CUTLASS 2.2 (#96)

Adds support for NVIDIA Ampere Architecture features. CUDA 11 Toolkit recommended.

examples/03_visualize_layout/CMakeLists.txt

@@ -1,4 +1,4 @@
-# Copyright (c) 2017-2019, NVIDIA CORPORATION.  All rights reserved.
+# Copyright (c) 2017-2020, 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:

examples/03_visualize_layout/options.h

@@ -1,5 +1,5 @@
 /***************************************************************************************************
- * Copyright (c) 2017-2019, NVIDIA CORPORATION.  All rights reserved.
+ * Copyright (c) 2017-2020, 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:

examples/03_visualize_layout/register_layout.cu

@@ -1,5 +1,5 @@
 /***************************************************************************************************
- * Copyright (c) 2017-2019, NVIDIA CORPORATION.  All rights reserved.
+ * Copyright (c) 2017-2020, 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:
@@ -34,6 +34,8 @@
 #include "cutlass/layout/pitch_linear.h"
 #include "cutlass/layout/tensor_op_multiplicand_sm70.h"
 #include "cutlass/layout/tensor_op_multiplicand_sm75.h"
+#include "cutlass/layout/tensor_op_multiplicand_sm80.h"
+
 #include "visualize_layout.h"
 #include "register_layout.h"
 
@@ -59,18 +61,40 @@ void RegisterLayouts(std::map<std::string, std::unique_ptr<VisualizeLayoutBase>
       // Integer matrix multiply.int4 8832  TN kblock128
       {"TensorOpMultiplicand<4,128>",
        new VisualizeLayout<cutlass::layout::TensorOpMultiplicand<4, 128>>},
+      // Integer matrix multiply.int4 16864 TN kblock256
+      {"TensorOpMultiplicand<4,256>",
+       new VisualizeLayout<cutlass::layout::TensorOpMultiplicand<4, 256>>},
       // Integer matrix multiply 8816  Interleaved-32
       {"TensorOpMultiplicand<8,32>",
        new VisualizeLayout<cutlass::layout::TensorOpMultiplicand<8, 32>>},
       // Integer matrix multiply 8816  TN kblock64
       {"TensorOpMultiplicand<8,64>",
        new VisualizeLayout<cutlass::layout::TensorOpMultiplicand<8, 64>>},
+      {"TensorOpMultiplicand<8,128>",
+       new VisualizeLayout<cutlass::layout::TensorOpMultiplicand<8, 128>>},
       // Matrix Multiply 1688  TN kblock32
       {"TensorOpMultiplicand<16,32>",
        new VisualizeLayout<cutlass::layout::TensorOpMultiplicand<16, 32>>},
       // Matrix multiply 1688  NT
       {"TensorOpMultiplicand<16,64>",
        new VisualizeLayout<cutlass::layout::TensorOpMultiplicand<16, 64>>},
+      // Matrix multiply 1688.TF32 TN kblock16
+      {"TensorOpMultiplicand<32,16>",
+       new VisualizeLayout<cutlass::layout::TensorOpMultiplicand<32, 16>>},
+      // Matrix multiply 1688.TF32 TN kblock32
+      {"TensorOpMultiplicand<32,32>",
+       new VisualizeLayout<cutlass::layout::TensorOpMultiplicand<32, 32>>},
+      // Matrix multiply 1688 NT
+      {"TensorOpMultiplicandCongruous<32,32>",
+       new VisualizeLayout<
+           cutlass::layout::TensorOpMultiplicandCongruous<32, 32>>},
+      // Matrix multiply 884 NT
+      {"TensorOpMultiplicandCongruous<64,16>",
+       new VisualizeLayout<
+           cutlass::layout::TensorOpMultiplicandCongruous<64, 16>>},
+      // Matrix multiply 884 TN
+      {"TensorOpMultiplicand64bCrosswise",
+       new VisualizeLayout<cutlass::layout::TensorOpMultiplicand64bCrosswise>},
       {"TensorOpMultiplicandCongruous<128,4>",
        new VisualizeLayout<
            cutlass::layout::TensorOpMultiplicandCongruous<128, 4>>},
@@ -82,7 +106,7 @@ void RegisterLayouts(std::map<std::string, std::unique_ptr<VisualizeLayoutBase>
            cutlass::layout::VoltaTensorOpMultiplicandCongruous<16>>},
       {"VoltaTensorOpMultiplicandCrosswise<16,32>",
        new VisualizeLayout<
-           cutlass::layout::VoltaTensorOpMultiplicandCrosswise<16, 32>>},
+           cutlass::layout::VoltaTensorOpMultiplicandCrosswise<16, 32>>}
   };
 
