merge develop

example/ck_tile/01_fmha/README.md

@@ -15,8 +15,7 @@ This will result in an executable `build/bin/tile_example_fmha_fwd`
 ## kernel
 The kernel template is `fmha_fwd_kernel.hpp`, this is the grid-wise op in old ck_tile's terminology. We put it here purposely, to demonstrate one can construct a kernel by using various internal component from ck_tile. We may still have an implementation under ck_tile's include path (in the future) for the kernel template.
 
-There are 3 template parameters for this kernel template.
-* `TilePartitioner` is used to map the workgroup to corresponding tile, `fmha_fwd_tile_partitioner.hpp` in this folder served as this purpose.
+There are 2 template parameters for this kernel template.
 * `FmhaPipeline` is one of the block_tile_pipeline(under `include/ck_tile/tile_program/block_tile_pipeline`) which is a performance critical component. Indeed, we did a lot of optimization and trials to optimize the pipeline and may still workout more performance pipeline and update into that folder. People only need to replace this pipeline type and would be able to enjoy the benefit of different performant implementations (stay tuned for updated pipeline(s)).
 * `EpiloguePipeline` will modify and store out the result in the last phase. People usually will do lot of post-fusion at this stage, so we also abstract this concept. Currently we didn't do much thing at the epilogue stage but leave the room for future possible support.
 

example/ck_tile/01_fmha/codegen/ops/fmha_bwd.py

@@ -506,6 +506,14 @@ def get_bwd_dq_dk_dv_blobs(kernel_filter : Optional[str], receipt, mask_impl) ->
                     cond &= deterministic == "f"
                     if not cond:
                         continue
+            if receipt == 4:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= bias in ['no', 'bias']
+                    cond &= dropout in ['no', 'dropout_wg32',  'dropout_wg16']
+                    cond &= dpad == dvpad
+                    cond &= deterministic == "f"
+                    if not cond:
+                        continue
             api_pool.register_dq_dk_dv_traits(k.api_trait())
             gen.append(k)
 
@@ -801,4 +809,4 @@ def list_blobs(file_path : Path, kernel_filter : Optional[str], receipt, mask_im
         _, kernels = get_bwd_dq_dk_dv_blobs(kernel_filter, receipt, mask_impl)
         for kernel in kernels:
             f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n")
-        f.write(str(file_path.parent / GEN_DIR / FMHA_BWD_API_FILENAME) + "\n")
\ No newline at end of file
+        f.write(str(file_path.parent / GEN_DIR / FMHA_BWD_API_FILENAME) + "\n")

example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py

@@ -29,11 +29,6 @@ K0_MAX_SUBMAX_MAP = {
     256: 256
 }
 
-TILE_PARTITIONER_MAP = {
-    "shb" : "ck_tile::FmhaFwdTilePartitioner_SHB",
-    "hbs" : "ck_tile::FmhaFwdTilePartitioner_HBS",
-}
-
 FMHA_FWD_KERNEL_HEADER = """// SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.\n
 // auto generated by generate.py
@@ -90,9 +85,7 @@ using fmha_epilogue_{F_idx} =
                                            {F_spad}, {F_dvpad}>>;
 
 using fmha_kernel_{F_idx} =
-    ck_tile::FmhaFwdKernel<{F_tile_partitioner}<fmha_shape_{F_idx}>,
-                  fmha_pipeline_{F_idx},
-                  fmha_epilogue_{F_idx}>;
+    ck_tile::FmhaFwdKernel<fmha_pipeline_{F_idx}, fmha_epilogue_{F_idx}>;
 
 using trait_{F_idx} = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode},{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}, {F_vlayout},
                         {F_pipeline_enum}, fmha_mask_{F_idx}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
@@ -329,12 +322,6 @@ class FmhaFwdKernel:
     F_pipeline      : FmhaFwdPipeline
     mask_impl       : str
 
-    def get_tp(self) -> str:
-        if self.F_mode == 'group':
-            return 'hbs'
-        else:
-            return 'shb'
-
     @property
     def template(self) -> str:
         kernel_body = str()
@@ -374,13 +361,12 @@ class FmhaFwdKernel:
                 F_pipeline_enum = PIPELINE_ENUM_MAP[self.F_pipeline.tag],
                 F_mask          = get_mask_map(self.mask_impl)[self.F_pipeline.F_mask],
                 F_mode          = MODE_MAP[self.F_mode],
-                F_pipeline      = PIPELINE_MAP[self.F_pipeline.tag],
-                F_tile_partitioner = TILE_PARTITIONER_MAP[self.get_tp()])
+                F_pipeline      = PIPELINE_MAP[self.F_pipeline.tag])
 
     @property
     def name(self) -> str:
         # TODO: we don't encode idx here
-        return f"fmha_fwd_d{self.F_hdim}_{self.F_dtype}_{self.F_mode}_{self.get_tp()}_" + \
+        return f"fmha_fwd_d{self.F_hdim}_{self.F_dtype}_{self.F_mode}_" + \
                 self.F_tile.name + '_' + self.F_pipeline.name
 
