@@ -566,6 +566,7 @@ struct vk_device_struct {
566566
567567 bool disable_fusion;
568568 bool disable_host_visible_vidmem;
569+ bool allow_sysmem_fallback;
569570
570571#ifdef GGML_VULKAN_MEMORY_DEBUG
571572 std::unique_ptr<vk_memory_logger> memory_logger;
@@ -1808,8 +1809,8 @@ static uint32_t find_properties(const vk::PhysicalDeviceMemoryProperties* mem_pr
18081809 return UINT32_MAX;
18091810}
18101811
1811- static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::MemoryPropertyFlags req_flags, vk::MemoryPropertyFlags fallback_flags = vk::MemoryPropertyFlags(0) ) {
1812- VK_LOG_DEBUG("ggml_vk_create_buffer(" << device->name << ", " << size << ", " << to_string(req_flags) << ", " << to_string(fallback_flags ) << ")");
1812+ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, const std::initializer_list< vk::MemoryPropertyFlags> & req_flags_list ) {
1813+ VK_LOG_DEBUG("ggml_vk_create_buffer(" << device->name << ", " << size << ", " << to_string(req_flags_list.begin()[0]) << ", " << to_string(req_flags_list.begin()[req_flags_list.size()-1] ) << ")");
18131814 if (size > device->max_memory_allocation_size) {
18141815 throw vk::OutOfDeviceMemoryError("Requested buffer size exceeds device memory allocation limit");
18151816 }
@@ -1836,42 +1837,27 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor
18361837
18371838 vk::PhysicalDeviceMemoryProperties mem_props = device->physical_device.getMemoryProperties();
18381839
1839- uint32_t memory_type_index = UINT32_MAX;
1840+ for (auto &req_flags : req_flags_list) {
1841+ uint32_t memory_type_index = find_properties(&mem_props, &mem_req, req_flags);
18401842
1841- memory_type_index = find_properties(&mem_props, &mem_req, req_flags);
1842- buf->memory_property_flags = req_flags;
1843+ if (memory_type_index == UINT32_MAX) {
1844+ continue;
1845+ }
1846+ buf->memory_property_flags = req_flags;
18431847
1844- if (memory_type_index == UINT32_MAX && fallback_flags) {
1845- memory_type_index = find_properties(&mem_props, &mem_req, fallback_flags);
1846- buf->memory_property_flags = fallback_flags;
1848+ try {
1849+ buf->device_memory = device->device.allocateMemory({ mem_req.size, memory_type_index });
1850+ break;
1851+ } catch (const vk::SystemError& e) {
1852+ // loop and retry
1853+ }
18471854 }
18481855
1849- if (memory_type_index == UINT32_MAX ) {
1856+ if (buf->device_memory == VK_NULL_HANDLE ) {
18501857 device->device.destroyBuffer(buf->buffer);
18511858 throw vk::OutOfDeviceMemoryError("No suitable memory type found");
18521859 }
18531860
1854- try {
1855- buf->device_memory = device->device.allocateMemory({ mem_req.size, memory_type_index });
1856- } catch (const vk::SystemError& e) {
1857- if (buf->memory_property_flags != fallback_flags) {
1858- // Try again with fallback flags
1859- memory_type_index = find_properties(&mem_props, &mem_req, fallback_flags);
1860- buf->memory_property_flags = fallback_flags;
1861-
1862- try {
1863- buf->device_memory = device->device.allocateMemory({ mem_req.size, memory_type_index });
1864- }
1865- catch (const vk::SystemError& e) {
1866- device->device.destroyBuffer(buf->buffer);
1867- throw e;
1868- }
1869- } else {
1870- // Out of Host/Device memory, clean up buffer
1871- device->device.destroyBuffer(buf->buffer);
1872- throw e;
1873- }
1874- }
18751861 buf->ptr = nullptr;
18761862
18771863 if (buf->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
@@ -1892,7 +1878,7 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor
18921878
18931879static vk_buffer ggml_vk_create_buffer_check(vk_device& device, size_t size, vk::MemoryPropertyFlags req_flags, vk::MemoryPropertyFlags fallback_flags = vk::MemoryPropertyFlags(0)) {
18941880 try {
1895- return ggml_vk_create_buffer(device, size, req_flags, fallback_flags);
1881+ return ggml_vk_create_buffer(device, size, { req_flags, fallback_flags} );
18961882 } catch (const vk::SystemError& e) {
18971883 std::cerr << "ggml_vulkan: Memory allocation of size " << size << " failed." << std::endl;
18981884 std::cerr << "ggml_vulkan: " << e.what() << std::endl;
@@ -1904,15 +1890,29 @@ static vk_buffer ggml_vk_create_buffer_device(vk_device& device, size_t size) {
19041890 vk_buffer buf;
19051891 try {
19061892 if (device->prefer_host_memory) {
