/*
* Copyright (C) 2015 The Android Open Source Project
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package com.example.android.rs.blasbenchmark;
import android.renderscript.*;
import android.util.Log;
import java.util.Random;
import java.lang.Math;
public class BNNMTest extends TestBase {
static {
System.loadLibrary("gemmdata");
}
native void getData(byte[] a, byte[] b, byte[] c);
ScriptIntrinsicBLAS mBLAS;
private Allocation matA;
private Allocation matB;
private Allocation matC;
private int m;
private int n;
private int k;
private int a_offset;
private int b_offset;
private int c_offset;
private int c_mult_int;
private int mTestSize;
BNNMTest(int testSize) {
mTestSize = testSize;
}
public void createTest() {
mBLAS = ScriptIntrinsicBLAS.create(mRS);
setTest();
}
private void setTest() {
switch (mTestSize) {
case 1:
setTestSmall();
break;
case 2:
setTestMedium();
break;
case 3:
setTestLarge();
break;
default:
break;
}
}
// In Java, the eight-bit 'byte' type is signed, but the API for the 8-bit
// matrix multiplication deals with unsigned bytes. This is a convenience
// function that converts arrays of unsigned ints to their equivalent
// representations as signed bytes. For example, the bit pattern 0xff is 255
// as an unsigned value, but -127 as a Java signed byte. So if you pass in an
// array of int[] {255} into this function, you'll get back byte[] {-127}.
private byte[] unsignedToSignedByte(int[] input) {
byte[] output = new byte[input.length];
for (int i = 0; i < input.length; ++i) {
output[i] = (byte)(input[i]);
}
return output;
}
private void addByteNoise(byte[] data, int count, float frequency, int maxDelta) {
Random rand = new Random();
for (int n = 0; n < count; ++n) {
if (rand.nextFloat() < frequency) {
final int originalValue = data[n];
final float direction = rand.nextFloat();
int delta = (int)(Math.ceil(rand.nextFloat() * maxDelta));
if (direction < 0.5f) {
delta = -delta;
}
int newValue = (originalValue + delta);
if (newValue < -127) {
newValue = -127;
}
if (newValue > 127) {
newValue = 127;
}
data[n] = (byte)(newValue);
}
}
}
private boolean testWithTolerance(byte[] c_byte, byte[] c_byte_output) {
// The testing procedure here is a bit complex, but the aim is to mimic the
// requirements we've empirically found running deep neural networks in real
// applications. We want to open the door to vendors using approximations that
// produce slightly different results for optimization's sake, but keep the
// precision loss within small enough bounds that we don't lose accuracy in
// the final result.
// After experimentation, we've found that we can tolerate around 5% of the
// output bytes being different by 1. Any larger differences are not tolerable
// and we can't get good results if the frequency of small differences is
// higher than 5%. This test tries to measure those properties on an example
// set of parameters that were captured from a real application.
// For example, if you uncommented this function that adds random noise to the
// results at a 3% specified frequency, the test should fail:
// AddByteNoise(c_byte_output, c_count, 0.03f, 1);
final boolean areSizesDifferent = (c_byte.length != c_byte_output.length);
final int c_count = Math.min(c_byte.length, c_byte_output.length);
int howManyDifferent = 0;
boolean areAnyTooDifferent = false;
for (int i = 0; i < c_count; i++) {
byte expectedValue = c_byte[i];
byte actualValue = c_byte_output[i];
int delta = (expectedValue - actualValue);
// First make sure that the difference is no more than one.
if ((delta < -1) || (delta > 1)) {
areAnyTooDifferent = true;
}
// If there is a difference, increment the counter to track it.
if (delta != 0) {
// Don't spam the logs if too many are different.
if (howManyDifferent < 50) {
android.util.Log.e("BNNM", "Mismatch at " + i +
": expected " + (expectedValue & 0xff) +
", got " + (actualValue & 0xff));
}
++howManyDifferent;
}
}
// We want no more than 2% of the values to show any differences, so work out
// what that means in absolute numbers.
final int percentThreshold = 2;
final int differenceThreshold = Math.max((percentThreshold * c_count) / 100, 1);
final boolean areTooManyDifferent = (howManyDifferent >= differenceThreshold);
if (areAnyTooDifferent) {
android.util.Log.e("BNNM", "Some outputs were too different.");
}
if (areTooManyDifferent) {
android.util.Log.e("BNNM", "There were too many small differences." +
" We can tolerate " + percentThreshold + "% (" +
differenceThreshold + "), but there were " + howManyDifferent);
}
return !(areAnyTooDifferent || areTooManyDifferent);
}
// This test multiplies a couple of small 8-bit matrices, and compares the
// results with hand-calculated expectations.
