leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 28 Packages Projects Releases Wiki Activity

Collection of changes to fix clang build. (#1200)

Browse Source 
* Remove unused variables

* Qualify calls to make_fragment_? from templated base class.

Fixes clang build error.

* Add missing `#include <cstdio>`

* Various changes to fix clang compile errors.

* More changes to fix clang build.

Remaining issues:

- `params` initializer of `CollectiveEpilogue`.
- `ops` initializer of `Sm90VisitorImplBase`.
- `__usAtomicCAS` needs to be added to clang upstream.

* Fix remaining clang build issues.

* Qualify `cute::rank()` calls.

* Qualify some more calls that are otherwise ambiguous between `cute` and `std` namespace.

* Double-escape special registers in inline asm.

* small change

---------

Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
This commit is contained in:
Christian Sigg
2023-12-08 14:42:12 -05:00
committed by GitHub
co-authored by Haicheng Wu
parent f4a0216601
commit e1483d5fa0
46 changed files with 308 additions and 273 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/cute/arch/cluster_sm90.hpp
+13 -13
View File
@@ -86,9 +86,9 @@ CUTE_DEVICE dim3 cluster_grid_dims()
{
#if defined(CUTE_ARCH_CLUSTER_SM90_ENABLED)
uint32_t x, y, z;
asm volatile("mov.u32 %0, %nclusterid.x;\n" : "=r"(x) : );
asm volatile("mov.u32 %0, %nclusterid.y;\n" : "=r"(y) : );
asm volatile("mov.u32 %0, %nclusterid.z;\n" : "=r"(z) : );
asm volatile("mov.u32 %0, %%nclusterid.x;\n" : "=r"(x) : );
asm volatile("mov.u32 %0, %%nclusterid.y;\n" : "=r"(y) : );
asm volatile("mov.u32 %0, %%nclusterid.z;\n" : "=r"(z) : );
return {x, y, z};
#elif defined(__CUDA_ARCH__)
// MSVC requires protecting use of gridDim with __CUDA_ARCH__.
@@ -105,9 +105,9 @@ CUTE_DEVICE dim3 cluster_id_in_grid()
{
#if defined(CUTE_ARCH_CLUSTER_SM90_ENABLED)
uint32_t x, y, z;
asm volatile("mov.u32 %0, %clusterid.x;\n" : "=r"(x) : );
asm volatile("mov.u32 %0, %clusterid.y;\n" : "=r"(y) : );
asm volatile("mov.u32 %0, %clusterid.z;\n" : "=r"(z) : );
asm volatile("mov.u32 %0, %%clusterid.x;\n" : "=r"(x) : );
asm volatile("mov.u32 %0, %%clusterid.y;\n" : "=r"(y) : );
asm volatile("mov.u32 %0, %%clusterid.z;\n" : "=r"(z) : );
return {x, y, z};
#elif defined(__CUDA_ARCH__)
// MSVC requires protecting use of blockIdx with __CUDA_ARCH__.
@@ -124,9 +124,9 @@ CUTE_DEVICE dim3 block_id_in_cluster()
{
#if defined(CUTE_ARCH_CLUSTER_SM90_ENABLED)
uint32_t x, y, z;
asm volatile("mov.u32 %0, %cluster_ctaid.x;\n" : "=r"(x) : );
asm volatile("mov.u32 %0, %cluster_ctaid.y;\n" : "=r"(y) : );
asm volatile("mov.u32 %0, %cluster_ctaid.z;\n" : "=r"(z) : );
asm volatile("mov.u32 %0, %%cluster_ctaid.x;\n" : "=r"(x) : );
asm volatile("mov.u32 %0, %%cluster_ctaid.y;\n" : "=r"(y) : );
asm volatile("mov.u32 %0, %%cluster_ctaid.z;\n" : "=r"(z) : );
return {x, y, z};
#else
return {0,0,0};
@@ -138,9 +138,9 @@ CUTE_DEVICE dim3 cluster_shape()
{
#if defined(CUTE_ARCH_CLUSTER_SM90_ENABLED)
uint32_t x, y, z;
asm volatile("mov.u32 %0, %cluster_nctaid.x;\n" : "=r"(x) : );
asm volatile("mov.u32 %0, %cluster_nctaid.y;\n" : "=r"(y) : );
asm volatile("mov.u32 %0, %cluster_nctaid.z;\n" : "=r"(z) : );
asm volatile("mov.u32 %0, %%cluster_nctaid.x;\n" : "=r"(x) : );
asm volatile("mov.u32 %0, %%cluster_nctaid.y;\n" : "=r"(y) : );
asm volatile("mov.u32 %0, %%cluster_nctaid.z;\n" : "=r"(z) : );
return {x, y, z};
#else
return {1,1,1};
@@ -152,7 +152,7 @@ CUTLASS_DEVICE uint32_t block_rank_in_cluster()
{
#if defined(CUTE_ARCH_CLUSTER_SM90_ENABLED)
uint32_t rank;
asm volatile("mov.u32 %0, %cluster_ctarank;\n" : "=r"(rank) :);
asm volatile("mov.u32 %0, %%cluster_ctarank;\n" : "=r"(rank) :);
return rank;
#else
return 0;
include/cute/config.hpp
+4 -3
View File
@@ -30,7 +30,7 @@
**************************************************************************************************/
#pragma once
#if defined(__CUDA_ARCH__) || defined(_NVHPC_CUDA)
#if defined(__CUDACC__) || defined(_NVHPC_CUDA)
# define CUTE_HOST_DEVICE __forceinline__ __host__ __device__
# define CUTE_DEVICE __forceinline__ __device__
# define CUTE_HOST __forceinline__ __host__
@@ -46,10 +46,11 @@
# define CUTE_HOST_RTC CUTE_HOST
#endif
#if !defined(__CUDACC_RTC__) && (defined(__CUDA_ARCH__) || defined(_NVHPC_CUDA))
#if !defined(__CUDACC_RTC__) && !defined(__clang__) && \
(defined(__CUDA_ARCH__) || defined(_NVHPC_CUDA))
# define CUTE_UNROLL #pragma unroll
# define CUTE_NO_UNROLL #pragma unroll 1
#elif defined(__CUDACC_RTC__)
#elif defined(__CUDACC_RTC__) || defined(__clang__)
# define CUTE_UNROLL _Pragma("unroll")
# define CUTE_NO_UNROLL _Pragma("unroll 1")
#else