Cycles: various fixes for HIP and compilation of HIP binaries

* Additional structs added to the hipew loader for device props
* Adds hipRTC functions to the loader for future usage
* Enables CPU+GPU usage for HIP
* Cleanup to the adaptive kernel compilation process
* Fix for kernel compilation failures with HIP with latest master

Ref T92393, D12958

4 files changed, 44 insertions, 43 deletions

diff --git a/intern/cycles/blender/addon/properties.py b/intern/cycles/blender/addon/properties.py
index 1d8ebe94694..2a51e0be2a4 100644
--- a/intern/cycles/blender/addon/properties.py
+++ b/intern/cycles/blender/addon/properties.py
@@ -1329,7 +1329,7 @@ class CyclesPreferences(bpy.types.AddonPreferences):
             elif entry.type == 'CPU':
                 cpu_devices.append(entry)
         # Extend all GPU devices with CPU.
-        if compute_device_type != 'CPU' and compute_device_type != 'HIP':
+        if compute_device_type != 'CPU':
             devices.extend(cpu_devices)
         return devices
 
diff --git a/intern/cycles/device/hip/device_impl.cpp b/intern/cycles/device/hip/device_impl.cpp
index 583ab8ae208..4ae714913ab 100644
--- a/intern/cycles/device/hip/device_impl.cpp
+++ b/intern/cycles/device/hip/device_impl.cpp
@@ -208,7 +208,7 @@ bool HIPDevice::use_adaptive_compilation()
   return DebugFlags().hip.adaptive_compile;
 }
 
-/* Common NVCC flags which stays the same regardless of shading model,
+/* Common HIPCC flags which stays the same regardless of shading model,
  * kernel sources md5 and only depends on compiler or compilation settings.
  */
 string HIPDevice::compile_kernel_get_common_cflags(const uint kernel_features)
@@ -239,11 +239,13 @@ string HIPDevice::compile_kernel(const uint kernel_features,
   int major, minor;
   hipDeviceGetAttribute(&major, hipDeviceAttributeComputeCapabilityMajor, hipDevId);
   hipDeviceGetAttribute(&minor, hipDeviceAttributeComputeCapabilityMinor, hipDevId);
+  hipDeviceProp_t props;
+  hipGetDeviceProperties(&props, hipDevId);
 
   /* Attempt to use kernel provided with Blender. */
   if (!use_adaptive_compilation()) {
     if (!force_ptx) {
-      const string fatbin = path_get(string_printf("lib/%s_sm_%d%d.cubin", name, major, minor));
+      const string fatbin = path_get(string_printf("lib/%s_%s.fatbin", name, props.gcnArchName));
       VLOG(1) << "Testing for pre-compiled kernel " << fatbin << ".";
       if (path_exists(fatbin)) {
         VLOG(1) << "Using precompiled kernel.";
@@ -283,17 +285,21 @@ string HIPDevice::compile_kernel(const uint kernel_features,
   const string kernel_md5 = util_md5_string(source_md5 + common_cflags);
 
   const char *const kernel_ext = "genco";
+  std::string options;
 #  ifdef _WIN32
-  const char *const options =
-      "save-temps -Wno-parentheses-equality -Wno-unused-value --hipcc-func-supp";
+  options.append("Wno-parentheses-equality -Wno-unused-value --hipcc-func-supp -ffast-math");
 #  else
-  const char *const options =
-      "save-temps -Wno-parentheses-equality -Wno-unused-value --hipcc-func-supp -O3 -ggdb";
+  options.append("Wno-parentheses-equality -Wno-unused-value --hipcc-func-supp -O3 -ffast-math");
 #  endif
+#  ifdef _DEBUG
+  options.append(" -save-temps");
+#  endif
+  options.append(" --amdgpu-target=").append(props.gcnArchName);
+
   const string include_path = source_path;
-  const char *const kernel_arch = force_ptx ? "compute" : "sm";
+  const char *const kernel_arch = props.gcnArchName;
   const string fatbin_file = string_printf(
-      "cycles_%s_%s_%d%d_%s", name, kernel_arch, major, minor, kernel_md5.c_str());
+      "cycles_%s_%s_%s", name, kernel_arch, kernel_md5.c_str());
   const string fatbin = path_cache_get(path_join("kernels", fatbin_file));
   VLOG(1) << "Testing for locally compiled kernel " << fatbin << ".";
   if (path_exists(fatbin)) {
@@ -350,7 +356,7 @@ string HIPDevice::compile_kernel(const uint kernel_features,
 
   string command = string_printf("%s -%s -I %s --%s %s -o \"%s\"",
                                  hipcc,
-                                 options,
+                                 options.c_str(),
                                  include_path.c_str(),
                                  kernel_ext,
                                  source_path.c_str(),
diff --git a/intern/cycles/kernel/CMakeLists.txt b/intern/cycles/kernel/CMakeLists.txt
index 7357c5804ed..6c87c9c32f2 100644
--- a/intern/cycles/kernel/CMakeLists.txt
+++ b/intern/cycles/kernel/CMakeLists.txt
@@ -487,9 +487,6 @@ endif()
 # HIP module
 
