minsukim-amd · Aug 7, 2026 · Aug 7, 2026 · Aug 7, 2026 · Aug 7, 2026 · Aug 7, 2026
@@ -24,13 +24,20 @@
 - changed-files:
   - any-glob-to-any-file: 'dnn-providers/hip-kernel-provider/**/*'
 
+# These labels help distinguish which subteams are responsible for reviews for changes made in rocKE
+
 "rocke-lib: attention":
 - changed-files:
   - any-glob-to-any-file: 'dnn-providers/hip-kernel-provider/rocke/library/**/*attention*'
   - any-glob-to-any-file: 'dnn-providers/hip-kernel-provider/rocke/library/**/*fmha*'
   - any-glob-to-any-file: 'dnn-providers/hip-kernel-provider/rocke/library/**/*sdpa*'
   - any-glob-to-any-file: 'dnn-providers/hip-kernel-provider/rocke/library/builders/**/attention/**/*'
 
+# This label serves to distiguish work being done in RocKE platform, for the reviewing team and internal tools
+"rocke: platform":
+- changed-files:
+  - any-glob-to-any-file: 'dnn-providers/hip-kernel-provider/rocke/platform/**/*'
+
 "project: hipdnn-integration-tests":
 - changed-files:
   - any-glob-to-any-file: 'dnn-providers/integration-tests/**/*'

@@ -139,13 +139,38 @@ if [[ ! -x "${TENSILELITE_CLIENT}" ]]; then
   exit 1
 fi
 
-# The fetched compiler artifact installs the host OpenMP runtime here.
-LLVM_HOST_RUNTIME_DIR="${ROCM_PATH}/lib/llvm/lib/x86_64-unknown-linux-gnu"
-if [[ ! -f "${LLVM_HOST_RUNTIME_DIR}/libomp.so" ]]; then
-  echo "OpenMP runtime not found: ${LLVM_HOST_RUNTIME_DIR}/libomp.so" >&2
+# The compiler artifact layout depends on whether TheRock enables LLVM's host
+# per-target runtime directories. Prefer the flat lib/llvm/lib layout, then the
+# matching host-triple directory when per-target runtime directories are used.
+LLVM_RUNTIME_ROOT="${ROCM_PATH}/lib/llvm/lib"
+if [[ ! -d "${LLVM_RUNTIME_ROOT}" ]]; then
+  echo "LLVM runtime directory not found: ${LLVM_RUNTIME_ROOT}" >&2
   exit 1
 fi
 
+LIBOMP_CANDIDATES=(
+  "${LLVM_RUNTIME_ROOT}/libomp.so"
+  "${LLVM_RUNTIME_ROOT}/$(uname -m)-unknown-linux-gnu/libomp.so"
+)
+LIBOMP_PATH=""
+for candidate in "${LIBOMP_CANDIDATES[@]}"; do
+  if [[ -e "${candidate}" ]]; then
+    LIBOMP_PATH="${candidate}"
+    break
+  fi
+done
+
+if [[ -z "${LIBOMP_PATH}" ]]; then
+  echo "OpenMP runtime not found at either expected path:" >&2
+  printf "  %s\n" "${LIBOMP_CANDIDATES[@]}" >&2
+  exit 1
+fi
+LLVM_HOST_RUNTIME_DIR="$(dirname "${LIBOMP_PATH}")"
+LLVM_RUNTIME_LIBRARY_PATH="${LLVM_RUNTIME_ROOT}"
+if [[ "${LLVM_HOST_RUNTIME_DIR}" != "${LLVM_RUNTIME_ROOT}" ]]; then
+  LLVM_RUNTIME_LIBRARY_PATH="${LLVM_RUNTIME_LIBRARY_PATH}:${LLVM_HOST_RUNTIME_DIR}"
+fi
+
 select_rocjitsu_target
 
