Merge remote-tracking branch 'origin/develop' into grouped_conv_3d_layout_fix

Dockerfile

 FROM ubuntu:20.04
-
+ARG DEBIAN_FRONTEND=noninteractive
 ARG ROCMVERSION=5.6
 ARG compiler_version=""
 ARG compiler_commit=""
@@ -9,23 +9,30 @@ RUN set -xe
 ARG DEB_ROCM_REPO=http://repo.radeon.com/rocm/apt/.apt_$ROCMVERSION/
 RUN useradd -rm -d /home/jenkins -s /bin/bash -u 1004 jenkins
 # Add rocm repository
+RUN chmod 1777 /tmp
 RUN apt-get update
-RUN apt-get install -y wget gnupg curl
-RUN --mount=type=ssh if [ "$ROCMVERSION" != "5.6"]; then \
+RUN apt-get install -y --allow-unauthenticated apt-utils wget gnupg2 curl
+RUN --mount=type=ssh if [ "$ROCMVERSION" != "5.6" ]; then \
         wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - && \
         sh -c "echo deb [arch=amd64] $DEB_ROCM_REPO ubuntu main > /etc/apt/sources.list.d/rocm.list"; \
-    else sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amd-nonfree-radeon_20.04-1_all.deb" && \
+    elif [ "$ROCMVERSION" = "5.6" ] && [ "$compiler_version" = "" ]; then \
+         sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amd-nonfree-radeon_20.04-1_all.deb" && \
          apt update && apt-get install -y ./amd-nonfree-radeon_20.04-1_all.deb && \
-         amdgpu-repo --amdgpu-build=1567752 --rocm-build=compute-rocm-dkms-no-npi-hipclang/11914 && \
-         DEBIAN_FRONTEND=noninteractive amdgpu-install -y --usecase=rocm ; \
+         amdgpu-repo --amdgpu-build=1567752 --rocm-build=compute-rocm-dkms-no-npi-hipclang/11914; \
+    elif [ "$ROCMVERSION" = "5.6" ] && [ "$compiler_version" = "rc3" ]; then \
+         sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amdgpu-install-internal_5.6-20.04-1_all.deb" && \
+         apt update && apt-get install -y ./amdgpu-install-internal_5.6-20.04-1_all.deb && \
+         sh -c 'echo deb [arch=amd64 trusted=yes] http://compute-artifactory.amd.com/artifactory/list/rocm-release-archive-20.04-deb/ 5.6 rel-45  > /etc/apt/sources.list.d/rocm-build.list' && \
+         amdgpu-repo --amdgpu-build=1602498; \
     fi
+RUN amdgpu-install -y --usecase=rocm --no-dkms
+
 RUN wget --no-check-certificate -qO - https://apt.kitware.com/keys/kitware-archive-latest.asc 2>/dev/null | apt-key add -
 RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list"
 RUN curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor -o /etc/apt/trusted.gpg.d/rocm-keyring.gpg
 
 # Install dependencies
 RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
-    apt-utils \
     build-essential \
     ccache \
     cmake \
@@ -38,16 +45,11 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
     libpthread-stubs0-dev \
     llvm-amdgpu \
     pkg-config \
-    python \
     python3 \
-    python-dev \
     python3-dev \
     python3-pip \
     sshpass \
     software-properties-common \
-    rocm-dev \
-    rocm-device-libs \
-    rocm-cmake \
     vim \
     nano \
     zlib1g-dev \

include/ck/utility/dynamic_buffer.hpp

@@ -19,7 +19,8 @@ namespace ck {
 template <AddressSpaceEnum BufferAddressSpace,
           typename T,
           typename ElementSpaceSize,
-          bool InvalidElementUseNumericalZeroValue>
+          bool InvalidElementUseNumericalZeroValue,
+          AmdBufferCoherenceEnum coherence = AmdBufferCoherenceEnum::DefaultCoherence>
 struct DynamicBuffer
 {
     using type = T;
@@ -77,13 +78,16 @@ struct DynamicBuffer
 
             if constexpr(InvalidElementUseNumericalZeroValue)
             {
-                return amd_buffer_load_invalid_element_return_zero<remove_cvref_t<T>, t_per_x>(
+                return amd_buffer_load_invalid_element_return_zero<remove_cvref_t<T>,
+                                                                   t_per_x,
+                                                                   coherence>(
                     p_data_, i, is_valid_element, element_space_size_);
             }
             else
             {
                 return amd_buffer_load_invalid_element_return_customized_value<remove_cvref_t<T>,
-                                                                               t_per_x>(
+                                                                               t_per_x,
+                                                                               coherence>(
                     p_data_, i, is_valid_element, element_space_size_, invalid_element_value_);
             }
         }
@@ -173,7 +177,7 @@ struct DynamicBuffer
         {
             constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
 
-            amd_buffer_store<remove_cvref_t<T>, t_per_x>(
+            amd_buffer_store<remove_cvref_t<T>, t_per_x, coherence>(
                 x, p_data_, i, is_valid_element, element_space_size_);
         }
         else if constexpr(GetAddressSpace() == AddressSpaceEnum::Lds &&
@@ -376,14 +380,19 @@ struct DynamicBuffer
     __host__ __device__ static constexpr bool IsDynamicBuffer() { return true; }
 };
 
-template <AddressSpaceEnum BufferAddressSpace, typename T, typename ElementSpaceSize>
+template <AddressSpaceEnum BufferAddressSpace,
+          AmdBufferCoherenceEnum coherence = AmdBufferCoherenceEnum::DefaultCoherence,
+          typename T,
+          typename ElementSpaceSize>
 __host__ __device__ constexpr auto make_dynamic_buffer(T* p, ElementSpaceSize element_space_size)
 {
-    return DynamicBuffer<BufferAddressSpace, T, ElementSpaceSize, true>{p, element_space_size};
+    return DynamicBuffer<BufferAddressSpace, T, ElementSpaceSize, true, coherence>{
+        p, element_space_size};
 }
 
 template <
     AddressSpaceEnum BufferAddressSpace,
+    AmdBufferCoherenceEnum coherence = AmdBufferCoherenceEnum::DefaultCoherence,
     typename T,
     typename ElementSpaceSize,
     typename X,
@@ -391,7 +400,7 @@ template <
 __host__ __device__ constexpr auto
 make_dynamic_buffer(T* p, ElementSpaceSize element_space_size, X invalid_element_value)
 {
-    return DynamicBuffer<BufferAddressSpace, T, ElementSpaceSize, false>{
+    return DynamicBuffer<BufferAddressSpace, T, ElementSpaceSize, false, coherence>{
         p, element_space_size, invalid_element_value};
 }