cutlass 1.3.1 (NVIDIA#46)

CUTLASS 1.3.1 patch resolves failing text with NVRTC.
mfkiwl · Apr 19, 2019 · fe3438a · fe3438a
1 parent 877bdca
commit fe3438a
Show file tree

Hide file tree

Showing 13 changed files with 133 additions and 79 deletions.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -1,5 +1,8 @@
 # NVIDIA CUTLASS Changelog
 
+## [1.3.1](https://github.com/NVIDIA/cutlass/releases/tag/v1.3.1) (2019-04-09)
+ * Corrected NVRTC unit tests.
+
 ## [1.3.0](https://github.com/NVIDIA/cutlass/releases/tag/v1.3.0) (2019-03-20)
  * Efficient GEMM kernel targeting Volta Tensor Cores via `mma.sync` instruction added in CUDA 10.1.
 

diff --git a/README.md b/README.md
@@ -2,7 +2,7 @@
 
 # CUTLASS 1.3
 
-_CUTLASS 1.3.0 - March 2019_
+_CUTLASS 1.3.1 - April 2019_
 
 CUTLASS is a collection of CUDA C++ template abstractions for implementing
 high-performance matrix-multiplication (GEMM) at all levels and scales within CUDA.
@@ -28,6 +28,10 @@ CUTLASS 1.3 is described in the [CUTLASS Documentation](CUTLASS.md) and the acco
 We describe the structure of an efficient GEMM in our talk at the
 [GPU Technology Conference 2018](http://on-demand.gputechconf.com/gtc/2018/presentation/s8854-cutlass-software-primitives-for-dense-linear-algebra-at-all-levels-and-scales-within-cuda.pdf).
 
+# What's New in CUTLASS 1.3.1
+_April 2019_
+* CUTLASS 1.3.1 corrected NVRTC unit tests..
+
 # What's New in CUTLASS 1.3
 _March 2019_
 * CUTLASS 1.3 includes an efficient GEMM implementation with the `mma.sync` instruction added in CUDA 10.1.

diff --git a/cutlass/cutlass.h b/cutlass/cutlass.h
@@ -34,7 +34,7 @@
 
 #define CUTLASS_MAJOR 1
 #define CUTLASS_MINOR 3
-#define CUTLASS_PATCH 0
+#define CUTLASS_PATCH 1
 #define CUTLASS_VERSION ((CUTLASS_MAJOR)*100 + (CUTLASS_MINOR)*10 + CUTLASS_PATCH)
 
 #ifdef __NVCC__
@@ -58,8 +58,13 @@
 
 // CUTLASS_PRAGMA_(UNROLL|NO_UNROLL) optimization directives for the CUDA compiler.
 #if defined(__CUDA_ARCH__)
+#ifdef __NVCC__
     #define CUTLASS_PRAGMA_UNROLL #pragma unroll
     #define CUTLASS_PRAGMA_NO_UNROLL #pragma unroll 1
+#elif defined(__CUDACC_RTC__)
+    #define CUTLASS_PRAGMA_UNROLL _Pragma("unroll")
+    #define CUTLASS_PRAGMA_NO_UNROLL _Pragma("unroll 1")
+#endif
 
     #define CUTLASS_GEMM_LOOP CUTLASS_PRAGMA_NO_UNROLL
 
@@ -80,6 +85,7 @@ template <typename T>
 struct DebugType {};
 
 template <typename T>
+CUTLASS_HOST_DEVICE
 void DebugTypeFunc(T const& t) {
   T::t;
 }

diff --git a/cutlass/gemm/gemm.h b/cutlass/gemm/gemm.h
@@ -33,7 +33,6 @@
 
 #include "cutlass/coord.h"
 #include "cutlass/util/platform.h"
-#include <cstdio>
 namespace cutlass {
 namespace gemm {
 
