Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

vulkan: support larger argsort (#17313)

Browse Source 
* vulkan: support larger argsort

This is an extension of the original bitonic sorting shader that puts the
temporary values in global memory and when more than 1024 threads are needed
it runs multiple workgroups and synchronizes through a pipelinebarrier.

To improve the memory access pattern, a copy of the float value is kept with
the index value. I've applied this same change to the original shared memory
version of the shader, which is still used when ncols <= 1024.

* Reduce the number of shader variants. Use smaller workgroups when doing a single pass, for a modest perf boost

* reduce loop overhead

* run multiple cols per invocation, to reduce barrier overhead
This commit is contained in:
Jeff Bolz
2025-11-19 17:25:50 +01:00
committed by GitHub
parent 2eba631b81
commit 1fa4551af0
5 changed files with 257 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp
+20 -21
View File
@@ -4,28 +4,27 @@
#include "types.glsl"
layout(constant_id = 0) const int BLOCK_SIZE = 1024;
layout(constant_id = 1) const int BLOCK_SIZE_LOG2 = 10;
layout(constant_id = 1) const int NCOLS_PADDED_LOG2 = 10;
#define ASC 0
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 (binding = 2) writeonly buffer D {int data_d[];};
layout (push_constant) uniform parameter {
uint ncols;
uint ncols_padded;
uint ncols_padded_log2;
uint nrows;
uint order;
uint outer_start;
uint outer_end;
uint inner_start;
uint inner_end;
} 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];
dst_row[idx0] = dst_row[idx1];
dst_row[idx1] = tmp;
}
shared ivec2 dst_row[BLOCK_SIZE];
void argsort(bool needs_bounds_check, const uint row) {
// bitonic sort
@@ -34,11 +33,10 @@ void argsort(bool needs_bounds_check, const uint row) {
const uint row_offset = row * p.ncols;
// initialize indices
dst_row[col] = col;
a_sh[col] = data_a[row_offset + col];
dst_row[col] = ivec2(col, floatBitsToInt(data_a[row_offset + col]));
barrier();
uint num_outer_loop_iters = BLOCK_SIZE_LOG2;
uint num_outer_loop_iters = NCOLS_PADDED_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++) {
@@ -47,14 +45,15 @@ void argsort(bool needs_bounds_check, const uint row) {
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;
ivec2 sh_idx_0 = dst_row[idx_0];
ivec2 sh_idx_1 = dst_row[idx_1];
bool idx_0_oob = needs_bounds_check ? sh_idx_0.x >= p.ncols : false;
bool idx_1_oob = needs_bounds_check ? sh_idx_1.x >= 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);
(!idx_1_oob && intBitsToFloat(sh_idx_0.y) > intBitsToFloat(sh_idx_1.y))) && (ixj > col)) {
dst_row[idx_0] = sh_idx_1;
dst_row[idx_1] = sh_idx_0;
}
barrier();
@@ -63,9 +62,9 @@ void argsort(bool needs_bounds_check, const uint row) {
if (col < p.ncols) {
if (p.order == ASC) {
data_d[row_offset + col] = dst_row[col];
data_d[row_offset + col] = dst_row[col].x;
} else {
data_d[row_offset + p.ncols - col - 1] = dst_row[col];
data_d[row_offset + p.ncols - col - 1] = dst_row[col].x;
}
}
}