added some of the performance query functionality

driver/common.h

@@ -487,6 +487,21 @@ typedef struct VkDescriptorPool_T
 	uint32_t freeAble;
 } _descriptorPool;
 
+typedef struct VkQuery_T
+{
+	uint32_t enabledCounters[VC4_PERFCNT_NUM_EVENTS];
+	uint64_t counterValues[DRM_VC4_MAX_PERF_COUNTERS];
+	uint32_t numEnabledCounters;
+	uint32_t perfmonIDs[2];
+} _query;
+
+typedef struct VkQueryPool_T
+{
+	VkQueryType type;
+	uint32_t queryCount;
+	_query* queryPool;
+} _queryPool;
+
 uint32_t getFormatBpp(VkFormat f);
 uint32_t packVec4IntoABGR8(const float rgba[4]);
 void createImageBO(_image* i);

driver/device.c

@@ -159,6 +159,53 @@ VKAPI_ATTR void VKAPI_CALL rpi_vkGetPhysicalDeviceQueueFamilyProperties(
 	*pQueueFamilyPropertyCount = elementsWritten;
 }
 
+VKAPI_ATTR VkResult VKAPI_CALL vkEnumeratePhysicalDeviceQueueFamilyPerformanceQueryCountersKHR(
+	VkPhysicalDevice                            physicalDevice,
+	uint32_t                                    queueFamilyIndex,
+	uint32_t*                                   pCounterCount,
+	VkPerformanceCounterKHR*                    pCounters,
+	VkPerformanceCounterDescriptionKHR*         pCounterDescriptions)
+{
+	assert(physicalDevice);
+	assert(pCounterCount);
+
+	if(!pCounters && !pCounterDescriptions)
+	{
+		*pCounterCount = numPerformanceCounterTypes;
+		return VK_SUCCESS;
+	}
+
+	int arraySize = *pCounterCount;
+	int elementsWritten = min(numPerformanceCounterTypes, arraySize);
+
+	for(int c = 0; c < elementsWritten; ++c)
+	{
+		pCounters[c] = performanceCounterTypes[c];
+		pCounterDescriptions[c] = performanceCounterDescriptions[c];
+	}
+
+	*pCounterCount = elementsWritten;
+
+	if(arraySize < numPerformanceCounterTypes)
+	{
+		return VK_INCOMPLETE;
+	}
+
+	return VK_SUCCESS;
+}
+
+VKAPI_ATTR void VKAPI_CALL vkGetPhysicalDeviceQueueFamilyPerformanceQueryPassesKHR(
+	VkPhysicalDevice                            physicalDevice,
+	const VkQueryPoolPerformanceCreateInfoKHR*  pPerformanceQueryCreateInfo,
+	uint32_t*                                   pNumPasses)
+{
+	assert(physicalDevice);
+	assert(pPerformanceQueryCreateInfo);
+	assert(pNumPasses);
+
+	*pNumPasses = pPerformanceQueryCreateInfo->counterIndexCount / DRM_VC4_MAX_PERF_COUNTERS;
+}
+
 /*
  * https://www.khronos.org/registry/vulkan/specs/1.1-extensions/html/vkspec.html#vkCreateDevice
  * vkCreateDevice verifies that extensions and features requested in the ppEnabledExtensionNames and pEnabledFeatures

driver/query.c

@@ -1,17 +1,60 @@
 #include "common.h"
 
+//TODO VkPerformanceQuerySubmitInfoKHR
+
+VKAPI_ATTR VkResult VKAPI_CALL vkAcquireProfilingLockKHR(
+	VkDevice                                    device,
+	const VkAcquireProfilingLockInfoKHR*        pInfo)
+{
+	//TODO
+	return VK_SUCCESS;
+}
+
+VKAPI_ATTR void VKAPI_CALL vkReleaseProfilingLockKHR(
+	VkDevice                                    device)
+{
+	//TODO
+}
+
 VKAPI_ATTR VkResult VKAPI_CALL rpi_vkCreateQueryPool(
 	VkDevice                                    device,
 	const VkQueryPoolCreateInfo*                pCreateInfo,
 	const VkAllocationCallbacks*                pAllocator,
 	VkQueryPool*                                pQueryPool)
 {
-	//TODO
+	assert(device);
+	assert(pQueryPool);
 
