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

CUTLASS 3.6.0 (#1850)

Browse Source 
* v3.6

* update changelog

* update readme

* fix typo

* fixing typos

* hopper gemm with weight prefetch

---------

Co-authored-by: yuzhai <yuzhai@nvidia.com>
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
This commit is contained in:
Yujia Zhai
2024-10-09 15:33:27 -04:00
committed by GitHub
co-authored by yuzhai Haicheng Wu
parent 0837a2a00a
commit cc3c29a81a
354 changed files with 105937 additions and 8197 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/unit/core/fast_numeric_conversion.cu
+3
View File
@@ -103,6 +103,7 @@ void run_test_integer_range_all() {
for (int i = 0; i < kN; ++i) {
source.host_view().at({0, i}) = Source(kIntSourceMin + (i % kIntRange));
}
source.sync_device();
@@ -116,12 +117,14 @@ void run_test_integer_range_all() {
// Verify conversion
bool passed = true;
for (int i = 0; i < kN; ++i) {
if(!(float(destination.host_view().at({0, i})) == float(source.host_view().at({0, i})))) {
passed = false;
break;
}
}
EXPECT_TRUE(passed) << " FastNumericArrayConverter failed";
// Print out results for the failed conversion.
test/unit/core/functional.cu
+12 -14
View File
@@ -497,20 +497,18 @@ __global__ void
test_cutlass_maximum(cutlass::half_t const* in1, cutlass::half_t const* in2, cutlass::half_t* out)
{
{
constexpr bool propagate_NaN = true;
cutlass::maximum<cutlass::half_t, propagate_NaN> op;
if (threadIdx.x == 0 && threadIdx.y == 0 && threadIdx.z == 0
&& blockIdx.x == 0 && blockIdx.y == 0 && blockIdx.z == 0) {
*out = op(*in1, *in2);
}
constexpr bool propagate_NaN = true;
cutlass::maximum<cutlass::half_t, propagate_NaN> op;
if (threadIdx.x == 0 && threadIdx.y == 0 && threadIdx.z == 0
&& blockIdx.x == 0 && blockIdx.y == 0 && blockIdx.z == 0) {
*out = op(*in1, *in2);
}
{
constexpr bool propagate_NaN = false;
cutlass::maximum<cutlass::half_t, propagate_NaN> op;
if (threadIdx.x == 0 && threadIdx.y == 0 && threadIdx.z == 0
&& blockIdx.x == 0 && blockIdx.y == 0 && blockIdx.z == 0) {
*out = op(*in1, *in2);
}
}
constexpr bool propagate_NaN = false;
cutlass::maximum<cutlass::half_t, propagate_NaN> op;
if (threadIdx.x == 0 && threadIdx.y == 0 && threadIdx.z == 0
&& blockIdx.x == 0 && blockIdx.y == 0 && blockIdx.z == 0) {
*out = op(*in1, *in2);
}
}
@@ -543,7 +541,7 @@ TEST(Functional, maximum_half_host_dont_propagate_NaN) {
EXPECT_EQ(result, y);
}
TEST(Function, maximum_half_device) {
TEST(FUnction, maximum_half_device) {
using Tensor = cutlass::HostTensor<cutlass::half_t, cutlass::layout::RowMajor>;
Tensor in1({1, 1});
test/unit/core/numeric_conversion.cu
+126
View File
@@ -217,6 +217,15 @@ TEST(NumericConversion, f32_to_fe4m3_rn) {
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, f32_to_fe4m3_rn_2_elements) {
int const kN = 2;
using Source = float;
const char source_name[] = "float";
using Destination = cutlass::float_e4m3_t;
const char dest_name[] = "float_e4m3_t";
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, f32_to_fe4m3_rn_array) {
int const kN = 27;
using Source = float;
@@ -235,6 +244,15 @@ TEST(NumericConversion, f32_to_fe5m2_rn) {
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, f32_to_fe5m2_rn_2_elements) {
int const kN = 2;
using Source = float;
const char source_name[] = "float";
using Destination = cutlass::float_e5m2_t;
const char dest_name[] = "float_e5m2_t";
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, f32_to_fe5m2_rn_array) {
int const kN = 27;
using Source = float;
@@ -253,6 +271,15 @@ TEST(NumericConversion, f16_to_fe4m3_rn) {
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, f16_to_fe4m3_rn_2_elements) {
int const kN = 2;
using Source = cutlass::half_t;
const char source_name[] = "half_t";
using Destination = cutlass::float_e4m3_t;
const char dest_name[] = "float_e4m3_t";
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, f16_to_fe4m3_rn_array) {
int const kN = 27;
using Source = cutlass::half_t;
@@ -271,6 +298,15 @@ TEST(NumericConversion, f16_to_fe5m2_rn) {
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, f16_to_fe5m2_rn_2_elements) {
int const kN = 27;
using Source = cutlass::half_t;
const char source_name[] = "half_t";
using Destination = cutlass::float_e5m2_t;
const char dest_name[] = "float_e5m2_t";
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, f16_to_fe5m2_rn_array) {
int const kN = 27;
using Source = cutlass::half_t;
@@ -289,6 +325,15 @@ TEST(NumericConversion, bf16_to_fe4m3_rn) {
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, bf16_to_fe4m3_rn_2_elements) {
