repo_name
stringlengths
1
62
dataset
stringclasses
1 value
lang
stringclasses
11 values
pr_id
int64
1
20.1k
owner
stringlengths
2
34
reviewer
stringlengths
2
39
diff_hunk
stringlengths
15
262k
code_review_comment
stringlengths
1
99.6k
intel-xpu-backend-for-triton
github_2023
cpp
1,859
intel
whitneywhtsang
@@ -387,9 +387,59 @@ createBlock2DReadWithAddressPayloadUpdate(TritonGEN::Matrix2DBlockLoadOp op, attrs); }; + auto createBlock2DReadGenISA = [&](Value ptr, + TritonGEN::Matrix2DBlockLoadOp op) { + assert(isa<LLVM::LLVMPointerType>(ptr.ge...
not used. ```suggestion ```
intel-xpu-backend-for-triton
github_2023
cpp
1,859
intel
whitneywhtsang
@@ -387,9 +387,59 @@ createBlock2DReadWithAddressPayloadUpdate(TritonGEN::Matrix2DBlockLoadOp op, attrs); }; + auto createBlock2DReadGenISA = [&](Value ptr, + TritonGEN::Matrix2DBlockLoadOp op) { + assert(isa<LLVM::LLVMPointerType>(ptr.ge...
[optional] ```suggestion auto vecType = dyn_cast<VectorType>(resType); assert(vecType && vecType.getShape().size() == 1 && "Expecting a 1D vector"); ```
intel-xpu-backend-for-triton
github_2023
cpp
1,859
intel
whitneywhtsang
@@ -387,9 +387,59 @@ createBlock2DReadWithAddressPayloadUpdate(TritonGEN::Matrix2DBlockLoadOp op, attrs); }; + auto createBlock2DReadGenISA = [&](Value ptr, + TritonGEN::Matrix2DBlockLoadOp op) { + assert(isa<LLVM::LLVMPointerType>(ptr.ge...
I would inline all these variables, as there is no extra info convey with them, and they are not used more than once.
intel-xpu-backend-for-triton
github_2023
cpp
1,835
intel
etiotto
@@ -164,7 +164,160 @@ LogicalResult TritonGEN::MatrixDPASOp::verify() { // gen.2Dblockload //===----------------------------------------------------------------------===// +static LogicalResult +verify2DBlockLoadHWRestriction(TritonGEN::Matrix2DBlockLoadOp op) { + VectorType resTy = op.getRes().getType(); + unsig...
I rather have this early exit (when both transpose and transform are true).
intel-xpu-backend-for-triton
github_2023
cpp
1,835
intel
etiotto
@@ -164,7 +164,160 @@ LogicalResult TritonGEN::MatrixDPASOp::verify() { // gen.2Dblockload //===----------------------------------------------------------------------===// +static LogicalResult +verify2DBlockLoadHWRestriction(TritonGEN::Matrix2DBlockLoadOp op) { + VectorType resTy = op.getRes().getType(); + unsig...
assert(op.getVnniTransform() && !op.getTranspose() && ...)
intel-xpu-backend-for-triton
github_2023
cpp
1,811
intel
LiyangLingIntel
@@ -425,54 +426,59 @@ class TritonIntelGPURewriteTensorPointerPass "Expecting LoadOp/StoreOp ptr in rewritedInfo"); auto info = rewritedInfo[ptr]; - // Load/store with tensor pointers implicitly will check the bound while - // accessing memory, so we should set `mask` and `other` (according to ...