@@ -84,6 +83,7 @@ void gemm_kernel_nolb(typename Gemm_::Params params) {
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 
+#if !defined(__CUDACC_RTC__)
 /// Partial specialization for launching the GEMM kernel with or without launch bounds
 template <typename Gemm, bool WithLaunchBounds>
 struct Launch {
@@ -152,7 +152,51 @@ struct Launch<Gemm, false> {
       smem_size,
       stream >>>(params);
   }
+
+  // Use device API to launch kernel
+  Launch(cudaError_t &result, CUfunction kernel,
+         typename Gemm::Params params, dim3 grid, dim3 block, CUstream stream = CU_STREAM_LEGACY) {
+    void* params_[] = {const_cast<void*>(reinterpret_cast<void const*>(&params))};
+
+    int smem_size = int(sizeof(typename Gemm::SharedStorage));
+    if (smem_size >= (48 << 10)) {
+
+      result = cudaFuncSetAttribute(
+        kernel,
+        cudaFuncAttributeMaxDynamicSharedMemorySize,
+        smem_size
+      );
+
+      if (result != cudaSuccess) {
+        return;
+      }
+
+      result = cudaFuncSetAttribute(
+        kernel,
+        cudaFuncAttributePreferredSharedMemoryCarveout,
+        100);
+
+      if (result != cudaSuccess) {
+        return;
+      }
+    }
+
+    CUresult launch_result = cuLaunchKernel(
+        kernel,
+        grid.x, grid.y, grid.z,
+        block.x, block.y, block.z,
+        smem_size, stream, params_, 0);
+
+    if (launch_result != CUDA_SUCCESS) {
+      result = cudaErrorLaunchFailure;
+      return;
+    }
+
+    result = cudaSuccess;
+    return;
+  }
 };
+#endif
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 
@@ -188,20 +232,13 @@ struct Gemm {
   static __host__ cudaError_t launch(CUfunction kernel,
                                      Params const& params,
                                      CUstream stream = CU_STREAM_LEGACY) {
+    cudaError_t result;
 
     // Launch the kernel.
-    void* params_[] = {const_cast<void*>(reinterpret_cast<void const*>(&params))};
-
-    CUresult result = cuLaunchKernel(
-        kernel,
-        params.grid.x, params.grid.y, params.grid.z,
-        params.block.x, params.block.y, params.block.z,
-        0, stream, params_, 0);
+    Launch<KernelClass, Traits::GemmConfig::kLaunchBounds>(
+      result, kernel, params, params.grid, params.block, stream);
 
-    if (result != CUDA_SUCCESS) {
-      return cudaErrorLaunchFailure;
-    }
-    return cudaSuccess;
+    return result;
   }
 
 #endif

diff --git a/cutlass/gemm/scalar_or_pointer.h b/cutlass/gemm/scalar_or_pointer.h
@@ -1,3 +1,4 @@
+
 /***************************************************************************************************
  * Copyright (c) 2017-2019, NVIDIA CORPORATION.  All rights reserved.
  *

diff --git a/cutlass/layout/thread/transform.h b/cutlass/layout/thread/transform.h
@@ -77,6 +77,7 @@ struct Copy {
   }
 };
 
+#if !defined(__CUDACC_RTC__) || defined(CUTLASS_NVRTC_HAS_FP16)
 template <int rank>
 struct Copy<half, half, rank, cutlass::MatrixLayout::RowMajor, cutlass::MatrixLayout::RowMajor> {
   CUTLASS_DEVICE void copy(cutlass::TensorView<half, rank, cutlass::MatrixLayout::RowMajor> dst,
@@ -140,6 +141,7 @@ struct Copy<half, half, 2, cutlass::MatrixLayout::RowMajor, cutlass::MatrixLayou
     }
   }
 };
+#endif
 
 /// igemm swizzle
 /// Transform a fragment.
@@ -239,6 +241,7 @@ struct Transform {
   }
 };
 
