llama.cpp/vulkan-shaders/rope_neox.comp
0cc4m 7c7836d9d4
Vulkan Shader Refactor, Memory Debugging Option (#7947)
* Refactor shaders, extract GLSL code from ggml_vk_generate_shaders.py into vulkan-shaders directory

* Improve debug log code

* Add memory debug output option

* Fix flake8

* Fix unnecessary high llama-3 VRAM use
2024-06-16 07:17:31 +02:00

38 lines
929 B
Plaintext

#version 450
#include "rope_head.comp"
void main() {
const uint col = gl_GlobalInvocationID.y * 2;
const uint row = gl_GlobalInvocationID.x;
if (col >= p.ncols) {
return;
}
if (col >= p.n_dims) {
const uint i = row*p.ncols + col;
data_d[i + 0] = data_a[i + 0];
data_d[i + 1] = data_a[i + 1];
return;
}
const uint i = row*p.ncols + col/2;
const uint i2 = row/p.p_delta_rows;
const float theta_base = data_pos[i2] * pow(p.theta_scale, col/2.0f);
const float freq_factor = p.has_ff != 0 ? data_ff[col/2] : 1.0f;
float cos_theta, sin_theta;
rope_yarn(theta_base / freq_factor, col, cos_theta, sin_theta);
const float x0 = float(data_a[i + 0]);
const float x1 = float(data_a[i + p.n_dims/2]);
data_d[i + 0] = D_TYPE(x0*cos_theta - x1*sin_theta);
data_d[i + p.n_dims/2] = D_TYPE(x0*sin_theta + x1*cos_theta);
}