```suggestion "Expecting StoreOp ptr in rewritedInfo"); ```
intel-xpu-backend-for-triton
github_2023
cpp
1,811
intel
LiyangLingIntel
@@ -425,54 +426,59 @@ class TritonIntelGPURewriteTensorPointerPass "Expecting LoadOp/StoreOp ptr in rewritedInfo");
```suggestion "Expecting LoadOp ptr in rewritedInfo"); ```
intel-xpu-backend-for-triton
github_2023
cpp
1,811
intel
whitneywhtsang
@@ -666,36 +672,36 @@ class TritonIntelGPURewriteTensorPointerPass Operation *rewriteOp(Operation *op, std::stack<Operation *> &eraser) { OpBuilder builder(op); - // Rewrite `make_tensor_ptr` and `advance` and make a tensor of pointers - // Rewriting functions return the next operation to visit, if ther...
would something like below works? ``` return TypeSwitch<Operation *, Operation *>(op) .Case<tt::MakeTensorPtrOp, tt::AdvanceOp, ...>([&](auto op) { return rewriteOp(builder, op, eraser); } ... ```
intel-xpu-backend-for-triton
github_2023
python
1,805
intel
chengjunlu
@@ -102,11 +93,13 @@ def extract_kernels(funcs): times = torch.tensor([sum([k.duration for k in ks]) * 1e-3 for ks in kernels], dtype=torch.float) if quantiles is not None: ret = torch.quantile(times, torch.tensor(quantiles, dtype=torch.float)).tolist() + # calculate average time exclude max a...
Is it better to align the samples for the std and cv to the avg? Then all the statistics are based on the same set of the samples.
intel-xpu-backend-for-triton
github_2023
python
1,805
intel
etiotto
@@ -53,17 +53,8 @@ def do_bench(fn, warmup=25, rep=100, grad_to_none=None, quantiles=None, fast_flu # Estimate the runtime of the function start_event = torch.xpu.Event(enable_timing=True) end_event = torch.xpu.Event(enable_timing=True) - start_event.record()
I'd like to understand the reason you are changing this code. What is the motivation/rationale ?
intel-xpu-backend-for-triton
github_2023
python
1,805
intel
etiotto
@@ -7,100 +7,15 @@ """ -# %%
What is the reason to remove the comments ?
intel-xpu-backend-for-triton
github_2023
python
1,805
intel
etiotto
@@ -102,11 +93,13 @@ def extract_kernels(funcs): times = torch.tensor([sum([k.duration for k in ks]) * 1e-3 for ks in kernels], dtype=torch.float) if quantiles is not None: ret = torch.quantile(times, torch.tensor(quantiles, dtype=torch.float)).tolist() + # calculate average time exclude max a...
why are we using the average rather than the mean ?
intel-xpu-backend-for-triton
github_2023
cpp
1,805
intel
whitneywhtsang
@@ -385,8 +385,97 @@ createBlock2DReadWithAddressPayloadUpdate(TritonGEN::Matrix2DBlockLoadOp op, attrs); }; + auto createBlock2DReadGenIsa = [&](Value ptr, + TritonGEN::Matrix2DBlockLoadOp op) { + assert(isa<LLVM::LLVMPointerType>(ptr.ge...
Can we do this in a separate PR with lit test?
intel-xpu-backend-for-triton
github_2023
python
1,805
intel
whitneywhtsang
@@ -64,6 +64,7 @@ def do_bench(fn, warmup=25, rep=100, grad_to_none=None, quantiles=None, fast_flu # compute number of warmup and repeat n_warmup = max(warmup, int(warmup / estimate_ms)) n_repeat = max(rep, int(rep / estimate_ms)) +
```suggestion ```
intel-xpu-backend-for-triton
github_2023
python
1,805
intel
whitneywhtsang
@@ -102,10 +103,17 @@ def extract_kernels(funcs): times = torch.tensor([sum([k.duration for k in ks]) * 1e-3 for ks in kernels], dtype=torch.float) if quantiles is not None: ret = torch.quantile(times, torch.tensor(quantiles, dtype=torch.float)).tolist() - # add coefficient of the variance. - ...
