Skip to content

webgpu: Batch submitted command encoders #4888

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
qjia7 merged 3 commits into from
Apr 2, 2021
Merged

webgpu: Batch submitted command encoders #4888

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
d2d36ca
webgpu: Use WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE
qjia7 Mar 31, 2021
c4d64c5
Address Kai's comment
qjia7 Apr 1, 2021
6ca8e21
Merge branch 'master' into encoder_count
qjia7 Apr 2, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 2 additions & 1 deletion tfjs-backend-webgpu/src/backend_webgpu.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -564,7 +564,8 @@ export class WebGPUBackend extends KernelBackend {
this.uniformDisposalQueue.push(uniformInfo);
}

if (env().get('WEBGPU_IMMEDIATE_EXECUTION_ENABLED')) {
if (env().get('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE') as number ===
Copy link
Contributor

@kainino0x kainino0x Apr 1, 2021

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Probably should be >= just in case multiple command buffers end up waiting for any reason, or the WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE is set to 0.

Copy link
Contributor Author

@qjia7 qjia7 Apr 1, 2021

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done

Copy link
Contributor

@kainino0x kainino0x Apr 1, 2021

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nit: The fact that these are command encoders and not command buffers is I think an accident of history and not strictly necessarily true in the future. I'd pick a more targeted name like
WEBGPU_DEFERRED_SUBMIT_BATCH_SIZE
or something

Copy link
Contributor Author

@qjia7 qjia7 Apr 1, 2021

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done

this.commandQueue.length) {
this.submitQueue();
}

Expand Down
172 changes: 89 additions & 83 deletions tfjs-backend-webgpu/src/backend_webgpu_test.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,9 +35,6 @@ describeWebGPU('backend webgpu cpu forwarding turned on', () => {
});

it('should not allocate GPU memory when CPU forwarding', async () => {
const savedFlag = tf.env().get('WEBGPU_IMMEDIATE_EXECUTION_ENABLED');
tf.env().set('WEBGPU_IMMEDIATE_EXECUTION_ENABLED', true);

const a = tf.tensor2d([2, 4, 6, 8], [2, 2]);
const b = tf.tensor2d([0.5, 0.5, 0.5, 0.5], [2, 2]);

Expand Down Expand Up @@ -65,43 +62,44 @@ describeWebGPU('backend webgpu cpu forwarding turned on', () => {

tf.test_util.expectArraysClose(
dData, new Float32Array([9, 12, 15, 19, 26, 33]));
tf.env().set('WEBGPU_IMMEDIATE_EXECUTION_ENABLED', savedFlag);
});
});

describeWebGPU('backend webgpu', () => {
it('should not leak memory in delayed mode', async () => {
const savedFlag = tf.env().get('WEBGPU_IMMEDIATE_EXECUTION_ENABLED');
tf.env().set('WEBGPU_IMMEDIATE_EXECUTION_ENABLED', false);
const a = tf.tensor2d([2, 4, 6, 8], [2, 2]);
const b = tf.tensor2d([0.5, 0.5, 0.5, 0.5], [2, 2]);

const c = tf.mul(a, b);

const startNumBytes = tf.memory().numBytes;
const startNumTensors = tf.memory().numTensors;
const startNumBytesInGPU = (tf.memory() as WebGPUMemoryInfo).numBytesInGPU;

const f = tf.tensor2d([1, 2, 3, 4, 5, 6], [2, 3]);
const d = tf.matMul(c, f);

const dData = await d.data();
const endNumBytes = tf.memory().numBytes;
const endNumTensors = tf.memory().numTensors;
const endNumBytesInGPU = (tf.memory() as WebGPUMemoryInfo).numBytesInGPU;

expect(endNumBytes - startNumBytes).toEqual(48);
expect(endNumTensors - startNumTensors).toEqual(2);
expect(endNumBytesInGPU - startNumBytesInGPU).toEqual(24);

