CUTLASS 2.4 (Implicit GEMM convolution) (#147)

CUTLASS 2.4 (Implicit GEMM Convolution) Co-authored-by: Manish Gupta <manigupta@nvidia.com>, Haicheng Wu <haichengw@nvidia.com>, Dustyn Blasig <dblasig@nvidia.com>, Andrew Kerr <akerr@nvidia.com>
2020-11-19 21:25:25 -08:00
parent c2b80ad4e4
commit 6615010cd0
224 changed files with 43939 additions and 1061 deletions
@@ -7,11 +7,15 @@
 ## Prerequisites

 CUTLASS requires:
- NVIDIA CUDA Toolkit (9.2 or later required, [11.0](https://developer.nvidia.com/cuda-toolkit) recommended)
+- NVIDIA CUDA Toolkit (9.2 or later required, [11.1](https://developer.nvidia.com/cuda-toolkit) recommended)
 - CMake 3.12+
 - host compiler supporting C++11 or greater (g++ 7.3.0 or Microsoft Visual Studio 2015 recommended)
 - Python 3.6+

+CUTLASS may be optionally compiled and linked with
+- cuBLAS
+- cuDNN v7.6 or later
+
 ## Initial build steps

 Construct a build directory and run CMake.
@@ -31,6 +35,23 @@ $ cmake .. -DCUTLASS_NVCC_ARCHS=80 -DCUTLASS_ENABLE_TESTS=OFF -DCUTLASS_UNITY_BU

 This reduces overall compilation time by excluding unit tests and enabling the unit build.

+You may reduce build times by compiling only certain operations by setting the `CUTLASS_LIBRARY_OPERATIONS` flag as shown below,
+executed from an empty `build/` directory. This only compiles 2-D convolution kernels.
+
+```bash
+$ cmake .. -DCUTLASS_NVCC_ARCHS=80 -DCUTLASS_LIBRARY_OPERATIONS=conv2d
+```
+
+You may also filter kernels by name by supplying a filter string with flag `CUTLASS_LIBRARY_KERNELS`. 
+
+```bash
+$ cmake .. -DCUTLASS_NVCC_ARCHS=80 -DCUTLASS_LIBRARY_KERNELS=s16816gemm,s16816fprop*128x128
+```
+
+You may explicitly exclude cuBLAS and cuDNN as dependencies with the following CMake flags.
+- `-DCUTLASS_ENABLE_CUBLAS=OFF`
+- `-DCUTLASS_ENABLE_CUDNN=OFF`
+

 ## Build and run the CUTLASS Profiler

@@ -39,7 +60,7 @@ From the `build/` directory created above, compile the the CUTLASS Profiler.
 $ make cutlass_profiler -j12
 ```

-Then execute the CUTLASS Profiler for a set of problem sizes.
+Then execute the CUTLASS Profiler computing GEMM, execute the following command.
 ```bash
 $ ./tools/profiler/cutlass_profiler --kernels=sgemm --m=4352 --n=4096 --k=4096

@@ -66,6 +87,45 @@ $ ./tools/profiler/cutlass_profiler --kernels=sgemm --m=4352 --n=4096 --k=4096
        Math: 13854.9 GFLOP/s
 ```

+To execute the CUTLASS Profiler for Convolution, run the following example.
+```bash
+$ ./tools/profiler/cutlass_profiler --kernels=s1688fprop --n=8 --h=224 --w=224 --c=128 --k=128 --r=3 --s=3 --pad_h=1 --pad_w=1
+```
+
+To execute all CUTLASS 2-D convolution operators, execute the following.
+```bash
+$ ./tools/profiler/cutlass_profiler --operation=conv2d--n=8 --h=224 --w=224 --c=128 --k=128 --r=3 --s=3
+
+
+=============================
+  Problem ID: 1
+
+        Provider: CUTLASS
+   OperationKind: conv2d
+       Operation: cutlass_simt_sfprop_optimized_128x128_8x2_nhwc
+
+          Status: Success
+    Verification: ON
+     Disposition: Passed
+
+reference_device: Passed
+
+       Arguments: --conv_kind=fprop --n=8 --h=224 --w=224 --c=128 --k=128 --r=3 --s=3 --p=224 --q=224 --pad_h=1 --pad_w=1  \
+                  --stride_h=1 --stride_w=1 --dilation_h=1 --dilation_w=1 --Activation=f32:nhwc --Filter=f32:nhwc --Output=f32:nhwc  \
+                  --conv_mode=cross --iterator_algorithm=optimized --alpha=1 --beta=0 --split_k_mode=serial --split_k_slices=1  \
+                  --eq_gemm_provider=none --op_class=simt --accum=f32 --cta_m=128 --cta_n=128 --cta_k=8 --stages=2 --warps_m=4  \
+                  --warps_n=2 --warps_k=1 --inst_m=1 --inst_n=1 --inst_k=1 --min_cc=50 --max_cc=1024
+
+           Bytes: 2055798784  bytes
+           FLOPs: 118482796544  flops
+
+         Runtime: 8.13237  ms
+          Memory: 235.431 GiB/s
+
+            Math: 14569.3 GFLOP/s
+
+```
+
 See [documentation for the CUTLASS Profiler](profiler.md) for more details.

 ## Build and run CUTLASS Unit Tests