vulkan : fix build for GGML_VULKAN_RUN_TESTS, add TFLOPS to log (#961)

AUTHORS

@@ -1,59 +1,91 @@
-# date: Tue Apr  9 20:16:51 EEST 2024
+# date: Thu Sep 26 09:19:50 CDT 2024
 # this file is auto-generated by scripts/gen-authors.sh
 
 0cc4m <[email protected]>
+65a <[email protected]>
 AT <[email protected]>
 Abhilash Majumder <[email protected]>
 Adam Tazi <[email protected]>
 AidanBeltonS <[email protected]>
+AidanBeltonS <[email protected]>
+Akarshan Biswas <[email protected]>
+Albert Jin <[email protected]>
+Alberto Cabrera Pérez <[email protected]>
+Alberto Cabrera Pérez <[email protected]>
 Alex Azarov <[email protected]>
+Alex O'Connell <[email protected]>
 Alex von Gluck IV <[email protected]>
 AmbientL <[email protected]>
 AmirAli Mirian <[email protected]>
 Ananta Bastola <[email protected]>
+Andreas (Andi) Kunar <[email protected]>
 Andrei <[email protected]>
 Arjun <[email protected]>
 Ashraful Islam <[email protected]>
 Astariul <[email protected]>
 AsukaMinato <[email protected]>
 Avi Lumelsky <[email protected]>
 Bart Pelle <[email protected]>
+Ben Ashbaugh <[email protected]>
 Borislav Stanimirov <[email protected]>
 Brad Ito <[email protected]>
+Brad Murray <[email protected]>
+Brian <[email protected]>
 Bryan Lozano <[email protected]>
 Carolinabanana <[email protected]>
+CarterLi999 <[email protected]>
 Cebtenzzre <[email protected]>
+Chen Xi <[email protected]>
+Chen Xi <[email protected]>
+Chris Elrod <[email protected]>
+Clint Herron <[email protected]>
+Conrad Kramer <[email protected]>
 Cordeiro <[email protected]>
 Cristiano Calcagno <[email protected]>
+DAN™ <[email protected]>
 Dan Forbes <[email protected]>
 Daniel Bevenius <[email protected]>
+Daniel Ziegenberg <[email protected]>
+Daniele <[email protected]>
 Daulet Zhanguzin <[email protected]>
+Dave <[email protected]>
+Dave Airlie <[email protected]>
+Dave Airlie <[email protected]>
 David Miller <[email protected]>
+DavidKorczynski <[email protected]>
 Davidson Francis <[email protected]>
+Dibakar Gope <[email protected]>
 Didzis Gosko <[email protected]>
 Diogo <[email protected]>
+Djip007 <[email protected]>
 Dr. Tom Murphy VII Ph.D <[email protected]>
 Ebey Abraham <[email protected]>
 Eldar Yusupov <[email protected]>
+Emmanuel Durand <[email protected]>
 Engininja2 <[email protected]>
 Erik Scholz <[email protected]>
 Ettore Di Giacinto <[email protected]>
 Eve <[email protected]>
 F1L1P <[email protected]>
+Faisal Zaghloul <[email protected]>
 FantasyGmm <[email protected]>
 Felix <[email protected]>
 Finn Voorhees <[email protected]>
 GainLee <[email protected]>
 George Hindle <[email protected]>
 Georgi Gerganov <[email protected]>
+Gilad S <[email protected]>
 Guillaume Wenzek <[email protected]>
 Halalaluyafail3 <[email protected]>
 Herman Semenov <[email protected]>
 Hirochika Matsumoto <[email protected]>
+Hong Bo PENG <[email protected]>
 Hugo Rosenkranz-Costa <[email protected]>
 Hyunsung Lee <[email protected]>
 IGUILIZ Salah-Eddine <[email protected]>
 Ian Bull <[email protected]>
+Ikko Eltociear Ashimine <[email protected]>
+Ivan Filipov <[email protected]>
 Ivan Stepanov <[email protected]>
 Ivan Zdane <[email protected]>
 Jack Mousseau <[email protected]>
@@ -62,35 +94,53 @@ JacobLinCool <[email protected]>
 Jakob Frick <[email protected]>
 Jan Ploski <[email protected]>
 Jared Van Bortel <[email protected]>
+Jeff Bolz <[email protected]>
 Jeffrey Quesnelle <[email protected]>
+Jeroen Mostert <[email protected]>
 Jiahao Li <[email protected]>
 JidongZhang-THU <[email protected]>
 Jiří Podivín <[email protected]>
 Jo Liss <[email protected]>
+Joe Todd <[email protected]>
 Johannes Gäßler <[email protected]>
 John Balis <[email protected]>
 Josh Bleecher Snyder <[email protected]>
 Judd <[email protected]>
+Justina Cho <[email protected]>
 Justine Tunney <[email protected]>
+Justine Tunney <[email protected]>
 Kawrakow <[email protected]>
+Kevin Gibbons <[email protected]>
 Konstantin Zhuravlyov <[email protected]>
 Kylin <[email protected]>
 LoganDark <[email protected]>
 LoganDark <[email protected]>
 LostRuins <[email protected]>
 Lukas Möller <[email protected]>
 M. Yusuf Sarıgöz <[email protected]>
+Mahesh Madhav <[email protected]>
 MaiHD <[email protected]>
+Mark Zhuang <[email protected]>
+Markus Tavenrath <[email protected]>
+Masaya, Kato <[email protected]>
