Merge branch 'develop' into add_fp16_wmma_conv_instance

example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp

@@ -3,7 +3,7 @@
 
 #include "common.hpp"
 
-#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp"
 
 using InDataType = BF16;
 // bf16 kernel use fp32 atomic add to accumulate Weight tensor into global memory
@@ -17,8 +17,20 @@ using OutElementOp = PassThrough;
 
 template <ck::index_t NDimSpatial>
 using DeviceConvBwdWeightInstance =
-    ck::tensor_operation::device::DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<
-        NDimSpatial,          // NDimSpatial
+    ck::tensor_operation::device::DeviceGroupedConvBwdWeight_Xdl_CShuffle<
+        NDimSpatial,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GNWC,
+                                      ck::tensor_layout::convolution::GNHWC,
+                                      ck::tensor_layout::convolution::GNDHWC>>,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GKXC,
+                                      ck::tensor_layout::convolution::GKYXC,
+                                      ck::tensor_layout::convolution::GKZYXC>>,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GNWK,
+                                      ck::tensor_layout::convolution::GNHWK,
+                                      ck::tensor_layout::convolution::GNDHWK>>,
         InDataType,           // InDataType
         WeiDataType,          // WeiDataType
         OutDataType,          // OutDataType

example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp

@@ -3,7 +3,7 @@
 
 #include "common.hpp"
 
-#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp"
 
 using InDataType  = F16;
 using WeiDataType = F16;
@@ -16,8 +16,20 @@ using OutElementOp = PassThrough;
 
 template <ck::index_t NDimSpatial>
 using DeviceConvBwdWeightInstance =
-    ck::tensor_operation::device::DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<
-        NDimSpatial,          // NDimSpatial
+    ck::tensor_operation::device::DeviceGroupedConvBwdWeight_Xdl_CShuffle<
+        NDimSpatial,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GNWC,
+                                      ck::tensor_layout::convolution::GNHWC,
+                                      ck::tensor_layout::convolution::GNDHWC>>,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GKXC,
+                                      ck::tensor_layout::convolution::GKYXC,
+                                      ck::tensor_layout::convolution::GKZYXC>>,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GNWK,
+                                      ck::tensor_layout::convolution::GNHWK,
+                                      ck::tensor_layout::convolution::GNDHWK>>,
         InDataType,           // InDataType
         WeiDataType,          // WeiDataType
         OutDataType,          // OutDataType

example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc

@@ -75,6 +75,8 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
     std::array<ck::index_t, NDimSpatial> input_spatial_lengths{};
     std::array<ck::index_t, NDimSpatial> filter_spatial_lengths{};
     std::array<ck::index_t, NDimSpatial> output_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> input_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> output_strides{};
     std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
     std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
     std::array<ck::index_t, NDimSpatial> input_left_pads{};
@@ -85,6 +87,8 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
     range_copy(conv_param.input_spatial_lengths_, begin(input_spatial_lengths));
     range_copy(conv_param.filter_spatial_lengths_, begin(filter_spatial_lengths));
     range_copy(conv_param.output_spatial_lengths_, begin(output_spatial_lengths));
+    range_copy(in_g_n_c_wis_desc.GetStrides(), begin(input_strides));
+    range_copy(out_g_n_k_wos_desc.GetStrides(), begin(output_strides));
     range_copy(conv_param.conv_filter_strides_, begin(conv_filter_strides));
     range_copy(conv_param.conv_filter_dilations_, begin(conv_filter_dilations));
     range_copy(conv_param.input_left_pads_, begin(input_left_pads));
@@ -103,6 +107,8 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
                                       input_spatial_lengths,
                                       filter_spatial_lengths,
                                       output_spatial_lengths,
+                                      input_strides,
+                                      output_strides,
                                       conv_filter_strides,
                                       conv_filter_dilations,
                                       input_left_pads,

example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp

@@ -17,7 +17,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
-#include "ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/utility/check_err.hpp"

example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp

@@ -17,7 +17,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
-#include "ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/utility/check_err.hpp"

