diff options
Diffstat (limited to 'mlir/test/Dialect/AMDGPU/invalid.mlir')
| -rw-r--r-- | mlir/test/Dialect/AMDGPU/invalid.mlir | 144 |
1 files changed, 111 insertions, 33 deletions
diff --git a/mlir/test/Dialect/AMDGPU/invalid.mlir b/mlir/test/Dialect/AMDGPU/invalid.mlir index 5784764..6308ea9 100644 --- a/mlir/test/Dialect/AMDGPU/invalid.mlir +++ b/mlir/test/Dialect/AMDGPU/invalid.mlir @@ -156,14 +156,6 @@ func.func @wmma_no_k_dim(%arg0 : vector<16xi8>, %arg1 : vector<8xi32>) -> vector // ----- -func.func @wmma_wrong_m_dim(%arg0 : vector<16xi8>, %arg1 : vector<8xi32>) -> vector<8xi32> { - // expected-error@+1 {{'amdgpu.wmma' op attribute 'm' failed to satisfy constraint: 32-bit signless integer attribute whose value is one of {16}}} - %0 = amdgpu.wmma 32x16x16 %arg0 * %arg0 + %arg1 : vector<16xi8>, vector<16xi8>, vector<8xi32> - func.return %0 : vector<8xi32> -} - -// ----- - func.func @wmma_wrong_n_dim(%arg0 : vector<16xi8>, %arg1 : vector<8xi32>) -> vector<8xi32> { // expected-error@+1 {{'amdgpu.wmma' op attribute 'n' failed to satisfy constraint: 32-bit signless integer attribute whose value is one of {16}}} %0 = amdgpu.wmma 16x32x16 %arg0 * %arg0 + %arg1 : vector<16xi8>, vector<16xi8>, vector<8xi32> @@ -173,14 +165,62 @@ func.func @wmma_wrong_n_dim(%arg0 : vector<16xi8>, %arg1 : vector<8xi32>) -> vec // ----- func.func @wmma_wrong_k_dim(%arg0 : vector<16xi8>, %arg1 : vector<8xi32>) -> vector<8xi32> { - // expected-error@+1 {{'amdgpu.wmma' op attribute 'k' failed to satisfy constraint: 32-bit signless integer attribute whose value is one of {16, 32}}} + // expected-error@+1 {{'amdgpu.wmma' op attribute 'k' failed to satisfy constraint: 32-bit signless integer attribute whose value is one of {4, 16, 32, 64, 128}}} %0 = amdgpu.wmma 16x16x24 %arg0 * %arg0 + %arg1 : vector<16xi8>, vector<16xi8>, vector<8xi32> func.return %0 : vector<8xi32> } // ----- -// Missinng `resetOffset` +func.func @wmma_source_length_mismatch(%arg0 : vector<8xf16>, %arg1 : vector<16xf16>, %arg2 : vector<8xf32>) -> vector<8xf32> { + // expected-error@+1 {{'amdgpu.wmma' op source vectors have different lengths}} + %0 = amdgpu.wmma 16x16x16 %arg0 * %arg1 + %arg2 : vector<8xf16>, vector<16xf16>, vector<8xf32> + func.return %0 : vector<8xf32> +} + +// ----- + +func.func @wmma_mismatched_float_types(%arg0 : vector<8xf16>, %arg1 : vector<8xbf16>, %arg2 : vector<8xf32>) -> vector<8xf32> { + // expected-error@+1 {{'amdgpu.wmma' op source element types must match (except for fp8/bf8)}} + %0 = amdgpu.wmma 16x16x16 %arg0 * %arg1 + %arg2 : vector<8xf16>, vector<8xbf16>, vector<8xf32> + func.return %0 : vector<8xf32> +} + +// ----- + +func.func @wmma_mismatched_int_types(%arg0 : vector<8xi8>, %arg1 : vector<8xi4>, %arg2 : vector<8xi32>) -> vector<8xi32> { + // expected-error@+1 {{'amdgpu.wmma' op source element types must match (except for fp8/bf8)}} + %0 = amdgpu.wmma 16x16x16 %arg0 * %arg1 + %arg2 : vector<8xi8>, vector<8xi4>, vector<8xi32> + func.return %0 : vector<8xi32> +} + +// ----- + +func.func @wmma_clamp_float(%arg0 : vector<8xf16>, %arg1 : vector<8xf32>) -> vector<8xf32> { + // expected-error@+1 {{'amdgpu.wmma' op clamp flag is not supported for float types}} + %0 = amdgpu.wmma 16x16x16 %arg0 * %arg0 + %arg1 {clamp} : vector<8xf16>, vector<8xf16>, vector<8xf32> + func.return %0 : vector<8xf32> +} + +// ----- + +func.func @wmma_unsignedA_float(%arg0 : vector<8xf16>, %arg1 : vector<8xf32>) -> vector<8xf32> { + // expected-error@+1 {{'amdgpu.wmma' op unsigned flags are not supported for float types}} + %0 = amdgpu.wmma 16x16x16 %arg0 * %arg0 + %arg1 {unsignedA} : vector<8xf16>, vector<8xf16>, vector<8xf32> + func.return %0 : vector<8xf32> +} + +// ----- + +func.func @wmma_unsignedB_float(%arg0 : vector<8xf16>, %arg1 : vector<8xf32>) -> vector<8xf32> { + // expected-error@+1 {{'amdgpu.wmma' op unsigned flags are not supported for float types}} + %0 = amdgpu.wmma 16x16x16 %arg0 * %arg0 + %arg1 {unsignedB} : vector<8xf16>, vector<8xf16>, vector<8xf32> + func.return %0 : vector<8xf32> +} + +// ----- + +// Missing `resetOffset` func.func @fat_raw_buffer_cast_stripped_offset(%m: memref<8xi32, strided<[1], offset: ?>, #gpu.address_space<global>>) -> memref<8xi32, #amdgpu.address_space<fat_raw_buffer>> { // expected-error@+1 {{'amdgpu.fat_raw_buffer_cast' op expected result type to be 'memref<8xi32, strided<[1], offset: ?>, #amdgpu.address_space<fat_raw_buffer>>' but got 'memref<8xi32, #amdgpu.address_space<fat_raw_buffer>>'}} %ret = amdgpu.fat_raw_buffer_cast %m : memref<8xi32, strided<[1], offset: ?>, #gpu.address_space<global>> to memref<8xi32, #amdgpu.address_space<fat_raw_buffer>> @@ -293,48 +333,86 @@ func.func @gather_to_lds_non_lds(%idx1 : index, %mem1 : memref<32xf16>, %mem2 : // ----- -func.func @amdgpu.scaled_ext_packed816_invalid_block_size_and_first_scale_byte_16(%v: vector<8xf8E5M2>, %scale: vector<4xf8E8M0FNU>) { - // expected-error@+1 {{'amdgpu.scaled_ext_packed816' op blockSize of 16 can only have firstScaleByte be 0 or 1.}} - %ret0 = amdgpu.scaled_ext_packed816 %v scale(%scale) blockSize(16) firstScaleLane(0) firstScaleByte(2) : vector<8xf8E5M2>, vector<4xf8E8M0FNU> -> vector<8xf16> - func.return +func.func @scaled_mfma_invalid_m(%arg0 : vector<4xf8E8M0FNU>, %arg1 : vector<32xf4E2M1FN>, %arg2 : vector<16xf32>) -> vector<16xf32> { + // expected-error@+1 {{'amdgpu.scaled_mfma' op attribute 'm' failed to satisfy constraint: 32-bit signless integer attribute whose value is one of {16, 32}}} + %0 = amdgpu.scaled_mfma 8x32x64 (%arg0[0] * %arg1) * (%arg0[1] * %arg1) + %arg2 : vector<4xf8E8M0FNU>, vector<32xf4E2M1FN>, vector<4xf8E8M0FNU>, vector<32xf4E2M1FN>, vector<16xf32> + func.return %0 : vector<16xf32> } // ----- -func.func @amdgpu.scaled_ext_packed816_invalid_block_size_and_first_scale_byte_32(%v: vector<8xf8E5M2>, %scale: vector<4xf8E8M0FNU>) { - // expected-error@+1 {{'amdgpu.scaled_ext_packed816' op blockSize of 32 can only have firstScaleByte be 0 or 2.