Julian
@schneckerstein.bsky.social
-10x Code-Person @symmetrybreak.bsky.social
Working on Misgiven and Tasty Grass Shader
Working on Misgiven and Tasty Grass Shader
But... I'm 30 and I love C?
April 5, 2025 at 12:54 PM
But... I'm 30 and I love C?
And on top:
Just because you got Volumetric Light, it doesn't mean that everything is shrouded by fog.
Just because you got Volumetric Light, it doesn't mean that everything is shrouded by fog.
March 19, 2025 at 10:30 AM
And on top:
Just because you got Volumetric Light, it doesn't mean that everything is shrouded by fog.
Just because you got Volumetric Light, it doesn't mean that everything is shrouded by fog.
endlich wieder twitter vibes hier
January 7, 2025 at 5:40 PM
endlich wieder twitter vibes hier
This concept is of course possible with Vector2/3/x as well. Take this example, where I calculate the magnitude/length of four 2D points at once.
Keep in mind: I barely scratched the surface here of what's possible, not to mention that SEE1 came out in 1999.
Happy Coding!
(5/5)
Keep in mind: I barely scratched the surface here of what's possible, not to mention that SEE1 came out in 1999.
Happy Coding!
(5/5)
![#include <xmmintrin.h>
void Distances_SIMD(void)
{
float points_X_arr[] = {16.0f, 00.0f, 10.0f, 43.0f};
float points_Y_arr[] = {00.0f, 16.0f, 10.0f, 08.0f};
__m128 points_X = _mm_load_ps(points_X_arr);
__m128 points_Y = _mm_load_ps(points_Y_arr);
__m128 xSqr = _mm_mul_ps(points_X, points_X);
__m128 ySqr = _mm_mul_ps(points_Y, points_Y);
__m128 length = _mm_sqrt_ps(_mm_add_ps(xSqr, ySqr));
float lengths[4];
_mm_store_ps(lengths, length);
for(int i = 0; i < 4; i++){
printf("Point = (%f,%f), Length = %f.\n", points_X_arr[i], points_Y_arr[i], length[i]);
}
}](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:uq2c4kut7jfnkegkypntvwrp/bafkreifjymp6xxfq2lyfwo3eyyfdcjpf3og3jldnykj625a3kumxhale4i@jpeg)
January 3, 2025 at 2:31 PM
This concept is of course possible with Vector2/3/x as well. Take this example, where I calculate the magnitude/length of four 2D points at once.
Keep in mind: I barely scratched the surface here of what's possible, not to mention that SEE1 came out in 1999.
Happy Coding!
(5/5)
Keep in mind: I barely scratched the surface here of what's possible, not to mention that SEE1 came out in 1999.
Happy Coding!
(5/5)
Now for thinking as SIMD as "multiple things at once" the code would look like this.
The key is here that we batched multiple calculation together, processing 4 circles at once.
(4/5)
The key is here that we batched multiple calculation together, processing 4 circles at once.
(4/5)
January 3, 2025 at 2:31 PM
Now for thinking as SIMD as "multiple things at once" the code would look like this.
The key is here that we batched multiple calculation together, processing 4 circles at once.
(4/5)
The key is here that we batched multiple calculation together, processing 4 circles at once.
(4/5)
Instead, you have to think about SIMD as processing multiple items at a once. For example: you have a list of circles and want to calculate their area. The basic version would process one circle at the time.
With thinking of SIMD as "fast Vector4" there would be no room for improvement. (3/5)
With thinking of SIMD as "fast Vector4" there would be no room for improvement. (3/5)
![void CircleArea_SIMD(void)
{
float circlesRadii[] = {1.0f, 2.0f, 0.5f, 4.0f};
__m128 circlesRadius = _mm_load_ps(circlesRadii);
const float pi = 3.141f;
__m128 pi_wide = _mm_load_ps1(&pi);
// caluclate: pi * radius * radius, but for 4 radii at once.
__m128 areas = _mm_mul_ps(pi_wide, _mm_mul_ps(circlesRadius, circlesRadius));
float areaScalar[4];
_mm_store_ps(areaScalar, areas);
for(int i = 0; i < 4; i++){
printf("Radius = %f, Area = %f.\n", circlesRadii[i], areaScalar[i]);
}
}](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:uq2c4kut7jfnkegkypntvwrp/bafkreicitpkilwyvuppg3q4ibxzti2fxak3lmjec7k7dl7m2ofkrlmxmtm@jpeg)
January 3, 2025 at 2:31 PM
Instead, you have to think about SIMD as processing multiple items at a once. For example: you have a list of circles and want to calculate their area. The basic version would process one circle at the time.
With thinking of SIMD as "fast Vector4" there would be no room for improvement. (3/5)
With thinking of SIMD as "fast Vector4" there would be no room for improvement. (3/5)
The key point is that many people get the *impression* that these wide SIMD vectors, like __m128 in SSE, are a direct replacement for a Vector4. When they then try to "SIMDfy" their Vector3/2 math, their code gets ugly and inefficient.
This is not how to SIMD.☝️ (2/5)
This is not how to SIMD.☝️ (2/5)
January 3, 2025 at 2:31 PM
The key point is that many people get the *impression* that these wide SIMD vectors, like __m128 in SSE, are a direct replacement for a Vector4. When they then try to "SIMDfy" their Vector3/2 math, their code gets ugly and inefficient.
This is not how to SIMD.☝️ (2/5)
This is not how to SIMD.☝️ (2/5)
Doc, I am so sorry.
December 19, 2024 at 9:12 PM
Doc, I am so sorry.