mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 36 Packages Projects Releases Wiki Activity

ggml-vulkan: add SGN operator, auto-generate Vulkan.csv and ops.md (#20219)

Browse Source
This commit is contained in:
Bertay Eren
2026-03-09 09:24:16 +03:00
committed by GitHub
parent b2f460bd3c
commit 0beb8db3a0
5 changed files with 51 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+9
View File
@@ -763,6 +763,7 @@ struct vk_device_struct {
vk_pipeline pipeline_ceil[2];
vk_pipeline pipeline_floor[2];
vk_pipeline pipeline_trunc[2];
vk_pipeline pipeline_sgn[2];
vk_pipeline pipeline_add1_f16_f16;
vk_pipeline pipeline_add1_f16_f32;
@@ -4393,6 +4394,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
CREATE_UNARY(ceil)
CREATE_UNARY(floor)
CREATE_UNARY(trunc)
CREATE_UNARY(sgn)
#undef CREATE_UNARY
#define CREATE_UNARY_RTE(name) \
@@ -9281,6 +9283,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
return ctx->device->pipeline_floor[dst->type == GGML_TYPE_F16];
case GGML_UNARY_OP_TRUNC:
return ctx->device->pipeline_trunc[dst->type == GGML_TYPE_F16];
case GGML_UNARY_OP_SGN:
return ctx->device->pipeline_sgn[dst->type == GGML_TYPE_F16];
default:
break;
}
@@ -12875,6 +12879,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
case GGML_UNARY_OP_CEIL:
case GGML_UNARY_OP_FLOOR:
case GGML_UNARY_OP_TRUNC:
case GGML_UNARY_OP_SGN:
ggml_vk_unary(ctx, compute_ctx, src0, node);
break;
case GGML_UNARY_OP_XIELU:
@@ -15004,6 +15009,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_UNARY_OP_CEIL:
case GGML_UNARY_OP_FLOOR:
case GGML_UNARY_OP_TRUNC:
case GGML_UNARY_OP_SGN:
return ggml_is_contiguous(op->src[0]) &&
(op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
(op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) &&
@@ -16170,6 +16176,9 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
case GGML_UNARY_OP_TRUNC:
tensor_clone = ggml_trunc(ggml_ctx, src_clone[0]);
break;
case GGML_UNARY_OP_SGN:
tensor_clone = ggml_sgn(ggml_ctx, src_clone[0]);
break;
default:
std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
GGML_ABORT("fatal error");