example/34_batchnorm/CMakeLists.txt

 add_example_executable(example_batchnorm_forward_training batchnorm_forward_training_nhwc.cpp)
+add_example_executable(example_batchnorm_forward_training_obsolete batchnorm_forward_training_nhwc_obsolete.cpp)
 add_example_executable(example_batchnorm_forward_inferring batchnorm_forward_inferring_nhwc.cpp)
 add_example_executable(example_batchnorm_backward batchnorm_backward_nhwc.cpp)

example/34_batchnorm/batchnorm_forward_training_nhwc.cpp

@@ -414,7 +414,7 @@ bool bnorm_fwd_nhwc_test(bool do_verification,
         (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
 
         y_dev.FromDevice(y.mData.data());
-        pass = pass && ck::utils::check_err(y, y_ref);
+        pass = pass && ck::utils::check_err(y, y_ref, "Incorrect normalized output values");
 
         if(updateMovingAverage)
         {
@@ -424,8 +424,12 @@ bool bnorm_fwd_nhwc_test(bool do_verification,
             resultRunningMean_dev.FromDevice(resultRunningMean.mData.data());
             resultRunningVariance_dev.FromDevice(resultRunningVariance.mData.data());
 
-            pass = pass && ck::utils::check_err(resultRunningMean, resultRunningMean_ref);
-            pass = pass && ck::utils::check_err(resultRunningVariance, resultRunningVariance_ref);
+            pass = pass && ck::utils::check_err(resultRunningMean,
+                                                resultRunningMean_ref,
+                                                "Incorrect running mean values");
+            pass = pass && ck::utils::check_err(resultRunningVariance,
+                                                resultRunningVariance_ref,
+                                                "Incorrect running variance values");
         };
 
         if(saveMeanAndInvVariance)
@@ -438,8 +442,11 @@ bool bnorm_fwd_nhwc_test(bool do_verification,
             resultSaveMean_dev.FromDevice(resultSaveMean.mData.data());
             resultSaveInvVariance_dev.FromDevice(resultSaveInvVariance.mData.data());
 
-            pass = pass && ck::utils::check_err(resultSaveMean, resultSaveMean_ref);
-            pass = pass && ck::utils::check_err(resultSaveInvVariance, resultSaveInvVariance_ref);
+            pass = pass && ck::utils::check_err(
+                               resultSaveMean, resultSaveMean_ref, "Incorrect saved mean values");
+            pass = pass && ck::utils::check_err(resultSaveInvVariance,
+                                                resultSaveInvVariance_ref,
+                                                "Incorrect saved invvariance values");
         };
     };
 

example/40_conv2d_fwd_quantization/CMakeLists.txt

+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
 list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
@@ -26,3 +27,4 @@ add_example_executable(example_conv2d_fwd_dl_bias_tanh_perlayer_quantization_int
 
 # Conv + bias + tanh perchannel quantization
 add_example_executable(example_conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8 conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp)
+endif()
\ No newline at end of file

example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp

@@ -2,7 +2,7 @@
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
-#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
 
 using InDataType           = int8_t;
 using WeiDataType          = int8_t;

example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp

@@ -2,7 +2,7 @@
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
-#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
 
 using InDataType   = int8_t;
 using WeiDataType  = int8_t;

example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp

@@ -2,7 +2,7 @@
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
-#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
 
 using InDataType           = int8_t;
 using WeiDataType          = int8_t;

example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp

@@ -2,7 +2,7 @@
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
-#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
 
 using InDataType   = int8_t;
 using WeiDataType  = int8_t;

example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perchannel_quantization_int8.cpp

@@ -2,7 +2,7 @@
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
-#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
 
 using InDataType           = int8_t;
 using WeiDataType          = int8_t;

example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perlayer_quantization_int8.cpp

