vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

vulkan: im2col and matmul optimizations for stable diffusion (#10942)

Browse source 
* tests: Add im2col perf tests

* vulkan: optimize im2col, more elements per thread

* vulkan: increase small tile size for NV_coopmat2

* vulkan: change im2col to 512 elements per workgroup
This commit is contained in:
Jeff Bolz 2024-12-29 03:16:34 -06:00 • committed by GitHub
parent fdd2188912
commit a813badbbd
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 66 additions and 29 deletions
Download patch file Download diff file Expand all files Collapse all files

73
ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp
View file

@ -2,6 +2,7 @@
#extension GL_EXT_shader_16bit_storage : require
#extension GL_EXT_spirv_intrinsics: enable
#extension GL_EXT_control_flow_attributes : require
#if RTE16
spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
@ -23,40 +24,64 @@ layout (push_constant) uniform parameter
#include "types.comp"
#define BLOCK_SIZE 256
layout(constant_id = 0) const uint BLOCK_SIZE = 32;
layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
const uint NUM_ITER = 512 / BLOCK_SIZE;
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
void main() {
const uint i = gl_GlobalInvocationID.x;
if (i >= p.pelements) {
return;
}
const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1);
const uint kx = i / ksize;
const uint kd = kx * ksize;
const uint ky = (i - kd) / p.OW;
const uint ix = i % p.OW;
const uint gidx = gl_GlobalInvocationID.x;
const uint oh = gl_GlobalInvocationID.y;
const uint batch = gl_GlobalInvocationID.z / p.IC;
const uint ic = gl_GlobalInvocationID.z % p.IC;
const uint iiw = ix * p.s0 + kx * p.d0 - p.p0;
const uint iih = oh * p.s1 + ky * p.d1 - p.p1;
const uint offset_dst =
((batch * p.OH + oh) * p.OW + ix) * p.CHW +
(ic * (p.KW * p.KH) + ky * p.KW + kx);
if (iih < 0 || iih >= p.IH || iiw < 0 || iiw >= p.IW) {
data_d[offset_dst] = D_TYPE(0.0f);
} else {
const uint offset_src = ic * p.offset_delta + batch * p.batch_offset;
data_d[offset_dst] = D_TYPE(data_a[offset_src + iih * p.IW + iiw]);
A_TYPE values[NUM_ITER];
uint offset_dst[NUM_ITER];
[[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) {
values[idx] = A_TYPE(0);
}
[[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) {
const uint i = gidx * NUM_ITER + idx;
const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1);
const uint kx = i / ksize;
const uint kd = kx * ksize;
const uint ky = (i - kd) / p.OW;
const uint ix = i % p.OW;
const uint iiw = ix * p.s0 + kx * p.d0 - p.p0;
const uint iih = oh * p.s1 + ky * p.d1 - p.p1;
offset_dst[idx] =
((batch * p.OH + oh) * p.OW + ix) * p.CHW +
(ic * (p.KW * p.KH) + ky * p.KW + kx);
if (i >= p.pelements) {
continue;
}
if (iih < p.IH && iiw < p.IW) {
const uint offset_src = ic * p.offset_delta + batch * p.batch_offset;
values[idx] = data_a[offset_src + iih * p.IW + iiw];
}
}
[[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) {
const uint i = gidx * NUM_ITER + idx;
if (i >= p.pelements) {
continue;
}
data_d[offset_dst[idx]] = D_TYPE(values[idx]);
}
}