Vulkan: adding GGML_OP_MULTI_ADD implementation (#582)

Co-authored-by: Iwan Kawrakow <[email protected]>

Author: ikawrakow

ggml/src/ggml-vulkan.cpp

@@ -447,6 +447,8 @@ struct vk_device_struct {
     vk_pipeline pipeline_fused_mul_silu[2];
     vk_pipeline pipeline_fused_mul_relu[2];
 
+    vk_pipeline pipeline_multi_add_f32;
+
     vk_pipeline pipeline_leaky_relu_f32;
     vk_pipeline pipeline_silu_back_f32;
     vk_pipeline pipeline_diag_mask_inf_f32;
@@ -683,6 +685,13 @@ struct vk_op_unary_push_constants {
 };
 static_assert(sizeof(vk_op_unary_push_constants) <= 128, "sizeof(vk_op_unary_push_constants) must be <= 128");
 
+struct vk_op_multiadd_push_constants {
+    uint32_t ne;
+    uint32_t ne0, ne1;
+    uint32_t nb0, nb01;
+    uint32_t nadd;
+};
+
 // See https://gmplib.org/~tege/divcnst-pldi94.pdf figure 4.1.
 // Precompute mp (m' in the paper) and L such that division
 // can be computed using a multiply (high 32b of 64b result)
@@ -2759,6 +2768,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_fused_mul_relu[0], "fused_mul_relu_f32", fused_mul_relu_f32_len, fused_mul_relu_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_fused_mul_relu[1], "fused_mul_relu_f16", fused_mul_relu_f16_len, fused_mul_relu_f16_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
+    ggml_vk_create_pipeline(device, device->pipeline_multi_add_f32, "multi_add_f32", multi_add_f32_len, multi_add_f32_data, "main", 2, sizeof(vk_op_multiadd_push_constants), {512, 1, 1}, {}, 1);
+
     ggml_vk_create_pipeline(device, device->pipeline_leaky_relu_f32, "leaky_relu_f32", leaky_relu_f32_len, leaky_relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_silu_back_f32, "silu_back_f32", silu_back_f32_len, silu_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
@@ -6451,6 +6462,12 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             }
             return nullptr;
         }
+    case GGML_OP_MULTI_ADD:
+        if (src0->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F32 ||
+            dst->ne[2] == 1 || dst->ne[3] == 1) {
+            return ctx->device->pipeline_multi_add_f32;
+        }
+        return nullptr;
     case GGML_OP_UNARY:
         if ((src0->type != GGML_TYPE_F32 && src0->type != GGML_TYPE_F16) ||
             (dst->type != GGML_TYPE_F32 && dst->type != GGML_TYPE_F16) ||
@@ -6588,6 +6605,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
     case GGML_OP_RMS_NORM:
     case GGML_OP_FUSED_RMS_NORM:
     case GGML_OP_IM2COL:
+    case GGML_OP_MULTI_ADD:
         return true;
     default:
         return false;
@@ -6889,6 +6907,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_CONCAT:
     case GGML_OP_UPSCALE:
     case GGML_OP_FUSED_MUL_UNARY:
+    case GGML_OP_MULTI_ADD:
     case GGML_OP_UNARY:
         {
             uint32_t ne = ggml_nelements(dst);
@@ -7278,6 +7297,12 @@ static void ggml_vk_fused_mul_unary(ggml_backend_vk_context * ctx, vk_context& s
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_FUSED_MUL_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
 }
 
+static void ggml_vk_multi_add(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    uint32_t nadd = (uint32_t)dst->op_params[0];
+    ggml_vk_op_f32<vk_op_multiadd_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_MULTI_ADD,
+            { (uint32_t)ggml_nelements(dst), (uint32_t)dst->ne[0], (uint32_t)dst->ne[1], (uint32_t)(dst->nb[1]/sizeof(float)), (uint32_t)(src0->nb[1]/sizeof(float)), nadd }, dryrun);
+}
+
 static void ggml_vk_diag_mask_inf(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     int32_t * op_params = (int32_t *)dst->op_params;
     ggml_vk_op_f32<vk_op_diag_mask_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] }, dryrun);
@@ -8463,6 +8488,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         }
         break;
     case GGML_OP_FUSED_MUL_UNARY:
+    case GGML_OP_MULTI_ADD:
     case GGML_OP_REPEAT:
     case GGML_OP_REPEAT_BACK:
     case GGML_OP_GET_ROWS:
@@ -8546,6 +8572,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         case GGML_OP_RMS_NORM_BACK:
         case GGML_OP_UNARY:
         case GGML_OP_FUSED_MUL_UNARY:
+        case GGML_OP_MULTI_ADD:
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_SOFT_MAX_BACK:
@@ -8662,6 +8689,9 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_FUSED_MUL_UNARY:
         ggml_vk_fused_mul_unary(ctx, compute_ctx, src0, src1, node, dryrun);
         break;
+    case GGML_OP_MULTI_ADD:
+        ggml_vk_multi_add(ctx, compute_ctx, src0, node, dryrun);
+        break;
     case GGML_OP_UNARY:
         switch (ggml_get_unary_op(node)) {
         case GGML_UNARY_OP_SILU:
@@ -8834,6 +8864,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
     case GGML_OP_REPEAT:
     case GGML_OP_REPEAT_BACK:
     case GGML_OP_FUSED_MUL_UNARY:
+    case GGML_OP_MULTI_ADD:
         buf = tensor->buffer;
 
