[Impeller] improve rrect blur performance on Android #37925

jonahwilliams · 2022-11-28T02:32:49Z

Improve the rrect blur performance by separating out the sigma > 0 and sigma <= 0 cases into distinct shaders. This change accounts for the vast majority of the performance improvement, which can be seen in the significant % of cycle reduction and removal of stack spilling on older GPUs in the malioc diff below.

Details

[Mali-T880]
Main shader
===========

Work registers: 4 (100% used at 100% occupancy)
Uniform registers: 1 (5% used)
- Stack spilling: true
+ Stack spilling: false

                                A      LS       T    Bound
- Total instruction cycles:   12.00    4.00    0.00        A
+ Total instruction cycles:    8.00    1.00    0.00        A
- Shortest path cycles:        2.64    3.00    0.00       LS
+ Shortest path cycles:        7.59    1.00    0.00        A
Longest path cycles:          N/A     N/A     N/A      N/A

A = Arithmetic, LS = Load/Store, T = Texture

Shader properties
=================

Has uniform computation: false



[Mali-T860]
Main shader
===========

Work registers: 4 (100% used at 100% occupancy)
Uniform registers: 1 (5% used)
- Stack spilling: true
+ Stack spilling: false

                                A      LS       T    Bound
- Total instruction cycles:   18.00    4.00    0.00        A
+ Total instruction cycles:   12.00    1.00    0.00        A
- Shortest path cycles:        4.00    3.00    0.00        A
+ Shortest path cycles:       11.50    1.00    0.00        A
Longest path cycles:          N/A     N/A     N/A      N/A

A = Arithmetic, LS = Load/Store, T = Texture

Shader properties
=================

Has uniform computation: false



[Mali-T830]
Main shader
===========

Work registers: 4 (100% used at 100% occupancy)
Uniform registers: 1 (5% used)
- Stack spilling: true
+ Stack spilling: false

                                A      LS       T    Bound
- Total instruction cycles:   18.00    4.00    0.00        A
+ Total instruction cycles:   12.00    1.00    0.00        A
- Shortest path cycles:        2.12    3.00    0.00       LS
+ Shortest path cycles:        5.00    1.00    0.00        A
Longest path cycles:          N/A     N/A     N/A      N/A

A = Arithmetic, LS = Load/Store, T = Texture

Shader properties
=================

Has uniform computation: false



[Mali-T820]
Main shader
===========

Work registers: 4 (100% used at 100% occupancy)
Uniform registers: 1 (5% used)
- Stack spilling: true
+ Stack spilling: false

                                A      LS       T    Bound
- Total instruction cycles:   36.00    4.00    0.00        A
+ Total instruction cycles:   24.00    1.00    0.00        A
- Shortest path cycles:        4.25    3.00    0.00        A
+ Shortest path cycles:       10.00    1.00    0.00        A
Longest path cycles:          N/A     N/A     N/A      N/A

A = Arithmetic, LS = Load/Store, T = Texture

Shader properties
=================

Has uniform computation: false



[Mali-T760]
Main shader
===========

Work registers: 4 (100% used at 100% occupancy)
Uniform registers: 1 (5% used)
- Stack spilling: true
+ Stack spilling: false

                                A      LS       T    Bound
- Total instruction cycles:   18.00    4.00    0.00        A
+ Total instruction cycles:   12.00    1.00    0.00        A
- Shortest path cycles:        4.00    3.00    0.00        A
+ Shortest path cycles:       11.50    1.00    0.00        A
Longest path cycles:          N/A     N/A     N/A      N/A

A = Arithmetic, LS = Load/Store, T = Texture

Shader properties
=================

Has uniform computation: false



[Mali-T720]
Main shader
===========

Work registers: 4 (100% used at 100% occupancy)
Uniform registers: 1 (5% used)
- Stack spilling: true
+ Stack spilling: false

                                A      LS       T    Bound
- Total instruction cycles:   36.00    4.00    0.00        A
+ Total instruction cycles:   24.00    1.00    0.00        A
- Shortest path cycles:        4.25    3.00    0.00        A
+ Shortest path cycles:       10.25    1.00    0.00        A
Longest path cycles:          N/A     N/A     N/A      N/A

