[Mlir-commits] [mlir] a5925ee - [MLIR][NVVM] Fix register mapping in `wgmma.mma_async`
Guray Ozen
llvmlistbot at llvm.org
Mon Aug 14 05:08:55 PDT 2023
Author: Guray Ozen
Date: 2023-08-14T14:08:49+02:00
New Revision: a5925eee5a8cbdf0ee8d77a744c245405987e3b6
URL: https://github.com/llvm/llvm-project/commit/a5925eee5a8cbdf0ee8d77a744c245405987e3b6
DIFF: https://github.com/llvm/llvm-project/commit/a5925eee5a8cbdf0ee8d77a744c245405987e3b6.diff
LOG: [MLIR][NVVM] Fix register mapping in `wgmma.mma_async`
WgmmaMmaAsync Op generates `wgmma.mma_async` PTX instruction that uses the same registers as read and write with mapping. Therefore, the registers count needs to be increased 2 times for the following registers.
This works changes this:
```
llvm.inline_asm has_side_effects asm_dialect = att "{wgmma.mma_async... {$0, $1, $2, $3, $4}, $5, $6, p", "=f,=f,=f,=f,0,1,2,3,l,l"
```
Into this one below. The only different is the number of registers ($8 and $9) that comes after read/write.
```
llvm.inline_asm has_side_effects asm_dialect = att "{wgmma.mma_async... {$0, $1, $2, $3, $4}, $8, $9, p", "=f,=f,=f,=f,0,1,2,3,l,l"
```
Reviewed By: qcolombet
Differential Revision: https://reviews.llvm.org/D157843
Added:
Modified:
mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp
mlir/test/Conversion/NVVMToLLVM/nvvm-to-llvm.mlir
Removed:
################################################################################
diff --git a/mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp b/mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp
index d6794881b442ee..08ecece19c0882 100644
--- a/mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp
+++ b/mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp
@@ -921,17 +921,17 @@ std::string NVVM::WgmmaMmaAsyncOp::getPtx() {
}
ss << "},";
- ss << " $" << (expectedOutputRegisters) << ","
- << " $" << (expectedOutputRegisters + 1) << ","
+ // Need to map read/write registers correctly.
+ regCnt = (regCnt * 2);
+ ss << " $" << (regCnt) << ","
+ << " $" << (regCnt + 1) << ","
<< " p";
if (!outputType.isInteger(32)) {
- ss << ", $" << (expectedOutputRegisters + 3) << ", $"
- << (expectedOutputRegisters + 4);
+ ss << ", $" << (regCnt + 3) << ", $" << (regCnt + 4);
}
// Don't add transpose parameters unless needed.
if (isF16) {
- ss << ", $" << (expectedOutputRegisters + 5) << ", $"
- << (expectedOutputRegisters + 6);
+ ss << ", $" << (regCnt + 5) << ", $" << (regCnt + 6);
}
ss << ";\n"
<< "}\n";
diff --git a/mlir/test/Conversion/NVVMToLLVM/nvvm-to-llvm.mlir b/mlir/test/Conversion/NVVMToLLVM/nvvm-to-llvm.mlir
index 3bcb41c18d4fb4..db2a8b2ab47d74 100644
--- a/mlir/test/Conversion/NVVMToLLVM/nvvm-to-llvm.mlir
+++ b/mlir/test/Conversion/NVVMToLLVM/nvvm-to-llvm.mlir
@@ -124,7 +124,7 @@ func.func @wgmma_f32_f16_f16(%descA : i64, %descB : i64) -> !mat64f32{
// CHECK: %[[V4:.*]] = llvm.extractvalue %[[RES]][4] : !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
// CHECK: %[[V11:.*]] = llvm.extractvalue %[[RES]][11] : !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
// CHECK: %[[V13:.*]] = llvm.extractvalue %[[RES]][13] : !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
- // CHECK: %[[RES1:.+]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $18, 0;\0Awgmma.mma_async.sync.aligned.m64n32k16.f32.f16.f16 {$0, $1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13, $14, $15}, $16, $17, p, $19, $20, $21, $22;\0A}\0A", "=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,l,l,n,n,n,n,n" %[[V0]], %{{.*}}, %{{.*}}, %{{.*}}, %[[V4]], %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %[[V11]], %{{.*}}, %[[V13]], %{{.*}}, %{{.*}}, %[[ARG0]], %[[ARG1]], %[[A1]], %[[A2]], %[[A3]], %[[A4]], %[[A5]] : (f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, i64, i64, i32, i32, i32, i32, i32) -> !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