         break;
@@ -9530,6 +9561,8 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const
                     return false;
             }
             break;
+        case GGML_OP_MULTI_ADD:
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32 && op->ne[2] == 1 && op->ne[3] == 1;
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             {
@@ -10266,6 +10299,8 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
         }
     } else if (tensor->op == GGML_OP_FUSED_MUL_UNARY) {
         tensor_clone = ggml_fused_mul_unary(ggml_ctx, src_clone[0], src_clone[1], (ggml_unary_op)tensor->op_params[0]);
+    } else if (tensor->op == GGML_OP_MULTI_ADD) {
+        tensor_clone = ggml_multi_add(ggml_ctx, src_clone[0], tensor->op_params[0]);
     } else if (tensor->op == GGML_OP_CPY || tensor->op == GGML_OP_DUP) {
         if (src1 == nullptr) {
             tensor_clone = ggml_dup(ggml_ctx, src_clone[0]);

ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -579,6 +579,8 @@ void process_shaders() {
     string_to_spv("fused_mul_relu_f16",   "fused_mul_relu.comp",   {{"A_TYPE", "float16_t"},   {"B_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
     string_to_spv("fused_mul_relu_f32",   "fused_mul_relu.comp",   {{"A_TYPE", "float"},       {"B_TYPE", "float"},     {"D_TYPE", "float"}});
 
+    string_to_spv("multi_add_f32", "multi_add.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+
     string_to_spv("gelu_f16",       "gelu.comp",        {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("gelu_f32",       "gelu.comp",        {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
     string_to_spv("gelu_quick_f16", "gelu_quick.comp",  {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});

src/llama.cpp

@@ -9870,28 +9870,28 @@ llm_expert_gating_func_type   gating_op,
         cb(cur, "ffn_moe_weighted", il);
     }
 
-#ifdef GGML_USE_VULKAN
-    // aggregate experts
-    ggml_tensor * moe_out = nullptr;
-    //ggml_tensor * first_expert = nullptr;
-    for (int i = 0; i < n_expert_used; ++i) {
-        ggml_tensor * cur_expert = ggml_view_2d(ctx, experts, n_embd, n_tokens,
-                experts->nb[2], i*experts->nb[1]);
-
-        if (i == 0) {
-            moe_out = cur_expert;
-        } else {
-            moe_out = ggml_add(ctx, moe_out, cur_expert);
-        }
-    }
-
-    if (n_expert_used == 1) {
-        // avoid returning a non-contiguous tensor
-        moe_out = ggml_cont(ctx, moe_out);
-    }
-
-    return moe_out;
-#else
+//#ifdef GGML_USE_VULKAN
+//    // aggregate experts
+//    ggml_tensor * moe_out = nullptr;
+//    //ggml_tensor * first_expert = nullptr;
+//    for (int i = 0; i < n_expert_used; ++i) {
+//        ggml_tensor * cur_expert = ggml_view_2d(ctx, experts, n_embd, n_tokens,
+//                experts->nb[2], i*experts->nb[1]);
+//
+//        if (i == 0) {
+//            moe_out = cur_expert;
+//        } else {
+//            moe_out = ggml_add(ctx, moe_out, cur_expert);
+//        }
+//    }
+//
+//    if (n_expert_used == 1) {
+//        // avoid returning a non-contiguous tensor
+//        moe_out = ggml_cont(ctx, moe_out);
+//    }
+//
+//    return moe_out;
+//#else
     if (n_expert_used == 1) {
         return ggml_cont(ctx, ggml_view_2d(ctx, experts, n_embd, n_tokens, experts->nb[2], 0));
     }
@@ -9900,7 +9900,7 @@ llm_expert_gating_func_type   gating_op,
                              ggml_view_2d(ctx, experts, n_embd, n_tokens, experts->nb[2], experts->nb[1]));
     }
     return ggml_multi_add(ctx, ggml_view_2d(ctx, experts, n_embd, n_tokens, experts->nb[2], 0), n_expert_used);
-#endif
+//#endif
 
 }