Enable HDIMDIFF64 (#84)

* Enable HDIMDIFF64

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>

* Add instantiations

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>

---------

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>

Author: MatthewBonanni

CMakeLists.txt

@@ -185,8 +185,8 @@ if (FA3_ENABLED AND ${CMAKE_CUDA_COMPILER_VERSION} GREATER_EQUAL 12.0)
         "hopper/instantiations/flash_fwd_hdim256_bf16*_sm90.cu")
     # Add these for hdim diff cases
     file(GLOB FA3_BF16_GEN_SRCS_
-        # "hopper/instantiations/flash_fwd_hdim64_256_bf16*_sm90.cu"
-        # "hopper/instantiations/flash_fwd_hdim64_512_bf16*_sm90.cu"
+        "hopper/instantiations/flash_fwd_hdim64_256_bf16*_sm90.cu"
+        "hopper/instantiations/flash_fwd_hdim64_512_bf16*_sm90.cu"
         "hopper/instantiations/flash_fwd_hdim192_128_bf16*_sm90.cu")
     list(APPEND FA3_BF16_GEN_SRCS ${FA3_BF16_GEN_SRCS_})
     file(GLOB FA3_BF16_GEN_SRCS_
@@ -202,8 +202,8 @@ if (FA3_ENABLED AND ${CMAKE_CUDA_COMPILER_VERSION} GREATER_EQUAL 12.0)
         "hopper/instantiations/flash_fwd_hdim256_fp16*_sm90.cu")
     # Add these for hdim diff cases
     file(GLOB FA3_FP16_GEN_SRCS_
-    #     "hopper/instantiations/flash_fwd_hdim64_256_fp16*_sm90.cu"
-    #     "hopper/instantiations/flash_fwd_hdim64_512_fp16*_sm90.cu"
+        "hopper/instantiations/flash_fwd_hdim64_256_fp16*_sm90.cu"
+        "hopper/instantiations/flash_fwd_hdim64_512_fp16*_sm90.cu"
         "hopper/instantiations/flash_fwd_hdim192_128_fp16*_sm90.cu")
     list(APPEND FA3_FP16_GEN_SRCS ${FA3_FP16_GEN_SRCS_})
     file(GLOB FA3_FP16_GEN_SRCS_
@@ -272,7 +272,7 @@ if (FA3_ENABLED AND ${CMAKE_CUDA_COMPILER_VERSION} GREATER_EQUAL 12.0)
         FLASHATTENTION_VARLEN_ONLY # Custom flag to save on binary size
         FLASHATTENTION_DISABLE_CLUSTER # disabled for varlen in any case
         # FLASHATTENTION_DISABLE_SM8x
-        FLASHATTENTION_DISABLE_HDIMDIFF64
+        # FLASHATTENTION_DISABLE_HDIMDIFF64
         # FLASHATTENTION_DISABLE_HDIMDIFF192
         CUTE_SM90_EXTENDED_MMA_SHAPES_ENABLED
         CUTLASS_ENABLE_GDC_FOR_SM90