Ck tile batched gemm example (#1615)

* [CK Tile] Batched GEMM Example * [CK Tile] Batched GEMM Example - minor refactor * [CK Tile] Batched GEMM Example - README update * [CK Tile] Batched Gemm Example - review changes - Added tensor data layours as input parameters - Changed structure of Host and Kernel args - Removed bug with invalid vector read on non-contiguous memory * [CK Tile] Batched Gemm Example - remove comment * [CK Tile] Batched Gemm Example - Add GTests part1 * [CK Tile] Batched Gemm Example - GTests part2 + review changes * [CK TILE] Batched GEMM post merge fixes * [CK Tile] Batched GEMM Example - fix pad views
2026-05-01 12:11:19 +00:00 · 2024-11-29 11:52:18 +01:00
parent bb652696e7
commit 78f0fea08e
15 changed files with 1098 additions and 2 deletions
--- a/include/ck_tile/host/reference/reference_gemm.hpp
+++ b/include/ck_tile/host/reference/reference_gemm.hpp
@@ -183,4 +183,116 @@ void reference_gemm_gpu(DeviceMem& a_device,

    return;
 }
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename LayoutA,
+          typename LayoutB,
+          typename LayoutC>
+void reference_batched_gemm_gpu(DeviceMem& a_device,
+                                DeviceMem& b_device,
+                                DeviceMem& c_device,
+                                index_t M,
+                                index_t N,
+                                index_t K,
+                                index_t stride_a,
+                                index_t stride_b,
+                                index_t stride_c,
+                                index_t batch_stride_A,
+                                index_t batch_stride_B,
+                                index_t batch_stride_C,
+                                index_t batch_count)
+{
+
+    ADataType* d_A;
+    BDataType* d_B;
+    CDataType* d_C;
+
+    hipError_t errA = hipMalloc(&d_A, batch_count * M * K * sizeof(ADataType));
+    hipError_t errB = hipMalloc(&d_B, batch_count * N * K * sizeof(BDataType));
+    hipError_t errC = hipMalloc(&d_C, batch_count * M * N * sizeof(CDataType));
+    if(errA != hipSuccess)
+    {
+        std::cerr << "Error allocating device memory for A: " << hipGetErrorString(errA)
+                  << std::endl;
+        return; // Early exit on error
+    }
+
+    if(errB != hipSuccess)
+    {
+        std::cerr << "Error allocating device memory for B: " << hipGetErrorString(errB)
+                  << std::endl;
+        return; // Early exit on error
+    }
+
+    if(errC != hipSuccess)
+    {
+        std::cerr << "Error allocating device memory for C: " << hipGetErrorString(errC)
+                  << std::endl;
+        return; // Early exit on error
+    }
+
+    errA = hipMemcpy(d_A,
+                     a_device.GetDeviceBuffer(),
+                     batch_count * M * K * sizeof(ADataType),
+                     hipMemcpyHostToDevice);
+    if(errA != hipSuccess)
+    {
+        std::cerr << "Error copying A to device: " << hipGetErrorString(errA) << std::endl;
+    }
+
+    errB = hipMemcpy(d_B,
+                     b_device.GetDeviceBuffer(),
+                     batch_count * N * K * sizeof(BDataType),
+                     hipMemcpyHostToDevice);
+    if(errB != hipSuccess)
+    {
+        std::cerr << "Error copying B to device: " << hipGetErrorString(errB) << std::endl;
+    }
+
+    int totalElements      = M * N;
+    int numThreadsPerBlock = 256; // Common choice for threads per block
+    int numBlocks          = (totalElements + numThreadsPerBlock - 1) / numThreadsPerBlock;
+
+    for(index_t batch_id = 0; batch_id < batch_count; ++batch_id)
+    {
+        ADataType* d_ATemp = d_A + batch_id * batch_stride_A;
+        BDataType* d_BTemp = d_B + batch_id * batch_stride_B;
+        CDataType* d_CTemp = d_C + batch_id * batch_stride_C;
+        naive_gemm_kernel<ADataType, BDataType, AccDataType, CDataType, LayoutA, LayoutB, LayoutC>
+            <<<numBlocks, numThreadsPerBlock>>>(
+                d_ATemp, d_BTemp, d_CTemp, M, N, K, stride_a, stride_b, stride_c);
+    }
+
+    errC = hipMemcpy(c_device.GetDeviceBuffer(),
+                     d_C,
+                     batch_count * M * N * sizeof(CDataType),
+                     hipMemcpyDeviceToHost);
+    if(errC != hipSuccess)
+    {
+        std::cerr << "Error copying C to device: " << hipGetErrorString(errC) << std::endl;
+    }
+
+    errA = hipFree(d_A);
+    if(errA != hipSuccess)
+    {
+        std::cerr << "Error free the A memory: " << hipGetErrorString(errA) << std::endl;
+    }
+
+    errB = hipFree(d_B);
+    if(errB != hipSuccess)
+    {
+        std::cerr << "Error free the B memory: " << hipGetErrorString(errB) << std::endl;
+    }
+
+    errC = hipFree(d_C);
+    if(errC != hipSuccess)
+    {
+        std::cerr << "Error free the C memory: " << hipGetErrorString(errC) << std::endl;
+    }
+
+    return;
+}
 } // namespace ck_tile