}} - %ret0 = amdgpu.scaled_ext_packed816 %v scale(%scale) blockSize(32) firstScaleLane(0) firstScaleByte(1) : vector<8xf8E5M2>, vector<4xf8E8M0FNU> -> vector<8xf16> - func.return +func.func @scaled_mfma_invalid_n(%arg0 : vector<4xf8E8M0FNU>, %arg1 : vector<32xf4E2M1FN>, %arg2 : vector<16xf32>) -> vector<16xf32> { + // expected-error@+1 {{'amdgpu.scaled_mfma' op attribute 'n' failed to satisfy constraint: 32-bit signless integer attribute whose value is one of {16, 32}}} + %0 = amdgpu.scaled_mfma 32x8x64 (%arg0[0] * %arg1) * (%arg0[1] * %arg1) + %arg2 : vector<4xf8E8M0FNU>, vector<32xf4E2M1FN>, vector<4xf8E8M0FNU>, vector<32xf4E2M1FN>, vector<16xf32> + func.return %0 : vector<16xf32> +} + +// ----- + +func.func @scaled_mfma_invalid_k(%arg0 : vector<4xf8E8M0FNU>, %arg1 : vector<32xf4E2M1FN>, %arg2 : vector<16xf32>) -> vector<16xf32> { + // expected-error@+1 {{'amdgpu.scaled_mfma' op attribute 'k' failed to satisfy constraint: 32-bit signless integer attribute whose value is one of {64, 128}}} + %0 = amdgpu.scaled_mfma 32x32x32 (%arg0[0] * %arg1) * (%arg0[1] * %arg1) + %arg2 : vector<4xf8E8M0FNU>, vector<32xf4E2M1FN>, vector<4xf8E8M0FNU>, vector<32xf4E2M1FN>, vector<16xf32> + func.return %0 : vector<16xf32> +} + +// ----- + +func.func @make_dma_base_invalid_addressspace(%idx: index, %mem: memref<8xi32>) { + // expected-error@+1 {{'amdgpu.make_dma_base' op lds memref must have workgroup address space attribute.}} + amdgpu.make_dma_base %mem[%idx], %mem[%idx] : memref<8xi32>, memref<8xi32> -> !amdgpu.tdm_base<i32> +} + +// ----- + +func.func @make_dma_base_invalid_addressspace(%idx: index, %smem : memref<8xi32, #gpu.address_space<workgroup>>) { + // expected-error@+1 {{'amdgpu.make_dma_base' op global memref must have global address space attribute.}} + amdgpu.make_dma_base %smem[%idx], %smem[%idx] : memref<8xi32, #gpu.address_space<workgroup>>, memref<8xi32, #gpu.address_space<workgroup>> -> !amdgpu.tdm_base<i32> } // ----- -func.func @amdgpu.scaled_ext_packed816_invalid_input_output_sizes(%v: vector<8xf8E5M2>, %scale: vector<4xf8E8M0FNU>) { - // expected-error@+1 {{'amdgpu.scaled_ext_packed816' op failed to verify that all of {source, res} have same shape}} - %ret0 = amdgpu.scaled_ext_packed816 %v scale(%scale) blockSize(16) firstScaleLane(0) firstScaleByte(0) : vector<8xf8E5M2>, vector<4xf8E8M0FNU> -> vector<16xf16> +func.func @make_dma_base_invalid_barrier(%base: !amdgpu.tdm_base<i32>, %barrier: memref<8xi32>, %idx: index) { + // expected-error@+1 {{'amdgpu.make_dma_descriptor' op atomic barrier address must be in LDS.}} + amdgpu.make_dma_descriptor %base globalSize [64, 64] globalStride [64, 1] sharedSize [64, 64] atomicBarrier(%barrier[%idx] : memref<8xi32>) : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor +} + +// ----- + +// CHECK-LABEL: func @make_dma_descriptor_invalid_empty_strides +// CHECK-SAME: (%[[BASE:.+]]: !amdgpu.tdm_base<i32>) +func.func @make_dma_descriptor_invalid_empty_strides(%base: !amdgpu.tdm_base<i32>) { + // expected-error@+1 {{'amdgpu.make_dma_descriptor' op strides must not be empty.