A = Arithmetic, LS = Load/Store, T = Texture


[Mali-G78AE]
Main shader
===========

Work registers: 32 (100% used at 100% occupancy)
- Uniform registers: 26 (40% used)
+ Uniform registers: 24 (37% used)
Stack spilling: false
- 16-bit arithmetic: 26%
+ 16-bit arithmetic: 29%

                                A      LS       V       T    Bound
- Total instruction cycles:    3.24    0.00    0.12    0.00        A
+ Total instruction cycles:    2.47    0.00    0.12    0.00        A
- Shortest path cycles:        0.20    0.00    0.12    0.00        A
+ Shortest path cycles:        2.47    0.00    0.12    0.00        A
- Longest path cycles:         3.11    0.00    0.12    0.00        A
+ Longest path cycles:         2.47    0.00    0.12    0.00        A

A = Arithmetic, LS = Load/Store, V = Varying, T = Texture

Shader properties
=================

Has uniform computation: true
Has side-effects: false
Modifies coverage: false
Uses late ZS test: false
Uses late ZS update: false
Reads color buffer: false



[Mali-G78]
Main shader
===========

Work registers: 32 (100% used at 100% occupancy)
- Uniform registers: 26 (40% used)
+ Uniform registers: 24 (37% used)
Stack spilling: false
- 16-bit arithmetic: 26%
+ 16-bit arithmetic: 29%

                                A      LS       V       T    Bound
- Total instruction cycles:    3.24    0.00    0.12    0.00        A
+ Total instruction cycles:    2.47    0.00    0.12    0.00        A
- Shortest path cycles:        0.20    0.00    0.12    0.00        A
+ Shortest path cycles:        2.47    0.00    0.12    0.00        A
- Longest path cycles:         3.11    0.00    0.12    0.00        A
+ Longest path cycles:         2.47    0.00    0.12    0.00        A

A = Arithmetic, LS = Load/Store, V = Varying, T = Texture

Shader properties
=================

Has uniform computation: true
Has side-effects: false
Modifies coverage: false
Uses late ZS test: false
Uses late ZS update: false
Reads color buffer: false



[Mali-G77]
Main shader
===========

Work registers: 32 (100% used at 100% occupancy)
- Uniform registers: 26 (40% used)
+ Uniform registers: 24 (37% used)
Stack spilling: false
- 16-bit arithmetic: 26%
+ 16-bit arithmetic: 29%

                                A      LS       V       T    Bound
- Total instruction cycles:    3.24    0.00    0.12    0.00        A
+ Total instruction cycles:    2.47    0.00    0.12    0.00        A
- Shortest path cycles:        0.20    0.00    0.12    0.00        A
+ Shortest path cycles:        2.47    0.00    0.12    0.00        A
- Longest path cycles:         3.11    0.00    0.12    0.00        A
+ Longest path cycles:         2.47    0.00    0.12    0.00        A

A = Arithmetic, LS = Load/Store, V = Varying, T = Texture

Shader properties
=================

Has uniform computation: true
Has side-effects: false
Modifies coverage: false
Uses late ZS test: false
Uses late ZS update: false
Reads color buffer: false



[Mali-G76]
Main shader
===========

- Work registers: 25 (78% used at 100% occupancy)
+ Work registers: 22 (68% used at 100% occupancy)
- Uniform registers: 22 (34% used)
+ Uniform registers: 16 (25% used)
Stack spilling: false
- 16-bit arithmetic: 35%
+ 16-bit arithmetic: 42%

                                A      LS       V       T    Bound
- Total instruction cycles:    2.53    0.00    0.12    0.00        A
+ Total instruction cycles:    1.67    0.00    0.12    0.00        A
- Shortest path cycles:        0.71    0.00    0.12    0.00        A
+ Shortest path cycles:        1.67    0.00    0.12    0.00        A
Longest path cycles:          N/A     N/A     N/A     N/A      N/A

A = Arithmetic, LS = Load/Store, V = Varying, T = Texture

Shader properties
=================

Has uniform computation: true
Has side-effects: false
Modifies coverage: false
Uses late ZS test: false
Uses late ZS update: false
Reads color buffer: false



