leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 25 Packages Projects Releases Wiki Activity

CUTLASS 3.1 (#915)

Browse Source 
Co-authored-by: Aniket Shivam <ashivam@nvidia.com>
This commit is contained in:
ANIKET SHIVAM
2023-04-14 23:19:34 -04:00
committed by GitHub
co-authored by Aniket Shivam
parent 9b8166e3f0
commit d572cc1aab
482 changed files with 37175 additions and 16410 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/profiler/src/device_context.cu
+14 -6
View File
@@ -76,7 +76,8 @@ DeviceAllocation *DeviceContext::allocate_tensor(
library::LayoutTypeID layout_id,
std::vector<int> const &extent,
std::vector<int64_t> const &stride,
int batch_count) {
int batch_count,
int seed_shift) {
DeviceAllocation *allocation =
allocate_tensor(name, type, layout_id, extent, stride, batch_count);
@@ -88,6 +89,12 @@ DeviceAllocation *DeviceContext::allocate_tensor(
if(!options.initialization.fix_data_distribution) {
// change data distribution based on bit width
switch(type) {
case library::NumericTypeID::kFE4M3:
data_distribution.set_uniform(-1, 1, 0);
break;
case library::NumericTypeID::kFE5M2:
data_distribution.set_uniform(-1, 1, 0);
break;
case library::NumericTypeID::kF16:
data_distribution.set_uniform(-3, 3, 0);
break;
@@ -118,12 +125,12 @@ DeviceAllocation *DeviceContext::allocate_tensor(
if (options.initialization.provider == library::Provider::kReferenceDevice) {
allocation->initialize_random_device(
options.initialization.seed,
options.initialization.seed + seed_shift,
data_distribution);
}
else if (options.initialization.provider == library::Provider::kReferenceHost) {
allocation->initialize_random_host(
options.initialization.seed,
options.initialization.seed + seed_shift,
data_distribution);
}
}
@@ -140,7 +147,8 @@ DeviceAllocation *DeviceContext::allocate_sparsemeta_tensor(
library::NumericTypeID type_a,
std::vector<int> const &extent,
std::vector<int64_t> const &stride,
int batch_count) {
int batch_count,
int seed_shift) {
DeviceAllocation *allocation =
allocate_tensor(name, type, layout_id, extent, stride, batch_count);
@@ -151,12 +159,12 @@ DeviceAllocation *DeviceContext::allocate_sparsemeta_tensor(
if (options.initialization.provider == library::Provider::kReferenceDevice) {
allocation->initialize_random_sparsemeta_device(
options.initialization.seed,
options.initialization.seed + seed_shift,
MetaSizeInBits);
}
else if (options.initialization.provider == library::Provider::kReferenceHost) {
allocation->initialize_random_sparsemeta_host(
options.initialization.seed,
options.initialization.seed + seed_shift,
MetaSizeInBits);
}
}