vulkan: optimize rms_norm, and allow the work to spread across multiple SMs

There are really two parts to this change:
(1) Some optimizations similar to what we have in soft_max, to unroll with
different numbers of iterations.
(2) A fusion optimization where we detect add followed by rms_norm, and make
the add shader atomically accumulate the values^2 into memory. Then the
rms_norm shader can just load that sum. This allows the rms_norm to be
parallelized across multiple workgroups, it just becomes a simple per-element
multiply.

The fusion optimization is currently only applied when the rms_norm is on a
single vector. This previously always ran on a single SM. It could apply more
broadly, but when there are other dimensions the work can already spread across
SMs, and there would be some complexity to tracking multiple atomic sums.

Author: jeffbolznv

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

@@ -372,6 +372,10 @@ struct vk_device_struct {
     bool subgroup_shuffle;
     bool multi_add;
 
+    bool atomic_float_add;
+    bool add_rms_fusion;
+    uint32_t atomic_binding_alignment;
+
     bool integer_dot_product;
 
     bool subgroup_size_control;
@@ -451,6 +455,8 @@ struct vk_device_struct {
     vk_pipeline pipeline_mul_norepeat[2][2][2];
     vk_pipeline pipeline_div[2][2][2];
     vk_pipeline pipeline_div_norepeat[2][2][2];
+    vk_pipeline pipeline_add_rms[2][2][2];
+    vk_pipeline pipeline_add_rms_norepeat[2][2][2];
 
     // indexed by num_additional_fused_ops == num_adds - 1
     vk_pipeline pipeline_multi_add[MAX_FUSED_ADDS];
@@ -1165,6 +1171,12 @@ class vk_perf_logger {
             timings[name].push_back(time);
             return;
         }
+        if (node->op == GGML_OP_RMS_NORM) {
+            std::string   name    = ggml_op_name(node->op);
+            name += "(" + std::to_string(node->ne[0]) + "," + std::to_string(node->ne[1]) + "," + std::to_string(node->ne[2]) + "," + std::to_string(node->ne[3]) + ")";
+            timings[name].push_back(time);
+            return;
+        }
         timings[ggml_op_name(node->op)].push_back(time);
     }
   private:
@@ -1179,10 +1191,13 @@ struct ggml_backend_vk_context {
 
     size_t semaphore_idx, event_idx;
     ggml_vk_garbage_collector gc;
-    size_t prealloc_size_x, prealloc_size_y, prealloc_size_split_k;
-    vk_buffer prealloc_x, prealloc_y, prealloc_split_k;
+    size_t prealloc_size_x, prealloc_size_y, prealloc_size_split_k, prealloc_size_atomic_add, prealloc_size_atomic_add_offset;
+    vk_buffer prealloc_x, prealloc_y, prealloc_split_k, prealloc_atomic_add;
     vk::Fence fence, almost_ready_fence;
     bool almost_ready_fence_pending {};
+    // Set before op_add and unset after op_rms_norm to indicate that the add should
+    // use atomics to accumulate the square of the vector components
+    bool do_add_rms_atomic;
 
     vk_buffer buffer_pool[MAX_VK_BUFFERS];
 
@@ -2921,8 +2936,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {0, 0}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_rms_norm_mul_f32, "rms_norm_mul_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {0, 1}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 4, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {0, 0}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_rms_norm_mul_f32, "rms_norm_mul_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 4, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {0, 1}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_rms_norm_back_f32, "rms_norm_back_f32", rms_norm_back_f32_len, rms_norm_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_l2_norm_f32, "l2_norm_f32", l2_norm_f32_len, l2_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
 
