Skip to content

Commit

Permalink
ggml-jni: add automation UT for ggml ops --- op_add,op_mul,op_mulmat
Browse files Browse the repository at this point in the history
  • Loading branch information
@zhouwg
zhouwg committed Apr 20, 2024
1 parent e8a1918 commit 2c9b8f6
Show file tree
Hide file tree
Showing 14 changed files with 382 additions and 108 deletions.
53 changes: 30 additions & 23 deletions cdeosplayer/cdeosplayer-lib/src/main/java/cdeos/media/player/CDEUtils.java
View file
Original file line number Diff line number Diff line change
Expand Up @@ -260,19 +260,20 @@ public class CDEUtils {
public static final int ASR_MODE_TRANSCRIPTION_RECORD = 3; // transcription + audio record

//keep sync with ggml-jni.h
public static final int BECHMARK_ASR = 0;
public static final int BECHMARK_MEMCPY = 1;
public static final int BECHMARK_MULMAT = 2;
public static final int BECHMARK_FULL = 3;
public static final int BENCHMARK_MATRIX = 4;
public static final int BENCHMARK_LLM = 5;
public static final int BENCHMARK_STABLEDIFFUSION= 6;
public static final int BENCHMARK_QNN_SAMPLE = 7;
public static final int BENCHMARK_QNN_SAVER = 8;
public static final int BENCHMARK_QNN_MATRIX = 9;
public static final int BENCHMARK_QNN_GGML = 10;
public static final int BENCHMARK_QNN_COMPLEX = 11;
public static final int BENCHMARK_QNN_GGML_OP = 12;
public static final int BENCHMARK_ASR = 0;
public static final int BENCHMARK_MEMCPY = 1;
public static final int BENCHMARK_MULMAT = 2;
public static final int BENCHMARK_FULL = 3;
//public static final int BENCHMARK_MATRIX = 4;//not used since 04-20-2024
public static final int BENCHMARK_LLM = 4;
public static final int BENCHMARK_STABLEDIFFUSION= 5;
public static final int BENCHMARK_QNN_SAMPLE = 6;
public static final int BENCHMARK_QNN_SAVER = 7;
public static final int BENCHMARK_QNN_MATRIX = 8;
public static final int BENCHMARK_QNN_GGML = 9;
public static final int BENCHMARK_QNN_COMPLEX = 10;
public static final int BENCHMARK_QNN_GGML_OP = 11;
public static final int BENCHMARK_QNN_AUTO_UT = 12;

//keep sync with ggml-qnn.h
public static final int QNN_BACKEND_CPU = 0;
Expand Down Expand Up @@ -3916,20 +3917,22 @@ public void onClick(DialogInterface dialog, int which) {

public static String getBenchmarkDesc(int benchmarkIndex) {
switch (benchmarkIndex) {
case BECHMARK_FULL:
return "GGML whisper_encode";
case BENCHMARK_FULL:
return "GGML whisper full";

case BECHMARK_MEMCPY:
return "GGML memcopy";
case BENCHMARK_MEMCPY:
return "GGML memcpy";

case BECHMARK_MULMAT:
case BENCHMARK_MULMAT:
return "GGML matrix multiply";

case BECHMARK_ASR:
return "GGML ASR inference";
case BENCHMARK_ASR:
return "GGML whisper ASR";

/*//not used since 04-20-2024
case BENCHMARK_MATRIX:
return "GGML matrix";
*/

case BENCHMARK_LLM:
return "GGML LLAMA";
Expand All @@ -3943,17 +3946,21 @@ public static String getBenchmarkDesc(int benchmarkIndex) {
case BENCHMARK_QNN_SAVER:
return "GGML QNN saver";


case BENCHMARK_QNN_MATRIX:
return "GGML QNN matrix manipulate";
return "GGML QNN matrix addition";

case BENCHMARK_QNN_GGML:
return "GGML QNN ggml";
return "GGML QNN mapping ggml tensor";

case BENCHMARK_QNN_COMPLEX:
return "GGML QNN complex graph";

case BENCHMARK_QNN_GGML_OP:
return "GGML QNN OP"; //for PoC-S49: implementation of other GGML OP(non-mulmat) using QNN API
return "GGML QNN OP UT"; //UT for PoC-S49: implementation of GGML OPs using QNN API

case BENCHMARK_QNN_AUTO_UT:
return "GGML QNN OP UT automation"; //automation UT for PoC-S49: implementation of GGML OPs using QNN API
}

