mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 31 Packages Projects Releases Wiki Activity

vulkan: optimize im2col (#21713)

Browse Source 
* vulkan: improve im2col memory write layout

* cap workgroups

* minimal device tuning

* use vendor_id instead of subgroup size
This commit is contained in:
Ruben Ortlam
2026-04-15 19:04:51 +02:00
committed by GitHub
parent 7e72b38bc1
commit b3d758750a
2 changed files with 46 additions and 63 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+9 -4
View File
@@ -1394,7 +1394,7 @@ struct vk_op_im2col_push_constants {
uint32_t IW; uint32_t IH;
uint32_t OW; uint32_t OH;
uint32_t KW; uint32_t KH;
uint32_t pelements;
uint32_t OH_batch;
uint32_t CHW;
int32_t s0; int32_t s1;
int32_t p0; int32_t p1;
@@ -10064,7 +10064,13 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
const uint32_t batch = src1->ne[is_2D ? 3 : 2];
elements = { OW * KW * KH, OH, batch * IC };
const uint32_t CHW = IC * KH * KW;
// Cap X workgroups to limit concurrent IC channel reads.
// The shader loops over X to cover the full CHW dimension.
// AMD prefers a lower limit
const uint32_t min_cap = ctx->device->vendor_id == VK_VENDOR_ID_AMD ? 512u : 4096u;
const uint32_t x_elements = std::min(CHW, std::max(min_cap, OW * KH * KW));
elements = { x_elements, OW, OH * batch };
elements[1] = std::min(elements[1], ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
elements[2] = std::min(elements[2], ctx->device->properties.limits.maxComputeWorkGroupCount[2]);
} break;
@@ -11727,7 +11733,6 @@ static void ggml_vk_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, co
const uint32_t offset_delta = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
const uint32_t batch_offset = src1->nb[is_2D ? 3 : 2] / 4; // nb is byte offset, src is type float32
const uint32_t pelements = OW * KW * KH;
const uint32_t batch = src1->ne[is_2D ? 3 : 2];
const ggml_backend_vk_buffer_context * d_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
@@ -11739,7 +11744,7 @@ static void ggml_vk_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, co
dst_addr,
batch_offset, offset_delta,
IC, IW, IH, OW, OH, KW, KH,
pelements,
OH * batch,
IC * KH * KW,
s0, s1, p0, p1, d0, d1, batch * IC
});