public void setTestSmall() {
// The A matrix is:
// | 1 | 4 |
// | 2 | 5 |
// | 3 | 6 |
byte[] a_byte = unsignedToSignedByte(new int[] {
1, 2, 3,
4, 5, 6,
});
final int a_rows = 3;
final int a_cols = 2;
a_offset = 0;
// The B matrix is:
// | -1 | -2 | -3 | -4 |
// | -5 | -6 | -7 | -8 |
// | -9 | -10 | -11 | -12 |
byte[] b_byte = unsignedToSignedByte(new int[] {
11, 7, 3,
10, 6, 2,
9, 5, 1,
8, 4, 0,
});
final int b_cols = 4;
b_offset = 12;
// EightBitGemm implements C = B.transposed() * A,
// so we expect to get these results:
// 1*-1 + 2*-5 + 3*-9 + 128 = 90
// 1*-2 + 2*-6 + 3*-10 + 128 = 84
// 1*-3 + 2*-7 + 3*-11 + 128 = 78
// 1*-4 + 2*-8 + 3*-12 + 128 = 72
// 4*-1 + 5*-5 + 6*-9 + 128 = 45
// 4*-2 + 5*-6 + 6*-10 + 128 = 30
// 4*-3 + 5*-7 + 6*-11 + 128 = 15
// 4*-4 + 5*-8 + 6*-12 + 128 = 0
// | 90 | 45 |
// | 84 | 30 |
// | 78 | 15 |
// | 72 | 0 |
c_offset = 128;
final int c_shift = 21;
c_mult_int = (1 << c_shift);
byte[] expected_data = unsignedToSignedByte(new int[] {
90, 84, 78, 72,
45, 30, 15, 0,
});
m = a_cols;
n = b_cols;
k = a_rows;
Type.Builder builder = new Type.Builder(mRS, Element.U8(mRS));
Type a_type = builder.setX(k).setY(m).create();
Type b_type = builder.setX(k).setY(n).create();
Type c_type = builder.setX(n).setY(m).create();
matA = Allocation.createTyped(mRS, a_type);
matB = Allocation.createTyped(mRS, b_type);
matC = Allocation.createTyped(mRS, c_type);
matA.copyFrom(a_byte);
matB.copyFrom(b_byte);
//During setup, do a sample run to see if the result is correct.
mBLAS.BNNM(matA, a_offset, matB, b_offset, matC, c_offset, c_mult_int);
int c_count = (m * n);
byte[] c_byte_output = new byte[c_count];
matC.copyTo(c_byte_output);
if (!testWithTolerance(expected_data, c_byte_output)) {
Log.e(TAG, "Result is not correct!");
throw new AssertionError("Result is not correct.");
}
}
// This test multiplies another two medium 8-bit matrices, and compares the
// results with the expected values. The data here is arbitrary.
public void setTestMedium() {
byte[] a_byte = unsignedToSignedByte(new int[] {
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23,
23, 22, 21, 20, 19, 18, 17, 16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1,
1, 23, 2, 22, 3, 21, 4, 20, 5, 19, 6, 18, 7, 17, 8, 16, 9, 15, 10, 14, 11, 13, 12,
23, 1, 22, 2, 21, 3, 20, 4, 19, 5, 18, 6, 17, 7, 16, 8, 15, 9, 14, 10, 13, 11, 12,
1, 1, 1, 1, 1, 1, 1, 1, 1, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12,
3, 1, 4, 1, 5, 8, 2, 3, 1, 14, 11, 15, 18, 12, 13, 11, 14, 11, 15, 18, 12, 13, 11,
8, 0, 5, 8, 1, 3, 7, 5, 7, 13, 10, 23, 13, 11, 17, 23, 12, 19, 17, 13, 14, 10, 19,
});
final int a_rows = 23;
final int a_cols = 7;
a_offset = 13;
byte[] b_byte = unsignedToSignedByte(new int[] {
0, 2, 4, 6, 8, 10, 1, 3, 5, 7, 9, 11, 0, 2, 4, 6, 8, 10, 1, 3, 5, 7, 9,
0, 20, 40, 60, 80, 10, 11, 13, 15, 17, 19, 21, 10, 12, 14, 6, 8, 10, 1, 3, 5, 7, 9,
1, 21, 41, 61, 81, 11, 12, 14, 16, 18, 20, 22, 11, 13, 15, 7, 9, 11, 2, 4, 6, 8, 9,
0, 19, 39, 59, 79, 9, 10, 12, 14, 16, 18, 20, 9, 11, 13, 5, 7, 9, 0, 2, 4, 6, 8,
2, 22, 42, 62, 82, 12, 13, 15, 17, 19, 21, 23, 12, 14, 16, 8, 9, 12, 3, 5, 7, 9, 9,
