metal : avoid using Metal's gpuAddress property (llama/16576)

* metal : avoid using Metal's gpuAddress property

* metal : fix rope kernels buffer check

diff --git a/src/ggml-metal/ggml-metal-device.m b/src/ggml-metal/ggml-metal-device.m
index c3fe8f4e9100236a538cda079959010f407acd75..553cf8f5f39ac8a6bdc58262ee2973468ce29407 100644 (file)
--- a/src/ggml-metal/ggml-metal-device.m
+++ b/src/ggml-metal/ggml-metal-device.m
@@ -7,6 +7,8 @@
 
 #include <Metal/Metal.h>
 
+#include <stdatomic.h>
+
 #ifndef TARGET_OS_VISION
 #define TARGET_OS_VISION 0
 #endif
@@ -22,6 +24,9 @@
 // overload of MTLGPUFamilyMetal3 (not available in some environments)
 static const NSInteger MTLGPUFamilyMetal3_GGML = 5001;
 
+// virtual address for GPU memory allocations
+static atomic_uintptr_t g_addr_device = 0x000000400ULL;
+
 #if !GGML_METAL_EMBED_LIBRARY
 // Here to assist with NSBundle Path Hack
 @interface GGMLMetalClass : NSObject
@@ -827,7 +832,7 @@ struct ggml_metal_buffer_wrapper {
 };
 
 struct ggml_metal_buffer {
-    void * all_data; // TODO: https://github.com/ggml-org/llama.cpp/pull/15985
+    void * all_data;
     size_t all_size;
 
     // if false, the Metal buffer data is allocated in private GPU memory and is not shared with the host
@@ -965,14 +970,15 @@ ggml_metal_buffer_t ggml_metal_buffer_init(ggml_metal_device_t dev, size_t size,
     if (shared) {
         res->all_data = ggml_metal_host_malloc(size_aligned);
         res->is_shared = true;
-        res->owned = true;
     } else {
-        // dummy, non-NULL value - we'll populate this after creating the Metal buffer below
-        res->all_data = (void *) 0x000000400ULL;
+        // use virtual address from g_addr_device counter
+        res->all_data = (void *) atomic_fetch_add_explicit(&g_addr_device, size_aligned, memory_order_relaxed);
         res->is_shared = false;
     }
     res->all_size = size_aligned;
 
+    res->owned = true;
+
     res->device = ggml_metal_device_get_obj(dev);
     res->queue  = ggml_metal_device_get_queue(dev);
 
@@ -983,15 +989,13 @@ ggml_metal_buffer_t ggml_metal_buffer_init(ggml_metal_device_t dev, size_t size,
         res->buffers[0].metal = nil;
 
         if (size_aligned > 0) {
-            if (props_dev->use_shared_buffers &&shared) {
+            if (props_dev->use_shared_buffers && shared) {
                 res->buffers[0].metal = [res->device newBufferWithBytesNoCopy:res->all_data
                                                                   length:size_aligned
                                                                  options:MTLResourceStorageModeShared
                                                              deallocator:nil];
             } else {
                 res->buffers[0].metal = [res->device newBufferWithLength:size_aligned options:MTLResourceStorageModePrivate];
-
-                res->all_data = (void *) (res->buffers[0].metal.gpuAddress);
             }
         }
 
@@ -1139,7 +1143,7 @@ bool ggml_metal_buffer_is_shared(ggml_metal_buffer_t buf) {
 
 void ggml_metal_buffer_memset_tensor(ggml_metal_buffer_t buf, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
     if (buf->is_shared) {
-        memset((char *)tensor->data + offset, value, size);
+        memset((char *) tensor->data + offset, value, size);
         return;
     }
 
@@ -1168,7 +1172,7 @@ void ggml_metal_buffer_memset_tensor(ggml_metal_buffer_t buf, struct ggml_tensor
 
 void ggml_metal_buffer_set_tensor(ggml_metal_buffer_t buf, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     if (buf->is_shared) {
-        memcpy((char *)tensor->data + offset, data, size);
+        memcpy((char *) tensor->data + offset, data, size);
         return;
     }
 
@@ -1223,7 +1227,7 @@ void ggml_metal_buffer_set_tensor(ggml_metal_buffer_t buf, struct ggml_tensor *
 
 void ggml_metal_buffer_get_tensor(ggml_metal_buffer_t buf, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     if (buf->is_shared) {
-        memcpy(data, (const char *)tensor->data + offset, size);
+        memcpy(data, (const char *) tensor->data + offset, size);
         return;
     }
 

diff --git a/src/ggml-metal/ggml-metal-impl.h b/src/ggml-metal/ggml-metal-impl.h
index a448c14f66b63616c960c712c6a4fb4f26b6ca07..fa2d82cefb40eb366fcfc66a74c0ae250fb2754f 100644 (file)
--- a/src/ggml-metal/ggml-metal-impl.h
+++ b/src/ggml-metal/ggml-metal-impl.h
@@ -251,6 +251,7 @@ typedef struct {
     int32_t  sect_1;
     int32_t  sect_2;
     int32_t  sect_3;
+    bool     src2;
 } ggml_metal_kargs_rope;
 
 typedef struct {

diff --git a/src/ggml-metal/ggml-metal-ops.cpp b/src/ggml-metal/ggml-metal-ops.cpp
index a61ea8fb5a7b37202daa13d78b1c0616f451a78a..784b7b77851e6873f9c8c4c97c417ceb4fc0211b 100644 (file)
--- a/src/ggml-metal/ggml-metal-ops.cpp
+++ b/src/ggml-metal/ggml-metal-ops.cpp
@@ -2969,6 +2969,7 @@ int ggml_metal_op_rope(ggml_metal_op_t ctx, int idx) {
         /* sect_1      =*/ sect_1,
         /* sect_2      =*/ sect_2,
         /* sect_3      =*/ sect_3,
+        /* src2        =*/ op->src[2] != nullptr,
     };
 
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_rope(lib, op);

diff --git a/src/ggml-metal/ggml-metal.metal b/src/ggml-metal/ggml-metal.metal
index 1029cf8f9a3ab00545433fa9d057b8ef3fcdacba..6d39ddcc634ef735e379a5710b410f2a17a25022 100644 (file)
--- a/src/ggml-metal/ggml-metal.metal
+++ b/src/ggml-metal/ggml-metal.metal
@@ -3748,7 +3748,7 @@ kernel void kernel_rope_norm(
 
             const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
 
-            const float freq_factor = src2 != src0 ? ((device const float *) src2)[ic] : 1.0f;
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
 
             rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
 
@@ -3801,7 +3801,7 @@ kernel void kernel_rope_neox(
 
             const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
 
-            const float freq_factor = src2 != src0 ? ((device const float *) src2)[ic] : 1.0f;
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
 
             rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
 
@@ -3872,7 +3872,7 @@ kernel void kernel_rope_multi(
 
             const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
 
-            const float freq_factor = src2 != src0 ? ((device const float *) src2)[ic] : 1.0f;
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
 
             rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
 
@@ -3939,7 +3939,7 @@ kernel void kernel_rope_vision(
             const float theta = theta_base * pow(args.freq_base, 2.0f * inv_ndims * p);
             // end of mrope
 
-            const float freq_factor = src2 != src0 ? ((device const float *) src2)[ic] : 1.0f;
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
 
             rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);