[MLIR][NVVM] Support for dense and sparse MMA with block scaling (#170566)

This change adds dense and sparse MMA with block scaling intrinsics to
MLIR -> NVVM IR -> NVPTX flow. NVVM and NVPTX implementation is based on
PTX ISA 9.0.

Author: kvederni

mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td

@@ -44,7 +44,7 @@ llvm.func @nvvm_tcgen05_mma_ashift(%d_tmem : !llvm.ptr<6>, %a_tmem: !llvm.ptr<6>
 llvm.func @nvvm_tcgen05_mma_mxf4nvf4_block_scale_default(%d_tmem : !llvm.ptr<6>, %a_tmem: !llvm.ptr<6>, %adesc: i64, %b_desc: i64, %idesc: i32, %enable_input_d: i1, %scalea: !llvm.ptr<6>, %scaleb: !llvm.ptr<6>) {
   // expected-error @below {{mxf4nvf4 requires block scale attribute}}
   nvvm.tcgen05.mma.block_scale %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d, %scalea, %scaleb
-  {kind = #nvvm.tcgen05_mma_block_scale_kind<mxf4nvf4>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, !llvm.ptr<6>, !llvm.ptr<6>)
+  {kind = #nvvm.block_scale_kind<mxf4nvf4>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, !llvm.ptr<6>, !llvm.ptr<6>)
   llvm.return
 }
 
@@ -54,7 +54,7 @@ llvm.func @nvvm_tcgen05_mma_mxf4nvf4_block_scale_default(%d_tmem : !llvm.ptr<6>,
 llvm.func @nvvm_tcgen05_mma_mxf4_block_scale_default(%d_tmem : !llvm.ptr<6>, %a_tmem: !llvm.ptr<6>, %adesc: i64, %b_desc: i64, %idesc: i32, %enable_input_d: i1, %scalea: !llvm.ptr<6>, %scaleb: !llvm.ptr<6>) {
   // expected-error @below {{mxf4 kind does not support block16 attribute}}
   nvvm.tcgen05.mma.block_scale %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d, %scalea, %scaleb
-  {kind = #nvvm.tcgen05_mma_block_scale_kind<mxf4>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>, ashift, blockScale = #nvvm.tcgen05_mma_block_scale<block16>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, !llvm.ptr<6>, !llvm.ptr<6>)
+  {kind = #nvvm.block_scale_kind<mxf4>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>, ashift, blockScale = #nvvm.tcgen05_mma_block_scale<block16>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, !llvm.ptr<6>, !llvm.ptr<6>)
   llvm.return
 }
 
@@ -104,7 +104,7 @@ llvm.func @nvvm_tcgen05_mma_sp_ashift(%d_tmem : !llvm.ptr<6>, %a_tmem: !llvm.ptr
 llvm.func @nvvm_tcgen05_mma_sp_mxf4nvf4_block_scale_default(%d_tmem : !llvm.ptr<6>, %a_tmem: !llvm.ptr<6>, %adesc: i64, %b_desc: i64, %idesc: i32, %enable_input_d: i1, %scalea: !llvm.ptr<6>, %scaleb: !llvm.ptr<6>, %spmetadata: !llvm.ptr<6>) {
   // expected-error @below {{mxf4nvf4 requires block scale attribute}}
   nvvm.tcgen05.mma.sp.block_scale %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d, %spmetadata, %scalea, %scaleb
-  {kind = #nvvm.tcgen05_mma_block_scale_kind<mxf4nvf4>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, !llvm.ptr<6>, !llvm.ptr<6>, !llvm.ptr<6>)
+  {kind = #nvvm.block_scale_kind<mxf4nvf4>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, !llvm.ptr<6>, !llvm.ptr<6>, !llvm.ptr<6>)
   llvm.return
 }
 
@@ -114,6 +114,6 @@ llvm.func @nvvm_tcgen05_mma_sp_mxf4nvf4_block_scale_default(%d_tmem : !llvm.ptr<
 llvm.func @nvvm_tcgen05_mma_sp_mxf4_block_scale_default(%d_tmem : !llvm.ptr<6>, %a_tmem: !llvm.ptr<6>, %adesc: i64, %b_desc: i64, %idesc: i32, %enable_input_d: i1, %scalea: !llvm.ptr<6>, %scaleb: !llvm.ptr<6>, %spmetadata: !llvm.ptr<6>) {
   // expected-error @below {{mxf4 kind does not support block16 attribute}}
   nvvm.tcgen05.mma.sp.block_scale %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d, %spmetadata, %scalea, %scaleb
-  {kind = #nvvm.tcgen05_mma_block_scale_kind<mxf4>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>, ashift, blockScale = #nvvm.tcgen05_mma_block_scale<block16>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, !llvm.ptr<6>, !llvm.ptr<6>, !llvm.ptr<6>)
+  {kind = #nvvm.block_scale_kind<mxf4>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>, ashift, blockScale = #nvvm.tcgen05_mma_block_scale<block16>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, !llvm.ptr<6>, !llvm.ptr<6>, !llvm.ptr<6>)
   llvm.return
 }