-	if(pCreateInfo->queryType == VK_QUERY_TYPE_OCCLUSION)
+	_queryPool* qp = ALLOCATE(sizeof(_queryPool), 1, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+
+	if(pCreateInfo->queryType == VK_QUERY_TYPE_OCCLUSION ||
+	   pCreateInfo->queryType == VK_QUERY_TYPE_PIPELINE_STATISTICS ||
+	   pCreateInfo->queryType == VK_QUERY_TYPE_TIMESTAMP)
 	{
 		UNSUPPORTED(VK_QUERY_TYPE_OCCLUSION);
 	}
+	else if(pCreateInfo->queryType == VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR)
+	{
+		qp->queryCount = pCreateInfo->queryCount;
+		qp->type = pCreateInfo->queryType;
+		qp->queryPool = ALLOCATE(sizeof(_query) * qp->queryCount, 1, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+
+		VkQueryPoolPerformanceCreateInfoKHR ci = *(VkQueryPoolPerformanceCreateInfoKHR*)pCreateInfo->pNext;
+
+		for(uint32_t c = 0; c < qp->queryCount; ++c)
+		{
+			assert(ci.counterIndexCount <= VC4_PERFCNT_NUM_EVENTS);
+			qp->queryPool[c].numEnabledCounters = ci.counterIndexCount;
+			memcpy(qp->queryPool[c].enabledCounters, ci.pCounterIndices, sizeof(uint32_t) * ci.counterIndexCount);
+
+			for(uint32_t d = 0; d < ci.counterIndexCount; d += DRM_VC4_MAX_PERF_COUNTERS)
+			{
+				qp->queryPool[c].perfmonIDs[d] = vc4_create_perfmon(controlFd, &qp->queryPool[c].enabledCounters[d], qp->queryPool[c].numEnabledCounters > DRM_VC4_MAX_PERF_COUNTERS ? DRM_VC4_MAX_PERF_COUNTERS : qp->queryPool[c].numEnabledCounters);
+			}
+		}
+
+		*pQueryPool = qp;
+	}
 
 	return VK_SUCCESS;
 }
@@ -30,7 +73,20 @@ VKAPI_ATTR void VKAPI_CALL rpi_vkDestroyQueryPool(
 	VkQueryPool                                 queryPool,
 	const VkAllocationCallbacks*                pAllocator)
 {
-	//TODO
+	assert(device);
+	assert(queryPool);
+
+	_queryPool* qp = queryPool;
+
+	for(uint32_t c = 0; c < qp->queryCount; ++c)
+	{
+		for(uint32_t d = 0; d < qp->queryPool[c].enabledCounters; d += DRM_VC4_MAX_PERF_COUNTERS)
+		{
+			vc4_destroy_perfmon(controlFd, qp->queryPool[c].perfmonIDs[d]);
+		}
+	}
+
+	FREE(qp->queryPool);
 }
 
 VKAPI_ATTR void VKAPI_CALL rpi_vkCmdEndQuery(
@@ -38,6 +94,9 @@ VKAPI_ATTR void VKAPI_CALL rpi_vkCmdEndQuery(
 	VkQueryPool                                 queryPool,
 	uint32_t                                    query)
 {
+	assert(commandBuffer);
+	assert(queryPool);
+
 	//TODO
 }
 
@@ -47,6 +106,9 @@ VKAPI_ATTR void VKAPI_CALL rpi_vkCmdBeginQuery(
 	uint32_t                                    query,
 	VkQueryControlFlags                         flags)
 {
+	assert(commandBuffer);
+	assert(queryPool);
+
 	//TODO
 }
 
@@ -73,7 +135,23 @@ VKAPI_ATTR VkResult VKAPI_CALL rpi_vkGetQueryPoolResults(
 	VkDeviceSize                                stride,
 	VkQueryResultFlags                          flags)
 {
-	//TODO
+	assert(device);
+	assert(queryPool);
+
+	//TODO flags
+
+	_queryPool* qp = queryPool;
+
+	for(uint32_t c = firstQuery; c < queryCount; ++c)
+	{
+		uint32_t counter = 0;
+		for(uint32_t d = 0; d < dataSize; d += stride, ++counter)
+		{
+			VkPerformanceCounterResultKHR* result = ((char*)pData) + d;
+			result->uint64 = qp->queryPool[c].counterValues[counter];
+		}
+	}
+
 	return VK_SUCCESS;
 }
 

driver/vkCaps.h

@@ -264,6 +264,10 @@ static VkExtensionProperties deviceExtensions[] =
 	{
 		.extensionName = "VK_KHR_driver_properties",
 		.specVersion = 1
+	},
+	{
+		.extensionName = "VK_KHR_performance_query",
+		.specVersion = 1
 	}
 };
 #define numDeviceExtensions (sizeof(deviceExtensions) / sizeof(VkExtensionProperties))
@@ -305,4 +309,254 @@ static VkFormat supportedFormats[] =
 };
 #define numSupportedFormats (sizeof(supportedFormats)/sizeof(VkFormat))
 