[Mali-G72]
Main shader
===========

- Work registers: 26 (81% used at 100% occupancy)
+ Work registers: 23 (71% used at 100% occupancy)
- Uniform registers: 22 (34% used)
+ Uniform registers: 16 (25% used)
Stack spilling: false
- 16-bit arithmetic: 35%
+ 16-bit arithmetic: 42%

                                A      LS       V       T    Bound
- Total instruction cycles:    5.07    0.00    0.25    0.00        A
+ Total instruction cycles:    3.40    0.00    0.25    0.00        A
- Shortest path cycles:        1.42    0.00    0.25    0.00        A
+ Shortest path cycles:        3.40    0.00    0.25    0.00        A
Longest path cycles:          N/A     N/A     N/A     N/A      N/A

A = Arithmetic, LS = Load/Store, V = Varying, T = Texture

Shader properties
=================

Has uniform computation: true
Has side-effects: false
Modifies coverage: false
Uses late ZS test: false
Uses late ZS update: false
Reads color buffer: false



[Mali-G715]
Main shader
===========

Work registers: 32 (100% used at 100% occupancy)
- Uniform registers: 26 (40% used)
+ Uniform registers: 24 (37% used)
Stack spilling: false
- 16-bit arithmetic: 26%
+ 16-bit arithmetic: 29%

                                A      LS       V       T    Bound
- Total instruction cycles:    1.10    0.00    0.03    0.00        A
+ Total instruction cycles:    0.85    0.00    0.03    0.00        A
- Shortest path cycles:        0.07    0.00    0.03    0.00        A
+ Shortest path cycles:        0.84    0.00    0.03    0.00        A
- Longest path cycles:         1.06    0.00    0.03    0.00        A
+ Longest path cycles:         0.85    0.00    0.03    0.00        A

A = Arithmetic, LS = Load/Store, V = Varying, T = Texture

Shader properties
=================

Has uniform computation: true
Has side-effects: false
Modifies coverage: false
Uses late ZS test: false
Uses late ZS update: false
Reads color buffer: false



[Mali-G710]
Main shader
===========

Work registers: 32 (100% used at 100% occupancy)
- Uniform registers: 26 (40% used)
+ Uniform registers: 24 (37% used)
Stack spilling: false
- 16-bit arithmetic: 26%
+ 16-bit arithmetic: 29%

                                A      LS       V       T    Bound
- Total instruction cycles:    2.20    0.00    0.06    0.00        A
+ Total instruction cycles:    1.70    0.00    0.06    0.00        A
- Shortest path cycles:        0.13    0.00    0.06    0.00        A
+ Shortest path cycles:        1.69    0.00    0.06    0.00        A
- Longest path cycles:         2.12    0.00    0.06    0.00        A
+ Longest path cycles:         1.70    0.00    0.06    0.00        A

A = Arithmetic, LS = Load/Store, V = Varying, T = Texture

Shader properties
=================

Has uniform computation: true
Has side-effects: false
Modifies coverage: false
Uses late ZS test: false
Uses late ZS update: false
Reads color buffer: false



[Mali-G71]
Main shader
===========

- Work registers: 29 (90% used at 100% occupancy)
+ Work registers: 24 (75% used at 100% occupancy)
- Uniform registers: 22 (34% used)
+ Uniform registers: 16 (25% used)
Stack spilling: false
- 16-bit arithmetic: 29%
+ 16-bit arithmetic: 37%

                                A      LS       V       T    Bound
- Total instruction cycles:    5.33    0.00    0.25    0.00        A
+ Total instruction cycles:    3.67    0.00    0.25    0.00        A
- Shortest path cycles:        1.25    0.00    0.25    0.00        A
+ Shortest path cycles:        3.67    0.00    0.25    0.00        A
Longest path cycles:          N/A     N/A     N/A     N/A      N/A

A = Arithmetic, LS = Load/Store, V = Varying, T = Texture

Shader properties
=================

Has uniform computation: true
Has side-effects: false
Modifies coverage: false
Uses late ZS test: false
Uses late ZS update: false
Reads color buffer: false



