-
Notifications
You must be signed in to change notification settings - Fork 160
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
The basic MLIR-based end-to-end GPU Tensor Core GEMM codegen. (#1199)
It is the first version that provides the flow of the naive end-to-end GPU GEMM codegen.
- Loading branch information
Showing
28 changed files
with
940 additions
and
201 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
9 changes: 9 additions & 0 deletions
9
...ler/mlir/disc/tests/disc-transform/data/default_schedule_matmul_nn_s_256x256x128_f16.mlir
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,9 @@ | ||
| module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 0 : i32}} { | ||
| func.func @main(%arg0: tensor<256x128xf16>, %arg1: tensor<128x256xf16>) -> (tensor<256x256xf16>) attributes {tf.entry_function = {inputs = "{{INPUTS}}", outputs = "{{OUTPUTS}}", input_placements="{{INPUT_PLACEMENTS}}", output_placements="{{OUTPUT_PLACEMENTS}}"}} { | ||
| %graph = tf_executor.graph { | ||
| %0:2 = tf_executor.island wraps "tf.MatMul"(%arg0, %arg1) {transpose_a = false, transpose_b = false} : (tensor<256x128xf16>, tensor<128x256xf16>) -> (tensor<256x256xf16>) | ||
| tf_executor.fetch %0 : tensor<256x256xf16> | ||
| } | ||
| return %graph : tensor<256x256xf16> | ||
| } | ||
| } |
9 changes: 9 additions & 0 deletions
9
tao_compiler/mlir/disc/tests/disc-transform/data/matmul_nn_s_256x256x128_f16.mlir
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,9 @@ | ||
| module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 0 : i32}} { | ||
| func.func @main(%arg0: tensor<256x128xf16>, %arg1: tensor<128x256xf16>) -> (tensor<256x256xf16>) attributes {tf.entry_function = {inputs = "{{INPUTS}}", outputs = "{{OUTPUTS}}", input_placements="{{INPUT_PLACEMENTS}}", output_placements="{{OUTPUT_PLACEMENTS}}"}} { | ||
| %graph = tf_executor.graph { | ||
| %0:2 = tf_executor.island wraps "tf.MatMul"(%arg0, %arg1) {transpose_a = false, transpose_b = false} : (tensor<256x128xf16>, tensor<128x256xf16>) -> (tensor<256x256xf16>) | ||
| tf_executor.fetch %0 : tensor<256x256xf16> | ||
| } | ||
| return %graph : tensor<256x256xf16> | ||
| } | ||
| } |
45 changes: 45 additions & 0 deletions
45
tao_compiler/mlir/disc/tests/disc-transform/data/matmul_nn_s_f16_gpu_schedule.mlir
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,45 @@ | ||
| transform.sequence failures(propagate) { | ||
| ^bb0(%arg0: !transform.any_op): | ||
| %0 = transform.structured.match attributes {disc.transform.name = "dot_general"} in %arg0 : (!transform.any_op) -> !transform.any_op | ||
| %1:2 = split_handle %0 : (!transform.any_op) -> (!transform.any_op, !transform.any_op) | ||
| %forall_op, %tiled_op = transform.structured.tile_to_forall_op %1#1 num_threads [] tile_sizes [128, 128](mapping = [#gpu.block<x>, #gpu.block<y>]) : (!transform.any_op) -> (!transform.any_op, !transform.any_op) | ||
| %fused_op, %new_containing_op = transform.structured.fuse_into_containing_op %1#0 into %forall_op : (!transform.any_op, !transform.any_op) -> (!transform.any_op, !transform.any_op) | ||
| %for_op, %splitted_op = transform.disc.split_reduction_serial %tiled_op by tile_sizes = [32] : (!transform.any_op) -> (!transform.any_op, !transform.any_op) | ||
| %promoted_dot, %lhs_alloc, %rhs_alloc = transform.disc.promote_dot_operands %for_op [0, 1] : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op) | ||