     @property
@@ -501,13 +487,20 @@ def get_fwd_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> Tuple[Fm
                 if kernel_filter != None:
                     if not fnmatch.fnmatch(k.name, kernel_filter):
                         continue
-                if receipt == 2:
+                if receipt in (2, 3):
                     cond = dtype in ['fp16', 'bf16']
                     cond &= pipeline.F_vlayout == 'row'
                     cond &= pipeline.F_bias in ['no', 'alibi']
                     cond &= pipeline.F_squant == 'f'
                     if not cond:
                         continue
+                if receipt == 4:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= pipeline.F_vlayout == 'row'
+                    cond &= pipeline.F_bias in ['no', 'bias']
+                    cond &= pipeline.F_squant == 'f'
+                    if not cond:
+                        continue
                 api_pool.register_traits(k.api_trait())
                 gen.append(k)
 

example/ck_tile/01_fmha/codegen/ops/fmha_fwd_appendkv.py

@@ -46,9 +46,7 @@ using fmha_pipeline_problem_{F_idx} = ck_tile::BlockFmhaFwdAppendKVPipelineProbl
 using fmha_pipeline_{F_idx} = ck_tile::BlockFmhaFwdAppendKVPipeline<
     fmha_pipeline_problem_{F_idx}>;
 
-using fmha_kernel_{F_idx} =
-    ck_tile::FmhaFwdAppendKVKernel<ck_tile::FmhaFwdAppendKVTilePartitioner<{F_bs}, {F_bsk}, {F_bd}, {F_bdv}>,
-                  fmha_pipeline_{F_idx}>;
+using fmha_kernel_{F_idx} = ck_tile::FmhaFwdAppendKVKernel<fmha_pipeline_{F_idx}>;
 
 using trait_{F_idx} = fmha_fwd_appendkv_traits_<{F_hdim}, {F_dtype}, {F_bs}, {F_bsk}, {F_bd}, {F_bdv}, {F_vlayout},
                         {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}, {F_rope}, {F_pagedkv}>;
@@ -355,4 +353,4 @@ def list_blobs(file_path : Path, kernel_filter : Optional[str], receipt, mask_im
         _, kernels = get_fwd_appendkv_blobs(kernel_filter, receipt, mask_impl)
         for kernel in kernels:
             f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n")
-        f.write(str(file_path.parent / GEN_DIR / FMHA_FWD_APPENDKV_API_FILENAME) + "\n")
\ No newline at end of file
+        f.write(str(file_path.parent / GEN_DIR / FMHA_FWD_APPENDKV_API_FILENAME) + "\n")

example/ck_tile/01_fmha/fmha_fwd.cpp

@@ -1131,15 +1131,16 @@ bool run(const ck_tile::ArgParser& arg_parser)
     {
         // NOTE: use gpu to do validation
         ck_tile::naive_attention_fwd_traits naive_t;
-        naive_t.q_type    = data_type;
-        naive_t.k_type    = data_type;
-        naive_t.v_type    = data_type;
-        naive_t.o_type    = data_type;
-        naive_t.q_layout  = i_perm == 1 ? "bhsd" : "bshd";
-        naive_t.k_layout  = i_perm == 1 ? "bhsd" : "bshd";
-        naive_t.v_layout  = i_perm == 1 ? "bhsd" : "bshd";
-        naive_t.o_layout  = o_perm == 1 ? "bhsd" : "bshd";
-        naive_t.variation = 0; // TODO?
+        naive_t.q_type     = data_type;
+        naive_t.k_type     = data_type;
+        naive_t.v_type     = data_type;
+        naive_t.o_type     = data_type;
+        naive_t.q_layout   = i_perm == 1 ? "bhsd" : "bshd";
+        naive_t.k_layout   = i_perm == 1 ? "bhsd" : "bshd";
+        naive_t.v_layout   = i_perm == 1 ? "bhsd" : "bshd";
+        naive_t.o_layout   = o_perm == 1 ? "bhsd" : "bshd";
+        naive_t.variation  = 0; // TODO?
+        naive_t.quant_algo = 0;
 
         ck_tile::DeviceMem o_naive_buf(o_host.get_element_space_size_in_bytes());
 

example/ck_tile/03_gemm/CMakeLists.txt

 add_executable(tile_example_gemm_basic EXCLUDE_FROM_ALL gemm_basic.cpp)
-add_executable(tile_example_universal_gemm EXCLUDE_FROM_ALL universal_gemm.cpp)
+add_executable(tile_example_gemm_universal EXCLUDE_FROM_ALL universal_gemm.cpp)
+target_compile_options(tile_example_gemm_universal PRIVATE
+  -mllvm -enable-noalias-to-md-conversion=0
+)