DarthAffe/stable-diffusion.cpp
1
0 Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-11-11 11:26:34 +00:00
Code Issues Packages Projects Releases Wiki Activity

perf(vulkan): accelerate MOSS normalization and SiTU staging (#2273)

Browse Source 
Co-authored-by: mirek190 <mirek190@gmail.com>
This commit is contained in:
leejetandmirek190 authored and GitHub committed 2026-11-05 03:01:31 +08:00
1 parent d1358945dc
commit 58bb3c1b1c
8 files changed
+358 -17
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/core/vulkan/rope.cpp
+20 -12
View File
@@ -72,7 +72,7 @@ namespace sd_vulkan_rope {
return false;
const auto* x = norm->src[0];
const auto* weight = mul->src[1];
return x && weight && x->type == GGML_TYPE_F32 && x->ne[0] == 128 && x->nb[0] == 4 &&
return x && weight && weight != norm && x->type == GGML_TYPE_F32 && x->ne[0] == 128 && x->nb[0] == 4 &&
norm->type == GGML_TYPE_F32 && mul->type == GGML_TYPE_F32 && ggml_are_same_shape(x, norm) && ggml_are_same_shape(norm, mul) &&
weight->type == GGML_TYPE_F32 && weight->ne[0] == 128 && ggml_nelements(weight) == 128 &&
ggml_is_contiguous(weight) && bounded(weight, info) && bounded(x, info) && (x->ne[1] + 3) / 4 <= info.max_group_count[0];
@@ -81,32 +81,35 @@ namespace sd_vulkan_rope {
ggml_ext_vulkan_device_info info;
api(device)->device_info(device, &info);
// Workspace writeback uses Vulkan's four-byte copy granularity.
return info.float16 && scale->op == GGML_OP_SCALE && scale->src[0] == rope && scale->type == GGML_TYPE_F32 &&
param(scale, 1) == 0.f && copy->op == GGML_OP_CPY && copy->src[0] == scale && copy->type == GGML_TYPE_F16 &&
ggml_is_contiguous(copy) && aligned(copy, 4) && ggml_are_same_shape(scale, copy);
return info.float16 && (!scale || (scale->op == GGML_OP_SCALE && scale->src[0] == rope && scale->type == GGML_TYPE_F32 && param(scale, 1) == 0.f)) &&
copy->op == GGML_OP_CPY && copy->src[0] == (scale ? scale : rope) && copy->type == GGML_TYPE_F16 &&
ggml_is_contiguous(copy) && aligned(copy, 4) && ggml_are_same_shape(rope, copy);
}
static int match(ggml_backend_dev_t device, const ggml_cgraph* graph, int first, void*) {
auto* g = const_cast<ggml_cgraph*>(graph);
if (first + 2 >= ggml_graph_n_nodes(g) || !fused_supports(device, ggml_graph_node(g, first), ggml_graph_node(g, first + 1), ggml_graph_node(g, first + 2)))
return 0;
return first + 4 < ggml_graph_n_nodes(g) && cast_supports(device, ggml_graph_node(g, first + 2), ggml_graph_node(g, first + 3), ggml_graph_node(g, first + 4)) ? 5 : 3;
if (first + 4 < ggml_graph_n_nodes(g) && cast_supports(device, ggml_graph_node(g, first + 2), ggml_graph_node(g, first + 3), ggml_graph_node(g, first + 4)))
return 5;
return first + 3 < ggml_graph_n_nodes(g) && cast_supports(device, ggml_graph_node(g, first + 2), nullptr, ggml_graph_node(g, first + 3)) ? 4 : 3;
}
static void optimize(ggml_backend_dev_t device, ggml_cgraph* graph, void*) {
for (int last = 2; last < ggml_graph_n_nodes(graph); ++last) {
const auto* output = ggml_graph_node(graph, last);
const auto* scale = output->op == GGML_OP_CPY ? output->src[0] : nullptr;
const auto* rope = scale && scale->op == GGML_OP_SCALE ? scale->src[0] : output;
const auto* source = output->op == GGML_OP_CPY ? output->src[0] : nullptr;
const auto* scale = source && source->op == GGML_OP_SCALE ? source : nullptr;
const auto* rope = source ? (scale ? scale->src[0] : source) : output;
if (!is_rope(rope))
continue;
const auto* mul = rope->src[0];
const auto* norm = mul->src[0];
if (!norm || !fused_supports(device, norm, mul, rope))
continue;
const int count = output == rope ? 3 : cast_supports(device, rope, scale, output) ? 5
const int count = output == rope ? 3 : cast_supports(device, rope, scale, output) ? (scale ? 5 : 4)
: 0;
if (!count)
continue;
const ggml_tensor* nodes[] = {norm, mul, rope, scale, output};
const ggml_tensor* nodes[] = {norm, mul, rope, scale ? scale : output, output};
int indices[5], next = count - 1;
for (int i = last; i >= std::max(0, last - 64) && next >= 0; --i)
if (ggml_graph_node(graph, i) == nodes[next])
@@ -170,6 +173,7 @@ namespace sd_vulkan_rope {
pc.dst_offset = buffers[2].shader_offset / uint32_t(ggml_type_size(output->type));
std::memcpy(&pc.interleaved, ggml_ext_op_attributes(rope), 4);
uint32_t groups[3] = {1, 1, 1};
const uint32_t apply_output_scale = count == 5;
ggml_ext_vulkan_pipeline pipeline = {};
pipeline.name = "sd.rope_apply.vulkan";
pipeline.spirv = sd_vulkan_shaders::rope_apply;
@@ -181,9 +185,13 @@ namespace sd_vulkan_rope {
groups[0] = std::min(n_groups, info.max_group_count[0]);
groups[1] = (n_groups + groups[0] - 1) / groups[0];
} else {
pipeline.name = count == 5 ? "sd.rms_rope_cast.vulkan" : "sd.rms_rope_apply.vulkan";
pipeline.spirv = count == 5 ? sd_vulkan_shaders::rms_rope_cast : sd_vulkan_shaders::rms_rope_apply;
pipeline.spirv_size = count == 5 ? sizeof(sd_vulkan_shaders::rms_rope_cast) : sizeof(sd_vulkan_shaders::rms_rope_apply);
pipeline.name = count > 3 ? "sd.rms_rope_cast.vulkan" : "sd.rms_rope_apply.vulkan";
pipeline.spirv = count > 3 ? sd_vulkan_shaders::rms_rope_cast : sd_vulkan_shaders::rms_rope_apply;
pipeline.spirv_size = count > 3 ? sizeof(sd_vulkan_shaders::rms_rope_cast) : sizeof(sd_vulkan_shaders::rms_rope_apply);
if (count > 3) {
pipeline.constants = &apply_output_scale;
pipeline.constant_count = 1;
}
pipeline.push_constant_size = sizeof(FusedParameters);
pipeline.full_subgroups = true;
pipeline.required_subgroup_size = info.subgroup_size_control ? 32 : 0;