```suggestion if (times.numel() > 2): # exclude max and min times times = torch.sort(times).values[1:-1] std = torch.std(times) mean = torch.mean(times) cv = std / mean ```
intel-xpu-backend-for-triton
github_2023
python
1,799
intel
whitneywhtsang
@@ -351,7 +351,6 @@ def matmul(a, b): [1, 512, 32768, 8192], # [1, 1024, 16384, 8192], # [1, 1024, 28672, 8192], # - [1, 3072, 4096, 3072], #
Let's not remove until we have streamk working for Triton, or else we are not checking if it works. Let's add a FIXME to remind ourselves here.
intel-xpu-backend-for-triton
github_2023
python
1,799
intel
LiyangLingIntel
@@ -388,15 +387,14 @@ def benchmark(B, M, N, K, provider): quantiles = [0.5, 0.0, 1.0] if provider == 'onednn': - ms, min_ms, max_ms, mean, cv = benchmark_suit.do_bench(lambda: torch.matmul(a, b), warmup=100, rep=100, - quantiles=quantiles...
How much impact the `fast_flush` has on the performance data?
intel-xpu-backend-for-triton
github_2023
others
1,345
intel
pbchekin
@@ -4,6 +4,10 @@ run-name: ${{ inputs.run_name }} on: workflow_dispatch: inputs: + enable_unskip: + description: Ignore all pytest.skip, disable as default
Nit: ```suggestion description: Ignore pytest.skip ```
intel-xpu-backend-for-triton
github_2023
others
1,345
intel
pbchekin
@@ -34,10 +34,10 @@ pytest() { mkdir -p "$CURRENT_SKIPLIST_DIR" # skip comments in the skiplist sed -e '/^#/d' "$TRITON_TEST_SKIPLIST_DIR/$TRITON_TEST_SUITE.txt" > "$CURRENT_SKIPLIST_DIR/$TRITON_TEST_SUITE.txt" - pytest_extra_args+=( - "--deselect-from-file=$CURRENT_SKIPLIST...
```suggestion if [[ $TEST_UNSKIP = false ]]; then pytest_extra_args+=( "--deselect-from-file=$CURRENT_SKIPLIST_DIR/$TRITON_TEST_SUITE.txt" "--select-fail-on-missing" ) fi ```
intel-xpu-backend-for-triton
github_2023
others
1,345
intel
pbchekin
@@ -34,10 +34,12 @@ pytest() { mkdir -p "$CURRENT_SKIPLIST_DIR" # skip comments in the skiplist sed -e '/^#/d' "$TRITON_TEST_SKIPLIST_DIR/$TRITON_TEST_SUITE.txt" > "$CURRENT_SKIPLIST_DIR/$TRITON_TEST_SUITE.txt" - pytest_extra_args+=( - "--deselect-from-file=$CURRENT_SKIPLIST...
```suggestion "--deselect-from-file=$CURRENT_SKIPLIST_DIR/$TRITON_TEST_SUITE.txt" ```
intel-xpu-backend-for-triton
github_2023
others
1,345
intel
pbchekin
@@ -34,10 +34,14 @@ pytest() { mkdir -p "$CURRENT_SKIPLIST_DIR" # skip comments in the skiplist sed -e '/^#/d' "$TRITON_TEST_SKIPLIST_DIR/$TRITON_TEST_SUITE.txt" > "$CURRENT_SKIPLIST_DIR/$TRITON_TEST_SUITE.txt" - pytest_extra_args+=( - "--deselect-from-file=$CURRENT_SKIPLIST...
IMHO timeout 15s is not enough for some tests. Also have you tested that pytest-timeout works well with pytest-xdist? I had several weird errors using them both and also see https://github.com/pytest-dev/pytest-xdist/issues/220. Potentially for this mode (TEST_UNSKIP) you need to disable xdist and run tests sequentiall...
intel-xpu-backend-for-triton
github_2023
others
1,345
intel
pbchekin
@@ -114,3 +119,4 @@ jobs: ignore_errors: ${{ inputs.ignore_errors || false }} skip_list: ${{ inputs.skip_list }} run_name: ${{ inputs.run_name }} + enable_unskip: ${{ (github.event_name == 'workflow_dispatch' && inputs.enable_unskip) || false }}
Nit: i think this can be simpler: ```suggestion enable_unskip: ${{ inputs.enable_unskip || false }} ```
intel-xpu-backend-for-triton
github_2023
python
1,345
intel
pbchekin
@@ -76,9 +76,25 @@ def parse_report(report_path: pathlib.Path) -> ReportStats: except FileNotFoundError: pass stats.fixme += len(testsuite_fixme_tests) - deselected = get_deselected(report_path) - stats.skipped += deselected - stats.total += deselected + + test_unskip = os.get...