 if [[ -z "${ROCJITSU_CONFIG}" || ! -f "${ROCJITSU_CONFIG}" ]]; then
@@ -161,7 +186,7 @@ mkdir -p "${RACE_REPORT_DIR}"
 # that use it to find the ROCm install root.
 export ROCM_PATH
 export PATH="${ROCM_PATH}/bin:${ROCM_PATH}/lib/llvm/bin:${PATH}"
-export LD_LIBRARY_PATH="${ROCM_PATH}/lib:${ROCM_PATH}/lib/rocm_sysdeps/lib:${ROCM_PATH}/lib/llvm/lib:${LLVM_HOST_RUNTIME_DIR}:${LD_LIBRARY_PATH:-}"
+export LD_LIBRARY_PATH="${ROCM_PATH}/lib:${ROCM_PATH}/lib/rocm_sysdeps/lib:${LLVM_RUNTIME_LIBRARY_PATH}:${LD_LIBRARY_PATH:-}"
 export PYTHONPATH="${TENSILELITE_ROOT}${PYTHONPATH:+:${PYTHONPATH}}"
 
 echo "ROCM_PATH=${ROCM_PATH}"
@@ -174,6 +199,7 @@ echo "HIPBLASLT_BENCH=${HIPBLASLT_BENCH}"
 echo "TENSILELITE_ROOT=${TENSILELITE_ROOT}"
 echo "TENSILE_DRIVER=${TENSILE_DRIVER}"
 echo "TENSILELITE_CLIENT=${TENSILELITE_CLIENT}"
+echo "LIBOMP_PATH=${LIBOMP_PATH}"
 echo "RACE_REPORT_DIR=${RACE_REPORT_DIR}"
 echo "PATH=${PATH}"
 echo "LD_LIBRARY_PATH=${LD_LIBRARY_PATH}"

@@ -120,6 +120,7 @@ jobs:
             -DROCM_LIBS_ENABLE_ROOT_CTEST=ON \
             -DENABLE_CLANG_FORMAT=OFF \
             -DHIPKERNELPROVIDER_ENABLE_ROCKE=ON \
+            -DHIPDNN_ENABLE_CUDNN_COMPATIBILITY=ON \
             -DENABLE_CLANG_TIDY=OFF
 
       - name: Build
@@ -280,6 +281,7 @@ jobs:
             -DENABLE_CLANG_FORMAT=OFF \
             -DENABLE_CLANG_TIDY=ON \
             -DHIPKERNELPROVIDER_ENABLE_ROCKE=ON \
+            -DHIPDNN_ENABLE_CUDNN_COMPATIBILITY=ON \
             -DCMAKE_CXX_CLANG_TIDY=/usr/bin/clang-tidy-20
 
       - name: Build

@@ -19,6 +19,7 @@ on:
 
 permissions:
   contents: read
+  actions: read
 
 concurrency:
   # For scheduled runs, cancel any in-progress nightly so only one runs at a time.

@@ -36,6 +36,7 @@ on:
 
 permissions:
   contents: read
+  actions: read
 
 concurrency:
   # Only allow one run at a time for 'pull_request' events.

@@ -30,6 +30,7 @@ on:
 
 permissions:
   contents: read
+  actions: read
 
 concurrency:
   # A PR number if a pull request and otherwise the commit hash. This cancels

@@ -64,6 +64,7 @@ on:
 