return "unknown";
Expand Down
2 changes: 1 addition & 1 deletion cdeosplayer/cdeosplayer-lib/src/main/java/org/ggml/ggmljava.java
View file
Original file line number Diff line number Diff line change
Expand Up @@ -119,7 +119,7 @@ public enum ggml_op {
/**
* @param modelPath /sdcard/kantv/ggml-xxxxxx.bin or /sdcard/kantv/xxxxxx.gguf or qualcomm's prebuilt dedicated model.so or ""
* @param audioPath /sdcard/kantv/jfk.wav
* @param nBenchType 0: asr(transcription) 1: memcpy 2: mulmat 3: full/whisper_encode 4: matrix 5: LLAMA 6: stable diffusion 7: QNN sample 8: QNN saver 9: QNN matrix 10: QNN GGML 11: QNN complex 12: QNN GGML OP
* @param nBenchType 0: whisper asr 1: memcpy 2: mulmat 3: whisper full 4: LLAMA 5: stable diffusion 6: QNN sample 7: QNN saver 8: QNN matrix 9: QNN GGML 10: QNN complex 11: QNN GGML OP(QNN UT) 12: QNN UT automation
* @param nThreadCounts 1 - 8
* @param nBackendType 0: CPU 1: GPU 2: DSP 3: ggml("fake" QNN backend, just for compare performance)
* @param nOpType type of matrix manipulate / GGML OP / type of various complex/complicated computation graph
Expand Down
2 changes: 1 addition & 1 deletion cdeosplayer/kantv/src/main/java/com/cdeos/kantv/ui/fragment/ASRResearchFragment.java
View file
Original file line number Diff line number Diff line change
Expand Up @@ -255,7 +255,7 @@ public void onNothingSelected(AdapterView<?> parent) {

}
});
spinnerBenchType.setSelection(CDEUtils.BENCHMARK_QNN_GGML_OP);
spinnerBenchType.setSelection(CDEUtils.BENCHMARK_ASR);

Spinner spinnerThreadsCounts = mActivity.findViewById(R.id.spinnerThreadCounts);
String[] arrayThreadCounts = getResources().getStringArray(R.array.threadCounts);
Expand Down
4 changes: 2 additions & 2 deletions cdeosplayer/kantv/src/main/res/layout/fragment_asr.xml
View file
Original file line number Diff line number Diff line change
Expand Up @@ -43,7 +43,7 @@
<TextView
android:layout_width="wrap_content"
android:layout_height="wrap_content"
android:text="Bench:" />
android:text="Bench" />

<Spinner
android:id="@+id/spinnerBenchType"
Expand All @@ -55,7 +55,7 @@
<TextView
android:layout_width="wrap_content"
android:layout_height="wrap_content"
android:text="Threads:" />
android:text="Threads" />

<Spinner
android:id="@+id/spinnerThreadCounts"
Expand Down
28 changes: 14 additions & 14 deletions cdeosplayer/kantv/src/main/res/values/arrays.xml
View file
Original file line number Diff line number Diff line change
Expand Up @@ -60,20 +60,20 @@
</string-array>

<string-array name="benchType">
<item>asr</item>
<item>mempcpy</item>
<item>mulmat</item>
<item>full</item>
<item>matrix</item>
<item>llama</item>
<item>stablediffusion</item> <!-- not work on Xiaomi 14 currently -->
<!-- step by step -->
<item>qnn-sample</item> <!-- "play with /say hello to" QNN Sample -->
<item>qnn-saver</item> <!-- study QNN SDK mechanism by QNN Saver -->
<item>qnn-matrix</item> <!-- offload a simple matrix addition operation to QNN -->
<item>qnn-ggml</item> <!-- mapping ggml tensor to QNN tensor -->
<item>qnn-complex</item><!-- complex computation graph in C/C++ or GGML, and then offload them to QNN -->
<item>qnn-ggml-op</item><!-- for PoC-S49: implementation of other GGML OP(non-mulmat) using QNN API -->
<item>whisper asr</item> <!-- whisper asr benchmark / asr validation -->
<item>memcpy</item> <!-- whisper memcopy benchmark -->
<item>whisper mulmat</item> <!-- whisper mulmat benchmark -->
<item>whisper full</item> <!-- whisper full benchmark -->
<item>llama</item> <!-- llama benchmark, not work currently -->
<item>stablediffusion</item> <!-- stable diffusion benchmark, not work currently -->
<!-- step by step for PoC: Add Qualcomm mobile SoC native backend for GGML, https://github.com/zhouwg/kantv/issues/121 -->
<item>qnn-sample</item> <!-- "play with /say hello to" QNN Sample -->
<item>qnn-saver</item> <!-- study QNN SDK mechanism by QNN Saver -->
<item>qnn-matrix</item> <!-- offload a simple matrix addition operation to QNN -->
<item>qnn-ggml</item> <!-- mapping ggml tensor to QNN tensor -->
<item>qnn-complex</item> <!-- complex computation graph in C/C++ or GGML, and then offload them to QNN -->
<item>qnn-ggml-op</item> <!-- UT for PoC-S49: implementation of other GGML OP(non-mulmat) using QNN API -->
<item>qnn-auto-ut</item> <!-- automation UT for PoC-S49:implementation of other GGML OP(non-mulmat) using QNN API -->
</string-array>

<string-array name="threadCounts">
Expand Down
Loading

0 comments on commit 2c9b8f6

Please sign in to comment.