+ // CHECK: %[[RES1:.+]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $18, 0;\0Awgmma.mma_async.sync.aligned.m64n32k16.f32.f16.f16 {$0, $1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13, $14, $15}, $32, $33, p, $35, $36, $37, $38;\0A}\0A", "=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,l,l,n,n,n,n,n" %[[V0]], %{{.*}}, %{{.*}}, %{{.*}}, %[[V4]], %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %[[V11]], %{{.*}}, %[[V13]], %{{.*}}, %{{.*}}, %[[ARG0]], %[[ARG1]], %[[A1]], %[[A2]], %[[A3]], %[[A4]], %[[A5]] : (f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, i64, i64, i32, i32, i32, i32, i32) -> !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
// CHECK: %[[C2:.*]] = llvm.mlir.constant(2 : i64) : i64
// CHECK: %[[DESCa:.+]] = llvm.add %[[ARG0]], %[[C2]] : i64
// CHECK: %[[DESCb:.+]] = llvm.add %[[ARG1]], %[[C2]] : i64
@@ -132,7 +132,7 @@ func.func @wgmma_f32_f16_f16(%descA : i64, %descB : i64) -> !mat64f32{
// CHECK: %[[V4_2:.*]] = llvm.extractvalue %[[RES1]][4] : !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
// CHECK: %[[V11_2:.*]] = llvm.extractvalue %[[RES1]][11] : !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
// CHECK: %[[V13_2:.*]] = llvm.extractvalue %[[RES1]][13] : !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
- // CHECK: %[[RES_2:.+]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $18, 0;\0Awgmma.mma_async.sync.aligned.m64n32k16.f32.f16.f16 {$0, $1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13, $14, $15}, $16, $17, p, $19, $20, $21, $22;\0A}\0A", "=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,l,l,n,n,n,n,n" %[[V0_2]], %{{.*}}, %{{.*}}, %{{.*}}, %[[V4_2]], %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %[[V11_2]], %{{.*}}, %[[V13_2]], %{{.*}}, %{{.*}}, %[[DESCa]], %[[DESCb]], %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}} : (f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, i64, i64, i32, i32, i32, i32, i32) -> !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
+ // CHECK: %[[RES_2:.+]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $18, 0;\0Awgmma.mma_async.sync.aligned.m64n32k16.f32.f16.f16 {$0, $1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13, $14, $15}, $32, $33, p, $35, $36, $37, $38;\0A}\0A", "=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,l,l,n,n,n,n,n" %[[V0_2]], %{{.*}}, %{{.*}}, %{{.*}}, %[[V4_2]], %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %[[V11_2]], %{{.*}}, %[[V13_2]], %{{.*}}, %{{.*}}, %[[DESCa]], %[[DESCb]], %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}} : (f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, i64, i64, i32, i32, i32, i32, i32) -> !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
%result = llvm.mlir.undef : !mat64f32
%result1 = nvvm.wgmma.mma_async
%descA, %descB,
@@ -168,17 +168,17 @@ func.func @wgmma_s32_s8_s8_satfinite(%descA : i64, %descB : i64) -> !mat16i32{
// CHECK: %[[V1:.*]] = llvm.extractvalue %[[RES]][1]
// CHECK: %[[V2:.*]] = llvm.extractvalue %[[RES]][2]
// CHECK: %[[V3:.*]] = llvm.extractvalue %[[RES]][3]
-// CHECK: %[[RES_2:.*]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $6, 0;\0Awgmma.mma_async.sync.aligned.m64n8k32.s32.s8.s8.satfinite {$0, $1, $2, $3}, $4, $5, p;\0A}\0A", "=r,=r,=r,=r,0,1,2,3,l,l,n" %[[V0]], %[[V1]], %[[V2]], %[[V3]], %[[ARG0]], %[[ARG1]], %[[A1]] : (i32, i32, i32, i32, i64, i64, i32) -> !llvm.struct<(i32, i32, i32, i32)>