 Mathijs de Bruin <[email protected]>
+Matt Stephenson <[email protected]>
+Max Krasnyansky <[email protected]>
 Mayank Kumar Pal <[email protected]>
 Meng, Hengyu <[email protected]>
+Mengqing Cao <[email protected]>
 Metal Whale <[email protected]>
 Michael Klimenko <[email protected]>
 Michael Podvitskiy <[email protected]>
 Michael Verrilli <[email protected]>
+Molly Sophia <[email protected]>
+Natsu <[email protected]>
+Neo Zhang <[email protected]>
 Neo Zhang Jianyu <[email protected]>
 Neuman Vong <[email protected]>
 Nevin <[email protected]>
+Nicholai Tukanov <[email protected]>
+Nico Bosshard <[email protected]>
 Nouamane Tazi <[email protected]>
 Olivier Chafik <[email protected]>
 Olivier Chafik <[email protected]>
@@ -101,6 +151,9 @@ Paul Tsochantaris <[email protected]>
 Philpax <[email protected]>
 Pierre Alexandre SCHEMBRI <[email protected]>
 Playdev <[email protected]>
+Przemysław Pawełczyk <[email protected]>
+R0CKSTAR <[email protected]>
+R0CKSTAR <[email protected]>
 Radoslav Gerganov <[email protected]>
 Radosław Gryta <[email protected]>
 Ravindra Marella <[email protected]>
@@ -109,15 +162,20 @@ Reinforce-II <[email protected]>
 Reza Rezvan <[email protected]>
 Rick G <[email protected]>
 RiverZhou <[email protected]>
+Ronsor <[email protected]>
+Rotem Dan <[email protected]>
 Ryan Hitchman <[email protected]>
+Salvatore Mesoraca <[email protected]>
 Sam Spilsbury <[email protected]>
 Sanchit Gandhi <[email protected]>
 Santtu Keskinen <[email protected]>
 Sergio López <[email protected]>
 Shijie <[email protected]>
 Siddharth Ramakrishnan <[email protected]>
+Sigbjørn Skjæret <[email protected]>
 Skyler Celestinian-Sterling <[email protected]>
 Slava Primenko <[email protected]>
+Srihari-mcw <[email protected]>
 Steward Garcia <[email protected]>
 Supreet Sethi <[email protected]>
 Takuya Takeuchi <[email protected]>
@@ -127,38 +185,61 @@ Tanmay Sachan <[email protected]>
 Timothy Cronin <[email protected]>
 Tom Bailey <[email protected]>
 Tom Jobbins <[email protected]>
+Tony Wasserka <[email protected]>
 Tyé singwa <[email protected]>
 UEXTM.com <[email protected]>
 WillCorticesAI <[email protected]>
+William Tambellini <[email protected]>
+William Tambellini <[email protected]>
 XiaotaoChen <[email protected]>
+Xuan Son Nguyen <[email protected]>
 Yavor Ivanov <[email protected]>
 YavorGIvanov <[email protected]>
+Yilong Guo <[email protected]>
+Yilong Guo <[email protected]>
+agray3 <[email protected]>
 apcameron <[email protected]>
 appvoid <[email protected]>
 ariez-xyz <[email protected]>
 automaticcat <[email protected]>
+bandoti <[email protected]>
 bmwl <[email protected]>
 bobqianic <[email protected]>
 bssrdf <[email protected]>
 chengchi <[email protected]>
 compilade <[email protected]>
+compilade <[email protected]>
 ddpasa <[email protected]>
 denersc <[email protected]>
 dscripka <[email protected]>
 fitzsim <[email protected]>
+fraxy-v <[email protected]>
 goerch <[email protected]>
+goldwaving <[email protected]>
 hidenorly <[email protected]>
+hipudding <[email protected]>
 hydai <[email protected]>
 jaeminSon <[email protected]>
+jdomke <[email protected]>
+jiez <[email protected]>
 johnson442 <[email protected]>
+junchao-loongson <[email protected]>
+k.h.lai <[email protected]>
 katsu560 <[email protected]>
 klosax <[email protected]>
+kunnis <[email protected]>
+l3utterfly <[email protected]>
 le.chang <[email protected]>
 leejet <[email protected]>
 leejet <[email protected]>
+liuwei-git <[email protected]>
+luoyu-intel <[email protected]>
 magicse <[email protected]>
+mashizora <[email protected]>
+matteo <[email protected]>
 ochafik <[email protected]>
 otaGran <[email protected]>
+pengxin99 <[email protected]>
 pikalover6 <[email protected]>
 postmasters <[email protected]>
 sjinzh <[email protected]>
@@ -168,8 +249,15 @@ snadampal <[email protected]>
 taher <[email protected]>
 texmex76 <[email protected]>
 the-crypt-keeper <[email protected]>
+thewh1teagle <[email protected]>
+ucag.li <[email protected]>
 ulatekh <[email protected]>
+wangshuai09 <[email protected]>
+woachk <[email protected]>
 yangyaofei <[email protected]>
+yuri@FreeBSD <yuri@FreeBSD>
+zhentaoyu <[email protected]>
 zhouwg <[email protected]>
+zhouwg <[email protected]>
 布客飞龙 <[email protected]>
 旺旺碎冰冰 <[email protected]>