 permissions:
   contents: read
+  actions: read
 
 concurrency:
   # A PR number if a pull request and otherwise the commit hash. This cancels
@@ -148,16 +149,17 @@ jobs:
       windows_amdgpu_families: ${{ inputs.windows_amdgpu_families || 'gfx110X' }}
       linux_test_labels: ${{ inputs.linux_test_labels || '' }}
       windows_test_labels: ${{ inputs.windows_test_labels || '' }}
-      # HARDCODED: Copy all stages except math-libs from TheRock baseline
-      # rocm-libraries only affects math-libs and ml-libs source sets
-      # Note: compiler-runtime excluded due to hardcoded CMake paths (fftw3, etc.)
-      # TODO(geomin12): Enable automatic stage reuse once dry-run mode issues resolved
-      prebuilt_stages: ${{ inputs.prebuilt_stages || 'runtime-tests,comm-libs,storage-libs,debug-tools,dctools-core,profiler-apps,media-libs,wsl-rocdxg' }}
-      baseline_run_id: ${{ inputs.baseline_run_id || '30989762651' }}
+      # Manual override for prebuilt stages (empty = use automatic stage reuse)
+      prebuilt_stages: ${{ inputs.prebuilt_stages || '' }}
+      baseline_run_id: ${{ inputs.baseline_run_id || '' }}
+      # Enable automatic stage reuse - TheRock finds commit-compatible baselines
+      # and determines which stages to rebuild based on changed files
+      stage_reuse_mode: reuse-stage
       repository: ROCm/TheRock
       # Concrete SHA resolved once by resolve_therock_ref; setup re-exports the
       # checked-out commit as needs.setup.outputs.ref for downstream jobs.
       ref: ${{ needs.resolve_therock_ref.outputs.therock_ref }}
+      # Pass external repo info for stage-impact analysis
       external_repo: '{"repository":"${{ github.repository }}","ref":"${{ github.sha }}"}'
 
   linux_build_and_test:

@@ -47,6 +47,7 @@ list(APPEND CMAKE_MODULE_PATH ${CMAKE_CURRENT_SOURCE_DIR}/cmake
 
 option(HIPKERNELPROVIDER_ENABLE_TESTS "Build tests" ON)
 option(HIPKERNELPROVIDER_ENABLE_ROCKE "Build rocKE engine and smoke tests" OFF)
+option(HIPDNN_ENABLE_KERNEL_INGESTOR "Enable the generic kernel ingestor" OFF)
 option(ENABLE_ASM_SDPA_ENGINE "Build the ASM SDPA engine" ON)
 option(ENABLE_HIP_MLOPS_ENGINE "Build the HIP kernel engine" ON)
 
@@ -194,6 +195,28 @@ if(HIPKERNELPROVIDER_ENABLE_ROCKE)
     set(ROCKE_INSTALL_PROVIDER_CTEST_SUBDIR "hip_kernel_provider")
     set(ROCKE_INSTALL_STANDALONE_CTEST OFF)
 
+    # TEMPORARY: do not install the rocKE Python CI tests (rocke_python_import_smoke,
+    # rocke_golden_static, rocke_pytest) in provider mode.
+    #
+    # These three CTest entries import the `rocke` package from the test artifact.
+    # ROCm/TheRock#6849 removed the "bin/hip_kernel_provider/rocke/**" artifact
+    # include on the assumption that the staged rocKE wheels replaced it, but the
+    # wheels are never staged (see below), so neither import source survives and
+    # every rocKE Python entry fails with `ModuleNotFoundError: No module named
+    # 'rocke'`. rocm-libraries builds this component with
+    # HIPKERNELPROVIDER_ENABLE_ROCKE=ON, so the breakage lands on every
+    # hipkernelprovider PR here while TheRock's own CI (flag defaults OFF) stays green.
+    #
+    # This matches what the ROCKE_INSTALL_PYTHON_TESTS option already documents for
+    # provider mode; nothing was actually setting it. The C++ rocKE smoke tests are
+    # unaffected and keep running.
+    #
+    # Revert this line once ROCm/TheRock#7145 has landed and the wheel-staging order
+    # bug in rocke/CMakeLists.txt is fixed (ROCKE_BUILD_PYENV / ROCKE_WHEEL_DIR are
+    # declared after add_subdirectory(platform), so the wheel install() never
+    # registers). Tracked by ROCm/rocm-libraries#10497.
+    set(ROCKE_INSTALL_PYTHON_TESTS OFF)
+
     add_subdirectory(rocke)
 else()
     message(STATUS "rocKE features disabled")

@@ -177,6 +177,32 @@ int rocke_conv_problem_k_gemm(const rocke_conv_problem_t* p);
  * int never hits. */
 long long rocke_conv_problem_flops(const rocke_conv_problem_t* p);
 