Nit: ```suggestion if test_unskip not in ("true", "false"): ```
intel-xpu-backend-for-triton
github_2023
python
1,345
intel
pbchekin
@@ -76,9 +76,25 @@ def parse_report(report_path: pathlib.Path) -> ReportStats: except FileNotFoundError: pass stats.fixme += len(testsuite_fixme_tests) - deselected = get_deselected(report_path) - stats.skipped += deselected - stats.total += deselected + + test_unskip = os.get...
Please use the same logic as in `get_deselected`.
intel-xpu-backend-for-triton
github_2023
python
1,345
intel
pbchekin
@@ -76,9 +76,25 @@ def parse_report(report_path: pathlib.Path) -> ReportStats: except FileNotFoundError: pass stats.fixme += len(testsuite_fixme_tests) - deselected = get_deselected(report_path) - stats.skipped += deselected - stats.total += deselected + + test_unskip = os.get...
Please do not hardcode the name of the file here. Instead, you can use a convention for example if `{name}-coredump.txt` exists then you can use it to adjust `skipped` and `total`. Move the processing to a function similar to `get_deselected`.
intel-xpu-backend-for-triton
github_2023
others
1,345
intel
pbchekin
@@ -34,10 +34,20 @@ pytest() { mkdir -p "$CURRENT_SKIPLIST_DIR" # skip comments in the skiplist sed -e '/^#/d' "$TRITON_TEST_SKIPLIST_DIR/$TRITON_TEST_SUITE.txt" > "$CURRENT_SKIPLIST_DIR/$TRITON_TEST_SUITE.txt" - pytest_extra_args+=( - "--deselect-from-file=$CURRENT_SKIPLIST...
Please do not hard code test suite.
intel-xpu-backend-for-triton
github_2023
others
1,345
intel
pbchekin
@@ -0,0 +1,28 @@ +#!/bin/bash +# set -x +# Generate a .txt recording all cases that cause coredump + +SCRIPTS_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )" +cd $SCRIPTS_DIR/../python/test/unit + +TRITON_TEST_SKIPLIST_DIR="${TRITON_TEST_SKIPLIST_DIR:-$SCRIPTS_DIR/skiplist/default}" +COREDUMP...
I don't like the idea to restart pytest after test crash, `pytest-xdist` automatically handles this, see https://pytest-xdist.readthedocs.io/en/stable/crash.html. The `pre-run.sh` script basically runs all core tests, then you run them again in the workflow, that is a waste of time and compute resources. Potentially yo...
intel-xpu-backend-for-triton
github_2023
others
1,345
intel
pbchekin
@@ -200,7 +205,7 @@ jobs: source ./scripts/pytest-utils.sh cd python/test/unit TRITON_TEST_SUITE=language \ - pytest -vvv -n 8 --device xpu language/ --ignore=language/test_line_info.py --ignore=language/test_subprocess.py + pytest -vvv -n 8 --max-worker-restart=50...
Move `--max-worker-restart=500` to `pytest-utils.sh` to the same place where you have `--timeout=500`.
intel-xpu-backend-for-triton
github_2023
others
1,345
intel
pbchekin
@@ -141,13 +146,15 @@ run_core_tests() { echo "****** Running Triton Core tests ******" echo "***************************************************" CORE_TEST_DIR=$TRITON_PROJ/python/test/unit + if [ ! -d "${CORE_TEST_DIR}" ]; then echo "Not found '${CORE_TEST_DIR}'. Build Triton please" ; ex...