src/ggml-vulkan.cpp

@@ -5008,6 +5008,8 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
         }
     }
 
+    ggml_pipeline_allocate_descriptor_sets(ctx->device);
+
     vk_buffer d_X = ggml_vk_create_buffer_check(ctx->device, sizeof(X_TYPE) * x_ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
     vk_buffer d_Y = ggml_vk_create_buffer_check(ctx->device, sizeof(Y_TYPE) * y_ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
     vk_buffer d_D = ggml_vk_create_buffer_check(ctx->device, sizeof(float) * d_ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
@@ -5124,7 +5126,9 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
 
     avg_err /= m * n;
 
-    std::cerr << "TEST " << shname << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time / num_it << "ms avg_err=" << avg_err << std::endl;
+    double tflops = 2.0*m*n*k*batch*num_it / (time / 1000.0) / (1000.0*1000.0*1000.0*1000.0);
+
+    std::cerr << "TEST " << shname << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time / num_it << "ms " << tflops << " TFLOPS avg_err=" << avg_err << std::endl;
 
     if (avg_err > 0.1) {
         std::cerr << "m = " << first_err_m << " n = " << first_err_n << " b = " << first_err_b << std::endl;
@@ -5246,12 +5250,14 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_
 
     ggml_pipeline_request_descriptor_sets(ctx->device, p, 1);
 
+    ggml_pipeline_allocate_descriptor_sets(ctx->device);
+
     ggml_vk_buffer_write(qx_buf, 0, qx, qx_sz);
 
     vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
     ggml_vk_ctx_begin(ctx->device, subctx);
     const std::vector<uint32_t> pc = { 1, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne };
-    ggml_vk_dispatch_pipeline(ctx, subctx, p, { { qx_buf, 0, qx_sz }, { x_buf, 0, x_sz_f16 } }, pc.size() * sizeof(int), pc.data(), { (uint32_t)ne, 1, 1});
+    ggml_vk_dispatch_pipeline(ctx, subctx, p, { vk_subbuffer{ qx_buf, 0, qx_sz }, vk_subbuffer{ x_buf, 0, x_sz_f16 } }, pc.size() * sizeof(int), pc.data(), { (uint32_t)ne, 1, 1});
     ggml_vk_ctx_end(subctx);
 
     auto begin = std::chrono::high_resolution_clock::now();
@@ -5378,6 +5384,8 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
         }
     }
 
+    ggml_pipeline_allocate_descriptor_sets(ctx->device);
+
     ggml_vk_buffer_write(qx_buf, 0, qx, qx_sz);
     ggml_vk_buffer_write(y_buf, 0, y, y_sz);
 
@@ -5445,7 +5453,9 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
 
     avg_err /= m * n;
 
-    std::cerr << "TEST MMQ " << shname << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time_ms / num_it << "ms avg_err=" << avg_err << std::endl;
+    double tflops = 2.0*m*n*k*batch*num_it / (time_ms / 1000.0) / (1000.0*1000.0*1000.0*1000.0);
+
+    std::cerr << "TEST MMQ " << shname << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time_ms / num_it << "ms " << tflops << " TFLOPS avg_err=" << avg_err << std::endl;
 
     if (avg_err > 0.01 || std::isnan(avg_err)) {
         std::cerr << "m = " << first_err_m << " n = " << first_err_n << " b = " << first_err_b << std::endl;
@@ -5497,9 +5507,6 @@ static ggml_tensor_extra_gpu * ggml_vk_tensor_create_extra(ggml_tensor * tensor)
 
 static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
 #if defined(GGML_VULKAN_RUN_TESTS)
-    ctx->staging = ggml_vk_create_buffer_check(ctx->device, 100ul * 1024ul * 1024ul,
-        vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
-        vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
     ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_F32);
     ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_0);
     ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_1);