[Mali-G68]
Main shader
===========

Work registers: 32 (100% used at 100% occupancy)
- Uniform registers: 26 (40% used)
+ Uniform registers: 24 (37% used)
Stack spilling: false
- 16-bit arithmetic: 26%
+ 16-bit arithmetic: 29%

                                A      LS       V       T    Bound
- Total instruction cycles:    3.24    0.00    0.12    0.00        A
+ Total instruction cycles:    2.47    0.00    0.12    0.00        A
- Shortest path cycles:        0.20    0.00    0.12    0.00        A
+ Shortest path cycles:        2.47    0.00    0.12    0.00        A
- Longest path cycles:         3.11    0.00    0.12    0.00        A
+ Longest path cycles:         2.47    0.00    0.12    0.00        A

A = Arithmetic, LS = Load/Store, V = Varying, T = Texture

Shader properties
=================

Has uniform computation: true
Has side-effects: false
Modifies coverage: false
Uses late ZS test: false
Uses late ZS update: false
Reads color buffer: false



[Mali-G615]
Main shader
===========

Work registers: 32 (100% used at 100% occupancy)
- Uniform registers: 26 (40% used)
+ Uniform registers: 24 (37% used)
Stack spilling: false
- 16-bit arithmetic: 26%
+ 16-bit arithmetic: 29%

                                A      LS       V       T    Bound
- Total instruction cycles:    1.10    0.00    0.03    0.00        A
+ Total instruction cycles:    0.85    0.00    0.03    0.00        A
- Shortest path cycles:        0.07    0.00    0.03    0.00        A
+ Shortest path cycles:        0.84    0.00    0.03    0.00        A
- Longest path cycles:         1.06    0.00    0.03    0.00        A
+ Longest path cycles:         0.85    0.00    0.03    0.00        A

A = Arithmetic, LS = Load/Store, V = Varying, T = Texture

Shader properties
=================

Has uniform computation: true
Has side-effects: false
Modifies coverage: false
Uses late ZS test: false
Uses late ZS update: false
Reads color buffer: false



[Mali-G610]
Main shader
===========

Work registers: 32 (100% used at 100% occupancy)
- Uniform registers: 26 (40% used)
+ Uniform registers: 24 (37% used)
Stack spilling: false
- 16-bit arithmetic: 26%
+ 16-bit arithmetic: 29%

                                A      LS       V       T    Bound
- Total instruction cycles:    2.20    0.00    0.06    0.00        A
+ Total instruction cycles:    1.70    0.00    0.06    0.00        A
- Shortest path cycles:        0.13    0.00    0.06    0.00        A
+ Shortest path cycles:        1.69    0.00    0.06    0.00        A
- Longest path cycles:         2.12    0.00    0.06    0.00        A
+ Longest path cycles:         1.70    0.00    0.06    0.00        A

A = Arithmetic, LS = Load/Store, V = Varying, T = Texture

Shader properties
=================

Has uniform computation: true
Has side-effects: false
Modifies coverage: false
Uses late ZS test: false
Uses late ZS update: false
Reads color buffer: false



[Mali-G57]
Main shader
===========

Work registers: 32 (100% used at 100% occupancy)
- Uniform registers: 26 (40% used)
+ Uniform registers: 24 (37% used)
Stack spilling: false
- 16-bit arithmetic: 26%
+ 16-bit arithmetic: 29%

                                A      LS       V       T    Bound
- Total instruction cycles:    3.24    0.00    0.12    0.00        A
+ Total instruction cycles:    2.47    0.00    0.12    0.00        A
- Shortest path cycles:        0.20    0.00    0.12    0.00        A
+ Shortest path cycles:        2.47    0.00    0.12    0.00        A
- Longest path cycles:         3.11    0.00    0.12    0.00        A
+ Longest path cycles:         2.47    0.00    0.12    0.00        A

A = Arithmetic, LS = Load/Store, V = Varying, T = Texture

Shader properties
=================

Has uniform computation: true
Has side-effects: false
Modifies coverage: false
Uses late ZS test: false
Uses late ZS update: false
Reads color buffer: false