tf.test_util.expectArraysClose(
dData, new Float32Array([9, 12, 15, 19, 26, 33]));
tf.env().set('WEBGPU_IMMEDIATE_EXECUTION_ENABLED', savedFlag);
});

it('should not leak memory in immediate mode', async () => {
const savedFlag = tf.env().get('WEBGPU_IMMEDIATE_EXECUTION_ENABLED');
tf.env().set('WEBGPU_IMMEDIATE_EXECUTION_ENABLED', true);
it('should not leak memory when command encoder count is larger than one',
async () => {
const savedFlag = tf.env().get('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE');
tf.env().set('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE', 15);
const a = tf.tensor2d([2, 4, 6, 8], [2, 2]);
const b = tf.tensor2d([0.5, 0.5, 0.5, 0.5], [2, 2]);

const c = tf.mul(a, b);

const startNumBytes = tf.memory().numBytes;
const startNumTensors = tf.memory().numTensors;
const startNumBytesInGPU =
(tf.memory() as WebGPUMemoryInfo).numBytesInGPU;

const f = tf.tensor2d([1, 2, 3, 4, 5, 6], [2, 3]);
const d = tf.matMul(c, f);

const dData = await d.data();
const endNumBytes = tf.memory().numBytes;
const endNumTensors = tf.memory().numTensors;
const endNumBytesInGPU = (tf.memory() as WebGPUMemoryInfo).numBytesInGPU;

expect(endNumBytes - startNumBytes).toEqual(48);
expect(endNumTensors - startNumTensors).toEqual(2);
expect(endNumBytesInGPU - startNumBytesInGPU).toEqual(24);

tf.test_util.expectArraysClose(
dData, new Float32Array([9, 12, 15, 19, 26, 33]));
tf.env().set('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE', savedFlag);
});

it('should not leak memory when command encoder count is one', async () => {
const savedFlag = tf.env().get('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE');
tf.env().set('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE', 1);
const a = tf.tensor2d([2, 4, 6, 8], [2, 2]);
const b = tf.tensor2d([0.5, 0.5, 0.5, 0.5], [2, 2]);

Expand All @@ -125,12 +123,12 @@ describeWebGPU('backend webgpu', () => {

tf.test_util.expectArraysClose(
dData, new Float32Array([9, 12, 15, 19, 26, 33]));
tf.env().set('WEBGPU_IMMEDIATE_EXECUTION_ENABLED', savedFlag);
tf.env().set('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE', savedFlag);
});

it('should recycle buffers in immediate mode', () => {
const savedFlag = tf.env().get('WEBGPU_IMMEDIATE_EXECUTION_ENABLED');
tf.env().set('WEBGPU_IMMEDIATE_EXECUTION_ENABLED', true);
it('should recycle buffers when command encoder count is one', () => {
const savedFlag = tf.env().get('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE');
tf.env().set('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE', 1);
const backend = tf.backend() as WebGPUBackend;
const bufferManager = backend.getBufferManager();
bufferManager.reset();
Expand Down Expand Up @@ -164,51 +162,59 @@ describeWebGPU('backend webgpu', () => {
const usedBuffersAfterSecondMatMul = bufferManager.getNumUsedBuffers();
expect(freeBuffersAfterSecondMatMul - freeBuffersAfterSecondMul).toEqual(0);
expect(usedBuffersAfterSecondMatMul - usedBuffersAfterSecondMul).toEqual(2);
tf.env().set('WEBGPU_IMMEDIATE_EXECUTION_ENABLED', savedFlag);
tf.env().set('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE', savedFlag);
});

it('should not recycle buffers in delayed mode', async () => {
const savedFlag = tf.env().get('WEBGPU_IMMEDIATE_EXECUTION_ENABLED');
tf.env().set('WEBGPU_IMMEDIATE_EXECUTION_ENABLED', false);
const backend = tf.backend() as WebGPUBackend;
const bufferManager = backend.getBufferManager();
bufferManager.reset();

