating lane
if and only if that lane has the same value in value as the current lane, and
a 0-bit for all other lanes.
__match_all returns a bit mask containing a 1-bit for every participating lane
if and only if they all have the same value in value as the current lane, and
a 0-bit for all other lanes. The predicate pred is set to true if and only if
all participating threads have the same value in value.
The _sync variants require a 64-bit unsigned integer mask argument that
specifies the lanes in the warp that will participate in cross-lane
communication with the calling lane. Each participating thread must have its own
bit set in its mask argument, and all active threads specified in any mask
argument must execute the same call with the same mask, otherwise the result is
undefined.
Warp shuffle functions#
The default width is warpSize (see Warp cross-lane functions). Half-float shuffles are not supported.
T __shfl (T var, int srcLane, int width=warpSize);
T __shfl_up (T var, unsigned int delta, int width=warpSize);
T __shfl_down (T var, unsigned int delta, int width=warpSize);
T __shfl_xor (T var, int laneMask, int width=warpSize);
T __shfl_sync (unsigned long long mask, T var, int srcLane, int width=warpSize);
T __shfl_up_sync (unsigned long long mask, T var, unsigned int delta, int width=warpSize);
T __shfl_down_sync (unsigned long long mask, T var, unsigned int delta, int width=warpSize);
T __shfl_xor_sync (unsigned long long mask, T var, int laneMask, int width=warpSize);
T can be a 32-bit integer type, 64-bit integer type or a single precision or
double precision floating point type.
The _sync variants require a 64-bit unsigned integer mask argument that
specifies the lanes in the warp that will participate in cross-lane
communication with the calling lane. Each participating thread must have its own
bit set in its mask argument, and all active threads specified in any mask
argument must execute the same call with the same mask, otherwise the result is
undefined.
Cooperative groups functions#
You can use cooperative groups to synchronize groups of threads. Cooperative groups also provide a way of communicating between groups of threads at a granularity that is different from the block.
HIP supports the following kernel language cooperative groups types and functions:
Function |
Supported in HIP |
Supported in CUDA |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
|
✓ |
✓ |
For further information, check Cooperative Groups API or Cooperative Groups how to.
Warp matrix functions#
Warp matrix functions allow a warp to cooperatively operate on small matrices that have elements spread over lanes in an unspecified manner.
HIP does not support kernel language warp matrix types or functions.
Function |
Supported in HIP |
Supported in CUDA |
|
✗ |
✓ |
|
✗ |
✓ |
|
✗ |
✓ |
|
✗ |
✓ |
|
✗ |
✓ |
Independent thread scheduling#
Certain architectures that support CUDA allow threads to progress independently of each other. This independent thread scheduling makes intra-warp synchronization possible.
HIP does not support this type of scheduling.
Profiler Counter Function#
The CUDA __prof_trigger() instruction is not supported.
Assert#
The assert function is supported in HIP. Assert function is used for debugging purpose, when the input expression equals to zero, the execution will be stopped.
void assert(int input)
There are two kinds of implementations for assert functions depending on the use sceneries,
- One is for the host version of assert, which is defined in assert.h,
- Another is the device version of assert, which is implemented in hip/hip_runtime.h.
Users need to include assert.h to use assert. For assert to work in both device and host functions, users need to include "hip/hip_runtime.h".
HIP provides the function abort() which can be used to terminate the application when terminal failures are detected. It is implemented using the __builtin_trap() function.
This function produces a similar effect of using asm("trap") in the CUDA code.
Note
In HIP, the function terminates the entire application, while in CUDA, asm("trap") only terminates the dispatch and the application continues to run.
printf#
printf function is supported in HIP.
The following is a simple example to print information in the kernel.
#include <hip/hip_runtime.h>
__global__ void run_printf() { printf("Hello World\n"); }
int main() {
run_printf<<<dim3(1), dim3(1), 0, 0>>>();
}
Device-Side Dynamic Global Memory Allocation#
Device-side dynamic global memory allocation is under development. HIP now includes a preliminary implementation of malloc and free that can be called from device functions.