+// CHECK: %[[RES_2:.*]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $6, 0;\0Awgmma.mma_async.sync.aligned.m64n8k32.s32.s8.s8.satfinite {$0, $1, $2, $3}, $8, $9, p;\0A}\0A", "=r,=r,=r,=r,0,1,2,3,l,l,n" %[[V0]], %[[V1]], %[[V2]], %[[V3]], %[[ARG0]], %[[ARG1]], %[[A1]] : (i32, i32, i32, i32, i64, i64, i32) -> !llvm.struct<(i32, i32, i32, i32)>
// CHECK: %[[V0_2:.*]] = llvm.extractvalue %[[RES_2]][0]
// CHECK: %[[V1_2:.*]] = llvm.extractvalue %[[RES_2]][1]
// CHECK: %[[V2_2:.*]] = llvm.extractvalue %[[RES_2]][2]
// CHECK: %[[V3_2:.*]] = llvm.extractvalue %[[RES_2]][3]
-// CHECK: %[[RES_3:.*]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $6, 0;\0Awgmma.mma_async.sync.aligned.m64n8k32.s32.s8.s8.satfinite {$0, $1, $2, $3}, $4, $5, p;\0A}\0A", "=r,=r,=r,=r,0,1,2,3,l,l,n" %[[V0_2]], %[[V1_2]], %[[V2_2]], %[[V3_2]], %[[ARG0]], %[[ARG1]], %{{.*}}
+// CHECK: %[[RES_3:.*]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $6, 0;\0Awgmma.mma_async.sync.aligned.m64n8k32.s32.s8.s8.satfinite {$0, $1, $2, $3}, $8, $9, p;\0A}\0A", "=r,=r,=r,=r,0,1,2,3,l,l,n" %[[V0_2]], %[[V1_2]], %[[V2_2]], %[[V3_2]], %[[ARG0]], %[[ARG1]], %{{.*}}
// CHECK: %[[V0_3:.*]] = llvm.extractvalue %[[RES_3]][0]
// CHECK: %[[V1_3:.*]] = llvm.extractvalue %[[RES_3]][1]
// CHECK: %[[V2_3:.*]] = llvm.extractvalue %[[RES_3]][2]
// CHECK: %[[V3_3:.*]] = llvm.extractvalue %[[RES_3]][3]
-// CHECK: %[[RES1:.*]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $6, 0;\0Awgmma.mma_async.sync.aligned.m64n8k32.s32.s8.s8.satfinite {$0, $1, $2, $3}, $4, $5, p;\0A}\0A", "=r,=r,=r,=r,0,1,2,3,l,l,n" %[[V0_3]], %[[V1_3]], %[[V2_3]], %[[V3_3]], %[[ARG0]], %[[ARG1]], %{{.*}}
+// CHECK: %[[RES1:.*]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $6, 0;\0Awgmma.mma_async.sync.aligned.m64n8k32.s32.s8.s8.satfinite {$0, $1, $2, $3}, $8, $9, p;\0A}\0A", "=r,=r,=r,=r,0,1,2,3,l,l,n" %[[V0_3]], %[[V1_3]], %[[V2_3]], %[[V3_3]], %[[ARG0]], %[[ARG1]], %{{.*}}
%result1 = nvvm.wgmma.mma_async %descA, %descB,
#nvvm.shape<m = 64, n = 8, k = 32>,
D [%result, #nvvm.wgmma_scale_out<one>, <satfinite>],
@@ -209,17 +209,17 @@ func.func @wgmma_s32_u8_u8(%descA : i64, %descB : i64) -> !mat16i32 {
// CHECK: %[[V1:.*]] = llvm.extractvalue %[[RES]][1]
// CHECK: %[[V2:.*]] = llvm.extractvalue %[[RES]][2]
// CHECK: %[[V3:.*]] = llvm.extractvalue %[[RES]][3]
-// CHECK: %[[RES_2:.*]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $6, 0;\0Awgmma.mma_async.sync.aligned.m64n8k32.s32.u8.u8 {$0, $1, $2, $3}, $4, $5, p;\0A}\0A", "=r,=r,=r,=r,0,1,2,3,l,l,n" %[[V0]], %[[V1]], %[[V2]], %[[V3]], %[[ARG0]], %[[ARG1]], %[[A1]] : (i32, i32, i32, i32, i64, i64, i32) -> !llvm.struct<(i32, i32, i32, i32)>
+// CHECK: %[[RES_2:.*]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $6, 0;\0Awgmma.mma_async.sync.aligned.m64n8k32.s32.u8.u8 {$0, $1, $2, $3}, $8, $9, p;\0A}\0A", "=r,=r,=r,=r,0,1,2,3,l,l,n" %[[V0]], %[[V1]], %[[V2]], %[[V3]], %[[ARG0]], %[[ARG1]], %[[A1]] : (i32, i32, i32, i32, i64, i64, i32) -> !llvm.struct<(i32, i32, i32, i32)>
// CHECK: %[[V0_2:.*]] = llvm.extractvalue %[[RES_2]][0]
// CHECK: %[[V1_2:.*]] = llvm.extractvalue %[[RES_2]][1]
// CHECK: %[[V2_2:.*]] = llvm.extractvalue %[[RES_2]][2]
// CHECK: %[[V3_2:.*]] = llvm.extractvalue %[[RES_2]][3]
-// CHECK: %[[RES_3:.*]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $6, 0;\0Awgmma.mma_async.sync.aligned.m64n8k32.s32.u8.u8 {$0, $1, $2, $3}, $4, $5, p;\0A}\0A", "=r,=r,=r,=r,0,1,2,3,l,l,n" %[[V0_2]], %[[V1_2]], %[[V2_2]], %[[V3_2]], %[[ARG0]], %[[ARG1]], %{{.*}}