+/* ConvProblem.is_pointwise property:
+ *   Y == 1 and X == 1 and sH == 1 and sW == 1 and pH == 0 and pW == 0
+ *   and (not is_3d or (Z == 1 and sD == 1 and pD == 0))
+ * For pointwise convolutions A, B, D are flat 2-D matrices and the
+ * coordinate-transform DAG reduces to multiply+add. */
+static inline bool rocke_conv_problem_is_pointwise(const rocke_conv_problem_t* p)
+{
+    if(p->Y != 1 || p->X != 1 || p->sH != 1 || p->sW != 1 || p->pH != 0 || p->pW != 0)
+        return false;
+    if(p->is_3d && (p->Z != 1 || p->sD != 1 || p->pD != 0))
+        return false;
+    return true;
+}
+
+/* ConvProblem.cpg property: C / groups (channels per group). */
+static inline int rocke_conv_problem_cpg(const rocke_conv_problem_t* p)
+{
+    return p->groups > 1 ? p->C / p->groups : p->C;
+}
+
+/* ConvProblem.kpg property: K / groups (output channels per group). */
+static inline int rocke_conv_problem_kpg(const rocke_conv_problem_t* p)
+{
+    return p->groups > 1 ? p->K / p->groups : p->K;
+}
+
 /* ConvProblem.short() ->
  *   2-D: f"N{N}H{Hi}W{Wi}C{C}_K{K}Y{Y}X{X}"
  *   3-D: f"N{N}D{Di}H{Hi}W{Wi}C{C}_K{K}Z{Z}Y{Y}X{X}"

@@ -225,13 +225,44 @@ typedef struct rocke_conv_build_ctx
     int load_vec; /* _choose_load_vec(spec)    */
 
     /* ---- coordinate-transform descriptors ---- */
-    rocke_tensor_descriptor_t* A_desc; /* make_a_descriptor(p, decompose_m)     */
-    rocke_tensor_descriptor_t* B_desc; /* make_b_descriptor(p)                  */
+    rocke_tensor_descriptor_t* A_desc; /* make_a_descriptor(p, decompose_m); NULL if pointwise */
+    rocke_tensor_descriptor_t* B_desc; /* make_b_descriptor(p);              NULL if pointwise */
     /* D_desc is built lazily inside the epilogue phase (Python builds it per
      * epilogue fn); held here so an epilogue_override and the two stock
      * epilogues share one instance. NULL until the epilogue phase populates. */
     rocke_tensor_descriptor_t* D_desc;
 
