pass-by-void-pointer for gridwise_dynamic_gemm_v1r2 (#38)

* pass-by-void-pointer for gridwise_dynamic_gemm_v1r2

* use pass-by-value by default

composable_kernel/include/driver/driver_dynamic_gemm_v1r2.hpp

@@ -167,6 +167,7 @@ __host__ float driver_dynamic_gemm_v1r2(const FloatAB* p_a_grid,
                   << c_m0_m10_m11_n0_n10_n11_grid_desc.GetLength(I5) << "}" << std::endl;
     }
 
+#if CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VALUE
     float ave_time = 0;
 
     if(has_main_k_block_loop && has_double_tail_k_block_loop)
@@ -279,6 +280,136 @@ __host__ float driver_dynamic_gemm_v1r2(const FloatAB* p_a_grid,
     }
 
     return ave_time;
+#elif CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VOID_POINTER
+    DeviceMem a_k_m0_m1_grid_desc_dev_buf(sizeof(AKM0M1GridDesc));
+    DeviceMem b_k_n0_n1_grid_desc_dev_buf(sizeof(BKN0N1GridDesc));
+    DeviceMem c_m0_m10_m11_n0_n10_n11_grid_desc_dev_buf(sizeof(CM0M10M11N0N10N11GridDesc));
+    DeviceMem c_blockid_to_m0_n0_block_cluster_adaptor_dev_buf(
+        sizeof(CBlockIdToM0N0BlockClusterAdaptor));
+
+    a_k_m0_m1_grid_desc_dev_buf.ToDevice(&a_k_m0_m1_grid_desc);
+    b_k_n0_n1_grid_desc_dev_buf.ToDevice(&b_k_n0_n1_grid_desc);
+    c_m0_m10_m11_n0_n10_n11_grid_desc_dev_buf.ToDevice(&c_m0_m10_m11_n0_n10_n11_grid_desc);
+    c_blockid_to_m0_n0_block_cluster_adaptor_dev_buf.ToDevice(
+        &c_blockid_to_m0_n0_block_cluster_adaptor);
+
+    float ave_time = 0;
+
+    if(has_main_k_block_loop && has_double_tail_k_block_loop)
+    {
+        const auto kernel =
+            kernel_dynamic_gemm_v1r2<GridwiseGemm,
+                                     FloatAB,
+                                     FloatC,
+                                     remove_reference_t<AKM0M1GridDesc>,
+                                     remove_reference_t<BKN0N1GridDesc>,
+                                     remove_reference_t<CM0M10M11N0N10N11GridDesc>,
+                                     remove_reference_t<CBlockIdToM0N0BlockClusterAdaptor>,
+                                     true,
+                                     true>;
+
+        ave_time = launch_and_time_kernel(
+            kernel,
+            nrepeat,
+            dim3(grid_size),
+            dim3(BlockSize),
+            0,
+            0,
+            p_a_grid,
+            p_b_grid,
+            p_c_grid,
+            (void __CONSTANT__*)a_k_m0_m1_grid_desc_dev_buf.GetDeviceBuffer(),
+            (void __CONSTANT__*)b_k_n0_n1_grid_desc_dev_buf.GetDeviceBuffer(),
+            (void __CONSTANT__*)c_m0_m10_m11_n0_n10_n11_grid_desc_dev_buf.GetDeviceBuffer(),
+            (void __CONSTANT__*)c_blockid_to_m0_n0_block_cluster_adaptor_dev_buf.GetDeviceBuffer());
+    }
+    else if(has_main_k_block_loop && !has_double_tail_k_block_loop)
+    {
+        const auto kernel =
+            kernel_dynamic_gemm_v1r2<GridwiseGemm,
+                                     FloatAB,
+                                     FloatC,
+                                     remove_reference_t<AKM0M1GridDesc>,
+                                     remove_reference_t<BKN0N1GridDesc>,
+                                     remove_reference_t<CM0M10M11N0N10N11GridDesc>,
+                                     remove_reference_t<CBlockIdToM0N0BlockClusterAdaptor>,