+// CHECK: %[[RES_3:.*]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $6, 0;\0Awgmma.mma_async.sync.aligned.m64n8k32.s32.u8.u8 {$0, $1, $2, $3}, $8, $9, p;\0A}\0A", "=r,=r,=r,=r,0,1,2,3,l,l,n" %[[V0_2]], %[[V1_2]], %[[V2_2]], %[[V3_2]], %[[ARG0]], %[[ARG1]], %{{.*}}
// CHECK: %[[V0_3:.*]] = llvm.extractvalue %[[RES_3]][0]
// CHECK: %[[V1_3:.*]] = llvm.extractvalue %[[RES_3]][1]
// CHECK: %[[V2_3:.*]] = llvm.extractvalue %[[RES_3]][2]
// CHECK: %[[V3_3:.*]] = llvm.extractvalue %[[RES_3]][3]
-// CHECK: %[[RES1:.*]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $6, 0;\0Awgmma.mma_async.sync.aligned.m64n8k32.s32.u8.u8 {$0, $1, $2, $3}, $4, $5, p;\0A}\0A", "=r,=r,=r,=r,0,1,2,3,l,l,n" %[[V0_3]], %[[V1_3]], %[[V2_3]], %[[V3_3]], %[[ARG0]], %[[ARG1]], %{{.*}}
+// CHECK: %[[RES1:.*]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $6, 0;\0Awgmma.mma_async.sync.aligned.m64n8k32.s32.u8.u8 {$0, $1, $2, $3}, $8, $9, p;\0A}\0A", "=r,=r,=r,=r,0,1,2,3,l,l,n" %[[V0_3]], %[[V1_3]], %[[V2_3]], %[[V3_3]], %[[ARG0]], %[[ARG1]], %{{.*}}
%result = llvm.mlir.undef : !mat16i32
%result1 = nvvm.wgmma.mma_async %descA, %descB,
#nvvm.shape<m = 64, n = 8, k = 32>,
@@ -252,8 +252,8 @@ func.func @wgmma_s32_u8_u8(%descA : i64, %descB : i64) -> !mat16i32 {
// CHECK-LABEL: @wgmma_f32_tf32_tf32
func.func @wgmma_f32_tf32_tf32(%descA : i64, %descB : i64) -> !mat32f32 {
- // CHECK: %[[RES:.+]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $34, 0;\0Awgmma.mma_async.sync.aligned.m64n64k8.f32.tf32.tf32 {$0, $1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13, $14, $15, $16, $17, $18, $19, $20, $21, $22, $23, $24, $25, $26, $27, $28, $29, $30, $31}, $32, $33, p, $35, $36;\0A}\0A", "=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,l,l,n,n,n"
- // CHECK: %[[RES_2:.+]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $34, 0;\0Awgmma.mma_async.sync.aligned.m64n64k8.f32.tf32.tf32 {$0, $1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13, $14, $15, $16, $17, $18, $19, $20, $21, $22, $23, $24, $25, $26, $27, $28, $29, $30, $31}, $32, $33, p, $35, $36;\0A}\0A", "=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,l,l,n,n,n"
+ // CHECK: %[[RES:.+]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $34, 0;\0Awgmma.mma_async.sync.aligned.m64n64k8.f32.tf32.tf32 {$0, $1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13, $14, $15, $16, $17, $18, $19, $20, $21, $22, $23, $24, $25, $26, $27, $28, $29, $30, $31}, $64, $65, p, $67, $68;\0A}\0A", "=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,l,l,n,n,n"
+ // CHECK: %[[RES_2:.+]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $34, 0;\0Awgmma.mma_async.sync.aligned.m64n64k8.f32.tf32.tf32 {$0, $1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13, $14, $15, $16, $17, $18, $19, $20, $21, $22, $23, $24, $25, $26, $27, $28, $29, $30, $31}, $64, $65, p, $67, $68;\0A}\0A", "=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,l,l,n,n,n"
%result = llvm.mlir.undef : !mat32f32
%result1 = nvvm.wgmma.mma_async %descA, %descB,
#nvvm.shape<m = 64, n = 64, k = 8>,
@@ -281,8 +281,8 @@ func.func @wgmma_f32_tf32_tf32(%descA : i64, %descB : i64) -> !mat32f32 {
// CHECK-LABEL: @wgmma_f32_e4m3_e4m3
func.func @wgmma_f32_e4m3_e4m3(%descA : i64, %descB : i64) -> !mat32f32 {
