P4 to Git Change 1312566 by lmoriche@lmoriche_opencl_dev on 2016/09/08 18:25:02

SWDEV-94610 - Make sure each kernarg segment sits on a different cache line (align the kernargs on cache lines at minimum). Minor misc cleanups.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.cpp#13 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rockernel.cpp#14 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rockernel.hpp#8 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocprogram.cpp#27 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#13 edit

rocclr/runtime/device/rocm/rocdevice.cpp

@@ -817,8 +817,18 @@ Device::populateOCLDeviceConstants()
     info_.maxSamplers_ = 16;
     info_.bufferFromImageSupport_ = CL_FALSE;
     info_.oclcVersion_ = "OpenCL C " OPENCL_VERSION_STR " ";
-    strcpy(info_.driverVersion_, "1.0 Provisional (hsa)");
-    info_.version_ = "OpenCL " OPENCL_VERSION_STR " ";
+
+    uint16_t major, minor;
+    if (hsa_agent_get_info(_bkendDevice, HSA_AGENT_INFO_VERSION_MAJOR, &major)
+            != HSA_STATUS_SUCCESS
+     || hsa_agent_get_info(_bkendDevice, HSA_AGENT_INFO_VERSION_MINOR, &minor)
+            != HSA_STATUS_SUCCESS) {
+        return false;
+    }
+    std::stringstream ss;
+    ss << major << "." << minor << " (hsa)";
+    strcpy(info_.driverVersion_, ss.str().c_str());
+    info_.version_ = "OpenCL " /*OPENCL_VERSION_STR*/"1.2" " ";
 
     info_.builtInKernels_ = "";
     info_.linkerAvailable_ = true;

rocclr/runtime/device/rocm/rockernel.cpp

@@ -689,16 +689,14 @@ Kernel::Kernel(
                const uint32_t workgroupGroupSegmentByteSize,
                const uint32_t workitemPrivateSegmentByteSize,
                const uint32_t kernargSegmentByteSize,
-               const uint32_t kernargSegmentAlignment,
-               uint extraArgsNum)
+               const uint32_t kernargSegmentAlignment)
     : device::Kernel(name),
       program_(prog),
       kernelCodeHandle_(kernelCodeHandle),
       workgroupGroupSegmentByteSize_(workgroupGroupSegmentByteSize),
       workitemPrivateSegmentByteSize_(workitemPrivateSegmentByteSize),
       kernargSegmentByteSize_(kernargSegmentByteSize),
-      kernargSegmentAlignment_(kernargSegmentAlignment),
-      extraArgumentsNum_(extraArgsNum) {}
+      kernargSegmentAlignment_(kernargSegmentAlignment) {}
 
 #if defined(WITH_LIGHTNING_COMPILER)
 bool Kernel::init_LC()

rocclr/runtime/device/rocm/rockernel.hpp

@@ -96,8 +96,7 @@ public:
         const uint32_t workgroupGroupSegmentByteSize,
         const uint32_t workitemPrivateSegmentByteSize,
         const uint32_t kernargSegmentByteSize,
-        const uint32_t kernargSegmentAlignment,
-        uint extraArgsNum);
+        const uint32_t kernargSegmentAlignment);
 
     const uint64_t& KernelCodeHandle() {
         return kernelCodeHandle_;
@@ -144,11 +143,6 @@ public:
         return NULL;
     }
 
-    //! Max number of possible extra (hidden) kernel arguments
-    static const uint MaxExtraArgumentsNum = 6;
-
-    uint extraArgumentsNum() const { return extraArgumentsNum_; }
-
     //! Return printf info array
     const std::vector<PrintfInfo>& printfInfo() const {return printf_;}
 
@@ -172,7 +166,6 @@ private:
     const uint32_t kernargSegmentByteSize_;
     const uint32_t kernargSegmentAlignment_;
     size_t kernelDirectiveOffset_;
-    const uint extraArgumentsNum_; // Number of arguments in Kernenv
     std::vector<PrintfInfo> printf_;
 };
 

rocclr/runtime/device/rocm/rocprogram.cpp

@@ -1093,9 +1093,6 @@ HSAILProgram::linkImpl_LC(amd::option::Options *options)
             return false;
         }
 
-        // for OpenCL default hidden kernel arguments assuming there is no printf
-        size_t numHiddenKernelArgs = 3; // FIXME_Wilkin
-
         Kernel *aKernel = new roc::Kernel(
             kernelName,
             this,
@@ -1105,10 +1102,8 @@ HSAILProgram::linkImpl_LC(amd::option::Options *options)
             // TODO: remove the workaround
             //   add 24 bytes for global offsets as workaround for LC reporting
             //   excluded the hidden arguments
-            amd::alignUp(kernargSegmentByteSize, sizeof(size_t)) + numHiddenKernelArgs * sizeof(size_t),
-            kernargSegmentAlignment,
-            numHiddenKernelArgs
-        );
+            amd::alignUp(kernargSegmentByteSize, sizeof(size_t)) + 3 * sizeof(size_t),
+            amd::alignUp(kernargSegmentAlignment,device().info().globalMemCacheLineSize_));
         if (!aKernel->init()) {
             return false;
         }
@@ -1450,11 +1445,30 @@ std::string
 HSAILProgram::hsailOptions(amd::option::Options* options)
 {
     std::string hsailOptions;
+
     //Set options for the standard device specific options
+
+    hsailOptions.append(" -D__AMD__");
+
+    int major, minor;
+    ::sscanf(device().info().version_, "OpenCL %d.%d ", &major, &minor);
+
+    std::stringstream ss;
+    ss << " -D__OPENCL_VERSION__=" << (major * 100 + minor * 10);
+    hsailOptions.append(ss.str());
+
+    if (device().info().imageSupport_ && options->oVariables->ImageSupport) {
+        hsailOptions.append(" -D__IMAGE_SUPPORT__");
+    }
+
     //This is just for legacy compiler code
     // All our devices support these options now
-    hsailOptions.append(" -DFP_FAST_FMAF");
-    hsailOptions.append(" -DFP_FAST_FMA");
+    if (options->oVariables->FastFMA) {
+        hsailOptions.append(" -DFP_FAST_FMA");
+    }
+    if (options->oVariables->FastFMAF) {
+        hsailOptions.append(" -DFP_FAST_FMAF");
+    }
 
     if (dev().deviceInfo().gfxipVersion_ < 900) {
         hsailOptions.append(" -cl-denorms-are-zero");

rocclr/runtime/device/rocm/rocvirtual.cpp

@@ -519,11 +519,6 @@ VirtualGPU::create(bool profilingEna)
     }
 
     device::BlitManager::Setup  blitSetup;
- #if defined(WITH_LIGHTNING_COMPILER)
-    // TODO: Wilkin - remove the setting of value_ after image kernels are available
-    blitSetup.value_ = 0x3724;  // disable the image related BLIT kernels for now
-#endif // defined(WITH_LIGHTNING_COMPILER)
-
     blitMgr_ = new KernelBlitManager(*this, blitSetup);
     if ((NULL == blitMgr_) || !blitMgr_->create(roc_device_)) {
         LogError("Could not create BlitManager!");