Revert "Grouped Gemm with looping over the tiles. (#788)" (#982)

This reverts commit a4f72a31.

test/grouped_gemm/test_grouped_gemm_ut_cases.inc

@@ -4,7 +4,7 @@ TEST_P(RRR_F16_F16_F16, TinyCases)
 {
     const std::vector<int> Ms{0, 1};
     constexpr int N = 768;
-    constexpr int K = 1088;
+    constexpr int K = 544;
 
     const std::vector<int> Ns(Ms.size(), N);
     const std::vector<int> Ks(Ms.size(), K);
@@ -17,9 +17,9 @@ TEST_P(RRR_F16_F16_F16, TinyCases)
 
 TEST_P(RRR_F16_F16_F16, SmallCases)
 {
-    const std::vector<int> Ms{2, 3, 4, 5};
+    const std::vector<int> Ms{2, 1, 3, 4, 5, 0};
     constexpr int N = 768;
-    constexpr int K = 1088;
+    constexpr int K = 544;
 
     const std::vector<int> Ns(Ms.size(), N);
     const std::vector<int> Ks(Ms.size(), K);
@@ -34,7 +34,7 @@ TEST_P(RRR_F16_F16_F16, MidCases)
 {
     const std::vector<int> Ms{167, 183, 177, 153, 139, 204};
     constexpr int N = 768;
-    constexpr int K = 1088;
+    constexpr int K = 544;
 
     const std::vector<int> Ns(Ms.size(), N);
     const std::vector<int> Ks(Ms.size(), K);
@@ -49,7 +49,7 @@ TEST_P(RRR_F16_F16_F16, Regular)
 {
     const std::vector<int> Ms{64, 128, 256};
     constexpr int N = 768;
-    constexpr int K = 640;
+    constexpr int K = 320;
 
     const std::vector<int> Ns(Ms.size(), N);
     const std::vector<int> Ks(Ms.size(), K);
@@ -79,7 +79,7 @@ TEST_P(RCR_F16_F16_F16, TinyCases)
 {
     const std::vector<int> Ms{0, 1};
     constexpr int N = 768;
-    constexpr int K = 1088;
+    constexpr int K = 544;
 
     const std::vector<int> Ns(Ms.size(), N);
     const std::vector<int> Ks(Ms.size(), K);
@@ -91,9 +91,9 @@ TEST_P(RCR_F16_F16_F16, TinyCases)
 
 TEST_P(RCR_F16_F16_F16, SmallCases)
 {
-    const std::vector<int> Ms{2, 3, 4, 5};
+    const std::vector<int> Ms{2, 1, 3, 4, 5, 0};
     constexpr int N = 768;
-    constexpr int K = 1088;
+    constexpr int K = 544;
 
     const std::vector<int> Ns(Ms.size(), N);
     const std::vector<int> Ks(Ms.size(), K);
@@ -123,7 +123,7 @@ TEST_P(RCR_F16_F16_F16, Regular)
 {
     const std::vector<int> Ms{32, 64, 128, 256};
     constexpr int N = 768;
-    constexpr int K = 640;
+    constexpr int K = 320;
 
     const std::vector<int> Ns(Ms.size(), N);
     const std::vector<int> Ks(Ms.size(), K);
@@ -151,9 +151,9 @@ TEST_P(RCR_F16_F16_F16, MNKPadded)
 
 TEST_P(RRR_F16_F16_F16_LargeK, TestLargeKBatch)
 {
-    const std::vector<int> Ms{127, 150, 188, 210};
+    const std::vector<int> Ms{188, 210};
     constexpr int N = 768;
-    constexpr int K = 8192;
+    constexpr int K = 4096;
 
     const std::vector<int> Ns(Ms.size(), N);
     const std::vector<int> Ks(Ms.size(), K);