- // CHECK: %[[RES:.+]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $34, 0;\0Awgmma.mma_async.sync.aligned.m64n64k32.f32.e4m3.e4m3 {$0, $1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13, $14, $15, $16, $17, $18, $19, $20, $21, $22, $23, $24, $25, $26, $27, $28, $29, $30, $31}, $32, $33, p, $35, $36;\0A}\0A", "=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,l,l,n,n,n"
- // CHECK: %[[RES_2:.+]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $34, 0;\0Awgmma.mma_async.sync.aligned.m64n64k32.f32.e4m3.e4m3 {$0, $1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13, $14, $15, $16, $17, $18, $19, $20, $21, $22, $23, $24, $25, $26, $27, $28, $29, $30, $31}, $32, $33, p, $35, $36;\0A}\0A", "=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,l,l,n,n,n"
+ // CHECK: %[[RES:.+]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $34, 0;\0Awgmma.mma_async.sync.aligned.m64n64k32.f32.e4m3.e4m3 {$0, $1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13, $14, $15, $16, $17, $18, $19, $20, $21, $22, $23, $24, $25, $26, $27, $28, $29, $30, $31}, $64, $65, p, $67, $68;\0A}\0A", "=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,l,l,n,n,n"
+ // CHECK: %[[RES_2:.+]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $34, 0;\0Awgmma.mma_async.sync.aligned.m64n64k32.f32.e4m3.e4m3 {$0, $1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13, $14, $15, $16, $17, $18, $19, $20, $21, $22, $23, $24, $25, $26, $27, $28, $29, $30, $31}, $64, $65, p, $67, $68;\0A}\0A", "=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,l,l,n,n,n"
%result = llvm.mlir.undef : !mat32f32
%result1 = nvvm.wgmma.mma_async %descA, %descB,
#nvvm.shape<m = 64, n = 64, k = 32>,
@@ -309,8 +309,8 @@ func.func @wgmma_f32_e4m3_e4m3(%descA : i64, %descB : i64) -> !mat32f32 {
// CHECK-LABEL: @wgmma_f32_e5m2_e4m3
func.func @wgmma_f32_e5m2_e4m3(%descA : i64, %descB : i64) -> !mat32f32 {
- // CHECK: %[[RES:.+]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $34, 0;\0Awgmma.mma_async.sync.aligned.m64n64k32.f32.e5m2.e4m3 {$0, $1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13, $14, $15, $16, $17, $18, $19, $20, $21, $22, $23, $24, $25, $26, $27, $28, $29, $30, $31}, $32, $33, p, $35, $36;\0A}\0A", "=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,l,l,n,n,n"
- // CHECK: %[[RES_2:.+]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $34, 0;\0Awgmma.mma_async.sync.aligned.m64n64k32.f32.e5m2.e4m3 {$0, $1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13, $14, $15, $16, $17, $18, $19, $20, $21, $22, $23, $24, $25, $26, $27, $28, $29, $30, $31}, $32, $33, p, $35, $36;\0A}\0A", "=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,l,l,n,n,n"
+ // CHECK: %[[RES:.+]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $34, 0;\0Awgmma.mma_async.sync.aligned.m64n64k32.f32.e5m2.e4m3 {$0, $1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13, $14, $15, $16, $17, $18, $19, $20, $21, $22, $23, $24, $25, $26, $27, $28, $29, $30, $31}, $64, $65, p, $67, $68;\0A}\0A", "=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,l,l,n,n,n"
+ // CHECK: %[[RES_2:.+]] = llvm.inline_asm has_side_effects asm_dialect = att "{\0A.reg .pred p;\0Asetp.ne.b32 p, $34, 0;\0Awgmma.mma_async.sync.aligned.m64n64k32.f32.e5m2.e4m3 {$0, $1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13, $14, $15, $16, $17, $18, $19, $20, $21, $22, $23, $24, $25, $26, $27, $28, $29, $30, $31}, $64, $65, p, $67, $68;\0A}\0A", "=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,=f,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,l,l,n,n,n"
%result = llvm.mlir.undef : !mat32f32
%result1 = nvvm.wgmma.mma_async %descA, %descB,
#nvvm.shape<m = 64, n = 64, k = 32>,
More information about the Mlir-commits
mailing list