Working example

shaders/fp32.comp

 #version 450
 
-layout(local_size_x = 1) in;
+layout(local_size_x = 1024) in;
 
 // The input-output tensors bind index is relative to index in parameter passed
-layout(set = 0, binding = 0) buffer buffer_in_data {
-float data[];
-};
+layout(set = 0, binding = 0) buffer buffer_in_data {float data[];};
 
-void main() {
-uint index = gl_GlobalInvocationID.x;
- // Iterating to simulate longer process
-for(int i = 0;
-i < 2048;
-i ++) {
-fma(data[index], data[index], data[index]);
-fma(data[index], - data[index], data[index]);
-fma(data[index], data[index], - data[index]);
-fma(data[index], data[index], data[index]);
+vec2 fp32_8(vec2 a, vec2 b, vec2 c) {
+    vec2 result = {0.f, 0.f};
+    for(int i = 0; i < 2048; i ++) {
+        result.x = fma(a.x, b.x, c.x);
+        result.x = fma(-a.y, b.y, result.x);
+        result.y = fma(a.x, b.y, result.x);
+        result.y = fma(a.y, b.x, result.y);
+    }
+    return result;
 }
+
+void main() {
+    uint index = gl_GlobalInvocationID.x;
+    // Iterating to simulate longer process
+    vec2 a = { data[index], data[index + 1] };
+    vec2 b = { 1.f, 2.f };
+    vec2 c = { 3.f, 4.f };
+    vec2 d = {0.f, 0.f};
+    
+    for(int i = 0; i < 2048; i ++) {
+        d = d+ fp32_8(a, b, c);
+    }
+
+    data[index] = d.x + d.y;
 }