+#if !defined(__CUDACC_RTC__) || defined(CUTLASS_NVRTC_HAS_FP16)
 template <typename Shape, int Rank, typename DstLayout, typename SrcLayout>
 struct Transform<Shape, Rank, half, DstLayout, half, SrcLayout> {
   typedef Fragment<half, ShapeCount<Shape>::kCount> DstFragment;
@@ -266,6 +269,7 @@ struct Transform<Shape, Rank, half, DstLayout, half, SrcLayout> {
     Transformer.copy(dstView, srcView);
   }
 };
+#endif
 
 template <typename Shape, int Rank, typename DstLayout, typename SrcLayout>
 struct Transform<Shape, Rank, int8_t, DstLayout, int8_t, SrcLayout> {

diff --git a/cutlass/tensor_view.h b/cutlass/tensor_view.h
@@ -36,8 +36,6 @@
 
 #pragma once
 
-#include <cmath>
-
 #include "cutlass/cutlass.h"
 #include "cutlass/tensor_ref.h"
 

diff --git a/cutlass/tile_iterator.h b/cutlass/tile_iterator.h
@@ -34,7 +34,6 @@
 #include "cutlass/load_store.h"
 #include "cutlass/predicate_vector.h"
 #include "cutlass/vector.h"
-#include <cstdio>
 
 namespace cutlass {
 

diff --git a/cutlass/util/performance_tuning.h b/cutlass/util/performance_tuning.h
diff --git a/cutlass/vector.h b/cutlass/vector.h
@@ -88,6 +88,8 @@ union Vector {
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 
+#if !defined(__CUDACC_RTC__) || defined(CUTLASS_NVRTC_HAS_FP16)
+
 template <>
 union Vector<half, 1> {
   /// The scalar type.
@@ -118,7 +120,6 @@ union Vector<half, 1> {
   }
 };
 
-#if !defined(__CUDACC_RTC__) || defined(CUTLASS_NVRTC_HAS_FP16)
 
 template <int kLanes_>
 union Vector<half, kLanes_> {

diff --git a/tools/nvrtc/CMakeLists.txt b/tools/nvrtc/CMakeLists.txt
@@ -54,7 +54,7 @@ if (CUTLASS_NVRTC_ENABLE)
   string(APPEND NVRTC_INCLUDES_STRINGS "char const *kCutlassHeaders[] = {\n")
   string(APPEND NVRTC_INCLUDES_NAMES "char const *kCutlassHeaderNames[] = {\n")
 
-  add_nvrtc_headers(${CMAKE_SOURCE_DIR} "${CUTLASS_CORE};${CUTLASS_GEMM};${CUTLASS_UTIL};${CUTLASS_DEVICE}")
+  add_nvrtc_headers(${CMAKE_SOURCE_DIR} "${CUTLASS_CORE};${CUTLASS_GEMM};${CUTLASS_UTIL};${CUTLASS_DEVICE};${CUTLASS_ARCH};${CUTLASS_LAYOUT_THREAD}")
   message("${CMAKE_CURRENT_SOURCE_DIR}/")
   add_nvrtc_headers("${CMAKE_CURRENT_SOURCE_DIR}/stdlib" "assert.h;stdint.h")
   if(CUTLASS_NVRTC_HAS_CUDA_FP16)

diff --git a/tools/test/unit/gemm/gemm_nvrtc.cu b/tools/test/unit/gemm/gemm_nvrtc.cu
@@ -43,6 +43,8 @@ TEST(Dgemm_nvrtc_64x32x8, dgemm_nvrtc_64x32x8_nt) {
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 
+#if (!defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 610))
+
 TEST(Igemm__nvrtc_128x128x32, igemm_nvrtc_256x256x64_tt) {
   typedef cutlass::gemm::IgemmTraits<cutlass::MatrixLayout::kRowMajor,
     cutlass::MatrixLayout::kRowMajor,
@@ -52,6 +54,8 @@ TEST(Igemm__nvrtc_128x128x32, igemm_nvrtc_256x256x64_tt) {
   run_gemm_nvrtc<IgemmTraits>(gemm_traits, 256, 256, 64);
 }
 
+#endif
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 
 TEST(Sgemm_nvrtc_128x128x8, sgemm_nvrtc_128x112x16_alpha2_beta1_nt) {