CUTLASS 2.0 (#62)
CUTLASS 2.0 Substantially refactored for - Better performance, particularly for native Turing Tensor Cores - Robust and durable templates spanning the design space - Encapsulated functionality embodying modern C++11 programming techniques - Optimized containers and data types for efficient, generic, portable device code Updates to: - Quick start guide - Documentation - Utilities - CUTLASS Profiler Native Turing Tensor Cores - Efficient GEMM kernels targeting Turing Tensor Cores - Mixed-precision floating point, 8-bit integer, 4-bit integer, and binarized operands Coverage of existing CUTLASS functionality: - GEMM kernels targeting CUDA and Tensor Cores in NVIDIA GPUs - Volta Tensor Cores through native mma.sync and through WMMA API - Optimizations such as parallel reductions, threadblock rasterization, and intra-threadblock reductions - Batched GEMM operations - Complex-valued GEMMs Note: this commit and all that follow require a host compiler supporting C++11 or greater.
This commit is contained in:
Andrew Kerr
@@ -0,0 +1,247 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * Redistributions in binary form must reproduce the above copyright notice, this list of
|
||||
* conditions and the following disclaimer in the documentation and/or other materials
|
||||
* provided with the distribution.
|
||||
* * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
|
||||
* to endorse or promote products derived from this software without specific prior written
|
||||
* permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
|
||||
* IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
|
||||
* FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
|
||||
* BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
|
||||
* OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
|
||||
* STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
/*! \file
|
||||
\brief Statically sized array of elements that accommodates all CUTLASS-supported numeric types
|
||||
and is safe to use in a union.
|
||||
*/
|
||||
|
||||
#include "../common/cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/array.h"
|
||||
#include "cutlass/util/device_memory.h"
|
||||
#pragma warning( disable : 4800)
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
namespace test {
|
||||
namespace core {
|
||||
|
||||
/// Each thread clears its array and writes to global memory. No PRMT instructions should
|
||||
/// be generated if Array<T, N> is a multiple of 32 bits.
|
||||
template <typename T, int N>
|
||||
__global__ void test_array_clear(cutlass::Array<T, N> *ptr) {
|
||||
|
||||
cutlass::Array<T, N> storage;
|
||||
|
||||
storage.clear();
|
||||
|
||||
ptr[threadIdx.x] = storage;
|
||||
}
|
||||
|
||||
/// Each thread writes its thread index into the elements of its array and then writes the result
|
||||
/// to global memory.
|
||||
template <typename T, int N>
|
||||
__global__ void test_array_threadid(cutlass::Array<T, N> *ptr) {
|
||||
|
||||
cutlass::Array<T, N> storage;
|
||||
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int i = 0; i < N; ++i) {
|
||||
storage.at(i) = T(int(threadIdx.x));
|
||||
}
|
||||
|
||||
ptr[threadIdx.x] = storage;
|
||||
}
|
||||
|
||||
/// Each thread writes its thread index into the elements of its array and then writes the result
|
||||
/// to global memory.
|
||||
template <typename T, int N>
|
||||
__global__ void test_array_sequence(cutlass::Array<T, N> *ptr) {
|
||||
|
||||
cutlass::Array<T, N> storage;
|
||||
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int i = 0; i < N; ++i) {
|
||||
storage.at(i) = T(i);
|
||||
}
|
||||
|
||||
ptr[threadIdx.x] = storage;
|
||||
}
|
||||
|
||||
} // namespace core
|
||||
} // namespace test
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
template <typename T, int N>
|
||||
class TestArray {
|
||||
public:
|
||||
|
||||
//
|
||||
// Data members
|
||||
//
|
||||
|
||||
/// Number of threads
|
||||
int const kThreads = 32;
|
||||
|
||||
typedef cutlass::Array<T, N> ArrayTy;