const a = tf.tensor2d([2, 4, 6, 8], [2, 2]);
const b = tf.tensor2d([0.5, 0.5, 0.5, 0.5], [2, 2]);

const c = tf.mul(a, b);
const freeBuffersAfterFirstMul = bufferManager.getNumFreeBuffers();
const usedBuffersAfterFirstMul = bufferManager.getNumUsedBuffers();

const f = tf.tensor2d([1, 2, 3, 4, 5, 6], [2, 3]);
tf.matMul(c, f);
const freeBuffersAfterFirstMatMul = bufferManager.getNumFreeBuffers();
const usedBuffersAfterFirstMatMul = bufferManager.getNumUsedBuffers();
expect(freeBuffersAfterFirstMatMul - freeBuffersAfterFirstMul).toEqual(0);
expect(usedBuffersAfterFirstMatMul - usedBuffersAfterFirstMul).toEqual(2);

const a2 = tf.tensor2d([2, 4, 6, 8], [2, 2]);
const b2 = tf.tensor2d([0.5, 0.5, 0.5, 0.5], [2, 2]);

const c2 = tf.mul(a2, b2);
const freeBuffersAfterSecondMul = bufferManager.getNumFreeBuffers();
const usedBuffersAfterSecondMul = bufferManager.getNumUsedBuffers();
expect(freeBuffersAfterSecondMul - freeBuffersAfterFirstMatMul).toEqual(0);
expect(usedBuffersAfterSecondMul - usedBuffersAfterFirstMatMul).toEqual(3);

const f2 = tf.tensor2d([1, 2, 3, 4, 5, 6], [2, 3]);
const c3 = tf.matMul(c2, f2);
const freeBuffersAfterSecondMatMul = bufferManager.getNumFreeBuffers();
const usedBuffersAfterSecondMatMul = bufferManager.getNumUsedBuffers();
expect(freeBuffersAfterSecondMatMul - freeBuffersAfterSecondMul).toEqual(0);
expect(usedBuffersAfterSecondMatMul - usedBuffersAfterSecondMul).toEqual(2);

// Tests happen within a tidy so we need to read a tensor at the end of a
// test in delayed mode in order to force flush the disposal queue.
await c3.data();
tf.env().set('WEBGPU_IMMEDIATE_EXECUTION_ENABLED', savedFlag);
});
it('should not recycle buffers when command encoder count is larger than one',
async () => {
const savedFlag = tf.env().get('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE');
tf.env().set('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE', 15);
const backend = tf.backend() as WebGPUBackend;
const bufferManager = backend.getBufferManager();
bufferManager.reset();

const a = tf.tensor2d([2, 4, 6, 8], [2, 2]);
const b = tf.tensor2d([0.5, 0.5, 0.5, 0.5], [2, 2]);

const c = tf.mul(a, b);
const freeBuffersAfterFirstMul = bufferManager.getNumFreeBuffers();
const usedBuffersAfterFirstMul = bufferManager.getNumUsedBuffers();

const f = tf.tensor2d([1, 2, 3, 4, 5, 6], [2, 3]);
tf.matMul(c, f);
const freeBuffersAfterFirstMatMul = bufferManager.getNumFreeBuffers();
const usedBuffersAfterFirstMatMul = bufferManager.getNumUsedBuffers();
expect(freeBuffersAfterFirstMatMul - freeBuffersAfterFirstMul)
.toEqual(0);
expect(usedBuffersAfterFirstMatMul - usedBuffersAfterFirstMul)
.toEqual(2);

const a2 = tf.tensor2d([2, 4, 6, 8], [2, 2]);
const b2 = tf.tensor2d([0.5, 0.5, 0.5, 0.5], [2, 2]);

const c2 = tf.mul(a2, b2);
const freeBuffersAfterSecondMul = bufferManager.getNumFreeBuffers();
const usedBuffersAfterSecondMul = bufferManager.getNumUsedBuffers();
expect(freeBuffersAfterSecondMul - freeBuffersAfterFirstMatMul)
.toEqual(0);
expect(usedBuffersAfterSecondMul - usedBuffersAfterFirstMatMul)
.toEqual(3);