| %forall_op_0, %tiled_op_1 = transform.structured.tile_to_forall_op %promoted_dot num_threads [] tile_sizes [64, 64](mapping = [#gpu.warp<x>, #gpu.warp<y>]) : (!transform.any_op) -> (!transform.any_op, !transform.any_op) | ||
| %for_op_2, %splitted_op_3 = transform.disc.split_reduction_serial %tiled_op_1 by tile_sizes = [32] : (!transform.any_op) -> (!transform.any_op, !transform.any_op) | ||
| %tiled_linalg_op, %loops:3 = transform.structured.tile %for_op_2[16, 8, 16] {interchange = [0, 1, 2]} : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op, !transform.any_op) | ||
| transform.disc.apply_licm %arg0 : !transform.any_op | ||
| transform.disc.apply_dce %arg0 : !transform.any_op | ||
| transform.disc.apply_cse %arg0 : !transform.any_op | ||
| %2 = transform.structured.match ops{["func.func"]} in %arg0 : (!transform.any_op) -> !transform.any_op | ||
| %3 = transform.disc.apply_patterns %2 {canonicalization} : (!transform.any_op) -> !transform.any_op | ||
| %4 = transform.structured.vectorize %3 {vectorize_padding} : (!transform.any_op) -> !transform.any_op | ||
| transform.disc.apply_dce %arg0 : !transform.any_op | ||
| transform.disc.apply_cse %arg0 : !transform.any_op | ||
| %5 = transform.disc.bufferize {target_gpu} %arg0 : (!transform.any_op) -> !transform.any_op | ||
| %6 = transform.structured.match ops{["func.func"]} in %5 : (!transform.any_op) -> !transform.any_op | ||
| transform.disc.erase_dealloc %6 : (!transform.any_op) -> () | ||
| %7 = transform.structured.match ops{["func.func"]} in %5 : (!transform.any_op) -> !transform.any_op | ||
| transform.disc.transfer_write_zero_to_scf %7 : (!transform.any_op) -> () | ||
| transform.disc.apply_dce %5 : !transform.any_op | ||
| transform.disc.apply_cse %5 : !transform.any_op | ||
| %8 = transform.structured.match ops{["scf.forall"]} attributes {mapping = [#gpu.block<x>, #gpu.block<y>]} in %5 : (!transform.any_op) -> !transform.any_op | ||
| %9 = transform.disc.forall_to_gpu_ctas %8 : (!transform.any_op) -> !transform.any_op | ||
| %10 = transform.structured.match ops{["scf.forall"]} attributes {mapping = [#gpu.warp<x>, #gpu.warp<y>]} in %5 : (!transform.any_op) -> !transform.any_op | ||
| transform.disc.forall_to_gpu_warps %10 : (!transform.any_op) -> () | ||
| transform.disc.apply_dce %5 : !transform.any_op | ||
| transform.disc.apply_cse %5 : !transform.any_op | ||
| %11 = transform.structured.match ops{["linalg.generic"]} in %5 : (!transform.any_op) -> !transform.any_op | ||
| transform.disc.gmem_to_smem %11 : (!transform.any_op) -> () | ||
| %12 = transform.structured.match ops{["func.func"]} in %5 : (!transform.any_op) -> !transform.any_op | ||
| transform.disc.vector.vector_to_mma_conversion %12 : (!transform.any_op) -> () | ||
| transform.disc.apply_licm %5 : !transform.any_op | ||
| transform.disc.apply_dce %5 : !transform.any_op | ||
| transform.disc.apply_cse %5 : !transform.any_op | ||
| %13 = transform.structured.match ops{["func.func"]} in %5 : (!transform.any_op) -> !transform.any_op | ||
| transform.disc.inline_and_convert_gpu_ids %13 : (!transform.any_op) -> () | ||
| transform.disc.apply_licm %5 : !transform.any_op | ||
| transform.disc.apply_dce %5 : !transform.any_op | ||
| transform.disc.apply_cse %5 : !transform.any_op | ||
| } |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.