llama.cpp/ggml/src/vulkan-shaders/argsort.comp
Salvatore Mesoraca 544f409b4b
vulkan : argsort barriers must be under uniform control flow (ggml/951)
a return before a barrier (that happens only in some threads in
a workgroup) leads to UB.
While the old code actually works on some devices,
it fails on some others (i.e. "smaller" GPUs).

BTW, I think it would be better to set specialization constants
when the graph is built, in that way the local workgroup
could be sized appropriately.
But it would take a lot of work.

Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com>
2024-09-29 21:15:37 +03:00

70 lines
2.0 KiB
Plaintext

#version 450
#include "types.comp"
#define BLOCK_SIZE 1024
#define ASC 0
layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) buffer D {int data_d[];};
layout (push_constant) uniform parameter {
uint ncols;
uint ncols_pad;
uint order;
} p;
shared int dst_row[BLOCK_SIZE];
void swap(uint idx0, uint idx1) {
int tmp = dst_row[idx0];
dst_row[idx0] = dst_row[idx1];
dst_row[idx1] = tmp;
}
void main() {
// bitonic sort
const int col = int(gl_LocalInvocationID.x);
const uint row = gl_WorkGroupID.y;
const uint row_offset = row * p.ncols;
// initialize indices
if (col < p.ncols_pad) {
dst_row[col] = col;
}
barrier();
for (uint k = 2; k <= p.ncols_pad; k *= 2) {
for (uint j = k / 2; j > 0; j /= 2) {
const uint ixj = col ^ j;
if (col < p.ncols_pad && ixj > col) {
if ((col & k) == 0) {
if (dst_row[col] >= p.ncols ||
(dst_row[ixj] < p.ncols && (p.order == ASC ?
data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]] :
data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]]))
) {
swap(col, ixj);
}
} else {
if (dst_row[ixj] >= p.ncols ||
(dst_row[col] < p.ncols && (p.order == ASC ?
data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]] :
data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]]))
) {
swap(col, ixj);
}
}
}
barrier();
}
}
if (col < p.ncols) {
data_d[row_offset + col] = dst_row[col];
}
}