@@ -2992,20 +3007,22 @@ static void ggml_vk_load_shaders(vk_device& device) {
     };
 
     bool rte = device->float_controls_rte_fp16;
-#define CREATE_BINARY(name, namemod, spec) \
+#define CREATE_BINARY(name, namemod, spec, bindings) \
     for (int s0 : {0,1}) for (int s1 : {0,1}) for (int d : {0,1}) \
         ggml_vk_create_pipeline(device, device->pipeline_ ## name ## namemod[s0][s1][d], \
                                 #name + get_suffix(s0, s1, d) + #namemod, name ## _len[s0][s1][d][rte], name ## _data[s0][s1][d][rte], \
-                                "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, spec, 1);
-
-    CREATE_BINARY(add, , {0})
-    CREATE_BINARY(add, _norepeat, {1})
-    CREATE_BINARY(sub, , {0})
-    CREATE_BINARY(sub, _norepeat, {1})
-    CREATE_BINARY(mul, , {0})
-    CREATE_BINARY(mul, _norepeat, {1})
-    CREATE_BINARY(div, , {0})
-    CREATE_BINARY(div, _norepeat, {1})
+                                "main", (bindings), sizeof(vk_op_binary_push_constants), {512, 1, 1}, spec, 1);
+
+    CREATE_BINARY(add, , {0}, 4)
+    CREATE_BINARY(add, _norepeat, {1}, 4)
+    CREATE_BINARY(sub, , {0}, 3)
+    CREATE_BINARY(sub, _norepeat, {1}, 3)
+    CREATE_BINARY(mul, , {0}, 3)
+    CREATE_BINARY(mul, _norepeat, {1}, 3)
+    CREATE_BINARY(div, , {0}, 3)
+    CREATE_BINARY(div, _norepeat, {1}, 3)
+    CREATE_BINARY(add_rms, , {0}, 4)
+    CREATE_BINARY(add_rms, _norepeat, {1}, 4)
 #undef CREATE_BINARY
 
     if (device->multi_add) {
@@ -3286,6 +3303,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->coopmat_support = false;
         device->integer_dot_product = false;
         bool bfloat16_support = false;
+        bool atomic_float_support = false;
 
         for (const auto& properties : ext_props) {
             if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) {
@@ -3325,6 +3343,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
                        !getenv("GGML_VK_DISABLE_BFLOAT16")) {
                 bfloat16_support = true;
 #endif
+            } else if (strcmp("VK_EXT_shader_atomic_float", properties.extensionName) == 0) {
+                atomic_float_support = true;
             }
         }
 
@@ -3541,6 +3561,14 @@ static vk_device ggml_vk_get_device(size_t idx) {
             device_extensions.push_back("VK_KHR_shader_integer_dot_product");
         }
 
+        VkPhysicalDeviceShaderAtomicFloatFeaturesEXT atomic_float_features {};
+        atomic_float_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_FLOAT_FEATURES_EXT;
+        if (atomic_float_support) {
+            last_struct->pNext = (VkBaseOutStructure *)&atomic_float_features;
+            last_struct = (VkBaseOutStructure *)&atomic_float_features;
+            device_extensions.push_back("VK_EXT_shader_atomic_float");
+        }
+
         vkGetPhysicalDeviceFeatures2(device->physical_device, &device_features2);
 
         device->fp16 = device->fp16 && vk12_features.shaderFloat16;
@@ -3552,6 +3580,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
 #endif
 
         device->pipeline_robustness = pl_robustness_features.pipelineRobustness;
+        device->atomic_float_add = atomic_float_features.shaderBufferFloat32AtomicAdd;
 
         device->multi_add = vk12_props.shaderRoundingModeRTEFloat16 &&
                             device->properties.limits.maxPushConstantsSize >= sizeof(vk_op_multi_add_push_constants) &&
@@ -3872,6 +3901,12 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         device->disable_fusion = getenv("GGML_VK_DISABLE_FUSION") != nullptr;
 
+        device->add_rms_fusion = !device->disable_fusion &&
+                                 device->subgroup_add &&
+                                 device->atomic_float_add;
+        device->atomic_binding_alignment =
+            std::max(4u, (uint32_t)device->properties.limits.minStorageBufferOffsetAlignment);
+
         return device;
     }
 