int const kN = 27;
using Source = cutlass::bfloat16_t;
const char source_name[] = "bfloat16_t";
using Destination = cutlass::float_e4m3_t;
const char dest_name[] = "float_e4m3_t";
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, bf16_to_fe4m3_rn_array) {
int const kN = 27;
using Source = cutlass::bfloat16_t;
@@ -307,6 +352,15 @@ TEST(NumericConversion, bf16_to_fe5m2_rn) {
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, bf16_to_fe5m2_rn_2_elements) {
int const kN = 27;
using Source = cutlass::bfloat16_t;
const char source_name[] = "bfloat16_t";
using Destination = cutlass::float_e5m2_t;
const char dest_name[] = "float_e5m2_t";
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, bf16_to_fe5m2_rn_array) {
int const kN = 27;
using Source = cutlass::bfloat16_t;
@@ -327,6 +381,15 @@ TEST(NumericConversion, fe4m3_to_fe5m2_rn) {
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, fe4m3_to_fe5m2_2_elements) {
int const kN = 27;
using Source = cutlass::float_e4m3_t;
const char source_name[] = "float_e4m3_t";
using Destination = cutlass::float_e5m2_t;
const char dest_name[] = "float_e5m2_t";
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, fe4m3_to_fe5m2_array) {
int const kN = 27;
using Source = cutlass::float_e4m3_t;
@@ -345,6 +408,15 @@ TEST(NumericConversion, fe5m2_to_fe4m3_rn) {
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, fe5m2_to_fe4m3_2_elements) {
int const kN = 27;
using Source = cutlass::float_e5m2_t;
const char source_name[] = "float_e5m2_t";
using Destination = cutlass::float_e4m3_t;
const char dest_name[] = "float_e4m3_t";
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, fe5m2_to_fe4m3_array) {
int const kN = 27;
using Source = cutlass::float_e5m2_t;
@@ -375,6 +447,15 @@ TEST(NumericConversion, f32x8_to_s8x8_rn) {
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, fe4m3_to_f32_2_elements) {
int const kN = 2;
using Source = cutlass::float_e4m3_t;
const char source_name[] = "float_e4m3_t";
using Destination = float;
const char dest_name[] = "float";
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, fe4m3_to_f32_array) {
int const kN = 27;
using Source = cutlass::float_e4m3_t;
@@ -384,6 +465,15 @@ TEST(NumericConversion, fe4m3_to_f32_array) {
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, fe5m2_to_f32_2_elements) {
int const kN = 2;
using Source = cutlass::float_e5m2_t;
const char source_name[] = "float_e5m2_t";
using Destination = float;
const char dest_name[] = "float";
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, fe5m2_to_f32_array) {
int const kN = 27;
using Source = cutlass::float_e5m2_t;
@@ -402,6 +492,15 @@ TEST(NumericConversion, fe4m3_to_f16_rn) {
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, fe4m3_to_f16_2_elements) {
int const kN = 2;
using Source = cutlass::float_e4m3_t;
const char source_name[] = "float_e4m3_t";
using Destination = cutlass::half_t;
const char dest_name[] = "half_t";
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, fe4m3_to_f16_array) {
int const kN = 27;
using Source = cutlass::float_e4m3_t;
@@ -420,6 +519,15 @@ TEST(NumericConversion, fe5m2_to_f16_rn) {
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, fe5m2_to_f16_2_elements) {
int const kN = 2;
using Source = cutlass::float_e5m2_t;
const char source_name[] = "float_e5m2_t";
using Destination = cutlass::half_t;
const char dest_name[] = "half_t";
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, fe5m2_to_f16_array) {
int const kN = 27;
using Source = cutlass::float_e5m2_t;
@@ -438,6 +546,15 @@ TEST(NumericConversion, fe4m3_to_bf16_rn) {
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, fe4m3_to_bf16_2_elements) {
int const kN = 2;
using Source = cutlass::float_e4m3_t;
const char source_name[] = "float_e4m3_t";
using Destination = cutlass::bfloat16_t;
const char dest_name[] = "bfloat16_t";
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, fe4m3_to_bf16_array) {
int const kN = 27;
using Source = cutlass::float_e4m3_t;
@@ -456,6 +573,15 @@ TEST(NumericConversion, fe5m2_to_bf16_rn) {
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, fe5m2_to_bf16_2_elements) {
int const kN = 2;
using Source = cutlass::float_e5m2_t;
const char source_name[] = "float_e5m2_t";
using Destination = cutlass::bfloat16_t;
const char dest_name[] = "bfloat16_t";
test::core::kernel::run_test<Destination, Source, kN>(dest_name, source_name);
}
TEST(NumericConversion, fe5m2_to_bf16_array) {
int const kN = 27;
using Source = cutlass::float_e5m2_t;