+                                     true,
+                                     false>;
+
+        ave_time = launch_and_time_kernel(
+            kernel,
+            nrepeat,
+            dim3(grid_size),
+            dim3(BlockSize),
+            0,
+            0,
+            p_a_grid,
+            p_b_grid,
+            p_c_grid,
+            (void __CONSTANT__*)a_k_m0_m1_grid_desc_dev_buf.GetDeviceBuffer(),
+            (void __CONSTANT__*)b_k_n0_n1_grid_desc_dev_buf.GetDeviceBuffer(),
+            (void __CONSTANT__*)c_m0_m10_m11_n0_n10_n11_grid_desc_dev_buf.GetDeviceBuffer(),
+            (void __CONSTANT__*)c_blockid_to_m0_n0_block_cluster_adaptor_dev_buf.GetDeviceBuffer());
+    }
+    else if(!has_main_k_block_loop && has_double_tail_k_block_loop)
+    {
+        const auto kernel =
+            kernel_dynamic_gemm_v1r2<GridwiseGemm,
+                                     FloatAB,
+                                     FloatC,
+                                     remove_reference_t<AKM0M1GridDesc>,
+                                     remove_reference_t<BKN0N1GridDesc>,
+                                     remove_reference_t<CM0M10M11N0N10N11GridDesc>,
+                                     remove_reference_t<CBlockIdToM0N0BlockClusterAdaptor>,
+                                     false,
+                                     true>;
+
+        ave_time = launch_and_time_kernel(
+            kernel,
+            nrepeat,
+            dim3(grid_size),
+            dim3(BlockSize),
+            0,
+            0,
+            p_a_grid,
+            p_b_grid,
+            p_c_grid,
+            (void __CONSTANT__*)a_k_m0_m1_grid_desc_dev_buf.GetDeviceBuffer(),
+            (void __CONSTANT__*)b_k_n0_n1_grid_desc_dev_buf.GetDeviceBuffer(),
+            (void __CONSTANT__*)c_m0_m10_m11_n0_n10_n11_grid_desc_dev_buf.GetDeviceBuffer(),
+            (void __CONSTANT__*)c_blockid_to_m0_n0_block_cluster_adaptor_dev_buf.GetDeviceBuffer());
+    }
+    else
+    {
+        const auto kernel =
+            kernel_dynamic_gemm_v1r2<GridwiseGemm,
+                                     FloatAB,
+                                     FloatC,
+                                     remove_reference_t<AKM0M1GridDesc>,
+                                     remove_reference_t<BKN0N1GridDesc>,
+                                     remove_reference_t<CM0M10M11N0N10N11GridDesc>,
+                                     remove_reference_t<CBlockIdToM0N0BlockClusterAdaptor>,
+                                     false,
+                                     false>;
+
+        ave_time = launch_and_time_kernel(
+            kernel,
+            nrepeat,
+            dim3(grid_size),
+            dim3(BlockSize),
+            0,
+            0,
+            p_a_grid,
+            p_b_grid,
+            p_c_grid,
+            (void __CONSTANT__*)a_k_m0_m1_grid_desc_dev_buf.GetDeviceBuffer(),
+            (void __CONSTANT__*)b_k_n0_n1_grid_desc_dev_buf.GetDeviceBuffer(),
+            (void __CONSTANT__*)c_m0_m10_m11_n0_n10_n11_grid_desc_dev_buf.GetDeviceBuffer(),
+            (void __CONSTANT__*)c_blockid_to_m0_n0_block_cluster_adaptor_dev_buf.GetDeviceBuffer());
+    }
+
+    return ave_time;
+#endif
 }
 