+static VkPerformanceCounterKHR performanceCounterTypes[] =
+{ //TODO UUID
+	{
+		.unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+		.scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+		.storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_CYCLES_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_CYCLES_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_CYCLES_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_CYCLES_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_CYCLES_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_CYCLES_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_CYCLES_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_CYCLES_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_CYCLES_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	},
+	{
+	  .unit = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
+	  .scope = VK_PERFORMANCE_COUNTER_SCOPE_RENDER_PASS_KHR,
+	  .storage = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
+	}
+};
+
+static VkPerformanceCounterDescriptionKHR performanceCounterDescriptions[] =
+{
+	{
+		.name = "FRONT_END_PIPELINE_VALID_PRIMS_NO_RENDER",
+	},
+	{
+		.name = "FRONT_END_PIPELINE_VALID_PRIMS_RENDER",
+	},
+	{
+		.name = "FRONT_END_PIPELINE_CLIPPED_QUADS",
+	},
+	{
+		.name = "FRONT_END_PIPELINE_VALID_QUADS",
+	},
+	{
+		.name = "TILE_BUFFER_QUADS_NOT_PASSING_STENCIL",
+	},
+	{
+		.name = "TILE_BUFFER_QUADS_NOT_PASSING_Z_AND_STENCIL",
+	},
+	{
+		.name = "TILE_BUFFER_QUADS_PASSING_Z_AND_STENCIL",
+	},
+	{
+		.name = "TILE_BUFFER_QUADS_ZERO_COVERAGE",
+	},
+	{
+		.name = "TILE_BUFFER_QUADS_NON_ZERO_COVERAGE",
+	},
+	{
+		.name = "TILE_BUFFER_QUADS_WRITTEN_TO_COLOR_BUF",
+	},
+	{
+		.name = "PLB_PRIMS_OUTSIDE_VIEWPORT",
+	},
+	{
+		.name = "PLB_PRIMS_NEED_CLIPPING",
+	},
+	{
+		.name = "PRIMITIVE_SETUP_ENGINE_PRIMS_REVERSED",
+	},
+	{
+		.name = "QUAD_PROCESSOR_UNIT_TOTAL_IDLE_CYCLES",
+	},
+	{
+		.name = "QUAD_PROCESSOR_UNIT_TOTAL_CLK_CYCLES_VERTEX_COORD_SHADING",
+	},
+	{
+		.name = "QUAD_PROCESSOR_UNIT_TOTAL_CLK_CYCLES_FRAGMENT_SHADING",
+	},
+	{
+		.name = "QUAD_PROCESSOR_UNIT_TOTAL_CLK_CYCLES_EXEC_VALID_INST",
+	},
+	{
+		.name = "QUAD_PROCESSOR_UNIT_TOTAL_CLK_CYCLES_WAITING_TMUS",
+	},
+	{
+		.name = "QUAD_PROCESSOR_UNIT_TOTAL_CLK_CYCLES_WAITING_SCOREBOARD",
+	},
+	{
+		.name = "QUAD_PROCESSOR_UNIT_TOTAL_CLK_CYCLES_WAITING_VARYINGS",
+	},
+	{
+		.name = "QUAD_PROCESSOR_UNIT_TOTAL_INST_CACHE_HIT",
+	},
+	{
+		.name = "QUAD_PROCESSOR_UNIT_TOTAL_INST_CACHE_MISS",
+	},
+	{
+		.name = "QUAD_PROCESSOR_UNIT_TOTAL_UNIFORM_CACHE_HIT",
+	},
+	{
+		.name = "QUAD_PROCESSOR_UNIT_TOTAL_UNIFORM_CACHE_MISS",
+	},
+	{
+		.name = "TEXTURE_MEMORY_LOOKUP_UNIT_TOTAL_TEXT_QUADS_PROCESSED",
+	},
+	{
+		.name = "TEXTURE_MEMORY_LOOKUP_UNIT_TOTAL_TEXT_CACHE_MISS",
+	},
+	{
+		.name = "VERTEX_PIPE_MEMORY_TOTAL_CLK_CYCLES_VERTEX_DMA_WRITE_STALLED",
+	},
+	{
+		.name = "VERTEX_PIPE_MEMORY_TOTAL_CLK_CYCLES_VERTEX_DMA_STALLED",
+	},
+	{
+		.name = "L2C_TOTAL_L2_CACHE_HIT",
+	},
+	{
+		.name = "L2C_TOTAL_L2_CACHE_MISS",
+	}
+};
+
+#define numPerformanceCounterTypes (sizeof(performanceCounterTypes)/sizeof(uint32_t))
+
 #define VK_DRIVER_VERSION VK_MAKE_VERSION(1, 1, 0)