@@ -6914,10 +6949,19 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         case GGML_OP_ADD:
         {
             if (ctx->num_additional_fused_ops > 0) {
-                return ctx->device->pipeline_multi_add[ctx->num_additional_fused_ops];
+                if (ctx->do_add_rms_atomic) {
+                    return ctx->device->pipeline_multi_add_rms[ctx->num_additional_fused_ops];
+                } else {
+                    return ctx->device->pipeline_multi_add[ctx->num_additional_fused_ops];
+                }
+            }
+            if (ctx->do_add_rms_atomic) {
+                auto pipelines = ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_rms_norepeat : ctx->device->pipeline_add_rms;
+                return pipelines[src0->type == GGML_TYPE_F16][src1->type == GGML_TYPE_F16][dst->type == GGML_TYPE_F16];
+            } else {
+                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];
             }
-            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];
         }
         case GGML_OP_SUB:
         {
@@ -7523,7 +7567,12 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
             }
         } break;
     case GGML_OP_RMS_NORM:
-        elements = { (uint32_t)ne01, (uint32_t)ne02, (uint32_t)ne03 };
+        if (ctx->do_add_rms_atomic) {
+            // Run one element per thread, 128 threads per workgroup
+            elements = { (uint32_t)CEIL_DIV(ne00, 128), 1, 1 };
+        } else {
+            elements = { (uint32_t)ne01, (uint32_t)ne02, (uint32_t)ne03 };
+        }
         break;
 
     case GGML_OP_SUM:
@@ -7671,7 +7720,17 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         }
     }
 
-    if (op == GGML_OP_GLU) {
+    if (op == GGML_OP_ADD || op == GGML_OP_RMS_NORM) {
+        vk_buffer d_A = ctx->prealloc_atomic_add ? ctx->prealloc_atomic_add : d_X;
+        size_t a_buf_offset = ctx->prealloc_atomic_add ? ctx->prealloc_size_atomic_add_offset : 0;
+        ggml_vk_sync_buffers(subctx);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
+            { vk_subbuffer{ d_X, x_buf_offset, x_sz },
+              vk_subbuffer{ d_Y, y_buf_offset, y_sz },
+              vk_subbuffer{ d_D, d_buf_offset, d_sz },
+              vk_subbuffer{ d_A, a_buf_offset, VK_WHOLE_SIZE },
+            }, pc, elements);
+    } else if (op == GGML_OP_GLU) {
         // Empty src1 is possible in glu, but the shader needs a buffer
         vk_subbuffer subbuf_y;
         if (use_src1) {
@@ -7884,7 +7943,7 @@ static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
-        0.0f, 0.0f, 0,
+        0.0f, 0.0f, ctx->do_add_rms_atomic,
     }, dryrun);
 }
 
@@ -8353,8 +8412,13 @@ static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
-        op_params[0], 0.0f, 0,
+        op_params[0], 0.0f, ctx->do_add_rms_atomic,
     }, dryrun);
+
+    if (ctx->do_add_rms_atomic) {
+        ctx->prealloc_size_atomic_add_offset += ctx->device->atomic_binding_alignment;
+        ctx->do_add_rms_atomic = false;
+    }
 }
 
 static void ggml_vk_rms_norm_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -9632,6 +9696,14 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
         }
         ctx->prealloc_split_k = ggml_vk_create_buffer_device(ctx->device, ctx->prealloc_size_split_k);
     }
+    if (ctx->prealloc_atomic_add == nullptr || (ctx->prealloc_size_atomic_add > 0 && ctx->prealloc_atomic_add->size < ctx->prealloc_size_atomic_add)) {
+        VK_LOG_MEMORY("ggml_vk_preallocate_buffers(atomic_add_size: " << ctx->prealloc_atomic_add << ")");
+        // Resize buffer
+        if (ctx->prealloc_atomic_add != nullptr) {
+            ggml_vk_destroy_buffer(ctx->prealloc_atomic_add);
+        }
+        ctx->prealloc_atomic_add = ggml_vk_create_buffer_device(ctx->device, ctx->prealloc_size_atomic_add);
+    }
 }
 