1907- buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent, vk::MemoryPropertyFlagBits::eDeviceLocal);
1893+ buf = ggml_vk_create_buffer(device, size, {vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent,
1894+ vk::MemoryPropertyFlagBits::eDeviceLocal});
19081895 } else if (device->uma) {
19091896 // Fall back to host memory type
1910- buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eDeviceLocal, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
1897+ buf = ggml_vk_create_buffer(device, size, {vk::MemoryPropertyFlagBits::eDeviceLocal,
1898+ vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent});
19111899 } else if (device->disable_host_visible_vidmem) {
1912- buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eDeviceLocal, vk::MemoryPropertyFlagBits::eDeviceLocal);
1900+ if (device->allow_sysmem_fallback) {
1901+ buf = ggml_vk_create_buffer(device, size, {vk::MemoryPropertyFlagBits::eDeviceLocal,
1902+ vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent});
1903+ } else {
1904+ buf = ggml_vk_create_buffer(device, size, {vk::MemoryPropertyFlagBits::eDeviceLocal});
1905+ }
19131906 } else {
19141907 // use rebar if available, otherwise fallback to device only visible memory
1915- buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eDeviceLocal | vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent, vk::MemoryPropertyFlagBits::eDeviceLocal);
1908+ if (device->allow_sysmem_fallback) {
1909+ buf = ggml_vk_create_buffer(device, size, {vk::MemoryPropertyFlagBits::eDeviceLocal | vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent,
1910+ vk::MemoryPropertyFlagBits::eDeviceLocal,
1911+ vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent});
1912+ } else {
1913+ buf = ggml_vk_create_buffer(device, size, {vk::MemoryPropertyFlagBits::eDeviceLocal | vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent,
1914+ vk::MemoryPropertyFlagBits::eDeviceLocal});
1915+ }
19161916 }
19171917 } catch (const vk::SystemError& e) {
19181918 std::cerr << "ggml_vulkan: Device memory allocation of size " << size << " failed." << std::endl;
@@ -3437,6 +3437,9 @@ static vk_device ggml_vk_get_device(size_t idx) {
34373437 const char* GGML_VK_DISABLE_HOST_VISIBLE_VIDMEM = getenv("GGML_VK_DISABLE_HOST_VISIBLE_VIDMEM");
34383438 device->disable_host_visible_vidmem = GGML_VK_DISABLE_HOST_VISIBLE_VIDMEM != nullptr;
34393439
3440+ const char* GGML_VK_ALLOW_SYSMEM_FALLBACK = getenv("GGML_VK_ALLOW_SYSMEM_FALLBACK");
3441+ device->allow_sysmem_fallback = GGML_VK_ALLOW_SYSMEM_FALLBACK != nullptr;
3442+
34403443 bool fp16_storage = false;
34413444 bool fp16_compute = false;
34423445 bool maintenance4_support = false;
@@ -4774,8 +4777,8 @@ static vk_buffer ggml_vk_create_buffer_temp(ggml_backend_vk_context * ctx, size_
47744777static void * ggml_vk_host_malloc(vk_device& device, size_t size) {
47754778 VK_LOG_MEMORY("ggml_vk_host_malloc(" << size << ")");
47764779 vk_buffer buf = ggml_vk_create_buffer(device, size,
4777- vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
4778- vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
4780+ { vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
4781+ vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent} );
47794782
47804783 if(!(buf->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible)) {
47814784 fprintf(stderr, "WARNING: failed to allocate %.2f MB of pinned memory\n",
@@ -9182,7 +9185,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
91829185 if (ctx->prealloc_split_k != nullptr) {
91839186 ggml_vk_destroy_buffer(ctx->prealloc_split_k);
91849187 }
9185- ctx->prealloc_split_k = ggml_vk_create_buffer_check(ctx->device, sizeof(float) * d_ne * split_k, vk::MemoryPropertyFlagBits::eDeviceLocal);
9188+ ctx->prealloc_split_k = ggml_vk_create_buffer_check(ctx->device, sizeof(float) * d_ne * split_k, { vk::MemoryPropertyFlagBits::eDeviceLocal} );
91869189 }
91879190 }
91889191
@@ -9192,9 +9195,9 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
91929195
91939196 ggml_pipeline_allocate_descriptor_sets(ctx);
91949197
9195- vk_buffer d_X = ggml_vk_create_buffer_check(ctx->device, sizeof(X_TYPE) * x_ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
