Commit 31424b7cf for llama.cpp

commit 31424b7cf14604ca65a9195123da1413f31c4472
Author: shaofeiqi <shaoqi@qti.qualcomm.com>
Date:   Fri Nov 6 21:44:05 2026 -0800

    opencl: add arange (#31829)

    Assisted-by: Claude Opus 5.5

    Co-authored-by: Li He <lih@qti.qualcomm.com>

diff --git a/ggml/src/ggml-opencl/CMakeLists.txt b/ggml/src/ggml-opencl/CMakeLists.txt
index 58a6d438e..0ef7f8fe6 100644
--- a/ggml/src/ggml-opencl/CMakeLists.txt
+++ b/ggml/src/ggml-opencl/CMakeLists.txt
@@ -71,6 +71,7 @@ set(GGML_OPENCL_KERNELS
     lightning_indexer
     tri
     fill
+    arange
     fwht
     clamp
     cpy
diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index e0ee5518b..1a4c03519 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -856,6 +856,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_sigmoid_f32_nc, kernel_sigmoid_f16_nc;
     cl_kernel kernel_tri;
     cl_kernel kernel_fill;
+    cl_kernel kernel_arange;
     cl_kernel kernel_clamp, kernel_clamp_nc;
     cl_kernel kernel_geglu, kernel_reglu, kernel_swiglu, kernel_swiglu_oai, kernel_swiglu_clamp, kernel_geglu_erf,
               kernel_geglu_quick, kernel_geglu_f16, kernel_reglu_f16, kernel_swiglu_f16, kernel_swiglu_clamp_f16,
@@ -2048,6 +2049,24 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
         CL_CHECK(clReleaseProgram(prog));
     }

+    // arange
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "arange.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("arange.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_arange = clCreateKernel(prog, "kernel_arange_f32", &err), err));
+        GGML_LOG_CONT(".");
+
+        CL_CHECK(clReleaseProgram(prog));
+    }
+
     // clamp
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -11792,6 +11811,8 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
             return op->type == GGML_TYPE_F32 && ggml_is_contiguous(op);
         case GGML_OP_FILL:
             return op->type == GGML_TYPE_F32 && ggml_is_contiguous(op);
+        case GGML_OP_ARANGE:
+            return op->type == GGML_TYPE_F32;
         case GGML_OP_CLAMP:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_SOFT_MAX:
@@ -19593,6 +19614,40 @@ static void ggml_cl_fill(ggml_backend_t backend, const ggml_tensor * src0, const
     backend_ctx->enqueue_ndrange_kernel(kernel, 1, global_work_size, local_work_size, dst);
 }

+static void ggml_cl_arange(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    UNUSED(src0);
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    float start;
+    float step;
+    memcpy(&start, (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&step,  (float *) dst->op_params + 2, sizeof(float));
+
+    const int n = dst->ne[0];
+
+    cl_kernel kernel = backend_ctx->kernel_arange;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),      &n));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(float),    &start));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(float),    &step));
+
+    size_t local_work_size[1] = { 64 };
+    size_t global_work_size[1] = { ((size_t)n + local_work_size[0] - 1) / local_work_size[0] * local_work_size[0] };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 1, global_work_size, local_work_size, dst);
+}
+
 static void ggml_cl_clamp(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -38649,6 +38704,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_fill;
             break;
+        case GGML_OP_ARANGE:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_arange;
+            break;
         case GGML_OP_CLAMP:
             if (!any_on_device) {
                 return false;
diff --git a/ggml/src/ggml-opencl/kernels/arange.cl b/ggml/src/ggml-opencl/kernels/arange.cl
new file mode 100644
index 000000000..53dc383e6
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/arange.cl
@@ -0,0 +1,14 @@
+__kernel void kernel_arange_f32(
+        __global float * dst,
+        ulong offsetd,
+        int n,
+        float start,
+        float step
+) {
+    dst = (global float*)((global char*)dst + offsetd);
+
+    const int i = get_global_id(0);
+    if (i < n) {
+        dst[i] = start + step * i;
+    }
+}