   for (auto layout : layout_pairs) {

examples/03_visualize_layout/register_layout.h

@@ -1,5 +1,5 @@
 /***************************************************************************************************
- * Copyright (c) 2017-2019, NVIDIA CORPORATION.  All rights reserved.
+ * Copyright (c) 2017-2020, 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:

examples/03_visualize_layout/visualize_layout.cpp

@@ -1,5 +1,5 @@
 /***************************************************************************************************
- * Copyright (c) 2017-2019, NVIDIA CORPORATION.  All rights reserved.
+ * Copyright (c) 2017-2020, 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:
@@ -65,14 +65,26 @@ void print_usage(std::ostream &out) {
          "--extent=64,64 --vectorize=32 --output-shape=256,4\n"
       << "$ 03_visualize_layout \"TensorOpMultiplicand<4,128>\" "
          "--extent=128,32 --vectorize=32 --output-shape=256,4\n"
+      << "$ 03_visualize_layout \"TensorOpMultiplicand<4,256>\" "
+         "--extent=256,16 --vectorize=32 --output-shape=256,4\n"
       << "$ 03_visualize_layout \"TensorOpMultiplicand<8,32>\" "
          "--extent=32,64 --vectorize=16 --output-shape=128,4\n"
       << "$ 03_visualize_layout \"TensorOpMultiplicand<8,64>\" "
          "--extent=64,32 --vectorize=16 --output-shape=128,4\n"
+      << "$ 03_visualize_layout \"TensorOpMultiplicand<8,128>\" "
+         "--extent=128,16 --vectorize=16 --output-shape=128,4\n"
       << "$ 03_visualize_layout \"TensorOpMultiplicand<16,32>\" "
          "--extent=32,32 --vectorize=8 --output-shape=64,4\n"
       << "$ 03_visualize_layout \"TensorOpMultiplicand<16,64>\" "
          "--extent=64,16 --vectorize=8 --output-shape=64,4\n"
+      << "$ 03_visualize_layout \"TensorOpMultiplicand<32,16>\" "
+         "--extent=16,32 --vectorize=4 --output-shape=32,4\n"
+      << "$ 03_visualize_layout \"TensorOpMultiplicand<32,32>\" "
+         "--extent=32,16 --vectorize=4 --output-shape=32,4\n"
+      << "$ 03_visualize_layout \"TensorOpMultiplicandCongruous<32,32>\" "
+         "--extent=32,16 --vectorize=4 --output-shape=32,4\n"
+      << "$ 03_visualize_layout \"TensorOpMultiplicandCongruous<64, 16>\" "
+         "--extent=16,16 --vectorize=2 --output-shape=16,4\n"
       << "$ 03_visualize_layout \"VoltaTensorOpMultiplicandCrosswise<16,32>\" "
          "--extent=32,64 --vectorize=4 --output-shape=64,4\n"
       << "$ 03_visualize_layout \"VotlaTensorOpMultiplicandCongruous<16>\" "

examples/03_visualize_layout/visualize_layout.h

@@ -1,5 +1,5 @@
 /***************************************************************************************************
- * Copyright (c) 2017-2019, NVIDIA CORPORATION.  All rights reserved.
+ * Copyright (c) 2017-2020, 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: