vulkan: fuse adds (#15252)

* vulkan: fuse adds

Fuse adds that have the same shape, which are common in MoE models.
It will currently fuse up to 6 adds, because we assume no more than
8 descriptors per dispatch. But this could be changed.

* check runtimeDescriptorArray feature

* disable multi_add for Intel due to likely driver bug

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 6315584996654d76e759e627b98da02e88919c60..20674bee588186d528c9fc96ada98296eb93a964 100644 (file)
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -103,6 +103,8 @@ static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }
 struct ggml_backend_vk_context;
 
 #define MAX_PARAMETER_COUNT 8
+// Max number of adds that can be fused without exceeding MAX_PARAMETER_COUNT.
+#define MAX_FUSED_ADDS (MAX_PARAMETER_COUNT - 2)
 
 struct vk_pipeline_struct {
     std::string name;
@@ -368,6 +370,7 @@ struct vk_device_struct {
     bool float_controls_rte_fp16;
     bool subgroup_add;
     bool subgroup_shuffle;
+    bool multi_add;
 
     bool integer_dot_product;
 
@@ -449,6 +452,9 @@ struct vk_device_struct {
     vk_pipeline pipeline_div[2][2][2];
     vk_pipeline pipeline_div_norepeat[2][2][2];
 
+    // indexed by num_additional_fused_ops == num_adds - 1
+    vk_pipeline pipeline_multi_add[MAX_FUSED_ADDS];
+
     vk_pipeline pipeline_add_id_f32;
 
     vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32;
@@ -801,6 +807,14 @@ struct vk_op_binary_push_constants {
     float param1; float param2; int32_t param3;
 };
 
+struct vk_op_multi_add_push_constants {
+    // shape for dst
+    uint32_t ne20; uint32_t ne21; uint32_t ne22; uint32_t ne23;
+
+    // strides for srcs+dst
+    uint32_t nb[8][4];
+};
+
 struct vk_op_add_id_push_constants {
     uint32_t ne0;
     uint32_t ne1;
@@ -2994,6 +3008,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
     CREATE_BINARY(div, _norepeat, {1})
 #undef CREATE_BINARY
 
+    if (device->multi_add) {
+        for (uint32_t i = 0; i < MAX_FUSED_ADDS; ++i) {
+            ggml_vk_create_pipeline(device, device->pipeline_multi_add[i], "multi_add_f32_" + std::to_string(i+1), multi_add_f32_len, multi_add_f32_data, "main", MAX_PARAMETER_COUNT, sizeof(vk_op_multi_add_push_constants), {512, 1, 1}, {i+2}, 1);
+        }
+    }
+
     ggml_vk_create_pipeline(device, device->pipeline_add_id_f32, "add_id_f32", add_id_f32_len, add_id_f32_data, "main", 4, sizeof(vk_op_add_id_push_constants), {1, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_acc_f32, "acc_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
@@ -3533,6 +3553,12 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         device->pipeline_robustness = pl_robustness_features.pipelineRobustness;
 
+        device->multi_add = vk12_props.shaderRoundingModeRTEFloat16 &&
+                            device->properties.limits.maxPushConstantsSize >= sizeof(vk_op_multi_add_push_constants) &&
+                            vk12_features.runtimeDescriptorArray &&
+                            device->vendor_id != VK_VENDOR_ID_INTEL &&
+                            getenv("GGML_VK_DISABLE_MULTI_ADD") == nullptr;
+
         if (device->subgroup_size_control) {
             device->subgroup_min_size = subgroup_size_control_props.minSubgroupSize;
             device->subgroup_max_size = subgroup_size_control_props.maxSubgroupSize;
@@ -6887,6 +6913,9 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         switch (op) {
         case GGML_OP_ADD:
         {
+            if (ctx->num_additional_fused_ops > 0) {
+                return ctx->device->pipeline_multi_add[ctx->num_additional_fused_ops];
+            }
             auto pipelines = ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_norepeat : ctx->device->pipeline_add;
             return pipelines[src0->type == GGML_TYPE_F16][src1->type == GGML_TYPE_F16][dst->type == GGML_TYPE_F16];
         }
@@ -7743,6 +7772,107 @@ static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const
     }, dryrun);
 }
 
+static void ggml_vk_multi_add(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_cgraph * cgraph, int node_idx, bool dryrun = false) {
+    const ggml_tensor *first_node = cgraph->nodes[node_idx];
+    const ggml_tensor *dst = cgraph->nodes[node_idx + ctx->num_additional_fused_ops];
+
+    // Make a list of all the tensors used by the op.
+    // Last element of the list is the dest tensor.
+    const ggml_tensor *tensors[MAX_PARAMETER_COUNT];
+    uint32_t num_srcs = ctx->num_additional_fused_ops + 2;
+    uint32_t num_tensors = num_srcs + 1;
+    GGML_ASSERT(num_tensors <= MAX_PARAMETER_COUNT);
+
+    tensors[0] = first_node->src[0];
+    tensors[1] = first_node->src[1];
+    for (int32_t i = 0; i < ctx->num_additional_fused_ops; ++i) {
+        // check whether the previous result is src[0] or src[1]
+        if (cgraph->nodes[node_idx + i] == cgraph->nodes[node_idx + i + 1]->src[0]) {
+            tensors[i+2] = cgraph->nodes[node_idx + i + 1]->src[1];
+        } else {
+            tensors[i+2] = cgraph->nodes[node_idx + i + 1]->src[0];
+        }
+    }
+    tensors[num_srcs] = dst;
+
+    vk_op_multi_add_push_constants pc;
+    pc.ne20 = (uint32_t)dst->ne[0];
+    pc.ne21 = (uint32_t)dst->ne[1];
+    pc.ne22 = (uint32_t)dst->ne[2];
+    pc.ne23 = (uint32_t)dst->ne[3];
+
+    for (uint32_t i = 0; i < num_tensors; ++i) {
+        const ggml_tensor *t = tensors[i];
+        pc.nb[i][0] = (uint32_t)t->nb[0] / sizeof(float);
+        pc.nb[i][1] = (uint32_t)t->nb[1] / sizeof(float);
+        pc.nb[i][2] = (uint32_t)t->nb[2] / sizeof(float);
+        pc.nb[i][3] = (uint32_t)t->nb[3] / sizeof(float);
+    }
+
+    vk_pipeline pipeline = ctx->device->pipeline_multi_add[ctx->num_additional_fused_ops];
+
+    if (pipeline == nullptr) {
+        std::cerr << "ggml_vulkan: Error: Missing multi_add";
+        GGML_ABORT("fatal error");
+    }
+
+    if (dryrun) {
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+        return;
+    }
+
+    ggml_backend_vk_buffer_context * buf_ctx[MAX_PARAMETER_COUNT];
+    vk_buffer buf[MAX_PARAMETER_COUNT];
+    size_t offset[MAX_PARAMETER_COUNT];
+    bool uma[MAX_PARAMETER_COUNT];
+
+    for (uint32_t i = 0; i < num_tensors; ++i) {
+        buf_ctx[i] = (ggml_backend_vk_buffer_context *)tensors[i]->buffer->context;
+        buf[i] = nullptr;
+        offset[i] = 0;
+        uma[i] = false;
+
+        if (ctx->device->uma) {
+            ggml_vk_host_get(ctx->device, tensors[i]->data, buf[i], offset[i]);
+            uma[i] = buf[i] != nullptr;
+        }
+        if (!uma[i]) {
+            buf[i] = buf_ctx[i]->dev_buffer;
+            offset[i] = vk_tensor_offset(tensors[i]) + tensors[i]->view_offs;
+        }
+        GGML_ASSERT(buf[i] != nullptr);
+    }
+    // If any remaining descriptors are unused, just point them at src[0]
+    for (uint32_t i = num_tensors; i < MAX_PARAMETER_COUNT; ++i) {
+        buf[i] = buf[0];
+        offset[i] = 0;
+    }
+
+    std::array<uint32_t, 3> elements;
+
+    uint32_t ne = ggml_nelements(dst);
+    if (ne > 262144) {
+        elements = { 512, 512, CEIL_DIV(ne, 262144) };
+    } else if (ne > 512) {
+        elements = { 512, CEIL_DIV(ne, 512), 1 };
+    } else {
+        elements = { ne, 1, 1 };
+    }
+
+    ggml_vk_sync_buffers(subctx);
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
+        {
+            vk_subbuffer{ buf[0], offset[0], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[1], offset[1], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[2], offset[2], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[3], offset[3], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[4], offset[4], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[5], offset[5], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[6], offset[6], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[7], offset[7], VK_WHOLE_SIZE },
+        }, pc, elements);
+}
+
 static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
@@ -9703,8 +9833,11 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
 
