From 0a72e4df949d889cc5ae492405f4ec9b9b829c11 Mon Sep 17 00:00:00 2001
From: Chao Liu <chao.liu2@amd.com>
Date: Fri, 16 Jul 2021 22:55:01 -0500
Subject: [PATCH] Change initialization method of tensor for iGEMM (#49)

* change init method
---
 driver/conv_bwd_data_driver_v2.cpp       | 50 +++++++++++++--------
 driver/conv_driver_v2.cpp                | 56 +++++++++++++-----------
 driver/conv_driver_v2_olc.cpp            | 56 +++++++++++++-----------
 driver/include/host_tensor_generator.hpp | 18 ++++----
 4 files changed, 101 insertions(+), 79 deletions(-)

diff --git a/driver/conv_bwd_data_driver_v2.cpp b/driver/conv_bwd_data_driver_v2.cpp
index cc3274c7eb..4ae9c5c749 100644
--- a/driver/conv_bwd_data_driver_v2.cpp
+++ b/driver/conv_bwd_data_driver_v2.cpp
@@ -179,26 +179,38 @@ int main(int argc, char* argv[])
 
     std::size_t num_thread = std::thread::hardware_concurrency();
 
-    if(do_verification)
+    switch(init_method)
     {
-        switch(init_method)
-        {
-        case 0:
-            wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
-            out.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
-            break;
-        case 1:
-            wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
-            out.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
-            break;
-        case 2:
-            wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
-            out.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
-            break;
-        default:
-            wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
-            out.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
-        }
+    case 0:
+        // no initialization
+        break;
+    case 1:
+        out.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
+        break;
+    case 2:
+        out.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
+        break;
+    case 3:
+        out.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
+        break;
+    case 4:
+        out.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
+        break;
+    case 5:
+        out.GenerateTensorValue(GeneratorTensor_3<float>{0.0, 1.0}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_3<float>{-0.5, 0.5}, num_thread);
+        break;
+    default:
+        out.GenerateTensorValue(GeneratorTensor_2{1, 5}, num_thread);
+
+        auto gen_wei = [](auto... is) {
+            return GeneratorTensor_2{1, 5}(is...) * GeneratorTensor_Checkboard{}(is...);
+        };
+        wei.GenerateTensorValue(gen_wei, num_thread);
     }
 
     auto f_make_for_device_nchw = [&]() {
diff --git a/driver/conv_driver_v2.cpp b/driver/conv_driver_v2.cpp
index 8f972ad37e..3b9fde9257 100644
--- a/driver/conv_driver_v2.cpp
+++ b/driver/conv_driver_v2.cpp
@@ -205,34 +205,38 @@ int main(int argc, char* argv[])
 
     std::size_t num_thread = std::thread::hardware_concurrency();
 
-    if(do_verification)
+    switch(init_method)
     {
-        switch(init_method)
-        {
-        case 0:
-            in.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
-            wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
-            break;
-        case 1:
-            in.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
-            wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
-            break;
-        case 2:
-            in.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
-            wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
-            break;
-        case 3:
-            in.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
-            wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
-            break;
-        default:
-            in.GenerateTensorValue(GeneratorTensor_2{1, 5}, num_thread);
+    case 0:
+        // no initialization
+        break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
+        break;
+    case 2:
+        in.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
+        break;
+    case 3:
+        in.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
+        break;
+    case 4:
+        in.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
+        break;
+    case 5:
+        in.GenerateTensorValue(GeneratorTensor_3<float>{0.0, 1.0}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_3<float>{-0.5, 0.5}, num_thread);
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_2{1, 5}, num_thread);
 