[Mali-G52]
Main shader
===========

- Work registers: 25 (78% used at 100% occupancy)
+ Work registers: 22 (68% used at 100% occupancy)
- Uniform registers: 22 (34% used)
+ Uniform registers: 16 (25% used)
Stack spilling: false
- 16-bit arithmetic: 35%
+ 16-bit arithmetic: 42%

                                A      LS       V       T    Bound
- Total instruction cycles:    2.53    0.00    0.12    0.00        A
+ Total instruction cycles:    1.67    0.00    0.12    0.00        A
- Shortest path cycles:        0.71    0.00    0.12    0.00        A
+ Shortest path cycles:        1.67    0.00    0.12    0.00        A
Longest path cycles:          N/A     N/A     N/A     N/A      N/A

A = Arithmetic, LS = Load/Store, V = Varying, T = Texture

Shader properties
=================

Has uniform computation: true
Has side-effects: false
Modifies coverage: false
Uses late ZS test: false
Uses late ZS update: false
Reads color buffer: false



[Mali-G510]
Main shader
===========

Work registers: 32 (100% used at 100% occupancy)
- Uniform registers: 26 (40% used)
+ Uniform registers: 24 (37% used)
Stack spilling: false
- 16-bit arithmetic: 26%
+ 16-bit arithmetic: 29%

                                A      LS       V       T    Bound
- Total instruction cycles:    2.94    0.00    0.06    0.00        A
+ Total instruction cycles:    2.27    0.00    0.06    0.00        A
- Shortest path cycles:        0.18    0.00    0.06    0.00        A
+ Shortest path cycles:        2.25    0.00    0.06    0.00        A
- Longest path cycles:         2.83    0.00    0.06    0.00        A
+ Longest path cycles:         2.27    0.00    0.06    0.00        A

A = Arithmetic, LS = Load/Store, V = Varying, T = Texture

Shader properties
=================

Has uniform computation: true
Has side-effects: false
Modifies coverage: false
Uses late ZS test: false
Uses late ZS update: false
Reads color buffer: false



[Mali-G51]
Main shader
===========

- Work registers: 26 (81% used at 100% occupancy)
+ Work registers: 23 (71% used at 100% occupancy)
- Uniform registers: 22 (34% used)
+ Uniform registers: 16 (25% used)
Stack spilling: false
- 16-bit arithmetic: 35%
+ 16-bit arithmetic: 42%

                                A      LS       V       T    Bound
- Total instruction cycles:    5.07    0.00    0.12    0.00        A
+ Total instruction cycles:    3.40    0.00    0.12    0.00        A
- Shortest path cycles:        1.42    0.00    0.12    0.00        A
+ Shortest path cycles:        3.40    0.00    0.12    0.00        A
Longest path cycles:          N/A     N/A     N/A     N/A      N/A

A = Arithmetic, LS = Load/Store, V = Varying, T = Texture

Shader properties
=================

Has uniform computation: true
Has side-effects: false
Modifies coverage: false
Uses late ZS test: false
Uses late ZS update: false
Reads color buffer: false



[Mali-G310]
Main shader
===========

Work registers: 32 (100% used at 100% occupancy)
- Uniform registers: 26 (40% used)
+ Uniform registers: 24 (37% used)
Stack spilling: false
- 16-bit arithmetic: 26%
+ 16-bit arithmetic: 29%

                                A      LS       V       T    Bound
- Total instruction cycles:    4.41    0.00    0.12    0.00        A
+ Total instruction cycles:    3.41    0.00    0.12    0.00        A
- Shortest path cycles:        0.27    0.00    0.12    0.00        A
+ Shortest path cycles:        3.38    0.00    0.12    0.00        A
- Longest path cycles:         4.24    0.00    0.12    0.00        A
+ Longest path cycles:         3.41    0.00    0.12    0.00        A

A = Arithmetic, LS = Load/Store, V = Varying, T = Texture

Shader properties
=================

Has uniform computation: true
Has side-effects: false
Modifies coverage: false
Uses late ZS test: false
Uses late ZS update: false
Reads color buffer: false



[Mali-G31]
Main shader
===========

- Work registers: 26 (81% used at 100% occupancy)
+ Work registers: 23 (71% used at 100% occupancy)
- Uniform registers: 22 (34% used)
+ Uniform registers: 16 (25% used)
Stack spilling: false
- 16-bit arithmetic: 35%
+ 16-bit arithmetic: 42%

