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

CUTLASS 2.3.0 final.

Browse Source
This commit is contained in:
akerr 2020-09-25 10:34:46 -07:00
parent c53f3339bb
commit 37a8f9e598
8 changed files with 15 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
include/cutlass/epilogue/threadblock/predicated_tile_iterator.h
View File

@ -372,12 +372,11 @@ public:
bool guard = row_guard && mask_.predicates[column];
cutlass::arch::global_store<AccessType, sizeof(AccessType)>(
frag_ptr[frag_row_idx * ThreadMap::Iterations::kColumn +
column],
(void *)&memory_pointer[column * ThreadMap::Delta::kColumn /
kElementsPerAccess],
guard);
if (guard) {
memory_pointer[column * ThreadMap::Delta::kColumn / kElementsPerAccess] =
frag_ptr[frag_row_idx * ThreadMap::Iterations::kColumn + column];
}
}
if (row + 1 < ThreadMap::Iterations::kRow) {
@ -691,8 +690,9 @@ public:
bool guard = col_guard && mask_.predicates[iteration_contiguous_];
cutlass::arch::global_store<AccessType, sizeof(AccessType)>(
*frag_ptr, (void *)memory_pointer, guard);
if (guard) {
*memory_pointer = *frag_ptr;
}
}
/// Overrides the internal iteration index

2
include/cutlass/gemm/kernel/gemm.h
View File

@ -224,7 +224,7 @@ struct Gemm {
// Broadcast the warp_id computed by lane 0 to ensure dependent code
// is compiled as warp-uniform.
int warp_idx = __shfl_sync(0x1f, threadIdx.x / 32, 0);
int warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
int lane_idx = threadIdx.x % 32;
//

2
include/cutlass/gemm/kernel/gemm_array.h
View File

@ -184,7 +184,7 @@ struct GemmArray {
// Broadcast the warp_id computed by lane 0 to ensure dependent code
// is compiled as warp-uniform.
int warp_idx = __shfl_sync(0x1f, threadIdx.x / 32, 0);
int warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
int lane_idx = threadIdx.x % 32;

2
include/cutlass/gemm/kernel/gemm_batched.h
View File

@ -196,7 +196,7 @@ struct GemmBatched {
// Broadcast the warp_id computed by lane 0 to ensure dependent code
// is compiled as warp-uniform.
int warp_idx = __shfl_sync(0x1f, threadIdx.x / 32, 0);
int warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
int lane_idx = threadIdx.x % 32;

2
include/cutlass/gemm/kernel/gemm_planar_complex.h
View File

@ -512,7 +512,7 @@ public:
// Broadcast the warp_id computed by lane 0 to ensure dependent code
// is compiled as warp-uniform.
int warp_idx = __shfl_sync(0x1f, threadIdx.x / 32, 0);
int warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
int lane_idx = threadIdx.x % 32;

2
include/cutlass/gemm/kernel/gemm_planar_complex_array.h
View File

@ -441,7 +441,7 @@ public:
// Broadcast the warp_id computed by lane 0 to ensure dependent code
// is compiled as warp-uniform.
int warp_idx = __shfl_sync(0x1f, threadIdx.x / 32, 0);
int warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
int lane_idx = threadIdx.x % 32;
//

2
include/cutlass/gemm/kernel/gemm_universal.h
View File

@ -402,7 +402,7 @@ public:
// Broadcast the warp_id computed by lane 0 to ensure dependent code
// is compiled as warp-uniform.
int warp_idx = __shfl_sync(0x1f, threadIdx.x / 32, 0);
int warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
int lane_idx = threadIdx.x % 32;

2
include/cutlass/gemm/kernel/sparse_gemm.h
View File

@ -269,7 +269,7 @@ struct SparseGemm {
// Broadcast the warp_id computed by lane 0 to ensure dependent code
// is compiled as warp-uniform.
int warp_idx = __shfl_sync(0x1f, threadIdx.x / 32, 0);
int warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
int lane_idx = threadIdx.x % 32;
//