ROCm 6: bring back (#2019)

**Summary**

Bring back ROCm 6.0. I have traced the crash in Blender (that caused our
previous revert of ROCm 6) to a compiler bug and found a workaround to
the issue.

Notably, we're the first distro to introduce built-in compatibility with
non-officially supported architectures. Previously, to run an
unofficially supported GPU (e.g. my `gfx1032`), I must manually export
the environment variable `HSA_OVERRIDE_GFX_VERSION=10.3.0` to emulate my
GPU as `gfx1030`, an officially supported model. With these set of
patches, which are either directly borrowed from or inspired by @cgmb's
similar patches for Debian (thank you!), `HSA_OVERRIDE_GFX_VERSION` is
not needed for the following GPUS:

- `gfx902`, `gfx909`, `gfx90c` -> masked as `gfx900`
- `gfx101*` (all RDNA1 GPUs) -> masked as `gfx1010`
- `gfx103*` (all RDNA2 GPUs) -> masked as `gfx1030`

This should cover the majority of uses cases of ROCm, therefore to some
extent making us the distro with the best out-of-the-box ROCm
experience.

I've also PGO'ed AMD's fork of LLVM that's used to compile HIP code. It
gives somewhere around a 13% performance improvement.

**Test Plan**

- Ran Fooocus (that uses `pytorch`) and generated a few images.
- Rendered all sorts of Blender scenes that uses cycles (including the
ones that have caused the same crash on Blender bug reports)
- Ran `llama.cpp` with Gemma-2b-instruct.

**Changes**

- **Reapply "rocm-device-libs: update to 6.0.0"**
- **Reapply "rocm-cmake: update to 6.0.0"**
- **rocm-cmake: update to 6.0.0**
- **rocm-device-libs: update to 6.0.0**
- **hsakmt: update to 6.0.0**
- **rocm-runtime: update to 6.0.0**
- **rocm-info: update to 6.0.0**
- **rocm-smi: update to 6.0.0**
- **rocm-compilersupport: update to 6.0.0**
- **rocm-clr: bring back after ROCm 6 revert**
- **rocm-llvm: PGO**
- **rocprim: update to 3.0.0**
- **rocsparse: update to 3.0.2**
- **rocfft: update to 1.0.25**
- **rocsolver: update to 3.24.0**
- **rocblas: update to 4.0.0**
- **hipblas: update to 2.0.0**
- **rocrand: update to v2.0.17 (ROCm 6.0.0)**
- **hiprand: update to v2.0.17 (ROCm 6.0.0)**
- **rocmlir: update to 6.0.0**
- **rocm-core: update to 6.0.0**
- **rocm-hipify: update to 6.0.0**
- **rccl: update to 2.15.5 (ROCm 6.0.0)**
- **half: adjust header file location**
- **hsa-amd-aqlprofile: update to 6.0.0**
- **rocprofiler: update to 6.0.0**
- **roctracer: update to 6.0.0**
- **composable-kernel: bring back and build for more dtypes**
- **miopen: update to 3.00.0 (ROCm 6)**
- **hipsparse: update to 3.0.0 (ROCm 6.0)**
- **hipsolver: update to 2.0.0 (ROCm 6)**
- **hipfft: update to 1.0.13 (ROCm 6.0.0)**
- **hipcub: update to 3.0.0 (ROCm 6.0)**
- **frugally-deep: bring back**
- **functional-plus: bring back**
- **pytorch: update to v2.2.0 and ROCm 6.0**
- **python-torchvision: update to 0.17.0**
- **python-torchvision: update to v2.2.0 and ROCm 6.0**
- **python-torchtext: update to v0.17.0**
- **blender: bring back ROCm/HIP support**

**Checklist**

- [x] Package was built and tested against unstable

packages/b/blender/abi_used_symbols

@@ -1505,13 +1505,11 @@ libjack.so.0:jack_transport_locate
 libjack.so.0:jack_transport_query
 libjack.so.0:jack_transport_start
 libjack.so.0:jack_transport_stop
-libjemalloc.so.2:_ZdaPv
-libjemalloc.so.2:_ZdlPvSt11align_val_t
-libjemalloc.so.2:_ZnwmRKSt9nothrow_t
+libjemalloc.so.2:_ZdlPv
 libjemalloc.so.2:calloc
+libjemalloc.so.2:free
 libjemalloc.so.2:malloc_usable_size
 libjemalloc.so.2:memalign
-libjemalloc.so.2:realloc
 libjpeg.so.8:jpeg_CreateCompress
 libjpeg.so.8:jpeg_CreateDecompress
 libjpeg.so.8:jpeg_abort_decompress
@@ -1831,8 +1829,8 @@ libosdGPU.so.3.6.0:_ZN10OpenSubdiv6v3_6_03Osd14GLVertexBufferD1Ev
 libosdGPU.so.3.6.0:_ZN10OpenSubdiv6v3_6_03Osd21GLSLPatchShaderSource25GetPatchBasisShaderSourceB5cxx11Ev
 libosdGPU.so.3.6.0:_ZNK10OpenSubdiv6v3_6_03Osd14GLVertexBuffer14GetNumVerticesEv
 liboslcomp.so.1.12:_ZN9OSL_v1_1211OSLCompiler7compileEN16OpenImageIO_v2_517basic_string_viewIcSt11char_traitsIcEEERKSt6vectorINSt7__cxx1112basic_stringIcS4_SaIcEEESaISA_EES5_
-liboslcomp.so.1.12:_ZN9OSL_v1_1211OSLCompilerC1EPN16OpenImageIO_v2_512ErrorHandlerE
 liboslcomp.so.1.12:_ZN9OSL_v1_1211OSLCompilerD1Ev
+liboslexec.so.1.12:_ZN9OSL_v1_1211OSLCompilerC1EPN16OpenImageIO_v2_512ErrorHandlerE
 liboslexec.so.1.12:_ZN9OSL_v1_1213ShadingSystem11get_contextEPNS_13PerThreadInfoEPN16OpenImageIO_v2_513TextureSystem9PerthreadE
 liboslexec.so.1.12:_ZN9OSL_v1_1213ShadingSystem14ConnectShadersEN16OpenImageIO_v2_517basic_string_viewIcSt11char_traitsIcEEES5_S5_S5_
 liboslexec.so.1.12:_ZN9OSL_v1_1213ShadingSystem14ShaderGroupEndEv
@@ -2474,9 +2472,12 @@ libstdc++.so.6:_ZTv0_n24_NSoD0Ev
 libstdc++.so.6:_ZTv0_n24_NSoD1Ev
 libstdc++.so.6:_ZTv0_n24_NSt13basic_fstreamIcSt11char_traitsIcEED0Ev
 libstdc++.so.6:_ZTv0_n24_NSt13basic_fstreamIcSt11char_traitsIcEED1Ev
+libstdc++.so.6:_ZdaPv
 libstdc++.so.6:_ZdlPv
+libstdc++.so.6:_ZdlPvSt11align_val_t
 libstdc++.so.6:_Znam
 libstdc++.so.6:_Znwm
+libstdc++.so.6:_ZnwmRKSt9nothrow_t
 libstdc++.so.6:__cxa_allocate_exception
 libstdc++.so.6:__cxa_bad_cast
 libstdc++.so.6:__cxa_bad_typeid