         break;
     case GGML_OP_ADD:
-        ggml_vk_add(ctx, compute_ctx, src0, src1, node, dryrun);
-
+        if (ctx->num_additional_fused_ops) {
+            ggml_vk_multi_add(ctx, compute_ctx, cgraph, node_idx, dryrun);
+        } else {
+            ggml_vk_add(ctx, compute_ctx, src0, src1, node, dryrun);
+        }
         break;
     case GGML_OP_SUB:
         ggml_vk_sub(ctx, compute_ctx, src0, src1, node, dryrun);
@@ -10586,6 +10719,58 @@ static bool ggml_vk_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, st
     return true;
 }
 
+static uint32_t ggml_vk_fuse_multi_add(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph, int node_idx) {
+
+    const ggml_tensor *first_node = cgraph->nodes[node_idx];
+    if (first_node->op != GGML_OP_ADD) {
+        return 0;
+    }
+
+    if (!ctx->device->multi_add) {
+        return 0;
+    }
+
+    int32_t num_adds = 1;
+    while (node_idx + num_adds < cgraph->n_nodes &&
+           cgraph->nodes[node_idx + num_adds]->op == GGML_OP_ADD &&
+           num_adds < MAX_FUSED_ADDS) {
+        num_adds++;
+    }
+
+    // The shader currently requires same shapes (but different strides are allowed),
+    // everything f32, and no misalignment
+    for (int32_t i = 0; i < num_adds; ++i) {
+        const ggml_tensor *next_node = cgraph->nodes[node_idx + i];
+        if (!ggml_are_same_shape(first_node, next_node->src[0]) ||
+            !ggml_are_same_shape(first_node, next_node->src[1]) ||
+            next_node->type != GGML_TYPE_F32 ||
+            next_node->src[0]->type != GGML_TYPE_F32 ||
+            next_node->src[1]->type != GGML_TYPE_F32 ||
+            get_misalign_bytes(ctx, next_node) ||
+            get_misalign_bytes(ctx, next_node->src[0]) ||
+            get_misalign_bytes(ctx, next_node->src[1])) {
+            num_adds = i;
+        }
+    }
+
+    // Verify we can fuse these
+    ggml_op adds[MAX_FUSED_ADDS];
+    for (int32_t i = 0; i < num_adds; ++i) {
+        adds[i] = GGML_OP_ADD;
+    }
+
+    // decrease num_adds if they can't all be fused
+    while (num_adds > 1 && !ggml_can_fuse(cgraph, node_idx, adds, num_adds)) {
+        num_adds--;
+    }
+
+    // a single add is not "fused", so just return zero
+    if (num_adds == 1) {
+        return 0;
+    }
+    return num_adds;
+}
+
 static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)");
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
@@ -10599,8 +10784,13 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
 
     uint64_t total_mat_mul_bytes = 0;
     for (int i = 0; i < cgraph->n_nodes; i++) {
-        if (!ctx->device->disable_fusion && ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
-            ctx->num_additional_fused_ops = 1;
+        if (!ctx->device->disable_fusion) {
+            uint32_t num_adds = ggml_vk_fuse_multi_add(ctx, cgraph, i);
+            if (num_adds) {
+                ctx->num_additional_fused_ops = num_adds - 1;
+            } else if (ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
+                ctx->num_additional_fused_ops = 1;
+            }
         }
         ggml_vk_build_graph(ctx, cgraph, i, nullptr, 0, true, false, false, false);
         if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
@@ -10675,8 +10865,13 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
             mul_mat_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]);
         }
 