Move `--max-worker-restart=500` to `pytest-utils.sh` to the same place where you have `--timeout=500`.
intel-xpu-backend-for-triton
github_2023
others
1,345
intel
pbchekin
@@ -34,10 +34,14 @@ pytest() { mkdir -p "$CURRENT_SKIPLIST_DIR" # skip comments in the skiplist sed -e '/^#/d' "$TRITON_TEST_SKIPLIST_DIR/$TRITON_TEST_SUITE.txt" > "$CURRENT_SKIPLIST_DIR/$TRITON_TEST_SUITE.txt" - pytest_extra_args+=( - "--deselect-from-file=$CURRENT_SKIPLIST...
```suggestion pytest_extra_args+=( "--timeout=500" "--max-worker-restart=500" ) ```
intel-xpu-backend-for-triton
github_2023
cpp
1,636
intel
victor-eds
@@ -846,6 +846,14 @@ struct TritonRaiseBlockPointer return failure(); } + // As masks are incompatible with block pointer load/store ops + // Masks must be handled before the operation can be rewritten. + // This will be done in a future PR (Issue #1784). + // In the meantime, operations with ...
`success()` or `failure()`?
intel-xpu-backend-for-triton
github_2023
others
1,761
intel
whitneywhtsang
@@ -0,0 +1,38 @@ +// RUN: triton-opt %s -split-input-file --intel-decompose-unsupported-conversions | FileCheck %s + +#dpas = #triton_intel_gpu.dpas<{repeatCount = 8, systolicDepth = 8, executionSize = 16, opsPerChan = 2, threadsPerWarp = 16, warpsPerCTA = [32, 1], repCluster = [1, 2], A = [8, 16], B = [16, 32], C = [8...
```suggestion module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 32 : i32, "triton_gpu.threads-per-warp" = 16 : i32, triton_intel_gpu.min_sg_size = 16 : i32, triton_intel_gpu.support_dpas, triton_intel_gpu.support_sg_2d_block} { ```
intel-xpu-backend-for-triton
github_2023
others
1,761
intel
whitneywhtsang
@@ -0,0 +1,38 @@ +// RUN: triton-opt %s -split-input-file --intel-decompose-unsupported-conversions | FileCheck %s + +#dpas = #triton_intel_gpu.dpas<{repeatCount = 8, systolicDepth = 8, executionSize = 16, opsPerChan = 2, threadsPerWarp = 16, warpsPerCTA = [32, 1], repCluster = [1, 2], A = [8, 16], B = [16, 32], C = [8...
Please add comment in the test case on why you don't expect local_alloc/load in this case.
intel-xpu-backend-for-triton
github_2023
cpp
1,761
intel
whitneywhtsang
@@ -133,12 +138,10 @@ struct DecomposeUnsupportedConversions srcBlocked.getOrder(), srcBlocked.getCTALayout(), srcType.getElementType()), sharedMemorySpace); - auto tmp = builder.create<triton::gpu::ConvertLayoutOp>( + auto tmp = builder.create<triton::gpu::L...
Please upstream this changes.
intel-xpu-backend-for-triton
github_2023
cpp
1,761
intel
etiotto
@@ -84,29 +86,32 @@ struct DecomposeUnsupportedConversions } }); /* -------------------------------- */ - // Replace `mma -> dot_op` with `mma -> blocked -> dot_op` + // Replace `dpas -> dot_op` with `dpas -> shared -> dot_op`
Why not doing the same as NVidia and go to blocked layout which is later on converted to shared layout and then to dot layout. I think your motivation is to just avoid the intermediate step of converting from a dpas layout to blocked layout, which makes sense at first sight. My concern is that deviating from the w...
