A Shader Compile Error Occurs with Two AOV Outputs in a Specific Shader Node Configuration #115226

New Issue

Takuma-Kudou · 2023-11-21T09:24:17+01:00

Takuma-Kudou commented

2023-11-21 09:24:17 +01:00

System Information
Operating system: Windows 11 Pro
Graphics card: NVIDIA GeForce RTX 3060

Blender Version
Broken: Blender 3.01 - Blender 4.0.1

Short description of error
The combination of a 3D noise texture's color output and a bump to generate a normal, connected to the Normal of a BSDF node, causes a shader compile error when the BSDF node output is connected to two AOV output nodes.

Exact steps for others to reproduce the error
Please refer to the attached blend file.

The key points are as follows:

Set the noise texture to 3D and connect its color output to the height of the bump.
Connect the bump's normal output to the normal input of the BSDF.
Configure the BSDF output to reach two AOV outputs."

**System Information** Operating system: Windows 11 Pro Graphics card: NVIDIA GeForce RTX 3060 **Blender Version** Broken: Blender 3.01 - Blender 4.0.1 **Short description of error** The combination of a 3D noise texture's color output and a bump to generate a normal, connected to the Normal of a BSDF node, causes a shader compile error when the BSDF node output is connected to two AOV output nodes. **Exact steps for others to reproduce the error** Please refer to the attached blend file. The key points are as follows: - Set the noise texture to 3D and connect its color output to the height of the bump. - Connect the bump's normal output to the normal input of the BSDF. - Configure the BSDF output to reach two AOV outputs." ![image](/attachments/2608d01a-8f29-40a5-bcaf-afcfafb6c780)

IMG_3809.jpeg

240 KiB

shader_compile_error.blend

816 KiB

Takuma-Kudou added the

labels 2023-11-21 09:24:18 +01:00

Pratik Borhade commented

2023-11-21 10:34:20 +01:00

Hi, thanks for the report. This shader is complied correctly in 3.3 and prior versions. Raising the priority for awareness. I'll bisect.

Pratik Borhade added

and removed

labels 2023-11-21 10:34:34 +01:00

Philipp Oeser commented

2023-11-22 14:28:51 +01:00

Broke between 909f47e0e1 and 7a6cdeb242

ERROR (gpu.shader): GPU_material_compile Linking: 
      | Fragment info
      | -------------
      | C7050: "out_aov" might be used before being initialized

Broke between 909f47e0e197 and 7a6cdeb24279 ``` ERROR (gpu.shader): GPU_material_compile Linking: | Fragment info | ------------- | C7050: "out_aov" might be used before being initialized ```

Miguel Pozo commented

2023-11-23 19:28:47 +01:00

Note that the relevant part of the error is this one:

      | Internal error: assembly compile error for fragment shader at offset 1632592:
      | -- error message --
      | line 65847, column 1:  error: too many instructions
      | -- internal assembly text --

Reverting a84c92fc73 seems to fix it, however, I really don't understand why.
While the generated code is still far from optimal, that commit only removes death code.

This is what the diff between the generated code after that commit looks like:

diff --git a/before.glsl b/after.glsl
index 84d04ad..e6ca3a6 100644
--- a/before.glsl.c
+++ b/after.glsl.c
@@ -127,52 +127,10 @@ node_noise_tex_fbm_3d(tmp10, node_tree.u12, node_tree.u13, node_tree.u14, node_t
 float tmp25;
 set_value(float_from_vec4(tmp22), tmp25);
 
-vec3 tmp71;
-differentiate_texco(var_attrs.v0, tmp71);
-
-vec3 tmp79;
-mapping_mat4(tmp71, node_tree.u73, node_tree.u74, node_tree.u75, node_tree.u76, node_tree.u77, node_tree.u78, tmp79);
-
-float cons89 = float(uintBitsToFloat(1065353216u));
-float tmp90;
-vec4 tmp91;
-node_noise_tex_fbm_3d(tmp79, node_tree.u81, node_tree.u82, node_tree.u83, node_tree.u84, node_tree.u85, node_tree.u86, node_tree.u87, node_tree.u88, cons89, tmp90, tmp91);
-
-float tmp94;
-set_value(float_from_vec4(tmp91), tmp94);
-
-vec3 tmp145;
-differentiate_texco(var_attrs.v0, tmp145);
-
-vec3 tmp153;
-mapping_mat4(tmp145, node_tree.u147, node_tree.u148, node_tree.u149, node_tree.u150, node_tree.u151, node_tree.u152, tmp153);
-
-float cons163 = float(uintBitsToFloat(1065353216u));
-float tmp164;
-vec4 tmp165;
-node_noise_tex_fbm_3d(tmp153, node_tree.u155, node_tree.u156, node_tree.u157, node_tree.u158, node_tree.u159, node_tree.u160, node_tree.u161, node_tree.u162, cons163, tmp164, tmp165);
-
-float tmp168;
-set_value(float_from_vec4(tmp165), tmp168);
-
 return tmp25;
 }
 