|
||||
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
|
||||
/// Ctor
|
||||
TestArray() {
|
||||
|
||||
}
|
||||
|
||||
/// Runs the test
|
||||
void run() {
|
||||
|
||||
/// Device memory containing output
|
||||
cutlass::device_memory::allocation< ArrayTy > output(kThreads);
|
||||
std::vector< ArrayTy > output_host(kThreads);
|
||||
|
||||
dim3 grid(1,1);
|
||||
dim3 block(kThreads, 1, 1);
|
||||
|
||||
test::core::test_array_clear<<< grid, block >>>(output.get());
|
||||
|
||||
cudaError_t result = cudaDeviceSynchronize();
|
||||
ASSERT_EQ(result, cudaSuccess) << "CUDA error: " << cudaGetErrorString(result);
|
||||
|
||||
//
|
||||
// Verify contains all zeros
|
||||
//
|
||||
|
||||
cutlass::device_memory::copy_to_host(output_host.data(), output.get(), kThreads);
|
||||
|
||||
result = cudaGetLastError();
|
||||
ASSERT_EQ(result, cudaSuccess) << "CUDA error: " << cudaGetErrorString(result);
|
||||
|
||||
char const *ptr_host = reinterpret_cast<char const *>(output_host.data());
|
||||
for (int i = 0; i < sizeof(ArrayTy) * kThreads; ++i) {
|
||||
EXPECT_FALSE(ptr_host[i]);
|
||||
}
|
||||
|
||||
//
|
||||
// Verify each element contains the low bits of the thread Id
|
||||
//
|
||||
|
||||
test::core::test_array_threadid<<< grid, block >>>(output.get());
|
||||
|
||||
result = cudaDeviceSynchronize();
|
||||
ASSERT_EQ(result, cudaSuccess) << "CUDA error: " << cudaGetErrorString(result);
|
||||
|
||||
cutlass::device_memory::copy_to_host(output_host.data(), output.get(), kThreads);
|
||||
|
||||
result = cudaGetLastError();
|
||||
ASSERT_EQ(result, cudaSuccess) << "CUDA error: " << cudaGetErrorString(result);
|
||||
|
||||
for (int i = 0; i < kThreads; ++i) {
|
||||
T tid = T(i);
|
||||
|
||||
ArrayTy thread = output_host.at(i);
|
||||
|
||||
// Element-wise access
|
||||
for (int j = 0; j < N; ++j) {
|
||||
EXPECT_TRUE(tid == thread[j]);
|
||||
}
|
||||
|
||||
// Iterator access
|
||||
for (auto it = thread.begin(); it != thread.end(); ++it) {
|
||||
EXPECT_TRUE(tid == *it);
|
||||
}
|
||||
|
||||
// Range-based for
|
||||
for (auto const & x : thread) {
|
||||
EXPECT_TRUE(tid == x);
|
||||
}
|
||||
}
|
||||
|
||||
//
|
||||
// Verify each element
|
||||
//
|
||||
|
||||
test::core::test_array_sequence<<< grid, block >>>(output.get());
|
||||
|
||||
result = cudaDeviceSynchronize();
|
||||
ASSERT_EQ(result, cudaSuccess) << "CUDA error: " << cudaGetErrorString(result);
|
||||
|
||||
cutlass::device_memory::copy_to_host(output_host.data(), output.get(), kThreads);
|
||||
|
||||
result = cudaGetLastError();
|
||||
ASSERT_EQ(result, cudaSuccess) << "CUDA error: " << cudaGetErrorString(result);
|
||||
|
||||
for (int i = 0; i < kThreads; ++i) {
|
||||
|
||||
ArrayTy thread = output_host.at(i);
|
||||
|
||||
// Element-wise access
|
||||
for (int j = 0; j < N; ++j) {
|
||||
T got = T(j);
|
||||
EXPECT_TRUE(got == thread[j]);
|
||||
}
|
||||
|
||||
// Iterator access
|
||||
int j = 0;
|
||||
for (auto it = thread.begin(); it != thread.end(); ++it, ++j) {
|
||||
T got = T(j);
|
||||
EXPECT_TRUE(got == *it);
|
||||
}
|
||||
|
||||
// Range-based for
|
||||
j = 0;
|
||||
for (auto const & x : thread) {
|
||||
T got = T(j);
|
||||
EXPECT_TRUE(got == x);
|
||||
++j;
|
||||
}
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(Array, Int8x16) {
|
||||
TestArray<int8_t, 16>().run();
|
||||
}
|
||||
|
||||
TEST(Array, Int32x4) {
|
||||
TestArray<int, 4>().run();
|
||||
}
|
||||
|
||||
#if __CUDA_ARCH__ >= 520
|
||||
TEST(Array, Float16x8) {
|
||||
TestArray<cutlass::half_t, 8>().run();
|
||||
}
|
||||
#endif
|
||||
|
||||
TEST(Array, Float32x4) {
|
||||
TestArray<float, 4>().run();
|
||||
}
|
||||
|
||||
TEST(Array, Int4x32) {
|
||||
TestArray<cutlass::int4b_t, 32>().run();
|
||||
}
|
||||
|
||||
TEST(Array, Uint4x32) {
|
||||
TestArray<cutlass::uint4b_t, 32>().run();
|
||||
}
|
||||
|
||||
TEST(Array, Bin1x128) {
|
||||
TestArray<cutlass::bin1_t, 128>().run();
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
Reference in New Issue
Block a user