Question Shader.SetTexture, which will perform better? Optimization ideas?

I have a shader with multiple passes, where every pass requires some textures and variables which I need to bind from a C# script.

I want to execute these passes using Unity and render the result into a RenderTexture.

As I have many shader passes I came across performance problems while executing the shader on mobile devices using Metal as graphic backend (I used the Profiler and found out my application mainly stucks on Gfx.WaitForPresent, while doing so, my render thread still spends time in Camera.Render, so when I understood this [https://forum.unity.com/threads/pro...nd-editorloop-too-high-with-vsync-off.620854/] correctly, my application is CPU bound, potentially because I use a lot of shader.SetTexture and shader.SetFloat calls) . Therefore I have a question regarding performance:

Let's assume a simple scenario where I have 4 RenderTextures a, b, c, d and two passes. In the first pass I want to blit from a and b to c and in the second pass I want to blit from c to d.
Which version is better performance-wise (bind every RenderTexture once to the shader using a separate sampler or bind only the needed RenderTextures on every pass)?
Also do you have any other ideas how I can potentially increase performance given the sample code (it's highly simplified, my actual shader is more complex for sure)? Is it worth to change from Graphic.Blit to CommandBuffers? Anything else which I can try?

Version A (bind every RenderTexture ONCE using SetTexture in Start method):
C#

Code (CSharp):

1. Start(){
2.  
3. RenderTexture a = new RenderTexture(...);
4. RenderTexture b = new RenderTexture(...);
5. RenderTexture c = new RenderTexture(...);
6. RenderTexture d = new RenderTexture(...);
7. Material shader = Shader.Find("Unlit/CustomShader");
8. shader.SetTexture("a", a);
9. shader.SetTexture("b", b);
10. shader.SetTexture("c", c);
11. }
12.  
13. Update(){
14. // Execute first pass
15. Graphics.Blit(null, c, shader, 0);
16. // Execute second passs
17. Graphics.Blit(null, d, shader, 1);
18. }

Shader:

Code (CSharp):

1. Shader "Unlit/CustomShader"
2. {
3. SubShader
4. {
5. CGINCLUDE
6. #include "UnityCG.cginc"
7. uniform sampler2D a, b, c, d;
8.  
9. struct appdata
10. {
11. float4 vertex : POSITION;
12. float2 uv : TEXCOORD0;
13. };
14.  
15. struct v2f
16. {
17. float4 vertex : SV_POSITION;
18. float2 uv : TEXCOORD0;
19. };
20.  
21. v2f vert (appdata v)
22. {
23. v2f o;
24. o.vertex = UnityObjectToClipPos(v.vertex);
25. o.uv = v.uv;
26. return o;
27. }
28.  
29. ENDCG
30. Pass {
31. NAME "PASSONE"
32. Blend One OneMinusSrcAlpha
33. CGPROGRAM
34. #pragma vertex vert
35. #pragma fragment frag
36.  
37. half4 frag (v2f i) : SV_Target
38. {
39. return 0.5 * (tex2D(a, i.uv) + tex2D(b, i.uv));
40. }
41. ENDCG
42. }
43. Pass {
44. NAME "PASSTWO"
45. Blend One OneMinusSrcAlpha
46. CGPROGRAM
47. #pragma vertex vert
48. #pragma fragment frag
49.  
50. half4 frag (v2f i) : SV_Target
51. {
52. return tex2D(c, i.uv);
53. }
54. ENDCG
55. }
56. }
57. }

Version B (bind only needed RenderTextures every Update using SetTexture, we now save one sampler2D but need to execute SetTexture way more often):
C#

Code (CSharp):

1. Start(){
2.  
3. RenderTexture a = new RenderTexture(...);
4. RenderTexture b = new RenderTexture(...);
5. RenderTexture c = new RenderTexture(...);
6. RenderTexture d = new RenderTexture(...);
7. Material shader = Shader.Find("Unlit/CustomShader");
8. }
9.  
10. Update(){
11. // Bind
12. shader.SetTexture("tex1", a);
13. shader.SetTexture("tex2", b);
14. // Execute first pass
15. Graphics.Blit(null, c, shader, 0);
16.  
17. // Bind
18. shader.SetTexture("tex1", c);
19. // Execute second passs
20. Graphics.Blit(null, d, shader, 1);
21. }

Shader:

Code (CSharp):

1. Shader "Unlit/CustomShader"
2. {
3. SubShader
4. {
5. CGINCLUDE
6. #include "UnityCG.cginc"
7. uniform sampler2D tex1, tex2;
8.  
9. struct appdata
10. {
11. float4 vertex : POSITION;
12. float2 uv : TEXCOORD0;
13. };
14.  
15. struct v2f
16. {
17. float4 vertex : SV_POSITION;
18. float2 uv : TEXCOORD0;
19. };
20.  
21. v2f vert (appdata v)
22. {
23. v2f o;
24. o.vertex = UnityObjectToClipPos(v.vertex);
25. o.uv = v.uv;
26. return o;
27. }
28.  
29. ENDCG
30. Pass {
31. NAME "PASSONE"
32. Blend One OneMinusSrcAlpha
33. CGPROGRAM
34. #pragma vertex vert
35. #pragma fragment frag
36.  
37. half4 frag (v2f i) : SV_Target
38. {
39. return 0.5 * (tex2D(tex1, i.uv) + tex2D(tex2, i.uv));
40. }
41. ENDCG
42. }
43. Pass {
44. NAME "PASSTWO"
45. Blend One OneMinusSrcAlpha
46. CGPROGRAM
47. #pragma vertex vert
48. #pragma fragment frag
49.  
50. half4 frag (v2f i) : SV_Target
51. {
52. return tex2D(tex1, i.uv);
53. }
54. ENDCG
55. }
56. }
57. }

 

Does version B create new material instances in Update?

Honestly, what you're describing seems like it might just be too demanding for mobile. How many screens worth of pixels are you trying to draw each frame?

 

@trickster721 Thanks for the reply!
No I don't think it will create a new material instance, I only create an instance once while initialize (Shader.Find("Unlit/CustomShader") as far as I know.

It's working on mobile with 60FPS but my smartphones get's warm when I use full HD (1920x1080), so I'm interested in optimizing the performance.