9196- vk_buffer d_Y = ggml_vk_create_buffer_check(ctx->device, sizeof(Y_TYPE) * y_ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
9197- vk_buffer d_D = ggml_vk_create_buffer_check(ctx->device, sizeof(float) * d_ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
9198+ vk_buffer d_X = ggml_vk_create_buffer_check(ctx->device, sizeof(X_TYPE) * x_ne, { vk::MemoryPropertyFlagBits::eDeviceLocal} );
9199+ vk_buffer d_Y = ggml_vk_create_buffer_check(ctx->device, sizeof(Y_TYPE) * y_ne, { vk::MemoryPropertyFlagBits::eDeviceLocal} );
9200+ vk_buffer d_D = ggml_vk_create_buffer_check(ctx->device, sizeof(float) * d_ne, { vk::MemoryPropertyFlagBits::eDeviceLocal} );
91989201
91999202 X_TYPE* x = (X_TYPE *) malloc(sizeof(X_TYPE) * x_ne);
92009203 Y_TYPE* y = (Y_TYPE *) malloc(sizeof(Y_TYPE) * y_ne);
@@ -9420,8 +9423,8 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_
94209423 const size_t qx_sz = ne * ggml_type_size(quant)/ggml_blck_size(quant);
94219424 float * x = (float *) malloc(x_sz);
94229425 void * qx = malloc(qx_sz);
9423- vk_buffer qx_buf = ggml_vk_create_buffer_check(ctx->device, qx_sz, vk::MemoryPropertyFlagBits::eDeviceLocal);
9424- vk_buffer x_buf = ggml_vk_create_buffer_check(ctx->device, x_sz_f16, vk::MemoryPropertyFlagBits::eDeviceLocal);
9426+ vk_buffer qx_buf = ggml_vk_create_buffer_check(ctx->device, qx_sz, { vk::MemoryPropertyFlagBits::eDeviceLocal} );
9427+ vk_buffer x_buf = ggml_vk_create_buffer_check(ctx->device, x_sz_f16, { vk::MemoryPropertyFlagBits::eDeviceLocal} );
94259428 float * x_ref = (float *) malloc(x_sz);
94269429 ggml_fp16_t * x_chk = (ggml_fp16_t *) malloc(x_sz_f16);
94279430
@@ -9526,8 +9529,8 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_
95269529// float * x = (float *) malloc(x_sz);
95279530// block_q8_1 * qx = (block_q8_1 *)malloc(qx_sz);
95289531// block_q8_1 * qx_res = (block_q8_1 *)malloc(qx_sz);
9529- // vk_buffer x_buf = ggml_vk_create_buffer_check(ctx->device, x_sz, vk::MemoryPropertyFlagBits::eDeviceLocal);
9530- // vk_buffer qx_buf = ggml_vk_create_buffer_check(ctx->device, qx_sz, vk::MemoryPropertyFlagBits::eDeviceLocal);
9532+ // vk_buffer x_buf = ggml_vk_create_buffer_check(ctx->device, x_sz, { vk::MemoryPropertyFlagBits::eDeviceLocal} );
9533+ // vk_buffer qx_buf = ggml_vk_create_buffer_check(ctx->device, qx_sz, { vk::MemoryPropertyFlagBits::eDeviceLocal} );
95319534//
95329535// for (size_t i = 0; i < ne; i++) {
95339536// x[i] = rand() / (float)RAND_MAX;
@@ -9674,10 +9677,10 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
96749677 float * x = (float *) malloc(x_sz);
96759678 float * y = (float *) malloc(y_sz);
96769679 void * qx = malloc(qx_sz);
9677- vk_buffer qx_buf = ggml_vk_create_buffer_check(ctx->device, qx_sz, vk::MemoryPropertyFlagBits::eDeviceLocal);
9678- vk_buffer y_buf = ggml_vk_create_buffer_check(ctx->device, y_sz, vk::MemoryPropertyFlagBits::eDeviceLocal);
9679- vk_buffer qy_buf = ggml_vk_create_buffer_check(ctx->device, qy_sz, vk::MemoryPropertyFlagBits::eDeviceLocal);
9680- vk_buffer d_buf = ggml_vk_create_buffer_check(ctx->device, d_sz, vk::MemoryPropertyFlagBits::eDeviceLocal);
9680+ vk_buffer qx_buf = ggml_vk_create_buffer_check(ctx->device, qx_sz, { vk::MemoryPropertyFlagBits::eDeviceLocal} );
9681+ vk_buffer y_buf = ggml_vk_create_buffer_check(ctx->device, y_sz, { vk::MemoryPropertyFlagBits::eDeviceLocal} );
9682+ vk_buffer qy_buf = ggml_vk_create_buffer_check(ctx->device, qy_sz, { vk::MemoryPropertyFlagBits::eDeviceLocal} );
9683+ vk_buffer d_buf = ggml_vk_create_buffer_check(ctx->device, d_sz, { vk::MemoryPropertyFlagBits::eDeviceLocal} );
96819684 float * d = (float *) malloc(d_sz);
96829685 float * d_chk = (float *) malloc(d_sz);
96839686
@@ -9704,7 +9707,7 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
97049707 if (ctx->prealloc_split_k != nullptr) {
97059708 ggml_vk_destroy_buffer(ctx->prealloc_split_k);
97069709 }
9707- ctx->prealloc_split_k = ggml_vk_create_buffer_check(ctx->device, sizeof(float) * d_ne * split_k, vk::MemoryPropertyFlagBits::eDeviceLocal);
9710+ ctx->prealloc_split_k = ggml_vk_create_buffer_check(ctx->device, sizeof(float) * d_ne * split_k, { vk::MemoryPropertyFlagBits::eDeviceLocal} );
97089711 }
97099712 }
97109713 if (mmq) {
0 commit comments