 float ntree_fn1() {
-vec3 tmp2;
-differentiate_texco(var_attrs.v0, tmp2);
-
-vec3 tmp10;
-mapping_mat4(tmp2, node_tree.u4, node_tree.u5, node_tree.u6, node_tree.u7, node_tree.u8, node_tree.u9, tmp10);
-
-float cons20 = float(uintBitsToFloat(1065353216u));
-float tmp21;
-vec4 tmp22;
-node_noise_tex_fbm_3d(tmp10, node_tree.u12, node_tree.u13, node_tree.u14, node_tree.u15, node_tree.u16, node_tree.u17, node_tree.u18, node_tree.u19, cons20, tmp21, tmp22);
-
-float tmp25;
-set_value(float_from_vec4(tmp22), tmp25);
-
 vec3 tmp71;
 differentiate_texco(var_attrs.v0, tmp71);
 
@@ -187,52 +145,10 @@ node_noise_tex_fbm_3d(tmp79, node_tree.u81, node_tree.u82, node_tree.u83, node_t
 float tmp94;
 set_value(float_from_vec4(tmp91), tmp94);
 
-vec3 tmp145;
-differentiate_texco(var_attrs.v0, tmp145);
-
-vec3 tmp153;
-mapping_mat4(tmp145, node_tree.u147, node_tree.u148, node_tree.u149, node_tree.u150, node_tree.u151, node_tree.u152, tmp153);
-
-float cons163 = float(uintBitsToFloat(1065353216u));
-float tmp164;
-vec4 tmp165;
-node_noise_tex_fbm_3d(tmp153, node_tree.u155, node_tree.u156, node_tree.u157, node_tree.u158, node_tree.u159, node_tree.u160, node_tree.u161, node_tree.u162, cons163, tmp164, tmp165);
-
-float tmp168;
-set_value(float_from_vec4(tmp165), tmp168);
-
 return tmp94;
 }
 
 float ntree_fn2() {
-vec3 tmp2;
-differentiate_texco(var_attrs.v0, tmp2);
-
-vec3 tmp10;
-mapping_mat4(tmp2, node_tree.u4, node_tree.u5, node_tree.u6, node_tree.u7, node_tree.u8, node_tree.u9, tmp10);
-
-float cons20 = float(uintBitsToFloat(1065353216u));
-float tmp21;
-vec4 tmp22;
-node_noise_tex_fbm_3d(tmp10, node_tree.u12, node_tree.u13, node_tree.u14, node_tree.u15, node_tree.u16, node_tree.u17, node_tree.u18, node_tree.u19, cons20, tmp21, tmp22);
-
-float tmp25;
-set_value(float_from_vec4(tmp22), tmp25);
-
-vec3 tmp71;
-differentiate_texco(var_attrs.v0, tmp71);
-
-vec3 tmp79;
-mapping_mat4(tmp71, node_tree.u73, node_tree.u74, node_tree.u75, node_tree.u76, node_tree.u77, node_tree.u78, tmp79);
-
-float cons89 = float(uintBitsToFloat(1065353216u));
-float tmp90;
-vec4 tmp91;
-node_noise_tex_fbm_3d(tmp79, node_tree.u81, node_tree.u82, node_tree.u83, node_tree.u84, node_tree.u85, node_tree.u86, node_tree.u87, node_tree.u88, cons89, tmp90, tmp91);
-
-float tmp94;
-set_value(float_from_vec4(tmp91), tmp94);
-
 vec3 tmp145;
 differentiate_texco(var_attrs.v0, tmp145);