-            auto gen_wei = [](auto... is) {
-                return GeneratorTensor_2{1, 5}(is...) * GeneratorTensor_Checkboard{}(is...);
-            };
-            wei.GenerateTensorValue(gen_wei, num_thread);
-        }
+        auto gen_wei = [](auto... is) {
+            return GeneratorTensor_2{1, 5}(is...) * GeneratorTensor_Checkboard{}(is...);
+        };
+        wei.GenerateTensorValue(gen_wei, num_thread);
     }
 
     auto f_make_for_device_nchw = [&]() {
diff --git a/driver/conv_driver_v2_olc.cpp b/driver/conv_driver_v2_olc.cpp
index 0b1c91c81b..d117dfdd31 100644
--- a/driver/conv_driver_v2_olc.cpp
+++ b/driver/conv_driver_v2_olc.cpp
@@ -152,34 +152,38 @@ int main(int argc, char* argv[])
 
     std::size_t num_thread = std::thread::hardware_concurrency();
 
-    if(do_verification)
+    switch(init_method)
     {
-        switch(init_method)
-        {
-        case 0:
-            in.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
-            wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
-            break;
-        case 1:
-            in.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
-            wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
-            break;
-        case 2:
-            in.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
-            wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
-            break;
-        case 3:
-            in.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
-            wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
-            break;
-        default:
-            in.GenerateTensorValue(GeneratorTensor_2{1, 5}, num_thread);
+    case 0:
+        // no initialization
+        break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
+        break;
+    case 2:
+        in.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
+        break;
+    case 3:
+        in.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
+        break;
+    case 4:
+        in.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
+        break;
+    case 5:
+        in.GenerateTensorValue(GeneratorTensor_3<float>{0.0, 1.0}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_3<float>{-0.5, 0.5}, num_thread);
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_2{1, 5}, num_thread);
 
-            auto gen_wei = [](auto... is) {
-                return GeneratorTensor_2{1, 5}(is...) * GeneratorTensor_Checkboard{}(is...);
-            };
-            wei.GenerateTensorValue(gen_wei, num_thread);
-        }
+        auto gen_wei = [](auto... is) {
+            return GeneratorTensor_2{1, 5}(is...) * GeneratorTensor_Checkboard{}(is...);
+        };
+        wei.GenerateTensorValue(gen_wei, num_thread);
     }
 
     auto f_make_for_device_nchw = [&]() {
diff --git a/driver/include/host_tensor_generator.hpp b/driver/include/host_tensor_generator.hpp
index a62045a182..98192e066f 100644
--- a/driver/include/host_tensor_generator.hpp
+++ b/driver/include/host_tensor_generator.hpp
@@ -9,7 +9,7 @@ struct GeneratorTensor_1
     int value = 1;
 
     template <typename... Is>
-    double operator()(Is... is)
+    float operator()(Is... is)
     {
         return value;
     }
@@ -21,29 +21,31 @@ struct GeneratorTensor_2
     int max_value = 1;
 
     template <typename... Is>
-    double operator()(Is...)
+    float operator()(Is...)
     {
         return (std::rand() % (max_value - min_value)) + min_value;
     }
 };
 
+template <typename T>
 struct GeneratorTensor_3
 {
+    T min_value = 0;
+    T max_value = 1;
+
     template <typename... Is>
-    double operator()(Is... is)
+    float operator()(Is...)
     {
-        std::array<ck::index_t, sizeof...(Is)> dims = {{static_cast<ck::index_t>(is)...}};
+        float tmp = float(std::rand()) / float(RAND_MAX);
 
-        auto f_acc = [](auto a, auto b) { return 10 * a + b; };
-
-        return std::accumulate(dims.begin(), dims.end(), ck::index_t(0), f_acc);
+        return min_value + tmp * (max_value - min_value);
     }
 };
 
 struct GeneratorTensor_Checkboard
 {
     template <typename... Ts>
-    double operator()(Ts... Xs) const
+    float operator()(Ts... Xs) const
     {
         std::array<ck::index_t, sizeof...(Ts)> dims = {{static_cast<ck::index_t>(Xs)...}};
         return std::accumulate(dims.begin(),