 static bool ggml_vk_compute_forward(ggml_backend_vk_context* ctx, ggml_cgraph * cgraph, ggml_tensor* tensor, int tensor_idx, bool use_fence, bool almost_ready);
@@ -9687,10 +9759,21 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
             return false;
         }
         break;
+    case GGML_OP_ADD:
+        if (node_idx + 1 < cgraph->n_nodes &&
+            cgraph->nodes[node_idx + 1]->op == GGML_OP_RMS_NORM &&
+            cgraph->nodes[node_idx + 1]->src[0] == cgraph->nodes[node_idx] &&
+            ggml_nrows(cgraph->nodes[node_idx + 1]) == 1 &&
+            ctx->device->add_rms_fusion) {
+            if (dryrun) {
+                ctx->prealloc_size_atomic_add += ctx->device->atomic_binding_alignment;
+            }
+            ctx->do_add_rms_atomic = true;
+        }
+        break;
     case GGML_OP_REPEAT:
     case GGML_OP_REPEAT_BACK:
     case GGML_OP_GET_ROWS:
-    case GGML_OP_ADD:
     case GGML_OP_ADD_ID:
     case GGML_OP_ACC:
     case GGML_OP_SUB:
@@ -9808,6 +9891,9 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
                 // do the only thing needed for the dryrun.
                 vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, src2, node, node->op);
                 ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+                if (node->op == GGML_OP_RMS_NORM) {
+                    ctx->do_add_rms_atomic = false;
+                }
                 return false;
             }
         default:
@@ -10782,6 +10868,10 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         vk_instance.pfn_vkQueueBeginDebugUtilsLabelEXT(ctx->device->compute_queue.queue, reinterpret_cast<VkDebugUtilsLabelEXT*>(&dul));
     }
 
+    ctx->prealloc_size_atomic_add = 0;
+    ctx->prealloc_size_atomic_add_offset = 0;
+    ctx->do_add_rms_atomic = false;
+
     uint64_t total_mat_mul_bytes = 0;
     for (int i = 0; i < cgraph->n_nodes; i++) {
         if (!ctx->device->disable_fusion) {
@@ -10847,6 +10937,18 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         compute_ctx->s->buffer.writeTimestamp(vk::PipelineStageFlagBits::eAllCommands, ctx->device->query_pool, 0);
     }
 
+    if (ctx->prealloc_size_atomic_add) {
+        if (ctx->compute_ctx.expired()) {
+            compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
+            ctx->compute_ctx = compute_ctx;
+            ggml_vk_ctx_begin(ctx->device, compute_ctx);
+        } else {
+            compute_ctx = ctx->compute_ctx.lock();
+        }
+        // initialize atomic sums to zero.
+        ggml_vk_buffer_memset_async(compute_ctx, ctx->prealloc_atomic_add, 0, 0, ctx->prealloc_size_atomic_add);
+    }
+
     // Submit after enough work has accumulated, to overlap CPU cmdbuffer generation with GPU execution.
     // Estimate the amount of matmul work by looking at the weight matrix size, and submit every 100MB
     // (and scaled down based on model size, so smaller models submit earlier).

ggml/src/ggml-vulkan/vulkan-shaders/add.comp

@@ -1,12 +1,19 @@
 #version 450
 
 #extension GL_EXT_shader_16bit_storage : require
+#if ADD_RMS
+#extension GL_EXT_shader_atomic_float : enable
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+#extension GL_KHR_shader_subgroup_basic : enable
+#endif
 
 #include "types.comp"
 #include "generic_binary_head.comp"
 
 const uint num_threads = 256;
 
+layout (binding = 3) buffer AtomBuf {float data_atom;};
+
 layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
 
 void main() {
@@ -15,15 +22,29 @@ void main() {
     // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
     const uint num_iter = 2;
 
+    FLOAT_TYPE sum_sq = 0;
+
     [[unroll]] for (uint i = 0; i < num_iter; ++i) {
         if (idx >= p.ne) {
             continue;
         }
         uint i00, i01, i02, i03;
         get_indices(idx, i00, i01, i02, i03);
 
-        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)]));
+        FLOAT_TYPE sum = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)]);
+        sum_sq += sum*sum;
+
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(sum);
 
         idx += num_threads;
     }
+
+#if ADD_RMS
+    if (p.param3 != 0) {
+        sum_sq = subgroupAdd(sum_sq);
+        if (sum_sq != 0 && gl_SubgroupInvocationID == 0) {
+            atomicAdd(data_atom, sum_sq);
+        }
+    }
+#endif
 }