const f2 = tf.tensor2d([1, 2, 3, 4, 5, 6], [2, 3]);
const c3 = tf.matMul(c2, f2);
const freeBuffersAfterSecondMatMul = bufferManager.getNumFreeBuffers();
const usedBuffersAfterSecondMatMul = bufferManager.getNumUsedBuffers();
expect(freeBuffersAfterSecondMatMul - freeBuffersAfterSecondMul)
.toEqual(0);
expect(usedBuffersAfterSecondMatMul - usedBuffersAfterSecondMul)
.toEqual(2);

// Tests happen within a tidy so we need to read a tensor at the end of a
// test if encoder count in queue is larger than one in order to force
// flush the disposal queue.
await c3.data();
tf.env().set('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE', savedFlag);
});

it('readSync should throw if tensors are on the GPU', async () => {
const a = tf.tensor2d([1, 2, 3, 4], [2, 2]);
Expand Down
4 changes: 2 additions & 2 deletions tfjs-backend-webgpu/src/flags_webgpu.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,8 +19,8 @@ import {env} from '@tensorflow/tfjs-core';

const ENV = env();

/** Whether we submit commands to the device queue immediately. */
ENV.registerFlag('WEBGPU_IMMEDIATE_EXECUTION_ENABLED', () => true);
/** The command encoder count submitted to the device queue. */
ENV.registerFlag('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE', () => 15);

/**
* Whether we forward execution to the CPU backend if tensors are small and
Expand Down
16 changes: 8 additions & 8 deletions tfjs-backend-webgpu/src/matmul_test.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,9 +20,9 @@ import {test_util} from '@tensorflow/tfjs-core';
import {describeWebGPU} from './test_util';

describeWebGPU('matmul', () => {
it('it works in delayed mode.', async () => {
const savedFlag = tf.env().get('WEBGPU_IMMEDIATE_EXECUTION_ENABLED');
tf.env().set('WEBGPU_IMMEDIATE_EXECUTION_ENABLED', false);
it('it works when command encoder count is larger than one.', async () => {
const savedFlag = tf.env().get('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE');
tf.env().set('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE', 15);
const a = tf.tensor2d([1, 2, 3, 4], [2, 2]);
const b = tf.tensor2d([1, 2, 3, 4, 5, 6], [2, 3]);

Expand All @@ -34,12 +34,12 @@ describeWebGPU('matmul', () => {
const dData = await d.data();
test_util.expectArraysClose(
dData, new Float32Array([0, 12, 7.5, 0, 6.5, 66]));
tf.env().set('WEBGPU_IMMEDIATE_EXECUTION_ENABLED', savedFlag);
tf.env().set('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE', savedFlag);
});

it('it works in immediate mode.', async () => {
const savedFlag = tf.env().get('WEBGPU_IMMEDIATE_EXECUTION_ENABLED');
tf.env().set('WEBGPU_IMMEDIATE_EXECUTION_ENABLED', true);
it('it works when command encoder count is one.', async () => {
const savedFlag = tf.env().get('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE');
tf.env().set('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE', 1);
const a = tf.tensor2d([1, 2, 3, 4], [2, 2]);
const b = tf.tensor2d([1, 2, 3, 4, 5, 6], [2, 3]);

Expand All @@ -51,7 +51,7 @@ describeWebGPU('matmul', () => {
const dData = await d.data();
test_util.expectArraysClose(
dData, new Float32Array([0, 12, 7.5, 0, 6.5, 66]));
tf.env().set('WEBGPU_IMMEDIATE_EXECUTION_ENABLED', savedFlag);
tf.env().set('WEBGPU_COMMAND_ENCODER_COUNT_IN_QUEUE', savedFlag);
});

// tslint:disable-next-line:max-line-length
Expand Down