forked from onnx/onnx-mlir
-
Notifications
You must be signed in to change notification settings - Fork 0
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
[NNPA] zDNN extension for splitting ztensors (onnx#2688)
* zDNN extension for splitting ztensors. Apply to zdnn matmul in this commit. Signed-off-by: Tung D. Le <[email protected]> --------- Signed-off-by: Tung D. Le <[email protected]>
- Loading branch information
Showing
8 changed files
with
632 additions
and
10 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
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 +1,149 @@ | ||
| /* | ||
| * SPDX-License-Identifier: Apache-2.0 | ||
| */ | ||
|
|
||
| //===-------------------------- MatMul.c ----------------------------------===// | ||
| // | ||
| // Copyright 2024 The IBM Research Authors. | ||
| // | ||
| // ============================================================================= | ||
| // | ||
| // A wrapper of zdnn_matmul_op for ztensor partition and parallelism. | ||
| // | ||
| //===----------------------------------------------------------------------===// | ||
|
|
||
| // Include pthreads (need special treatment on z/OS). | ||
| #ifdef __MVS__ | ||
| #define _OPEN_THREADS | ||
| #endif | ||
| #include <pthread.h> | ||
|
|
||
| #include <assert.h> | ||
| #include <math.h> | ||
| #include <stdio.h> | ||
| #include <stdlib.h> | ||
| #include <sys/time.h> | ||
|
|
||
| #include "zDNNExtension.h" | ||
| #include "zdnn.h" | ||
|
|
||
| #ifdef __cplusplus | ||
| extern "C" { | ||
| #endif | ||
|
|
||
| static inline zdnn_status call_zdnn_matmul_op(const zdnn_ztensor *inputA, | ||
| const zdnn_ztensor *inputB, const zdnn_ztensor *inputC, int opType, | ||
| zdnn_ztensor *output, bool isBcast) { | ||
| if (isBcast) | ||
| return zdnn_matmul_bcast_op( | ||
| inputA, inputB, inputC, (zdnn_matmul_bcast_ops)opType, output); | ||
| return zdnn_matmul_op( | ||
| inputA, inputB, inputC, (zdnn_matmul_ops)opType, output); | ||
| } | ||
|
|
||
| static zdnn_status zdnn_matmul_op_common(const zdnn_ztensor *inputA, | ||
| const zdnn_ztensor *inputB, const zdnn_ztensor *inputC, int opType, | ||
| zdnn_ztensor *output, bool isBcast) { | ||
| // Verify that e4, e3, e1 do not exceed the maximum dimension size. Thus, we | ||
| // will split e2 safely. | ||
| OrigShape origShapeOfA; | ||
| getOrigShape(inputA, &origShapeOfA); | ||
| uint32_t maxDimSize = zdnn_get_nnpa_max_dim_idx_size(); | ||
| if ((origShapeOfA.e4 > maxDimSize) || (origShapeOfA.e3 > maxDimSize) || | ||
| (origShapeOfA.e1 > maxDimSize)) { | ||
| printf("[MatMul] The 1st tensor dimension exceeds maximum dimension index " | ||
| "size (MDIS) of %d: e4 = %d, e3 = %d, e1 = %d.\n", | ||
| maxDimSize, origShapeOfA.e4, origShapeOfA.e3, origShapeOfA.e1); | ||
| return ZDNN_EXCEEDS_MDIS; | ||
| } | ||
|
|
||
| // For a MatMul of (M,N)*(N,P), | ||
| // We split M that is e2 in (e4, e3, e2, e1). | ||
| SplitInfo splitInfoA, splitInfoY; | ||
| splitInfoA.axis = 2; | ||
| splitInfoY.axis = 2; | ||
| splitInfoA.chunkSize = OMZTensorSplitSize; | ||
| splitInfoY.chunkSize = OMZTensorSplitSize; | ||