 } // namespace ck

composable_kernel/include/tensor_operation/gridwise_dynamic_gemm_v1r2.hpp

@@ -12,6 +12,7 @@
 
 namespace ck {
 
+#if CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VALUE
 template <typename GridwiseGemm,
           typename FloatAB,
           typename FloatC,
@@ -50,6 +51,63 @@ __global__ void
                       integral_constant<bool, HasMainKBlockLoop>{},
                       integral_constant<bool, HasDoubleTailKBlockLoop>{});
 }
+#elif CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VOID_POINTER
+// pass tensor descriptor by __CONSTANT__ void pointer
+// __CONSTANT__ is needed to inform compiler void pointers in the kernel signature are pointing to
+// non-modifiable parameter address space, so compiler can enable corresponding optimization
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AKM0M1GridDesc,
+          typename BKN0N1GridDesc,
+          typename CM0M10M11N0N10N11GridDesc,
+          typename CBlockIdToM0N0BlockClusterAdaptor,
+          bool HasMainKBlockLoop,
+          bool HasDoubleTailKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_dynamic_gemm_v1r2(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const void __CONSTANT__* p_a_k_m0_m1_grid_desc,
+            const void __CONSTANT__* p_b_k_n0_n1_grid_desc,
+            const void __CONSTANT__* p_c_m0_m10_m11_n0_n10_n11_grid_desc,
+            const void __CONSTANT__* p_c_blockid_to_m0_n0_block_cluster_adaptor)
+{
+    // first cast void __CONSTANT__ void* to void*
+    // second cast void* to Desc*
+    // the copy constructor of tensor descriptor doesn't take address_space(4)
+    const auto a_k_m0_m1_grid_desc =
+        *reinterpret_cast<const AKM0M1GridDesc*>((const void*)p_a_k_m0_m1_grid_desc);
+    const auto b_k_n0_n1_grid_desc =
+        *reinterpret_cast<const BKN0N1GridDesc*>((const void*)p_b_k_n0_n1_grid_desc);
+    const auto c_m0_m10_m11_n0_n10_n11_grid_desc =
+        *reinterpret_cast<const CM0M10M11N0N10N11GridDesc*>(
+            (const void*)p_c_m0_m10_m11_n0_n10_n11_grid_desc);
+    const auto c_blockid_to_m0_n0_block_cluster_adaptor =
+        *reinterpret_cast<const CBlockIdToM0N0BlockClusterAdaptor*>(
+            (const void*)p_c_blockid_to_m0_n0_block_cluster_adaptor);
+
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
+
+    __shared__ FloatAB p_shared_block[shared_block_size];
+
+    GridwiseGemm::Run(p_a_grid,
+                      p_b_grid,
+                      p_c_grid,
+                      p_shared_block,
+                      a_k_m0_m1_grid_desc,
+                      b_k_n0_n1_grid_desc,
+                      c_m0_m10_m11_n0_n10_n11_grid_desc,
+                      c_blockid_to_m0_n0_block_cluster_adaptor,
+                      integral_constant<bool, HasMainKBlockLoop>{},
+                      integral_constant<bool, HasDoubleTailKBlockLoop>{});
+}
+#endif
 
 template <index_t BlockSize,
           typename FloatAB,

composable_kernel/include/utility/config.amd.hpp.in

@@ -14,9 +14,9 @@
 #define CK_DEVICE_BACKEND_AMD 1
 
 // GPU ID
-#if 1
+#if 0
 #define CK_AMD_GPU_GFX906 1
-#elif 0
+#elif 1
 #define CK_AMD_GPU_GFX908 1
 #elif 1
 #define CK_AMD_GPU_GFX1030 1

driver/src/conv_driver_v2.cpp

@@ -19,8 +19,8 @@
 
 #define USE_DYNAMIC_MODE 1
 #define USE_CONV_FWD_V4R4_NCHW 1
-#define USE_CONV_FWD_V4R4_NHWC 1
-#define USE_CONV_FWD_V4R5_NCHW 1
+#define USE_CONV_FWD_V4R4_NHWC 0
+#define USE_CONV_FWD_V4R5_NCHW 0
 #define USE_CONV_FWD_V5R1_NCHW 0
 
 enum ConvForwardAlgo