Exercise non zero base offsets within the same physical device

Components: Vulkan

VK-GL-CTS issue: 110

Affects: dEQP-VK.compute.device_group.dispatch_base

Change-Id: I205812017cb62070fd273521ae83704e051b2801
(cherry picked from commit 45e18a75add921a401b1e7ffbc85b3a51e943b7e)

diff --git a/external/vulkancts/modules/vulkan/compute/vktComputeBasicComputeShaderTests.cpp b/external/vulkancts/modules/vulkan/compute/vktComputeBasicComputeShaderTests.cpp
index 5b25575dd5e4f993f5624c387c7babc122ae3fb7..72cc3351a744920b56e5c25ae45cab8f2d5f0623 100644 (file)
--- a/external/vulkancts/modules/vulkan/compute/vktComputeBasicComputeShaderTests.cpp
+++ b/external/vulkancts/modules/vulkan/compute/vktComputeBasicComputeShaderTests.cpp
@@ -2491,6 +2491,7 @@ public:
                                                                                                                                const tcu::IVec3&       worksize,
                                                                                                                                const tcu::IVec3&       splitsize);
 
+       bool                                                    isInputVectorValid                      (const tcu::IVec3& small, const tcu::IVec3& big);
        tcu::TestStatus                                 iterate                                         (void);
 
 private:
@@ -2558,11 +2559,36 @@ DispatchBaseTestInstance::DispatchBaseTestInstance (Context& context,
        , m_workSize                    (worksize)
        , m_splitWorkSize               (splitsize)
 {
-       if (m_splitWorkSize.x() > m_workSize.x() ||
-               m_splitWorkSize.y() > m_workSize.y() ||
-               m_splitWorkSize.z() > m_workSize.z() ||
-               (multiplyComponents(m_splitWorkSize) >= multiplyComponents(m_workSize)))
-               TCU_THROW(TestError, "Split work group size too big.");
+       // For easy work distribution across physical devices:
+       // WorkSize should be a multiple of SplitWorkSize only in the X component
+       if ((!isInputVectorValid(m_splitWorkSize, m_workSize)) ||
+               (m_workSize.x() <= m_splitWorkSize.x()) ||
+               (m_workSize.y() != m_splitWorkSize.y()) ||
+               (m_workSize.z() != m_splitWorkSize.z()))
+               TCU_THROW(TestError, "Invalid Input.");
+
+       // For easy work distribution within the same physical device:
+       // SplitWorkSize should be a multiple of localSize in Y or Z component
+       if ((!isInputVectorValid(m_localSize, m_splitWorkSize)) ||
+               (m_localSize.x() != m_splitWorkSize.x()) ||
+               (m_localSize.y() >= m_splitWorkSize.y()) ||
+               (m_localSize.z() >= m_splitWorkSize.z()))
+               TCU_THROW(TestError, "Invalid Input.");
+
+       if ((multiplyComponents(m_workSize) / multiplyComponents(m_splitWorkSize)) < (deInt32) m_numPhysDevices)
+               TCU_THROW(TestError, "Not enough work to distribute across all physical devices.");
+
+       deUint32 totalWork = multiplyComponents(m_workSize) * multiplyComponents(m_localSize);
+       if ((totalWork > numValues) || (numValues % totalWork != 0))
+               TCU_THROW(TestError, "Buffer too small/not aligned to cover all values.");
+}
+
+bool DispatchBaseTestInstance::isInputVectorValid(const tcu::IVec3& small, const tcu::IVec3& big)
+{
+       if (((big.x() < small.x()) || (big.y() < small.y()) || (big.z() < small.z())) ||
+               ((big.x() % small.x() != 0) || (big.y() % small.y() != 0) || (big.z() % small.z() != 0)))
+               return false;
+       return true;
 }
 
 tcu::TestStatus DispatchBaseTestInstance::iterate (void)
@@ -2650,19 +2676,30 @@ tcu::TestStatus DispatchBaseTestInstance::iterate (void)
 
        vk.cmdPipelineBarrier(*cmdBuffer, VK_PIPELINE_STAGE_HOST_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, (VkDependencyFlags)0, 0, (const VkMemoryBarrier*)DE_NULL, 1, &hostWriteBarrier, 0, (const VkImageMemoryBarrier*)DE_NULL);
 
