Syncing up from internal repo to enable MI300. (#690)

* enable gfx940 * switch between intrinsic mfma routines on mi100/200 and mi300 * fix mfma_int8 on MI300 * disable 2 int8 examples on MI300 * Update cmake-ck-dev.sh * restore gitignore file * modify Jenkinsfile to the internal repo --------- Co-authored-by: Jing Zhang <jizha@amd.com> Co-authored-by: zjing14 <zhangjing14@gmail.com>
2026-05-03 21:21:22 +00:00 · 2023-04-28 16:22:59 -07:00
parent 54c90aae13
commit 4feebedd41
47 changed files with 228 additions and 72 deletions
--- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp
@@ -66,7 +66,8 @@ __global__ void
            const ReduceGridDescriptor_MBlock_MPerBlock reduce_grid_desc_mblock_mperblock,
            const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];

    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
--- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp
@@ -54,7 +54,8 @@ __global__ void
            const ReduceGridDescriptor_MBlock_MPerBlock reduce_grid_desc_mblock_mperblock,
            const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940))
    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];

    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
--- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp
@@ -44,7 +44,8 @@ __global__ void
                                        c_grid_desc_mblock_mperblock_nblock_nperblock,
                                    const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];

    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
--- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp
@@ -57,7 +57,8 @@ __global__ void
            const C0GridDescriptor_NBlock_NPerBlock c0_grid_desc_nblock_nperblock,
            const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];

    // TODO ANT: separate into MMA + Epilogue
--- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp
@@ -165,7 +165,8 @@ __global__ void
                                      const CElementwiseOperation c_element_op,
                                      const CBlockClusterAdaptor c_block_cluster_adaptor)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];

    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
--- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp
@@ -44,7 +44,8 @@ __global__ void
            const CElementwiseOperation c_element_op,
            const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];

    GridwiseGemm::template Run<HasMainK0BlockLoop>(p_a_grid,
--- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
@@ -43,7 +43,8 @@ __global__ void
            const CElementwiseOperation c_element_op,
            const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];

    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
--- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp
@@ -42,7 +42,8 @@ __global__ void
                                const CElementwiseOperation c_element_op,
                                const CBlockClusterAdaptor c_block_cluster_adaptor)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
    constexpr index_t shared_block_size =
        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);

--- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
@@ -30,7 +30,8 @@ __global__ void
        kernel_gemm_xdlops_v2r4r2_simplified(typename GridwiseGemm::Argument karg,
                                             const Block2CTileMap& b2c_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
    constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte();

    __shared__ uint8_t p_shared[shared_size];
--- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp
@@ -46,7 +46,8 @@ __global__ void
            const CElementwiseOperation c_element_op,
            const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];

    GridwiseGemm::template Run<HasMainK0BlockLoop>(
--- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp
@@ -49,7 +49,8 @@ __global__ void
            const CElementwiseOperation c_element_op,
            const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];

    GridwiseGemm::template Run<HasMainKBlockLoop>(
--- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp
@@ -53,7 +53,8 @@ __global__ void
            const CElementwiseOperation c_element_op,
            const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];

    GridwiseGemm::template Run<HasMainKBlockLoop>(