packages/b/blender/package.yml

@@ -1,6 +1,6 @@
 name       : blender
 version    : 4.1.0
-release    : 83
+release    : 84
 source     :
     - https://download.blender.org/source/blender-4.1.0.tar.xz : dc002d82e3c340c93855c6684730d01801b669a29b1cc6b75eeb990ba69e723f
     - https://download.blender.org/demo/test/BMW27.blend.zip : 08170de260488c14855a13db81c8d5ad844bc6162c3db13e1d9ba8094f55fe2b
@@ -52,14 +52,17 @@ builddeps  :
     - opensubdiv-devel
     - openvdb-devel
     - potrace-devel
-    - rocm-hip-devel
+    - rocm-hip
 rundeps    :
     - opencollada
-    - rocm-runtime
+    - rocm-clr
 clang      : yes
 environment: |
     export ROCM_PATH=/usr
     export HIP_PATH=/usr
+    # ROCm's LLVM fork has a bug that causes memory access
+    # fault when rendering the classroom example with cycles.
+    export HIP_CLANG_PATH=/usr/bin
     export DEVICE_LIB_PATH=/usr/lib64/amdgcn/bitcode
 setup      : |
     %cmake_ninja \
@@ -91,7 +94,7 @@ setup      : |
         -DWITH_SYSTEM_LZO=ON \
         -DWITH_USD=ON \
         -DWITH_X11_XINPUT=ON \
-        -DCYCLES_HIP_BINARIES_ARCH="%AMDGPUTARGETS%;gfx90c;gfx902;gfx1011;gfx1012;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1103"
+        -DCYCLES_HIP_BINARIES_ARCH="gfx803;gfx900;gfx902;gfx904;gfx906;gfx908;gfx90a;gfx90c;gfx1010;gfx1011;gfx1012;gfx1013;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1103"
 build      : |
     %ninja_build
 install    : |

packages/b/blender/pspec_x86_64.xml

@@ -730,6 +730,7 @@
             <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/lib/kernel_gfx1010.fatbin</Path>
             <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/lib/kernel_gfx1011.fatbin</Path>
             <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/lib/kernel_gfx1012.fatbin</Path>
+            <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/lib/kernel_gfx1013.fatbin</Path>
             <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/lib/kernel_gfx1030.fatbin</Path>
             <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/lib/kernel_gfx1031.fatbin</Path>
             <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/lib/kernel_gfx1032.fatbin</Path>
@@ -744,10 +745,10 @@
             <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/lib/kernel_gfx803.fatbin</Path>
             <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/lib/kernel_gfx900.fatbin</Path>
             <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/lib/kernel_gfx902.fatbin</Path>
-            <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/lib/kernel_gfx906:xnack-.fatbin</Path>
-            <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/lib/kernel_gfx908:xnack-.fatbin</Path>
-            <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/lib/kernel_gfx90a:xnack+.fatbin</Path>
-            <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/lib/kernel_gfx90a:xnack-.fatbin</Path>
+            <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/lib/kernel_gfx904.fatbin</Path>
+            <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/lib/kernel_gfx906.fatbin</Path>
+            <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/lib/kernel_gfx908.fatbin</Path>
+            <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/lib/kernel_gfx90a.fatbin</Path>
             <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/lib/kernel_gfx90c.fatbin</Path>
             <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/license/Apache2-license.txt</Path>
             <Path fileType="data">/usr/share/blender/4.1/scripts/addons/cycles/license/BSD-3-Clause-license.txt</Path>
@@ -3749,7 +3750,7 @@
         </Files>
     </Package>
     <History>
-        <Update release="83">
+        <Update release="84">
             <Date>2024-04-01</Date>
             <Version>4.1.0</Version>
             <Comment>Packaging update</Comment>

packages/c/composable-kernel/abi_libs

@@ -0,0 +1,6 @@
+libck_device_contraction_operations.so.1
+libck_device_conv_operations.so.1
+libck_device_gemm_operations.so.1
+libck_device_other_operations.so.1
+libck_device_reduction_operations.so.1
+libck_utility.so.1