+    /* ---- pointwise fast-path (Python p.is_pointwise) ---- *
+     * When true, A_desc / B_desc / D_desc are NULL and flat multiply+add
+     * arithmetic is used instead of the coordinate-transform DAG.
+     * Mirrors: if p.is_pointwise: A_desc = None; B_desc = None ...
+     *
+     * Forward conv: c_M_pw=p.M, c_C_pw=p.cpg, c_K_pw=p.kpg.
+     * Wgrad:        c_M_pw=wg_M, c_C_pw=p.cpg, c_K_pw=p.kpg,
+     *               c_wgK_pw=wg_K, c_wgN_pw=wg_N. */
+    bool is_pointwise;
+    int c_M_pw; /* forward: p.M;   wgrad: wg_M */
+    int c_C_pw; /* forward: p.cpg; wgrad: p.cpg (input channels per group) */
+    int c_K_pw; /* forward: p.kpg; wgrad: p.kpg (output channels per group) */
+    int c_wgK_pw; /* wgrad only: wg_K (= N*Ho*Wo, reduction dim); 0 for forward */
+    int c_wgN_pw; /* wgrad only: wg_N (= Y*X*C, N dim); 0 for forward */
+
+    /* ---- pointwise IR constants (emitted before buffer resources, matching Python) ---- *
+     * Python emits _c_C_ir, _c_K_ir, _c_M_ir, _always_valid (in that order)
+     * before make_buffer_resource calls. We pre-emit and cache them so the
+     * descriptor callbacks reuse rather than re-emit them. */
+    rocke_value_t* ir_c_C_pw; /* const_i32(cpg)    — first  (fwd) / _c_K_ir=kpg (wgrad) */
+    rocke_value_t* ir_c_K_pw; /* const_i32(kpg)    — second (fwd) / _c_C_ir=cpg (wgrad) */
+    rocke_value_t* ir_c_M_pw; /* const_i32(M/wg_M) — third  (fwd) / _c_wgM_ir  (wgrad) */
+    rocke_value_t* ir_always_valid; /* const_i32(1)  — fourth (fwd) / _c_wgN_ir  (wgrad) */
+    rocke_value_t* ir_c_wgN_pw; /* wgrad only: const_i32(wg_N) — fifth */
+
+    /* ---- optional B-descriptor override (wgrad only) ---- *
+     * When non-NULL, rocke_conv_emit_load_phase calls this instead of
+     * rocke_conv_b_descriptor. Used by wgrad to provide a different pointwise
+     * formula for the X (input) B-tile descriptor. */
+    rocke_loads_descriptor_fn b_descriptor_fn;
+
     /* ---- buffer resources (CK-Tile views over A/B/D) ---- */
     rocke_conv_buffer_resource_t a_buf_rsrc; /* make_buffer_resource(b, A, A_bytes) */
     rocke_conv_buffer_resource_t b_buf_rsrc; /* make_buffer_resource(b, Bp, B_bytes)*/
@@ -436,11 +467,11 @@ void rocke_conv_emit_direct_epilogue(rocke_ir_builder_t* b,
                                      rocke_value_t* const* accs,
                                      int num_accs,
                                      const rocke_warp_grid_t* grid,
-                                     rocke_value_t* d_rsrc);
+                                     rocke_value_t* d_rsrc,
+                                     rocke_value_t* ir_c_K_pw);
 
 /* _emit_direct_epilogue_wmma(b, spec, op, accs[n], warp_m_idx, warp_n_idx, lane,
- * block_m_off, block_n_off, d_rsrc, c0): WMMA per-lane fp16 store via the op's
- * c_layout map + the D descriptor. */
+ * block_m_off, block_n_off, d_rsrc, c0, ir_c_K_pw): WMMA per-lane fp16 store. */
 void rocke_conv_emit_direct_epilogue_wmma(rocke_ir_builder_t* b,
                                           const rocke_implicit_gemm_conv_spec_t* spec,
                                           const rocke_mmaop_t* op,
@@ -452,16 +483,18 @@ void rocke_conv_emit_direct_epilogue_wmma(rocke_ir_builder_t* b,
                                           rocke_value_t* block_m_off,
                                           rocke_value_t* block_n_off,
                                           rocke_value_t* d_rsrc,
-                                          rocke_value_t* c0);
+                                          rocke_value_t* c0,
+                                          rocke_value_t* ir_c_K_pw);
 
-/* _emit_cshuffle_epilogue(b, spec, accs[n], grid, d_rsrc): LDS-staged cshuffle
- * store via CShuffleEpilogue.from_grid + the D descriptor addr_fn. */
+/* _emit_cshuffle_epilogue(b, spec, accs[n], grid, d_rsrc, ir_c_K_pw): LDS-staged
+ * cshuffle store via CShuffleEpilogue.from_grid + the D descriptor addr_fn. */
 void rocke_conv_emit_cshuffle_epilogue(rocke_ir_builder_t* b,
                                        const rocke_implicit_gemm_conv_spec_t* spec,
                                        rocke_value_t* const* accs,
                                        int num_accs,
                                        const rocke_warp_grid_t* grid,
-                                       rocke_value_t* d_rsrc);
+                                       rocke_value_t* d_rsrc,
+                                       rocke_value_t* ir_c_K_pw);
 
 /* ----- driver-internal ctx population (the build prologue, lines 787-1032) ----
  * Splitting the long prologue out of the public entry keeps the glue TU small;