-       // Split the workload across all physical devices based on m_splitWorkSize
+       // Split the workload across all physical devices based on m_splitWorkSize.x()
        for (deUint32 physDevIdx = 0; physDevIdx < m_numPhysDevices; physDevIdx++)
        {
-               deUint32 baseGroupX = m_splitWorkSize.x() == m_workSize.x() ? 0 : physDevIdx * m_splitWorkSize.x();
-               deUint32 baseGroupY = m_splitWorkSize.y() == m_workSize.y() ? 0 : physDevIdx * m_splitWorkSize.y();
-               deUint32 baseGroupZ = m_splitWorkSize.z() == m_workSize.z() ? 0 : physDevIdx * m_splitWorkSize.z();
+               deUint32 baseGroupX = physDevIdx * m_splitWorkSize.x();
+               deUint32 baseGroupY = 0;
+               deUint32 baseGroupZ = 0;
 
-               deUint32 groupCountX = ((physDevIdx == (m_numPhysDevices - 1)) && (m_splitWorkSize.x() != m_workSize.x())) ? m_workSize.x() - baseGroupX : m_splitWorkSize.x();
-               deUint32 groupCountY = ((physDevIdx == (m_numPhysDevices - 1)) && (m_splitWorkSize.y() != m_workSize.y())) ? m_workSize.y() - baseGroupY : m_splitWorkSize.y();
-               deUint32 groupCountZ = ((physDevIdx == (m_numPhysDevices - 1)) && (m_splitWorkSize.z() != m_workSize.z())) ? m_workSize.z() - baseGroupZ : m_splitWorkSize.z();
+               // Split the workload within the physical device based on m_localSize.y() and m_localSize.z()
+               for (deInt32 localIdxY = 0; localIdxY < (m_splitWorkSize.y() / m_localSize.y()); localIdxY++)
+               {
+                       for (deInt32 localIdxZ = 0; localIdxZ < (m_splitWorkSize.z() / m_localSize.z()); localIdxZ++)
+                       {
+                               deUint32 offsetX = baseGroupX;
+                               deUint32 offsetY = baseGroupY + localIdxY * m_localSize.y();
+                               deUint32 offsetZ = baseGroupZ + localIdxZ * m_localSize.z();
+
+                               deUint32 localSizeX = (physDevIdx == (m_numPhysDevices - 1)) ? m_workSize.x() - baseGroupX : m_localSize.x();
+                               deUint32 localSizeY = m_localSize.y();
+                               deUint32 localSizeZ = m_localSize.z();
 
-               totalWorkloadSize += (groupCountX * groupCountY * groupCountZ);
-               vk.cmdDispatchBase(*cmdBuffer, baseGroupX, baseGroupY, baseGroupZ, groupCountX, groupCountY, groupCountZ);
+                               totalWorkloadSize += (localSizeX * localSizeY * localSizeZ);
+                               vk.cmdDispatchBase(*cmdBuffer, offsetX, offsetY, offsetZ, localSizeX, localSizeY, localSizeZ);
+                       }
+               }
        }
 
        vk.cmdPipelineBarrier(*cmdBuffer, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_HOST_BIT, (VkDependencyFlags)0, 0, (const VkMemoryBarrier*)DE_NULL, 1, &shaderWriteBarrier, 0, (const VkImageMemoryBarrier*)DE_NULL);
@@ -3074,8 +3111,8 @@ tcu::TestCaseGroup* createBasicDeviceGroupComputeShaderTests (tcu::TestContext&
 {
        de::MovePtr<tcu::TestCaseGroup> deviceGroupComputeTests(new tcu::TestCaseGroup(testCtx, "device_group", "Basic device group compute tests"));
 
-       deviceGroupComputeTests->addChild(new DispatchBaseTest(testCtx, "dispatch_base",        "Compute shader with base groups", 4096,        tcu::IVec3(4,1,2),      tcu::IVec3(32,4,2), tcu::IVec3(4,4,2)));
-       deviceGroupComputeTests->addChild(new DeviceIndexTest(testCtx,  "device_index",         "Compute shader using deviceIndex in SPIRV", 96,        tcu::IVec3(3,2,1),      tcu::IVec3(2,4,1)));
+       deviceGroupComputeTests->addChild(new DispatchBaseTest(testCtx, "dispatch_base",        "Compute shader with base groups",                              32768,  tcu::IVec3(4,2,4),      tcu::IVec3(16,8,8),     tcu::IVec3(4,8,8)));
+       deviceGroupComputeTests->addChild(new DeviceIndexTest(testCtx,  "device_index",         "Compute shader using deviceIndex in SPIRV",    96,             tcu::IVec3(3,2,1),      tcu::IVec3(2,4,1)));
 
        return deviceGroupComputeTests.release();