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

More updates for 3.1 (#958)

Browse Source 
* Updates for 3.1

* Minor change

* doc link fix

* Minor updates
This commit is contained in:
ANIKET SHIVAM
2023-05-24 10:17:16 -04:00
committed by GitHub
parent 13f413493a
commit f079619f5e
48 changed files with 1611 additions and 1858 deletions
Download Patch File Download Diff File Expand all files Collapse all files
media/docs/cutlass_3x_backwards_compatibility.md
+2 -2
View File
@@ -156,11 +156,11 @@ static bool
can_implement(Arguments const& args);
// Returns a dim3 representing the threadblock shape.
static constexpr dim3
static dim3
get_block_shape();
// Returns a dim3 representing the grid shape in terms of threadblocks.
static constexpr dim3
static dim3
get_grid_shape(Params const& params);
```
media/docs/profiler.md
+16 -4
View File
@@ -353,8 +353,9 @@ Hopper architecture and beyond so as to indicate new features of the kernel with
To best illustrate this naming convention, we will walk through the meaning of each of the components
in a GEMM kernel used by the profiler:
```
cutlass3x_sm90_tensorop_s64x128x16gemm_f16_f16_f32_f32_128x128x64_2x1x1_0_ntn_align8
cutlass3x_sm90_tensorop_s64x128x16gemm_f16_f16_f32_f16_f32_128x128x64_2x1x1_0_ntn_align8
```
The components within this name are as follows:
@@ -366,8 +367,7 @@ The components within this name are as follows:
* `s`: indicates that the Tensor Core instruction being used accumulates in single precision
(as opposed to `h`, which indicates half precision)
* `64x128x16gemm`: indicates that the shape of the Tensor Core instruction being used (MxNxK) is 64x128x16
* `f16_f16_f32_f16`: indicates that the data types for operands A, B, and C are each `f16`
(half precision) and that accumulation is performed using `f32` (single precision)
* `f16_f16_f32_f16_f16`: indicates that the data types for operands A, B, Accumulator, C and D (in that order).
* `128x128x64`: indicates that the thread block shape used in the GEMM (MxNxK) is 128x128x64
* `2x1x1`: indicates that the cluster shape being used is 2x1x1
* `0`: indicates that the kernel uses the CollectiveBuilder's automatic stage calculation to determine the
@@ -382,12 +382,24 @@ Note that in some special cases where the input A/B types do not match that of t
instruction's, the MMA facing input type is added to the instruction string as well.
```
cutlass3x_sm90_tensorop_s64x128x8tf32gemm_f32_f32_f32_f32_128x128x32_2x1x1_0_tnn_align4
cutlass3x_sm90_tensorop_s64x128x8tf32gemm_f32_f32_f32_f32_f32_128x128x32_2x1x1_0_tnn_align4
```
* `s64x128x8tf32gemm`: indicates that the MMA consumes inputs in `tf32` format, and therefore
the kernel performs rounding of the `f32` values in global memory while loading them into shared memory.
For custom mainloop or epilogue schedules, details of the opted-in schedule are appended to the end of the
kernel name. For example,
```
cutlass3x_sm90_tensorop_h64x128x16gemm_f16_f16_f16_void_f16_128x128x64_1x1x1_0_nnn_align8_warpspecialized_cooperative_epi_tma
```
* `warpspecialized_cooperative`: Mainloop employs a persistent warp-specialized mainloop and kernel schedule.
* `epi_tma`: Kernel epilogue employs TMA based vectorization.
* `f16_f16_f16_void_f16`: In this case, C type is set to `void`, indicating that residual matrix support
is disabled.
# Convolution
The CUTLASS Profiler is capable of executing 2-D and 3-D convolution problems for forwards and backwards