-        if (!ctx->device->disable_fusion && ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
-            ctx->num_additional_fused_ops = 1;
+        if (!ctx->device->disable_fusion) {
+            uint32_t num_adds = ggml_vk_fuse_multi_add(ctx, cgraph, i);
+            if (num_adds) {
+                ctx->num_additional_fused_ops = num_adds - 1;
+            } else if (ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
+                ctx->num_additional_fused_ops = 1;
+            }
         }
 
         // Signal the almost_ready fence when the graph is mostly complete (< 20% remaining)

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp b/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp
index 4b4316cf3d9f263e140b4dfb655158c446e6d596..750e785753954b76d9b494e650ff2b5644ea8f55 100644 (file)
--- a/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp
@@ -2,6 +2,7 @@
 #extension GL_EXT_control_flow_attributes : require
 
 #include "rte.comp"
+#include "utils.comp"
 
 layout (push_constant) uniform parameter
 {
@@ -28,25 +29,9 @@ uint get_aoffset() { return p.misalign_offsets >> 16; }
 uint get_boffset() { return (p.misalign_offsets >> 8) & 0xFF; }
 uint get_doffset() { return p.misalign_offsets & 0xFF; }
 
-// mod and div are expensive and coordinates/dimensions are often power of 2 or equal to 1
-uint fastmod(uint a, uint b) {
-    if ((b & (b-1)) == 0) {
-        return a & (b-1);
-    }
-    return a % b;
-}
-
-uint fastdiv(uint a, uint b) {
-    return (a < b) ? 0 : (a / b);
-}
 
 void get_indices(uint idx, out uint i00, out uint i01, out uint i02, out uint i03) {
-    i03 = fastdiv(idx, (p.ne02*p.ne01*p.ne00));
-    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
-    i02 = fastdiv((idx - i03_offset), (p.ne01*p.ne00));
-    const uint i02_offset = i02*p.ne01*p.ne00;
-    i01 = (idx - i03_offset - i02_offset) / p.ne00;
-    i00 = idx - i03_offset - i02_offset - i01*p.ne00;
+    get_indices(idx, i00, i01, i02, i03, p.ne00, p.ne01, p.ne02, p.ne03);
 }
 
 uint src0_idx(uint i00, uint i01, uint i02, uint i03) {

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp b/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp
new file mode 100644 (file)
index 0000000..0c7acb7
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp
@@ -0,0 +1,68 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+#extension GL_EXT_nonuniform_qualifier : enable
+#extension GL_EXT_control_flow_attributes : require
+
+#include "rte.comp"
+#include "types.comp"
+#include "utils.comp"
+
+layout (push_constant) uniform parameter2
+{
+    // shape for dst
+    uint ne20; uint ne21; uint ne22; uint ne23;
+
+    // strides for srcs+dst
+    uint nb[8][4];
+} p;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];} a[];
+layout (binding = 0) writeonly buffer D {D_TYPE data_d[];} d[];
+
+layout(constant_id = 0) const uint num_srcs = 2;
+
+uint src_idx(uint s, uint i00, uint i01, uint i02, uint i03) {
+    return i03*p.nb[s][3] + i02*p.nb[s][2] + i01*p.nb[s][1] + i00*p.nb[s][0];
+}
+
+uint dst_idx(uint i00, uint i01, uint i02, uint i03) {
+    uint nb20 = p.nb[num_srcs][0];
+    uint nb21 = p.nb[num_srcs][1];
+    uint nb22 = p.nb[num_srcs][2];
+    uint nb23 = p.nb[num_srcs][3];
+    return i03*nb23 + i02*nb22 + i01*nb21 + i00*nb20;
+}
+
+uint get_idx() {
+    return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+}
+
+const uint num_threads = 256;
+
+layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    uint idx = get_idx();
+
+    uint ne = p.ne20 * p.ne21 * p.ne22 * p.ne23;
+
+    // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
+    const uint num_iter = 2;
+
+    [[unroll]] for (uint i = 0; i < num_iter; ++i) {
+        if (idx >= ne) {
+            continue;
+        }
+        uint i00, i01, i02, i03;
+        get_indices(idx, i00, i01, i02, i03, p.ne20, p.ne21, p.ne22, p.ne23);
+
+        FLOAT_TYPE sum = FLOAT_TYPE(0);
+        [[unroll]] for (uint s = 0; s < num_srcs; ++s) {
+            sum += FLOAT_TYPE(a[s].data_a[src_idx(s, i00, i01, i02, i03)]);
+        }
+        d[num_srcs].data_d[dst_idx(i00, i01, i02, i03)] = D_TYPE(sum);
+
+        idx += num_threads;
+    }
+}

