Add mini weight kernel

config/igemm_fwd_gtc_gfx1030_nchwc_fp16x8_fsr.config

@@ -0,0 +1,83 @@
+[codegen]
+arch = 'gfx1030'
+code_object = 'cov3'
+mode = 'flat'
+
+#########################################################################################
+#--------------------------- 16x512x256
+[igemm_fwd_gtc]
+gemm_m_per_block         = 16
+gemm_n_per_block         = 512
+gemm_k_per_block         = 256
+lanegroup_tile_m         = 8
+lanegroup_wave_m         = 1
+lanegroup_repeat_m       = 2
+lanegroup_tile_n         = 8
+lanegroup_wave_n         = 8
+lanegroup_repeat_n       = 1
+tensor_a_thread_lengths  = [1, 1, 1,  8]       # 1xCEx1xK/Vec-c
+tensor_a_cluster_lengths = [1,32, 1, 16]       # 1xCEx1xK
+tensor_b_thread_lengths  = [1, 1, 1,  8]       # 1xCExNB0xVec-c
+tensor_b_cluster_lengths = [1, 1, 1,512]       # 1xCEx1xNB1
+direction                = "fwd"
+precision                = "fp16"
+tensor_layout            = 'nchwc_cyxkc'
+nxb                      = 0
+nxe                      = 1
+wavefront_size           = 64
+cumode                   = 0
+vector_c                 = 8
+mini_weights             = 1
+tensor_b_pass_through    = 1
+
+#--------------------------- 16x256x128
+[igemm_fwd_gtc]
+gemm_m_per_block         = 16
+gemm_n_per_block         = 256
+gemm_k_per_block         = 128
+lanegroup_tile_m         = 8
+lanegroup_wave_m         = 1
+lanegroup_repeat_m       = 2
+lanegroup_tile_n         = 8
+lanegroup_wave_n         = 8
+lanegroup_repeat_n       = 1
+tensor_a_thread_lengths  = [1, 1, 1,  8]       # 1xCEx1xK/Vec-c
+tensor_a_cluster_lengths = [1,16, 1, 16]       # 1xCEx1xK
+tensor_b_thread_lengths  = [1, 1, 1,  8]       # 1xCExNB0xVec-c
+tensor_b_cluster_lengths = [1, 1, 1,256]       # 1xCEx1xNB1
+direction                = "fwd"
+precision                = "fp16"
+tensor_layout            = 'nchwc_cyxkc'
+nxb                      = 0
+nxe                      = 1
+wavefront_size           = 32
+cumode                   = 0
+vector_c                 = 8
+mini_weights             = 1
+tensor_b_pass_through    = 1
+
+#--------------------------- 16x64x32
+[igemm_fwd_gtc]
+gemm_m_per_block         = 16
+gemm_n_per_block         = 64
+gemm_k_per_block         = 32
+lanegroup_tile_m         = 8
+lanegroup_wave_m         = 1
+lanegroup_repeat_m       = 2
+lanegroup_tile_n         = 8
+lanegroup_wave_n         = 8
+lanegroup_repeat_n       = 1
+tensor_a_thread_lengths  = [1, 1, 1,  8]       # 1xCEx1xK/Vec-c
+tensor_a_cluster_lengths = [1, 4, 1, 16]       # 1xCEx1xK
+tensor_b_thread_lengths  = [1, 1, 1,  8]       # 1xCExNB0xVec-c
+tensor_b_cluster_lengths = [1, 1, 1, 64]       # 1xCEx1xNB1
+direction                = "fwd"
+precision                = "fp16"
+tensor_layout            = 'nchwc_cyxkc'
+nxb                      = 0
+nxe                      = 1
+wavefront_size           = 64
+cumode                   = 0
+vector_c                 = 8
+mini_weights             = 1
+tensor_b_pass_through    = 1

python/igemm/igemm_base.py

@@ -235,6 +235,10 @@ def __init__(self, tunable_dict):
         self.vector_c                           = utility_dict_with_default_t(tunable_dict)('vector_c', 1)
         self.wavefront_size                     = utility_dict_with_default_t(tunable_dict)('wavefront_size', 64)
         self.cumode                             = utility_dict_with_default_t(tunable_dict)('cumode', 0)
+
+        self.mini_weights                       = utility_dict_with_default_t(tunable_dict)('mini_weights', 0)
+        if self.mini_weights == 1:
+            self.tensor_b_pass_through = 1
 
         assert type(self.tensor_a_thread_lengths) is list and type(self.tensor_a_cluster_lengths) is list
         assert type(self.tensor_b_thread_lengths) is list and type(self.tensor_b_cluster_lengths) is list
@@ -383,7 +387,8 @@ def _unmerge_x1_from_e(unroll_k, nxe):
             gemm_msg = f"gemm_m_per_block:{self.gemm_m_per_block} - {self.wave_tile_m}x{self.wave_step_m}x{self.wave_repeat_m}, gemm_n_per_block:{self.gemm_n_per_block} - {self.wave_tile_n}x{self.wave_step_n}x{self.wave_repeat_n}, gemm_k_per_block:{self.gemm_k_per_block}"
 
         assert self.num_global_load_a * self.block_size == self.gemm_m_per_block * self.gemm_k_per_block, gemm_msg
-        assert self.num_global_load_b * self.block_size == self.gemm_n_per_block * self.gemm_k_per_block, gemm_msg
+        if self.mini_weights != 1:
+            assert self.num_global_load_b * self.block_size == self.gemm_n_per_block * self.gemm_k_per_block, gemm_msg
 
         # LDS size
         self.lds_pad_m, self.lds_pad_n = self.get_lds_pad() # LDS pad
@@ -409,6 +414,11 @@ def _unmerge_x1_from_e(unroll_k, nxe):
             self.lds_total                      = self.lds_buffer_num * self.lds_single
         # print(f"lds_a:{self.lds_a}, lds_b:{self.lds_b}, lds_a_np2:{self.lds_a_np2}, lds_b_np2:{self.lds_b_np2}, lds_single:{self.lds_single}, lds_total:{self.lds_total}")
         # TODO: LDS size check
+
+        if self.mini_weights == 1:
+            self.lds_single            = 8 * 1024
+            self.lds_total             = 8 * 1024
+            self.lds_buffer_num        = 1
 
         # some parameter not in modular_conv
         if self.fma_type in (IGEMM_GTC_TUNABLE_FMA_TYPE_MAC, IGEMM_GTC_TUNABLE_FMA_TYPE_DLOPS):