__launch_bounds__#
GPU multiprocessors have a fixed pool of resources (primarily registers and shared memory) which are shared by the actively running warps. Using more resources can increase IPC of the kernel but reduces the resources available for other warps and limits the number of warps that can be simultaneously running. Thus GPUs have a complex relationship between resource usage and performance.
__launch_bounds__ allows the application to provide usage hints that influence the resources (primarily registers) used by the generated code. It is a function attribute that must be attached to a __global__ function:
__global__ void __launch_bounds__(MAX_THREADS_PER_BLOCK, MIN_WARPS_PER_EXECUTION_UNIT)
MyKernel(hipGridLaunch lp, ...)
...
__launch_bounds__ supports two parameters:
- MAX_THREADS_PER_BLOCK - The programmers guarantees that kernel will be launched with threads less than MAX_THREADS_PER_BLOCK. (On NVCC this maps to the .maxntid PTX directive). If no launch_bounds is specified, MAX_THREADS_PER_BLOCK is the maximum block size supported by the device (typically 1024 or larger). Specifying MAX_THREADS_PER_BLOCK less than the maximum effectively allows the compiler to use more resources than a default unconstrained compilation that supports all possible block sizes at launch time.
The threads-per-block is the product of (blockDim.x * blockDim.y * blockDim.z).
- MIN_WARPS_PER_EXECUTION_UNIT - directs the compiler to minimize resource usage so that the requested number of warps can be simultaneously active on a multi-processor. Since active warps compete for the same fixed pool of resources, the compiler must reduce resources required by each warp(primarily registers). MIN_WARPS_PER_EXECUTION_UNIT is optional and defaults to 1 if not specified. Specifying a MIN_WARPS_PER_EXECUTION_UNIT greater than the default 1 effectively constrains the compiler’s resource usage.
When launch kernel with HIP APIs, for example, hipModuleLaunchKernel(), HIP will do validation to make sure input kernel dimension size is not larger than specified launch_bounds.
In case exceeded, HIP would return launch failure, if AMD_LOG_LEVEL is set with proper value (for details, please refer to docs/markdown/hip_logging.md), detail information will be shown in the error log message, including
launch parameters of kernel dim size, launch bounds, and the name of the faulting kernel. It’s helpful to figure out which is the faulting kernel, besides, the kernel dim size and launch bounds values will also assist in debugging such failures.
Compiler Impact#
The compiler uses these parameters as follows: - The compiler uses the hints only to manage register usage, and does not automatically reduce shared memory or other resources. - Compilation fails if compiler cannot generate a kernel which meets the requirements of the specified launch bounds. - From MAX_THREADS_PER_BLOCK, the compiler derives the maximum number of warps/block that can be used at launch time. Values of MAX_THREADS_PER_BLOCK less than the default allows the compiler to use a larger pool of registers : each warp uses registers, and this hint constrains the launch to a warps/block size which is less than maximum. - From MIN_WARPS_PER_EXECUTION_UNIT, the compiler derives a maximum number of registers that can be used by the kernel (to meet the required #simultaneous active blocks). If MIN_WARPS_PER_EXECUTION_UNIT is 1, then the kernel can use all registers supported by the multiprocessor. - The compiler ensures that the registers used in the kernel is less than both allowed maximums, typically by spilling registers (to shared or global memory), or by using more instructions. - The compiler may use heuristics to increase register usage, or may simply be able to avoid spilling. The MAX_THREADS_PER_BLOCK is particularly useful in this cases, since it allows the compiler to use more registers and avoid situations where the compiler constrains the register usage (potentially spilling) to meet the requirements of a large block size that is never used at launch time.
CU and EU Definitions#
A compute unit (CU) is responsible for executing the waves of a work-group. It is composed of one or more execution units (EU) which are responsible for executing waves. An EU can have enough resources to maintain the state of more than one executing wave. This allows an EU to hide latency by switching between waves in a similar way to symmetric multithreading on a CPU. In order to allow the state for multiple waves to fit on an EU, the resources used by a single wave have to be limited. Limiting such resources can allow greater latency hiding, but can result in having to spill some register state to memory. This attribute allows an advanced developer to tune the number of waves that are capable of fitting within the resources of an EU. It can be used to ensure at least a certain number will fit to help hide latency, and can also be used to ensure no more than a certain number will fit to limit cache thrashing.
