vulkan: Support unaligned tensors for ROPE (ggml-org#22637)

Author: jeffbolznv

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -1354,6 +1354,8 @@ struct vk_op_rope_push_constants {
     uint32_t nb11;
     uint32_t nb12;
     uint32_t nb13;
+    uint32_t a_offset;
+    uint32_t d_offset;
 };
 static_assert(sizeof(vk_op_rope_push_constants) <= 128, "sizeof(vk_op_rope_push_constants) must be <= 128");
 
@@ -10126,6 +10128,15 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
     GGML_UNUSED(src3);
 }
 
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_rope_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
+    p.a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
+    p.d_offset = get_misalign_bytes(ctx, dst)  / ggml_type_size(dst->type);
+
+    GGML_UNUSED(src1);
+    GGML_UNUSED(src2);
+    GGML_UNUSED(src3);
+}
+
 template<typename PC>
 static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst, ggml_op op, PC&& pc) {
     VK_LOG_DEBUG("ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
@@ -11270,6 +11281,7 @@ static vk_op_rope_push_constants ggml_vk_make_rope_constants(const ggml_tensor *
         (uint32_t)src0->ne[2],
         nb01, nb02, nb03,
         nb11, nb12, nb13,
+        0, 0, // a_offset, d_offset filled in by init_pushconst_tensor_offsets
     };
 
     return rope;
@@ -11365,6 +11377,11 @@ static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx,
             GGML_ASSERT(buf[i] != nullptr);
         }
 
+        // a_offset is unused (the fused path reads from shared memory), but the rope/set_rows dst can be misaligned.
+        // Round the binding offset down to the storage buffer alignment; the in-element shift goes in pc.rope.d_offset.
+        pc.rope.d_offset = get_misalign_bytes(ctx, tensors[5]) / ggml_type_size(tensors[5]->type);
+        offset[5] &= ~(size_t(ctx->device->properties.limits.minStorageBufferOffsetAlignment) - 1);
+
         std::array<uint32_t, 3> elements;
         elements = { (uint32_t)rms->src[0]->ne[1], (uint32_t)rms->src[0]->ne[2], (uint32_t)rms->src[0]->ne[3] };
 

ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl

@@ -9,7 +9,7 @@ uint rope_a_coord(const uint i0, const uint i01, const uint i02, const uint i03,
     // Per-row offset in shared memory
     const uint ix = i0;
 #else
-    const uint ix = i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i0;
+    const uint ix = p.a_offset + i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i0;
 #endif
     return ix;
 }
@@ -48,6 +48,7 @@ void rope_norm(const uint i0, const uint i1, const uint i2, const uint i3, rope_
         idst = i1*p.nb11 + i0;
         idst += rope_data_i[i2].x * p.set_rows_stride;
     }
+    idst += p.d_offset;
 
     if (i0 >= p.n_dims) {
         rope_data_d[idst + 0] = ROPE_D_TYPE(rope_data_a[ix + 0]);
@@ -84,6 +85,7 @@ void rope_neox(const uint i0, const uint i1, const uint i2, const uint i3, rope_
         idst = i1*p.nb11 + i0/2;
         idst += rope_data_i[i2].x * p.set_rows_stride;
     }
+    idst += p.d_offset;
 
     if (i0 >= p.n_dims) {
         rope_data_d[idst + i0/2 + 0] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 0]);
@@ -121,6 +123,7 @@ void rope_multi(const uint i0, const uint i1, const uint i2, const uint i3, rope
         idst = i1*p.nb11 + i0/2;
         idst += rope_data_i[i2].x * p.set_rows_stride;
     }
+    idst += p.d_offset;
 
     if (i0 >= p.n_dims) {
         rope_data_d[idst + i0/2 + 0] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 0]);
@@ -176,7 +179,7 @@ void rope_vision(const uint i0, const uint i1, const uint i2, const uint i3, rop
         return;
     }
 
-    const uint idst = i0/2 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
+    const uint idst = p.d_offset + i0/2 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
     const uint ix = rope_a_coord(i0/2, i1, i2, i3, p);
 
     const int sect_dims = p.sections[0] + p.sections[1];

ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl

@@ -26,6 +26,9 @@ struct rope_params {
     uint nb11;
     uint nb12;
     uint nb13;
+
+    uint a_offset;
+    uint d_offset;
 };
 
 #endif // !defined(GGML_ROPE_PARAMS)

tests/test-backend-ops.cpp

@@ -4851,6 +4851,21 @@ struct test_rope : public test_case {
 
             a = ggml_view_4d(ctx, a, ne_a[0], ne_a[1], ne_a[2], ne_a[3], a->nb[1], a->nb[2], a->nb[3], 0);
             ggml_set_name(a, "view_of_a");
+        } else if (v == 2) {
+            // second-half slice along dim 0 (mimics build_rope_2d in clip.cpp).
+            // The non-zero view offset (ne_a[0] * elem_size) often produces a
+            // non-aligned buffer offset, which exercises backends' alignment paths.
+            auto ne = ne_a; ne[0] *= 2;
+            a = ggml_new_tensor(ctx, type, 4, ne.data());
+            if (forward) {
+                ggml_set_param(a);
+            }
+            ggml_set_name(a, "a");
+
+            a = ggml_view_4d(ctx, a, ne_a[0], ne_a[1], ne_a[2], ne_a[3],
+                             a->nb[1], a->nb[2], a->nb[3],
+                             ne_a[0] * ggml_element_size(a));
+            ggml_set_name(a, "view_of_a");
         } else {
             a = ggml_new_tensor(ctx, type, 4, ne_a.data());
             if (forward) {
@@ -4913,8 +4928,6 @@ struct test_rope : public test_case {
             } else {
                 out = ggml_rope_ext_back(ctx, a, pos, freq, n_dims, mode, 0, 10000.0f, fs, ef, af, 1.0f, 1.0f);
             }
-
-            // TODO: add test with a non-contiguous view as input ; this case is needed for build_rope_2d in clip.cpp
         }
         ggml_set_name(out, "out");
 
@@ -8687,6 +8700,13 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
 
                                 test_cases.emplace_back(new test_rope(type, { 64, 128, 2, 1},  64, GGML_ROPE_TYPE_NEOX, 512, fs, ef, af, ff, v, fw)); // neox (falcon 40B)
                             }
+
+                            // build_rope_2d-style: ROPE on a non-contiguous view
+                            // that starts at a non-zero offset along dim 0
+                            // (e.g. gemma4v vision second-half view).
+                            for (int rmode : { GGML_ROPE_TYPE_NORMAL, GGML_ROPE_TYPE_NEOX, GGML_ROPE_TYPE_MROPE, GGML_ROPE_TYPE_IMROPE, GGML_ROPE_TYPE_VISION }) {
+                                test_cases.emplace_back(new test_rope(type, { 36, 16, 2457, 1}, 36, rmode, 512, fs, ef, af, ff, 2, fw));
+                            }
                         }
 
                         all = false;