                                A      LS       V       T    Bound
- Total instruction cycles:    7.60    0.00    0.12    0.00        A
+ Total instruction cycles:    5.10    0.00    0.12    0.00        A
- Shortest path cycles:        2.12    0.00    0.12    0.00        A
+ Shortest path cycles:        5.10    0.00    0.12    0.00        A
Longest path cycles:          N/A     N/A     N/A     N/A      N/A

A = Arithmetic, LS = Load/Store, V = Varying, T = Texture

Shader properties
=================

Has uniform computation: true
Has side-effects: false
Modifies coverage: false
Uses late ZS test: false
Uses late ZS update: false
Reads color buffer: false



[Immortalis-G715]
Main shader
===========

Work registers: 32 (100% used at 100% occupancy)
- Uniform registers: 26 (40% used)
+ Uniform registers: 24 (37% used)
Stack spilling: false
- 16-bit arithmetic: 26%
+ 16-bit arithmetic: 29%

                                A      LS       V       T    Bound
- Total instruction cycles:    1.10    0.00    0.03    0.00        A
+ Total instruction cycles:    0.85    0.00    0.03    0.00        A
- Shortest path cycles:        0.07    0.00    0.03    0.00        A
+ Shortest path cycles:        0.84    0.00    0.03    0.00        A
- Longest path cycles:         1.06    0.00    0.03    0.00        A
+ Longest path cycles:         0.85    0.00    0.03    0.00        A

A = Arithmetic, LS = Load/Store, V = Varying, T = Texture

Shader properties
=================

Has uniform computation: true
Has side-effects: false
Modifies coverage: false
Uses late ZS test: false
Uses late ZS update: false
Reads color buffer: false

jonahwilliams · 2022-11-28T02:37:31Z

impeller/compiler/shader_lib/impeller/gaussian.glsl

 float IPGaussian(float x, float sigma) {
-  float variance = sigma * sigma;
-  return exp(-0.5 * x * x / variance) / (kSqrtTwoPi * sigma);
+  float variance = pow(sigma, 2.0);


These changes seem to give a minor improvement to register usage but also seeminly reduce the % of 16bit computation since the shaders produced by impellerc seem to always convert mediump to highp.

Overall, marginal win - could be bigger if we didn't lose mediump tho

jonahwilliams · 2022-11-30T23:01:46Z

We're going to hold off on this until we have better support for shader variants in the engine code.

jonahwilliams added 4 commits November 27, 2022 18:24

[Impeller] improve rrect blur performance on low end Android

c244fb5

++

482297a

++

cce965d

++

48271cf

jonahwilliams commented Nov 28, 2022

View reviewed changes

jonahwilliams mentioned this pull request Nov 28, 2022

[Impeller] using glsl built-in functions (pow, max) seems to force highp when compiled with impellerc flutter/flutter#116114

Closed

chinmaygarde assigned jonahwilliams Nov 29, 2022

chinmaygarde added the e: impeller label Nov 29, 2022

jonahwilliams closed this Nov 30, 2022

jonahwilliams mentioned this pull request Jan 27, 2023

[Impeller] support static shader variants using specialization constants flutter/flutter#119357

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[Impeller] improve rrect blur performance on Android #37925

[Impeller] improve rrect blur performance on Android #37925

Uh oh!

jonahwilliams commented Nov 28, 2022 •

edited

Loading

Uh oh!

jonahwilliams Nov 28, 2022

Uh oh!

jonahwilliams Nov 28, 2022

Uh oh!

jonahwilliams commented Nov 30, 2022

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

[Impeller] improve rrect blur performance on Android #37925

[Impeller] improve rrect blur performance on Android #37925

Uh oh!

Conversation

jonahwilliams commented Nov 28, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

jonahwilliams Nov 28, 2022

Choose a reason for hiding this comment

Uh oh!

jonahwilliams Nov 28, 2022

Choose a reason for hiding this comment

Uh oh!

jonahwilliams commented Nov 30, 2022

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

jonahwilliams commented Nov 28, 2022 •

edited

Loading