mirror of
https://github.com/ggerganov/llama.cpp.git
synced 2025-01-01 00:39:00 +01:00
5fd89a70ea
* Optimize Vulkan REPEAT performance * Use Vulkan GLSL fused multiply-add instruction where possible * Add GGML_VULKAN_PERF option to output performance data per operator * Rework and fix Vulkan descriptor set and descriptor pool handling * Fix float32 concat f16 shader validation error * Add Vulkan GROUP_NORM eps parameter * Fix validation error with transfer queue memory barrier flags * Remove trailing whitespaces
40 lines
1.3 KiB
Plaintext
40 lines
1.3 KiB
Plaintext
#version 450
|
|
|
|
#include "types.comp"
|
|
#include "generic_binary_head.comp"
|
|
|
|
void main() {
|
|
const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
|
|
const int dim = p.param3;
|
|
|
|
if (idx >= p.ne) {
|
|
return;
|
|
}
|
|
|
|
const uint i3 = idx / (p.ne22*p.ne21*p.ne20);
|
|
const uint i3_offset = i3 * p.ne22*p.ne21*p.ne20;
|
|
const uint i2 = (idx - i3_offset) / (p.ne21*p.ne20);
|
|
const uint i2_offset = i2*p.ne21*p.ne20;
|
|
const uint i1 = (idx - i3_offset - i2_offset) / p.ne20;
|
|
const uint i0 = idx - i3_offset - i2_offset - i1*p.ne20;
|
|
|
|
uint o[4] = {0, 0, 0, 0};
|
|
o[dim] = dim == 0 ? p.ne00 : (dim == 1 ? p.ne01 : (dim == 2 ? p.ne02 : p.ne03));
|
|
|
|
const uint src0_idx = i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0*p.nb00;
|
|
const uint src1_idx = (i3 - o[3])*p.nb13 + (i2 - o[2])*p.nb12 + (i1 - o[1])*p.nb11 + (i0 - o[0])*p.nb10;
|
|
const uint dst_idx = i3*p.nb23 + i2*p.nb22 + i1*p.nb21 + i0*p.nb20;
|
|
|
|
const bool is_src0 = i0 < p.ne00 && i1 < p.ne01 && i2 < p.ne02 && i3 < p.ne03;
|
|
|
|
#ifndef OPTIMIZATION_ERROR_WORKAROUND
|
|
data_d[p.d_offset + dst_idx] = D_TYPE(is_src0 ? data_a[src0_idx] : data_b[src1_idx]);
|
|
#else
|
|
if (is_src0) {
|
|
data_d[p.d_offset + dst_idx] = data_a[src0_idx];
|
|
} else {
|
|
data_d[p.d_offset + dst_idx] = data_b[src1_idx];
|
|
}
|
|
#endif
|
|
}
|