Note that the relevant part of the error is this one: ``` | Internal error: assembly compile error for fragment shader at offset 1632592: | -- error message -- | line 65847, column 1: error: too many instructions | -- internal assembly text -- ``` Reverting a84c92fc73 seems to fix it, however, I really don't understand why. While the generated code is still far from optimal, that commit only removes death code. This is what the diff between the generated code after that commit looks like: ```diff diff --git a/before.glsl b/after.glsl index 84d04ad..e6ca3a6 100644 --- a/before.glsl.c +++ b/after.glsl.c @@ -127,52 +127,10 @@ node_noise_tex_fbm_3d(tmp10, node_tree.u12, node_tree.u13, node_tree.u14, node_t float tmp25; set_value(float_from_vec4(tmp22), tmp25); -vec3 tmp71; -differentiate_texco(var_attrs.v0, tmp71); - -vec3 tmp79; -mapping_mat4(tmp71, node_tree.u73, node_tree.u74, node_tree.u75, node_tree.u76, node_tree.u77, node_tree.u78, tmp79); - -float cons89 = float(uintBitsToFloat(1065353216u)); -float tmp90; -vec4 tmp91; -node_noise_tex_fbm_3d(tmp79, node_tree.u81, node_tree.u82, node_tree.u83, node_tree.u84, node_tree.u85, node_tree.u86, node_tree.u87, node_tree.u88, cons89, tmp90, tmp91); - -float tmp94; -set_value(float_from_vec4(tmp91), tmp94); - -vec3 tmp145; -differentiate_texco(var_attrs.v0, tmp145); - -vec3 tmp153; -mapping_mat4(tmp145, node_tree.u147, node_tree.u148, node_tree.u149, node_tree.u150, node_tree.u151, node_tree.u152, tmp153); - -float cons163 = float(uintBitsToFloat(1065353216u)); -float tmp164; -vec4 tmp165; -node_noise_tex_fbm_3d(tmp153, node_tree.u155, node_tree.u156, node_tree.u157, node_tree.u158, node_tree.u159, node_tree.u160, node_tree.u161, node_tree.u162, cons163, tmp164, tmp165); - -float tmp168; -set_value(float_from_vec4(tmp165), tmp168); - return tmp25; } float ntree_fn1() { -vec3 tmp2; -differentiate_texco(var_attrs.v0, tmp2); - -vec3 tmp10; -mapping_mat4(tmp2, node_tree.u4, node_tree.u5, node_tree.u6, node_tree.u7, node_tree.u8, node_tree.u9, tmp10); - -float cons20 = float(uintBitsToFloat(1065353216u)); -float tmp21; -vec4 tmp22; -node_noise_tex_fbm_3d(tmp10, node_tree.u12, node_tree.u13, node_tree.u14, node_tree.u15, node_tree.u16, node_tree.u17, node_tree.u18, node_tree.u19, cons20, tmp21, tmp22); - -float tmp25; -set_value(float_from_vec4(tmp22), tmp25); - vec3 tmp71; differentiate_texco(var_attrs.v0, tmp71); @@ -187,52 +145,10 @@ node_noise_tex_fbm_3d(tmp79, node_tree.u81, node_tree.u82, node_tree.u83, node_t float tmp94; set_value(float_from_vec4(tmp91), tmp94); -vec3 tmp145; -differentiate_texco(var_attrs.v0, tmp145); - -vec3 tmp153; -mapping_mat4(tmp145, node_tree.u147, node_tree.u148, node_tree.u149, node_tree.u150, node_tree.u151, node_tree.u152, tmp153); - -float cons163 = float(uintBitsToFloat(1065353216u)); -float tmp164; -vec4 tmp165; -node_noise_tex_fbm_3d(tmp153, node_tree.u155, node_tree.u156, node_tree.u157, node_tree.u158, node_tree.u159, node_tree.u160, node_tree.u161, node_tree.u162, cons163, tmp164, tmp165); - -float tmp168; -set_value(float_from_vec4(tmp165), tmp168); - return tmp94; } float ntree_fn2() { -vec3 tmp2; -differentiate_texco(var_attrs.v0, tmp2); - -vec3 tmp10; -mapping_mat4(tmp2, node_tree.u4, node_tree.u5, node_tree.u6, node_tree.u7, node_tree.u8, node_tree.u9, tmp10); - -float cons20 = float(uintBitsToFloat(1065353216u)); -float tmp21; -vec4 tmp22; -node_noise_tex_fbm_3d(tmp10, node_tree.u12, node_tree.u13, node_tree.u14, node_tree.u15, node_tree.u16, node_tree.u17, node_tree.u18, node_tree.u19, cons20, tmp21, tmp22); - -float tmp25; -set_value(float_from_vec4(tmp22), tmp25); - -vec3 tmp71; -differentiate_texco(var_attrs.v0, tmp71); - -vec3 tmp79; -mapping_mat4(tmp71, node_tree.u73, node_tree.u74, node_tree.u75, node_tree.u76, node_tree.u77, node_tree.u78, tmp79); - -float cons89 = float(uintBitsToFloat(1065353216u)); -float tmp90; -vec4 tmp91; -node_noise_tex_fbm_3d(tmp79, node_tree.u81, node_tree.u82, node_tree.u83, node_tree.u84, node_tree.u85, node_tree.u86, node_tree.u87, node_tree.u88, cons89, tmp90, tmp91); - -float tmp94; -set_value(float_from_vec4(tmp91), tmp94); - vec3 tmp145; differentiate_texco(var_attrs.v0, tmp145); ```

after.glsl

6.7 KiB

before.glsl

9.7 KiB

Miguel Pozo referenced this issue from a commit

2023-12-22 17:32:35 +01:00

Fix #115226: Shader Nodes: Don't execute the same node twice

Miguel Pozo closed this issue

2023-12-22 17:32:36 +01:00

Blender Bot added

and removed

labels 2023-12-22 17:32:38 +01:00

Miguel Pozo commented

2023-12-22 17:40:22 +01:00

I still don't know why the Nvidia compiler freaks out after removing dead code, but I've pushed an improvement for AOVs that further optimizes the generated code and seems to do the trick. 🤷‍♂️

Miguel Pozo referenced this issue

2024-01-08 18:46:22 +01:00

GPU: Print shader source on compilation error #116470

Sign in to join this conversation.

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

A Shader Compile Error Occurs with Two AOV Outputs in a Specific Shader Node Configuration #115226