opencl: skip the Adreno KQ/KQV image kernels for multi-stream batches (#26189)

The Adreno KQ/KQV image1d kernels (ggml_cl_mul_mat_kq_kqv_adreno) ignore
dim 3 entirely: the sub-buffer covers only nb02*ne02 bytes and the kernel
receives no ne03/ne13/nb03/nb13 arguments. With the unified KV cache,
multi-sequence batches (e.g. llama-perplexity with its default -b 2048,
n_seq=4, or a multi-slot llama-server) present KQ/KQV as 4D tensors with
ne3 = n_stream, so every stream past the first reads the first stream's
K/V and produces garbage. Flash attention masks the bug where it is
enabled; devices where FA is declined (e.g. Adreno 740) hit it with
default settings.

Route ne03/ne13 > 1 to the general path, which handles dim 3, and honor
view_offs when creating the sub-buffers (currently always 0 for tensors
reaching this function, but the function would silently misread any
future view).

Llama-3.2-1B-Instruct Q4_0, wiki.test.raw, 8 chunks, -ngl 99:
- Adreno 740, default:            PPL 1817.64 -> 15.61
- Adreno 740, -fa 0:              PPL 1941.64 -> 15.61
- Adreno 840, -fa 0:              PPL 1943.90 -> 15.50
- single-stream (-b 512) results unchanged (15.6090)
- test-backend-ops -o MUL_MAT on 740: identical before/after (909 OK,
  12 pre-existing q6_K failures)

diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index a05d18ee30af36e219a3da58c9deb9842e39cd84..d07b8fe41a31fb20c14233e54f799d356ae85c4d 100644 (file)
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -15675,7 +15675,7 @@ static void ggml_cl_mul_mat_kq_kqv_adreno(ggml_backend_t backend, const ggml_ten
     // <--------------------------------------------> //
     extra0 = src0->view_src ? (ggml_tensor_extra_cl *)src0->view_src->extra : (ggml_tensor_extra_cl *)src0->extra;
 
-    region.origin = (extra0->offset);
+    region.origin = (extra0->offset + src0->view_offs);
     if (nb01 > nb02) {
         // KQ
         region.size = nb01 * ne01;
@@ -15691,7 +15691,7 @@ static void ggml_cl_mul_mat_kq_kqv_adreno(ggml_backend_t backend, const ggml_ten
 
     // create sub-buffer for B
     // <--------------------------------------------> //
-    region.origin = (extra1->offset);
+    region.origin = (extra1->offset + src1->view_offs);
     region.size = nb10 * ne10 * ne11 * ne12;
     B_sub_buffer = clCreateSubBuffer((extra1->data_device), 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &status);
     CL_CHECK(status);
@@ -15712,7 +15712,7 @@ static void ggml_cl_mul_mat_kq_kqv_adreno(ggml_backend_t backend, const ggml_ten
 
     // create sub-buffer for output C
     // <--------------------------------------------> //
-    region.origin = (extrad->offset);
+    region.origin = (extrad->offset + dst->view_offs);
     region.size = ne0 * ne1 * dst->ne[2] * dst->nb[0]; // size of C in bytes
     D_sub_buffer = clCreateSubBuffer((extrad->data_device), 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &status);
     CL_CHECK(status);
@@ -18591,6 +18591,8 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     if(src0t == GGML_TYPE_F16 && src1t == GGML_TYPE_F32){
         if (ne01 >= 64 && ne1 >= 32 && ne00 >= 16 && (ne12 % ne02) == 0  &&
+            // the KQ/KQV image kernels do not handle dim 3 (multi-stream batches)
+            ne03 == 1 && ne13 == 1 &&
             // dst is wrapped with image1d_buffer, the size limit applies, also src0
             (ne0 * ne1 * dst->ne[2] * dst->nb[0] / 4 <= backend_ctx->image_max_buffer_size)) {
             // For KQ