intel-xpu-backend-for-triton
github_2023
cpp
1,761
intel
whitneywhtsang
@@ -84,23 +86,24 @@ struct DecomposeUnsupportedConversions } }); /* -------------------------------- */ - // Replace `mma -> dot_op` with `mma -> blocked -> dot_op` + // Replace `dpas -> dot_op` with `dpas -> shared -> dot_op`
```suggestion // Replace `dpas -> dot_op` with `dpas -> blocked -> dot_op` ```
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
FMarno
@@ -24,6 +24,59 @@ namespace { constexpr unsigned offsetBitwidth = 32; constexpr unsigned shapeAndStridesBitwidth = 64; +std::optional<int64_t> getIntAttr(const OpFoldResult ofr) { + if (ofr.is<Attribute>() && isa<IntegerAttr>(ofr.get<Attribute>())) + return dyn_cast<IntegerAttr>(ofr.get<Attribute>()).getInt();
```suggestion return cast<IntegerAttr>(ofr.get<Attribute>()).getInt(); ```
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
FMarno
@@ -24,6 +24,59 @@ namespace { constexpr unsigned offsetBitwidth = 32; constexpr unsigned shapeAndStridesBitwidth = 64; +std::optional<int64_t> getIntAttr(const OpFoldResult ofr) { + if (ofr.is<Attribute>() && isa<IntegerAttr>(ofr.get<Attribute>())) + return dyn_cast<IntegerAttr>(ofr.get<Attribute>()).getInt();...
```suggestion auto val = cast<Value>(results[0]); ```
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
FMarno
@@ -24,6 +24,59 @@ namespace { constexpr unsigned offsetBitwidth = 32; constexpr unsigned shapeAndStridesBitwidth = 64; +std::optional<int64_t> getIntAttr(const OpFoldResult ofr) { + if (ofr.is<Attribute>() && isa<IntegerAttr>(ofr.get<Attribute>())) + return dyn_cast<IntegerAttr>(ofr.get<Attribute>()).getInt();...
```suggestion return getIntAttr(constOp.getValue()); ```
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
FMarno
@@ -24,6 +24,59 @@ namespace { constexpr unsigned offsetBitwidth = 32; constexpr unsigned shapeAndStridesBitwidth = 64; +std::optional<int64_t> getIntAttr(const OpFoldResult ofr) { + if (ofr.is<Attribute>() && isa<IntegerAttr>(ofr.get<Attribute>())) + return dyn_cast<IntegerAttr>(ofr.get<Attribute>()).getInt();...
```suggestion return intVal.has_value() && intVal.value() == 0; ```
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
FMarno
@@ -69,9 +122,13 @@ struct PtrState { // When PtrState describes a non-block pointer, shape field indicates how // address wraps around. As a result, a constant 0 indicates no wrap around // (i.e. modulo) for the dimension. - if (auto intOp = shape[dim].getDefiningOp<arith::ConstantIntOp>()) { - ...
is `results` used?
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
FMarno
@@ -69,9 +122,13 @@ struct PtrState { // When PtrState describes a non-block pointer, shape field indicates how // address wraps around. As a result, a constant 0 indicates no wrap around // (i.e. modulo) for the dimension. - if (auto intOp = shape[dim].getDefiningOp<arith::ConstantIntOp>()) { - ...
```suggestion return !hasConstZero(shape[dim]); ``` double check that one please!
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
FMarno
@@ -112,11 +169,72 @@ struct PtrState { sizes.push_back(lhsState.sizes[i]); } + // AddPtr where both lhs and rhs containing modulo operators not supported + if (lhsState.hasModulo() && rhsState.hasModulo()) { + op->emitRemark( + "TritonRaiseBlockPointer: do not support adding two point...
```suggestion assert(!(lhsState.hasModulo() || rhsState.hasModulo()) || (lhsState.getRank() <= 2) && "description"); ``` NIT: The conditional is only for debug code
intel-xpu-backend-for-triton
github_2023
others
1,570
intel
FMarno
@@ -236,3 +236,144 @@ tt.func @test_addptr_broadcast_rank_3(%arg0 : !tt.ptr<f32>) -> tensor<128x2x128x %3 = tt.load %2 : tensor<128x2x128x!tt.ptr<f32>> tt.return %3 : tensor<128x2x128xf32> } + + +// CHECK: tt.func public @wrap_side_by_side_masked([[PARAM_0_:%.+]]: !tt.ptr<f32>, [[PARAM_1_:%.+]]: !tt.ptr<...
