[runtime][HIP] Retire ROCm HAL backend (iree-org#17029)

This patch retires the ROCm HAL backend given the new HIP HAL backend is
functionally complete (iree-org#15790).
The new backend has been tested and proved to give on par or better
performance compared to the ROCm backend.

The documentation still preserves the `ROCm` term to denote the AMD
compute software while using `HIP` specifically for the runtime.

CMakeLists.txt

@@ -580,22 +580,6 @@ include(iree_plugin_register)
 set(IREE_PACKAGE_ROOT_DIR "${CMAKE_CURRENT_LIST_DIR}")
 set(IREE_PACKAGE_ROOT_PREFIX "iree")
 
-#-------------------------------------------------------------------------------
-# Experimental ROCM HAL driver
-# Enable with: -DIREE_EXTERNAL_HAL_DRIVERS=rocm
-#-------------------------------------------------------------------------------
-
-iree_register_external_hal_driver(
-  NAME
-    rocm
-  SOURCE_DIR
-    "${CMAKE_CURRENT_SOURCE_DIR}/experimental/rocm"
-  DRIVER_TARGET
-    iree::experimental::rocm::registration
-  REGISTER_FN
-    iree_hal_rocm_driver_module_register
-)
-
 #-------------------------------------------------------------------------------
 # Experimental WebGPU HAL driver
 # Enable with: -DIREE_EXTERNAL_HAL_DRIVERS=webgpu

build_tools/cmake/test_riscv.sh

@@ -36,7 +36,6 @@ declare -a label_exclude_args=(
   "^driver=vulkan$"
   "^driver=metal$"
   "^driver=cuda$"
-  "^driver=rocm$"
   "^driver=hip$"
   "^vulkan_uses_vk_khr_shader_float16_int8$"
   "^requires-filesystem$"

build_tools/pkgci/build_linux_packages.sh

@@ -188,7 +188,7 @@ function build_wheel() {
 function build_iree_runtime() {
   # We install the needed build deps below for the tools.
   IREE_RUNTIME_BUILD_TRACY=ON IREE_RUNTIME_BUILD_TRACY_TOOLS=ON \
-  IREE_EXTERNAL_HAL_DRIVERS="rocm" \
+  IREE_HAL_DRIVER_HIP=ON \
   build_wheel runtime/
 }
 

compiler/plugins/target/ROCM/builtins/ukernel/test/CMakeLists.txt

@@ -16,6 +16,6 @@ iree_lit_test_suite(
     FileCheck
     iree-compile
   LABELS
-    "driver=rocm"
+    "driver=hip"
     "hostonly"
 )

docs/website/docs/guides/deployment-configurations/gpu-rocm.md

@@ -45,14 +45,14 @@ the IREE compiler, then enable the ROCm compiler target with the
 
 ### Get the IREE runtime
 
-Next you will need to get an IREE runtime that includes the ROCm HAL driver.
+Next you will need to get an IREE runtime that includes the HIP HAL driver.
 
 #### :material-hammer-wrench: Build the runtime from source
 
 Please make sure you have followed the
 [Getting started](../../building-from-source/getting-started.md) page to build
-IREE from source, then enable the experimental ROCm HAL driver with the
-`IREE_EXTERNAL_HAL_DRIVERS=rocm` option.
+IREE from source, then enable the HIP HAL driver with the `IREE_HAL_DRIVER_HIP`
+option.
 
 ## Compile and run a program model
 
@@ -108,7 +108,7 @@ Run the following command:
 
 ``` shell hl_lines="2"
 iree-run-module \
-    --device=rocm \
+    --device=hip \
     --module=mobilenet_rocm.vmfb \
     --function=predict \
     --input="1x224x224x3xf32=0"

experimental/regression_suite/tests/pregenerated/test_llama2.py

@@ -223,15 +223,15 @@ def test_step_a100_vulkan_stripped(llama2_7b_f16qi4_a100_vulkan_vmfb):
 def test_step_rdna3_rocm_stripped(llama2_7b_f16qi4_stripped_rdna3_rocm_vmfb):
     iree_benchmark_module(
         llama2_7b_f16qi4_stripped_rdna3_rocm_vmfb,
-        device="rocm",
+        device="hip",
         function="first_vicuna_forward",
         args=[
             "--input=1x1xi64",
         ],
     )
     iree_benchmark_module(
         llama2_7b_f16qi4_stripped_rdna3_rocm_vmfb,
-        device="rocm",
+        device="hip",
         function="second_vicuna_forward",
         args=[
             "--input=1x1xi64",

experimental/regression_suite/tests/pregenerated/test_ukernel.py

@@ -156,7 +156,7 @@ def test_correctness_gfx90a_rocm(
 ):
     iree_run_module(
         argmax_ukernel_gfx90a_rocm_vmfb,
-        device="rocm",
+        device="hip",
         function="argmax_3d_dyn_f16i32",
         args=[
             f"--input=@{argmax_input_f16.path}",
@@ -165,7 +165,7 @@ def test_correctness_gfx90a_rocm(
     )
     iree_run_module(
         argmax_ukernel_gfx90a_rocm_vmfb,
-        device="rocm",
+        device="hip",
         function="argmax_3d_dyn_f16i64",
         args=[
             f"--input=@{argmax_input_f16.path}",
@@ -175,7 +175,7 @@ def test_correctness_gfx90a_rocm(
 
     iree_run_module(
         argmax_ukernel_gfx90a_rocm_vmfb,
-        device="rocm",
+        device="hip",
         function="argmax_3d_dyn_f32i32",
         args=[
             f"--input=@{argmax_input_f32.path}",
@@ -184,7 +184,7 @@ def test_correctness_gfx90a_rocm(
     )
     iree_run_module(
         argmax_ukernel_gfx90a_rocm_vmfb,
-        device="rocm",
+        device="hip",
         function="argmax_3d_dyn_f32i64",
         args=[
             f"--input=@{argmax_input_f32.path}",
@@ -205,7 +205,7 @@ def test_correctness_gfx940_rocm(
 ):
     iree_run_module(
         argmax_ukernel_gfx940_rocm_vmfb,
-        device="rocm",
+        device="hip",
         function="argmax_3d_dyn_f16i32",
         args=[
             f"--input=@{argmax_input_f16.path}",
@@ -214,7 +214,7 @@ def test_correctness_gfx940_rocm(
     )
     iree_run_module(
         argmax_ukernel_gfx940_rocm_vmfb,
-        device="rocm",
+        device="hip",
         function="argmax_3d_dyn_f16i64",
         args=[
             f"--input=@{argmax_input_f16.path}",
@@ -224,7 +224,7 @@ def test_correctness_gfx940_rocm(
 
     iree_run_module(
         argmax_ukernel_gfx940_rocm_vmfb,
-        device="rocm",
+        device="hip",
         function="argmax_3d_dyn_f32i32",
         args=[
             f"--input=@{argmax_input_f32.path}",
@@ -233,7 +233,7 @@ def test_correctness_gfx940_rocm(
     )
     iree_run_module(
         argmax_ukernel_gfx940_rocm_vmfb,
-        device="rocm",
+        device="hip",
         function="argmax_3d_dyn_f32i64",
         args=[
             f"--input=@{argmax_input_f32.path}",