|
|
||
| // Dim is small or ztensor split is disabled. | ||
| if (!OMZTensorSplitEnabled || !initSplitInfo(inputA, &splitInfoA)) { | ||
| if (OMZTensorSplitDebug) | ||
| printf("[MatMul] Not split zTensor ...\n"); | ||
| return call_zdnn_matmul_op(inputA, inputB, inputC, opType, output, isBcast); | ||
| } | ||
|
|
||
| // Split input A. | ||
| if (OMZTensorSplitDebug) | ||
| printf("[MatMul] Split the 1st ztensor along e2 into %d chunks of %d " | ||
| "elements \n", | ||
| splitInfoA.numOfChunks, splitInfoA.chunkSize); | ||
| initSplitInfo(output, &splitInfoY); | ||
|
|
||
| double splitTime = 0.; | ||
| double mmTime = 0.; | ||
| double mergeTime = 0.; | ||
| clock_t start_time, end_time; | ||
|
|
||
| // Split input A into chunks. | ||
| if (OMZTensorSplitDebug) | ||
| start_time = clock(); | ||
| splitZTensor(inputA, &splitInfoA, /*copyData=*/true); | ||
| splitZTensor(output, &splitInfoY, /*copyData=*/false); | ||
| if (OMZTensorSplitDebug) { | ||
| end_time = clock(); | ||
| splitTime = ((float)(end_time - start_time) / (float)CLOCKS_PER_SEC) * 1000; | ||
| } | ||
|
|
||
| // Call zdnn_matmul_op on each chunk. | ||
| if (OMZTensorSplitDebug) | ||
| start_time = clock(); | ||
| for (uint32_t i = 0; i < splitInfoA.numOfChunks; ++i) { | ||
| zdnn_status status = call_zdnn_matmul_op(splitInfoA.tensors + i, inputB, | ||
| inputC, opType, splitInfoY.tensors + i, isBcast); | ||
| assert(status == ZDNN_OK); | ||
| } | ||
| if (OMZTensorSplitDebug) { | ||
| end_time = clock(); | ||
| mmTime = ((float)(end_time - start_time) / (float)CLOCKS_PER_SEC) * 1000; | ||
| } | ||
|
|
||
| // Merging the chunks into the output. | ||
| if (OMZTensorSplitDebug) | ||
| start_time = clock(); | ||
| mergeZTensors(&splitInfoY, output); | ||
| if (OMZTensorSplitDebug) { | ||
| end_time = clock(); | ||
| mergeTime = ((float)(end_time - start_time) / (float)CLOCKS_PER_SEC) * 1000; | ||
| } | ||
|
|
||
| freeSplitInfoBuffer(&splitInfoA); | ||
| freeSplitInfoBuffer(&splitInfoY); | ||
|
|
||
| if (OMZTensorSplitDebug) | ||
| printf("[MatMul] split, %f, mm, %f, merge, %f (milliseconds)\n", splitTime, | ||
| mmTime, mergeTime); | ||
|
|
||
| return ZDNN_OK; | ||
| } | ||
|
|
||
| zdnn_status zdnn_matmul_op_ext(const zdnn_ztensor *inputA, | ||
| const zdnn_ztensor *inputB, const zdnn_ztensor *inputC, int opType, | ||
| zdnn_ztensor *output) { | ||
| return zdnn_matmul_op_common( | ||
| inputA, inputB, inputC, opType, output, /*isBcast=*/false); | ||
| } | ||
|
|
||
| zdnn_status zdnn_matmul_bcast_op_ext(const zdnn_ztensor *inputA, | ||
| const zdnn_ztensor *inputB, const zdnn_ztensor *inputC, int opType, | ||
| zdnn_ztensor *output) { | ||
| zdnn_status status = zdnn_matmul_op_common( | ||
| inputA, inputB, inputC, opType, output, /*isBcast=*/true); | ||
| // Compiler does not check the return result at this moment. Thus, check it | ||
| // here. | ||
| assert(status == ZDNN_OK && "Failed to execute MatMul on NNPA"); | ||
| return status; | ||
| } | ||
|
|
||
| #ifdef __cplusplus | ||
| } | ||
| #endif |
Oops, something went wrong.