0, 18, 38, 58, 78, 8, 9, 11, 13, 15, 17, 19, 8, 10, 12, 4, 6, 8, 0, 1, 3, 5, 7,
3, 23, 43, 63, 83, 13, 14, 16, 18, 20, 22, 24, 13, 15, 17, 9, 9, 13, 4, 6, 8, 9, 9,
0, 17, 37, 57, 77, 7, 8, 10, 12, 14, 16, 18, 7, 9, 11, 3, 5, 7, 0, 0, 2, 4, 6,
10, 20, 30, 40, 50, 1, 2, 3, 4, 5, 11, 12, 13, 14, 15, 21, 22, 23, 24, 25, 1, 2, 3,
});
final int b_cols = 9;
b_offset = 23;
c_offset = 2121;
final int c_shift = 21;
c_mult_int = 132359;
byte[] expected_data = unsignedToSignedByte(new int[] {
167, 53, 51, 54, 49, 55, 46,
56, 116, 153, 232, 232, 234, 231,
236, 232, 237, 174, 168, 131, 130,
132, 129, 133, 128, 133, 134, 151,
154, 152, 156, 151, 158, 150, 160,
156, 255, 113, 106, 120, 98, 127,
91, 134, 178, 231, 102, 97, 107,
92, 111, 87, 116, 164, 187, 76,
73, 78, 70, 81, 67, 83, 139,
});
m = a_cols;
n = b_cols;
k = a_rows;
Type.Builder builder = new Type.Builder(mRS, Element.U8(mRS));
Type a_type = builder.setX(k).setY(m).create();
Type b_type = builder.setX(k).setY(n).create();
Type c_type = builder.setX(n).setY(m).create();
matA = Allocation.createTyped(mRS, a_type);
matB = Allocation.createTyped(mRS, b_type);
matC = Allocation.createTyped(mRS, c_type);
matA.copyFrom(a_byte);
matB.copyFrom(b_byte);
//During setup, do a sample run to see if the result is correct.
mBLAS.BNNM(matA, a_offset, matB, b_offset, matC, c_offset, c_mult_int);
int c_count = (m * n);
byte[] c_byte_output = new byte[c_count];
matC.copyTo(c_byte_output);
if (!testWithTolerance(expected_data, c_byte_output)) {
Log.e(TAG, "Result is not correct!");
throw new AssertionError("Result is not correct.");
}
}
// This test takes a large set of real data captured from a convolutional
// neural network solving a computer vision problem, and runs it through the
// eight-bit matrix multiply. We test the results to make sure they're close
// enough to be usable.
public void setTestLarge() {
m = 256;
n = 192;
k = 1152;
a_offset = 0;
b_offset = 84;
c_mult_int = 3401;
c_offset = 74980;
int a_count = (m * k);
int b_count = (n * k);
int c_count = (m * n);
byte[] a_byte = new byte[a_count];
byte[] b_byte = new byte[b_count];
byte[] c_byte = new byte[c_count];
getData(a_byte, b_byte, c_byte);
Type.Builder builder = new Type.Builder(mRS, Element.U8(mRS));
Type a_type = builder.setX(k).setY(m).create();
Type b_type = builder.setX(k).setY(n).create();
Type c_type = builder.setX(n).setY(m).create();
matA = Allocation.createTyped(mRS, a_type);
matB = Allocation.createTyped(mRS, b_type);
matC = Allocation.createTyped(mRS, c_type);
matA.copyFrom(a_byte);
matB.copyFrom(b_byte);
//During setup, do a sample run to see if the result is correct.
mBLAS.BNNM(matA, a_offset, matB, b_offset, matC, c_offset, c_mult_int);
byte[] c_byte_output = new byte[c_count];
matC.copyTo(c_byte_output);
if (!testWithTolerance(c_byte, c_byte_output)) {
Log.e(TAG, "Result is not correct!");
throw new AssertionError("Result is not correct.");
}
}
public void runTest() {
mBLAS.BNNM(matA, a_offset, matB, b_offset, matC, c_offset, c_mult_int);
}
public String getTestInfo() {
return "8Bit GEMM Test: m=" + m + ", n=" + n + ", k=" + k;
}
}