 if(WITH_CYCLES_HIP_BINARIES AND WITH_CYCLES_DEVICE_HIP)
-  # 64 bit only
-  set(HIP_BITS 64)
-
   # build for each arch
   set(hip_sources device/hip/kernel.cpp
     ${SRC_HEADERS}
@@ -504,32 +501,41 @@ if(WITH_CYCLES_HIP_BINARIES AND WITH_CYCLES_DEVICE_HIP)
   set(hip_fatbins)
 
   macro(CYCLES_HIP_KERNEL_ADD arch prev_arch name flags sources experimental)
-    if(${arch} MATCHES "compute_.*")
-      set(format "ptx")
-    else()
-      set(format "fatbin")
-    endif()
+    set(format "fatbin")
     set(hip_file ${name}_${arch}.${format})
-
     set(kernel_sources ${sources})
     if(NOT ${prev_arch} STREQUAL "none")
-      if(${prev_arch} MATCHES "compute_.*")
-        set(kernel_sources ${kernel_sources} ${name}_${prev_arch}.ptx)
-      else()
-        set(kernel_sources ${kernel_sources} ${name}_${prev_arch}.fatbin)
-      endif()
+      set(kernel_sources ${kernel_sources} ${name}_${prev_arch}.fatbin)
     endif()
 
     set(hip_kernel_src "/device/hip/${name}.cpp")
 
-    set(hip_flags ${flags}
+    if(WIN32)
+      set(hip_command ${CMAKE_COMMAND})
+      set(hip_flags
+        -E env "HIP_PATH=${HIP_ROOT_DIR}" "PATH=${HIP_PERL_PATH}"
+        ${HIP_HIPCC_EXECUTABLE}.bat)
+    else()
+      set(hip_command ${HIP_HIPCC_EXECUTABLE})
+      set(hip_flags)
+    endif()
+
+    set(hip_flags
+      ${hip_flags}
+      --amdgpu-target=${arch}
+      ${HIP_HIPCC_FLAGS}
+      --genco
+      ${CMAKE_CURRENT_SOURCE_DIR}${hip_kernel_src}
+      ${flags}
       -D CCL_NAMESPACE_BEGIN=
       -D CCL_NAMESPACE_END=
       -D HIPCC
-      -m ${HIP_BITS}
       -I ${CMAKE_CURRENT_SOURCE_DIR}/..
       -I ${CMAKE_CURRENT_SOURCE_DIR}/device/hip
-      --use_fast_math
+      -Wno-parentheses-equality
+      -Wno-unused-value
+      --hipcc-func-supp
+      -ffast-math
       -o ${CMAKE_CURRENT_BINARY_DIR}/${hip_file})
 
     if(${experimental})
@@ -541,20 +547,9 @@ if(WITH_CYCLES_HIP_BINARIES AND WITH_CYCLES_DEVICE_HIP)
       set(hip_flags ${hip_flags} -D __KERNEL_DEBUG__)
     endif()
 
-    if(WITH_NANOVDB)
-      set(hip_flags ${hip_flags}
-        -D WITH_NANOVDB
-        -I "${NANOVDB_INCLUDE_DIR}")
-    endif()
-
-    add_custom_command(
-      OUTPUT ${hip_file}
-      COMMAND ${HIP_HIPCC_EXECUTABLE}
-          -arch=${arch}
-          ${HIP_HIPCC_FLAGS}
-          --${format}
-          ${CMAKE_CURRENT_SOURCE_DIR}${hip_kernel_src}
-          ${hip_flags}
+    add_custom_target(
+      ${hip_file}
+      COMMAND ${hip_command} ${hip_flags}
       DEPENDS ${kernel_sources})
     delayed_install("${CMAKE_CURRENT_BINARY_DIR}" "${hip_file}" ${CYCLES_INSTALL_PATH}/lib)
     list(APPEND hip_fatbins ${hip_file})
diff --git a/intern/cycles/kernel/device/hip/globals.h b/intern/cycles/kernel/device/hip/globals.h
index 39978ae7899..28e1cc4282f 100644
--- a/intern/cycles/kernel/device/hip/globals.h
+++ b/intern/cycles/kernel/device/hip/globals.h
@@ -27,10 +27,10 @@ CCL_NAMESPACE_BEGIN
 
 /* Not actually used, just a NULL pointer that gets passed everywhere, which we
  * hope gets optimized out by the compiler. */
-struct KernelGlobals {
-  /* NOTE: Keep the size in sync with SHADOW_STACK_MAX_HITS. */
+struct KernelGlobalsGPU {
   int unused[1];
 };
+typedef ccl_global const KernelGlobalsGPU *ccl_restrict KernelGlobals;
 
 /* Global scene data and textures */
 __constant__ KernelData __data;