CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-11 10:36:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: Optimize argsort (#15354)

Browse Source 
- Launch an appropriate number of invocations (next larger power of two).
32 invocations is common and the barrier is much cheaper there.
- Specialize for "needs bounds checking" vs not.
- Make the code less branchy and [[unroll]] the loops. In the final code,
I see no branches inside the main loop (only predicated stores) when
needs_bounds_check is false.
- Always sort ascending, then apply the ascending vs descending option when
doing the final stores to memory.
- Copy the values into shared memory, makes them slightly cheaper to access.
This commit is contained in:
Jeff Bolz
2025-08-17 03:41:45 -05:00
committed by GitHub
parent 65349f26f2
commit de5627910d
3 changed files with 51 additions and 42 deletions
Download Patch File Download Diff File Expand all files Collapse all files

68
ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp
View File

@@ -1,22 +1,24 @@
#version 450
#extension GL_EXT_control_flow_attributes : enable
#include "types.comp"
#define BLOCK_SIZE 1024
layout(constant_id = 0) const int BLOCK_SIZE = 1024;
layout(constant_id = 1) const int BLOCK_SIZE_LOG2 = 10;
#define ASC 0
layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
layout(local_size_x_id = 0, 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];
shared A_TYPE a_sh[BLOCK_SIZE];
void swap(uint idx0, uint idx1) {
int tmp = dst_row[idx0];
@@ -24,7 +26,7 @@ void swap(uint idx0, uint idx1) {
dst_row[idx1] = tmp;
}
void main() {
void argsort(bool needs_bounds_check) {
// bitonic sort
const int col = int(gl_LocalInvocationID.x);
const uint row = gl_WorkGroupID.y;
@@ -32,38 +34,46 @@ void main() {
const uint row_offset = row * p.ncols;
// initialize indices
if (col < p.ncols_pad) {
dst_row[col] = col;
}
dst_row[col] = col;
a_sh[col] = data_a[row_offset + 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);
}
}
uint num_outer_loop_iters = BLOCK_SIZE_LOG2;
[[unroll]] for (uint k = 2, outer_idx = 0; outer_idx < num_outer_loop_iters; k *= 2, outer_idx++) {
uint num_inner_loop_iters = outer_idx + 1;
[[unroll]] for (uint j = k / 2, inner_idx = 0; inner_idx < num_inner_loop_iters; j /= 2, inner_idx++) {
const int ixj = int(col ^ j);
int idx_0 = (col & k) == 0 ? col : ixj;
int idx_1 = (col & k) == 0 ? ixj : col;
int sh_idx_0 = dst_row[idx_0];
int sh_idx_1 = dst_row[idx_1];
bool idx_0_oob = needs_bounds_check ? sh_idx_0 >= p.ncols : false;
bool idx_1_oob = needs_bounds_check ? sh_idx_1 >= p.ncols : false;
if ((idx_0_oob ||
(!idx_1_oob && a_sh[sh_idx_0] > a_sh[sh_idx_1])) && (ixj > col)) {
swap(idx_0, idx_1);
}
barrier();
}
}
if (col < p.ncols) {
data_d[row_offset + col] = dst_row[col];
if (p.order == ASC) {
data_d[row_offset + col] = dst_row[col];
} else {
data_d[row_offset + p.ncols - col - 1] = dst_row[col];
}
}
}
void main() {
if (p.ncols == BLOCK_SIZE) {
argsort(false);
} else {
argsort(true);
}
}