@@ -2,7 +2,7 @@
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
-#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
 
 using InDataType  = int8_t;
 using WeiDataType = int8_t;

example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp16.cpp

@@ -8,7 +8,7 @@
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
-#include "ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/utility/check_err.hpp"

example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp32.cpp

@@ -8,7 +8,7 @@
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
-#include "ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/utility/check_err.hpp"

include/ck/ck.hpp

@@ -27,6 +27,21 @@
 #define CK_WAVELET_MIN_BLOCK_PER_CU 2
 #endif
 
+// kernel attribute: amdgpu_waves_per_eu()
+#ifdef CK_USE_WAVES_PER_EU
+// for 1-wave kernels, control arguments of amdgpu_waves_per_eu() attribute
+#ifndef CK_MIN_WAVES_PER_EU
+#define CK_MIN_WAVES_PER_EU 0
+#endif
+
+#ifndef CK_MAX_WAVES_PER_EU
+#define CK_MAX_WAVES_PER_EU 0
+#endif
+
+#else
+#define CK_USE_WAVES_PER_EU 0
+#endif
+
 // buffer resource
 #ifndef __HIP_DEVICE_COMPILE__ // for host code
 #define CK_BUFFER_RESOURCE_3RD_DWORD -1
@@ -148,6 +163,10 @@
 #define CK_EXPERIMENTAL_INTER_WAVE_INSTANCES 1
 // experimental feature: add instances using pipeline v2
 #define CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES 1
+// experimental feature: optimize pipeline v2 by IGLP strategy (value=ID of strategy)
+#ifndef CK_EXPERIMENTAL_PIPELINE_V2_IGLP_OPT
+#define CK_EXPERIMENTAL_PIPELINE_V2_IGLP_OPT 0
+#endif
 
 // hack: have underlying assumption that need to be satsified, otherwise it's a bug
 // hack for forcing register to keep idx_diff_low_const in SGPR. idx_diff_low_const must be

include/ck/tensor_operation/gpu/block/blockwise_welford.hpp

@@ -4,7 +4,7 @@
 #pragma once
 
 #include "ck/tensor_description/cluster_descriptor.hpp"
-#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/get_shift.hpp"
 
 namespace ck {
 
@@ -35,10 +35,11 @@ struct BlockwiseWelford
     static constexpr auto thread_cluster_desc =
         make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
 
+    template <typename CountDataType>
     __device__ static inline void
-    Merge(T& mean_a, T& var_a, int& count_a, T mean_b, T var_b, int count_b)
+    Merge(T& mean_a, T& var_a, CountDataType& count_a, T mean_b, T var_b, CountDataType count_b)
     {
-        int count            = count_a + count_b;
+        CountDataType count  = count_a + count_b;
         T count_b_over_count = count == 0 ? type_convert<T>(0) : type_convert<T>(count_b) / count;
         T delta              = mean_b - mean_a;
         mean_a += delta * count_b_over_count;
@@ -46,11 +47,12 @@ struct BlockwiseWelford
         count_a = count;
     }
 
-    __device__ static void Run(T& mean_value, T& var_value, int& count)
+    template <typename CountDataType>
+    __device__ static void Run(T& mean_value, T& var_value, CountDataType& count)
     {
         __shared__ T mean_block_buf[BlockSize];
         __shared__ T var_block_buf[BlockSize];
-        __shared__ int count_block_buf[BlockSize];
+        __shared__ CountDataType count_block_buf[BlockSize];
 
         constexpr auto cluster_len_shift = get_shift<BufferLength_K>();
 
@@ -78,11 +80,11 @@ struct BlockwiseWelford
 
                 T mean1              = mean_block_buf[offset1];
                 T var1               = var_block_buf[offset1];
-                int count1 = count_block_buf[offset1];
+                CountDataType count1 = count_block_buf[offset1];
 
                 T mean2              = mean_block_buf[offset2];
                 T var2               = var_block_buf[offset2];
-                int count2 = count_block_buf[offset2];
+                CountDataType count2 = count_block_buf[offset2];
 
                 Merge(mean1, var1, count1, mean2, var2, count2);
 

include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp

@@ -4,7 +4,7 @@
 #pragma once
 
 #include "ck/tensor_description/cluster_descriptor.hpp"
-#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/get_shift.hpp"
 #include "ck/utility/reduction_functions_accumulate.hpp"
 
 namespace ck {