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/utils.comp b/ggml/src/ggml-vulkan/vulkan-shaders/utils.comp
new file mode 100644 (file)
index 0000000..dc4a1e6
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/utils.comp
@@ -0,0 +1,25 @@
+#ifndef UTILS_COMP
+#define UTILS_COMP
+
+// mod and div are expensive and coordinates/dimensions are often power of 2 or equal to 1
+uint fastmod(uint a, uint b) {
+    if ((b & (b-1)) == 0) {
+        return a & (b-1);
+    }
+    return a % b;
+}
+
+uint fastdiv(uint a, uint b) {
+    return (a < b) ? 0 : (a / b);
+}
+
+void get_indices(uint idx, out uint i00, out uint i01, out uint i02, out uint i03, uint ne00, uint ne01, uint ne02, uint ne03) {
+    i03 = fastdiv(idx, (ne02*ne01*ne00));
+    const uint i03_offset = i03 * ne02*ne01*ne00;
+    i02 = fastdiv((idx - i03_offset), (ne01*ne00));
+    const uint i02_offset = i02*ne01*ne00;
+    i01 = (idx - i03_offset - i02_offset) / ne00;
+    i00 = idx - i03_offset - i02_offset - i01*ne00;
+}
+
+#endif // UTILS_COMP

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index 68933d19f2ec56a65380358359cf0296826ab681..67a9ee422bf4751f20003f5a0d90700659f57e66 100644 (file)
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -677,6 +677,8 @@ void process_shaders() {
 
     string_to_spv("add_id_f32", "add_id.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
 
+    string_to_spv("multi_add_f32", "multi_add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}});
+
     for (auto &c : compiles) {
         c.wait();
     }

diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 39547f0649e8e64955185da67ef47a3dcdc17496..0df5ebb56504ebd636bb3806015b205f9488103b 100644 (file)
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -2491,12 +2491,12 @@ struct test_bin_bcast : public test_case {
         : op(op), type(type), ne(ne), nr(nr), nf(nf) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
-        GGML_ASSERT(nf <= 8);
+        GGML_ASSERT(nf <= 16);
 
         ggml_tensor * a = ggml_new_tensor_4d(ctx, type, ne[0]*nr[0], ne[1]*nr[1], ne[2]*nr[2], ne[3]*nr[3]);
         ggml_set_name(a, "a");
 
-        ggml_tensor * b[8];
+        ggml_tensor * b[16];
         for (int i = 0; i < nf; ++i) {
             b[i] = ggml_new_tensor(ctx, type, 4, ne.data());
             ggml_set_name(b[i], (std::string("b") + std::to_string(i)).c_str());
@@ -5658,6 +5658,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {10, 5, 4, 3}, {1, 1, 2, 2}, 6));
     test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {10, 5, 4, 3}, {1, 2, 2, 2}, 7));
     test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {16, 5, 4, 3}, {2, 2, 2, 2}, 8));
+    test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {16, 5, 4, 3}, {1, 1, 1, 1}, 16));
 
     test_cases.emplace_back(new test_add1());
     test_cases.emplace_back(new test_scale());