Maybe the constant declaration should be check-dag since the order is not important?
intel-xpu-backend-for-triton
github_2023
others
1,570
intel
FMarno
@@ -236,3 +236,144 @@ tt.func @test_addptr_broadcast_rank_3(%arg0 : !tt.ptr<f32>) -> tensor<128x2x128x %3 = tt.load %2 : tensor<128x2x128x!tt.ptr<f32>> tt.return %3 : tensor<128x2x128xf32> } + + +// CHECK: tt.func public @wrap_side_by_side_masked([[PARAM_0_:%.+]]: !tt.ptr<f32>, [[PARAM_1_:%.+]]: !tt.ptr<...
```suggestion %c0_i32 = arith.constant 0 : i32 %c1_i32 = arith.constant 1 : i32 %c2_i32 = arith.constant 2 : i32 %c4_i32 = arith.constant 4 : i32 %cst_0 = arith.constant dense<2> : tensor<4x1xi32> %cst_1 = arith.constant dense<6> : tensor<4xi32> %cst_2 = arith.constant dense<2> : tens...
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
victor-eds
@@ -24,6 +24,59 @@ namespace { constexpr unsigned offsetBitwidth = 32; constexpr unsigned shapeAndStridesBitwidth = 64; +std::optional<int64_t> getIntAttr(const OpFoldResult ofr) { + if (ofr.is<Attribute>() && isa<IntegerAttr>(ofr.get<Attribute>())) + return dyn_cast<IntegerAttr>(ofr.get<Attribute>()).getInt();...
Can we asssert on `op` having exactly 1 result?
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
victor-eds
@@ -24,6 +24,59 @@ namespace { constexpr unsigned offsetBitwidth = 32; constexpr unsigned shapeAndStridesBitwidth = 64; +std::optional<int64_t> getIntAttr(const OpFoldResult ofr) { + if (ofr.is<Attribute>() && isa<IntegerAttr>(ofr.get<Attribute>())) + return dyn_cast<IntegerAttr>(ofr.get<Attribute>()).getInt();...
```suggestion if (results.empty()) { ```
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
victor-eds
@@ -24,6 +24,59 @@ namespace { constexpr unsigned offsetBitwidth = 32; constexpr unsigned shapeAndStridesBitwidth = 64; +std::optional<int64_t> getIntAttr(const OpFoldResult ofr) { + if (ofr.is<Attribute>() && isa<IntegerAttr>(ofr.get<Attribute>())) + return dyn_cast<IntegerAttr>(ofr.get<Attribute>()).getInt();...
Are we actually getting any constant value after this point? I'd expect this to be folded if result is `arith.constant`
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
victor-eds
@@ -24,6 +24,59 @@ namespace { constexpr unsigned offsetBitwidth = 32; constexpr unsigned shapeAndStridesBitwidth = 64; +std::optional<int64_t> getIntAttr(const OpFoldResult ofr) {
```suggestion std::optional<int64_t> getIntAttr(OpFoldResult ofr) { ```
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
victor-eds
@@ -24,6 +24,59 @@ namespace { constexpr unsigned offsetBitwidth = 32; constexpr unsigned shapeAndStridesBitwidth = 64; +std::optional<int64_t> getIntAttr(const OpFoldResult ofr) { + if (ofr.is<Attribute>() && isa<IntegerAttr>(ofr.get<Attribute>())) + return dyn_cast<IntegerAttr>(ofr.get<Attribute>()).getInt();...
```suggestion bool hasConstZero(Value val) { ```
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
victor-eds
@@ -112,11 +169,72 @@ struct PtrState { sizes.push_back(lhsState.sizes[i]); } + // AddPtr where both lhs and rhs containing modulo operators not supported + if (lhsState.hasModulo() && rhsState.hasModulo()) { + op->emitRemark( + "TritonRaiseBlockPointer: do not support adding two point...
This will work if both have modulo, right?