}} + amdgpu.make_dma_descriptor %base globalSize [0, 1] globalStride [] sharedSize [1, 0] : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor func.return } // ----- -func.func @scaled_mfma_invalid_m(%arg0 : vector<4xf8E8M0FNU>, %arg1 : vector<32xf4E2M1FN>, %arg2 : vector<16xf32>) -> vector<16xf32> { - // expected-error@+1 {{'amdgpu.scaled_mfma' op attribute 'm' failed to satisfy constraint: 32-bit signless integer attribute whose value is one of {16, 32}}} - %0 = amdgpu.scaled_mfma 8x32x64 (%arg0[0] * %arg1) * (%arg0[1] * %arg1) + %arg2 : vector<4xf8E8M0FNU>, vector<32xf4E2M1FN>, vector<4xf8E8M0FNU>, vector<32xf4E2M1FN>, vector<16xf32> - func.return %0 : vector<16xf32> +// CHECK-LABEL: func @make_dma_descriptor_invalid_innermost_stride +// CHECK-SAME: (%[[BASE:.+]]: !amdgpu.tdm_base<i32>) +func.func @make_dma_descriptor_invalid_innermost_stride(%base: !amdgpu.tdm_base<i32>) { + // expected-error@+1 {{'amdgpu.make_dma_descriptor' op strides for the innermost dimension must be 1.}} + amdgpu.make_dma_descriptor %base globalSize [2, 2] globalStride [1, 2] sharedSize [1, 0] : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor + func.return } // ----- -func.func @scaled_mfma_invalid_n(%arg0 : vector<4xf8E8M0FNU>, %arg1 : vector<32xf4E2M1FN>, %arg2 : vector<16xf32>) -> vector<16xf32> { - // expected-error@+1 {{'amdgpu.scaled_mfma' op attribute 'n' failed to satisfy constraint: 32-bit signless integer attribute whose value is one of {16, 32}}} - %0 = amdgpu.scaled_mfma 32x8x64 (%arg0[0] * %arg1) * (%arg0[1] * %arg1) + %arg2 : vector<4xf8E8M0FNU>, vector<32xf4E2M1FN>, vector<4xf8E8M0FNU>, vector<32xf4E2M1FN>, vector<16xf32> - func.return %0 : vector<16xf32> +// CHECK-LABEL: func @make_dma_descriptor_invalid_size_and_stride_sizes +// CHECK-SAME: (%[[BASE:.+]]: !amdgpu.tdm_base<i32>) +func.func @make_dma_descriptor_invalid_size_and_stride_sizes(%base: !amdgpu.tdm_base<i32>) { + // expected-error@+1 {{'amdgpu.make_dma_descriptor' op strides and sizes must have same rank.}} + amdgpu.make_dma_descriptor %base globalSize [1, 1, 1] globalStride [1, 1] sharedSize [1, 0] : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor + func.return } // ----- -func.func @scaled_mfma_invalid_k(%arg0 : vector<4xf8E8M0FNU>, %arg1 : vector<32xf4E2M1FN>, %arg2 : vector<16xf32>) -> vector<16xf32> { - // expected-error@+1 {{'amdgpu.scaled_mfma' op attribute 'k' failed to satisfy constraint: 32-bit signless integer attribute whose value is one of {64, 128}}} - %0 = amdgpu.scaled_mfma 32x32x32 (%arg0[0] * %arg1) * (%arg0[1] * %arg1) + %arg2 : vector<4xf8E8M0FNU>, vector<32xf4E2M1FN>, vector<4xf8E8M0FNU>, vector<32xf4E2M1FN>, vector<16xf32> - func.return %0 : vector<16xf32> +// CHECK-LABEL: func @make_dma_descriptor_invalid_shared_and_global_rank +// CHECK-SAME: (%[[BASE:.+]]: !amdgpu.tdm_base<i32>) +func.func @make_dma_descriptor_invalid_shared_and_global_rank(%base: !amdgpu.tdm_base<i32>) { + // expected-error@+1 {{'amdgpu.make_dma_descriptor' op tensor must have same rank as tile.}} + amdgpu.make_dma_descriptor %base globalSize [4, 4] globalStride [1, 1] sharedSize [1, 2, 3] : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor + func.return } + |