Porting from CUDA __launch_bounds#
CUDA defines a __launch_bounds which is also designed to control occupancy:
__launch_bounds(MAX_THREADS_PER_BLOCK, MIN_BLOCKS_PER_MULTIPROCESSOR)
The second parameter
__launch_boundsparameters must be converted to the format used __hip_launch_bounds, which uses warps and execution-units rather than blocks and multi-processors (this conversion is performed automatically by HIPIFY tools).
MIN_WARPS_PER_EXECUTION_UNIT = (MIN_BLOCKS_PER_MULTIPROCESSOR * MAX_THREADS_PER_BLOCK) / 32
The key differences in the interface are:
- Warps (rather than blocks):
The developer is trying to tell the compiler to control resource utilization to guarantee some amount of active Warps/EU for latency hiding. Specifying active warps in terms of blocks appears to hide the micro-architectural details of the warp size, but makes the interface more confusing since the developer ultimately needs to compute the number of warps to obtain the desired level of control.
- Execution Units (rather than multiprocessor):
The use of execution units rather than multiprocessors provides support for architectures with multiple execution units/multi-processor. For example, the AMD GCN architecture has 4 execution units per multiprocessor. The hipDeviceProps has a field executionUnitsPerMultiprocessor.
Platform-specific coding techniques such as #ifdef can be used to specify different launch_bounds for NVCC and HIP-Clang platforms, if desired.
maxregcount#
Unlike NVCC, HIP-Clang does not support the --maxregcount option. Instead, users are encouraged to use the hip_launch_bounds directive since the parameters are more intuitive and portable than
micro-architecture details like registers, and also the directive allows per-kernel control rather than an entire file. hip_launch_bounds works on both HIP-Clang and NVCC targets.
Asynchronous Functions#
The supported asynchronous functions reference are located on the following pages:
Register Keyword#
The register keyword is deprecated in C++, and is silently ignored by both NVCC and HIP-Clang. You can pass the option -Wdeprecated-register the compiler warning message.
Pragma Unroll#
Unroll with a bounds that is known at compile-time is supported. For example:
#pragma unroll 16 /* hint to compiler to unroll next loop by 16 */
for (int i=0; i<16; i++) ...
#pragma unroll 1 /* tell compiler to never unroll the loop */
for (int i=0; i<16; i++) ...
#pragma unroll /* hint to compiler to completely unroll next loop. */
for (int i=0; i<16; i++) ...
In-Line Assembly#
GCN ISA In-line assembly is supported.
There are some usage limitations in ROCm compiler for inline asm support, please refer to Inline ASM statements for details.
Users can get related background resources on how to use inline assembly for any usage of inline assembly features.
A short example program including an inline assembly statement can be found at inline asm tutorial.
For further usage of special AMD GPU hardware features that are available through assembly, please refer to the ISA manual for AMDGPU usage, in which AMD GCN is listed from gfx906 to RDNA 3.5.
C++ Support#
The following C++ features are not supported:
Run-time-type information (RTTI)
Try/catch
Partially supported features:
Virtual functions
Virtual functions are not supported if objects containing virtual function tables are passed between GPU’s of different offload arch’s, e.g. between gfx906 and gfx1030. Otherwise virtual functions are supported.
Kernel Compilation#
hipcc now supports compiling C++/HIP kernels to binary code objects.
The file format for binary is .co which means Code Object. The following command builds the code object using hipcc.
hipcc --genco --offload-arch=[TARGET GPU] [INPUT FILE] -o [OUTPUT FILE]
[TARGET GPU] = GPU architecture
[INPUT FILE] = Name of the file containing kernels
[OUTPUT FILE] = Name of the generated code object file
Note
When using binary code objects is that the number of arguments to the kernel is different on HIP-Clang and NVCC path. Refer to the HIP module_api sample for differences in the arguments to be passed to the kernel.
gfx-arch-specific-kernel#
Clang defined ‘__gfx*__’ macros can be used to execute gfx arch specific codes inside the kernel. Refer to the sample in HIP 14_gpu_arch sample.