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
victor-eds
@@ -112,11 +169,72 @@ struct PtrState { sizes.push_back(lhsState.sizes[i]); } + // AddPtr where both lhs and rhs containing modulo operators not supported + if (lhsState.hasModulo() && rhsState.hasModulo()) { + op->emitRemark( + "TritonRaiseBlockPointer: do not support adding two point...
Does this correspond with "skipping"?
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
victor-eds
@@ -112,11 +169,72 @@ struct PtrState { sizes.push_back(lhsState.sizes[i]); } + // AddPtr where both lhs and rhs containing modulo operators not supported + if (lhsState.hasModulo() && rhsState.hasModulo()) { + op->emitRemark( + "TritonRaiseBlockPointer: do not support adding two point...
Can we have this out of the loop as this will only happen in the first iteration and break out of the loop?
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
victor-eds
@@ -377,6 +510,82 @@ struct TritonRaiseBlockPointer IRMapping ptrMap; }; +template <typename OpTy, typename = std::enable_if_t<llvm::is_one_of< + OpTy, arith::RemSIOp, arith::RemUIOp>::value>> +LogicalResult TritonRaiseBlockPointer::visitAddPointerRemOperand( + OpTy remOp, PtrState ...
NIT: `failed(...)`
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
victor-eds
@@ -377,6 +510,82 @@ struct TritonRaiseBlockPointer IRMapping ptrMap; }; +template <typename OpTy, typename = std::enable_if_t<llvm::is_one_of< + OpTy, arith::RemSIOp, arith::RemUIOp>::value>> +LogicalResult TritonRaiseBlockPointer::visitAddPointerRemOperand( + OpTy remOp, PtrState ...
NIT: `failed(...)`
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
victor-eds
@@ -377,6 +510,82 @@ struct TritonRaiseBlockPointer IRMapping ptrMap; }; +template <typename OpTy, typename = std::enable_if_t<llvm::is_one_of< + OpTy, arith::RemSIOp, arith::RemUIOp>::value>> +LogicalResult TritonRaiseBlockPointer::visitAddPointerRemOperand( + OpTy remOp, PtrState ...
I would not worry much about this case, as I guess the canonicalizer should simplify this beforehand, but cool to check
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
victor-eds
@@ -377,6 +510,82 @@ struct TritonRaiseBlockPointer IRMapping ptrMap; }; +template <typename OpTy, typename = std::enable_if_t<llvm::is_one_of< + OpTy, arith::RemSIOp, arith::RemUIOp>::value>> +LogicalResult TritonRaiseBlockPointer::visitAddPointerRemOperand( + OpTy remOp, PtrState ...
If this can cause issues, can we maybe change the order in an earlier pass?
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
victor-eds
@@ -377,6 +510,82 @@ struct TritonRaiseBlockPointer IRMapping ptrMap; }; +template <typename OpTy, typename = std::enable_if_t<llvm::is_one_of< + OpTy, arith::RemSIOp, arith::RemUIOp>::value>> +LogicalResult TritonRaiseBlockPointer::visitAddPointerRemOperand( + OpTy remOp, PtrState ...
Format
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
victor-eds
@@ -112,11 +169,72 @@ struct PtrState { sizes.push_back(lhsState.sizes[i]); } + // AddPtr where both lhs and rhs containing modulo operators not supported + if (lhsState.hasModulo() && rhsState.hasModulo()) { + op->emitRemark( + "TritonRaiseBlockPointer: do not support adding two point...
Format
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
etiotto
@@ -377,6 +478,81 @@ struct TritonRaiseBlockPointer IRMapping ptrMap; }; +template <typename OpTy, typename = std::enable_if_t<llvm::is_one_of< + OpTy, arith::RemSIOp, arith::RemUIOp>::value>> +LogicalResult TritonRaiseBlockPointer::visitAddPointerRemOperand( + OpTy remOp, PtrState ...
Change to switch stmt
intel-xpu-backend-for-triton
github_2023
cpp
1,570
intel
whitneywhtsang
@@ -113,11 +155,55 @@ struct PtrState { sizes.push_back(lhsState.sizes[i]); } + // AddPtr where both lhs and rhs containing modulo operators not supported + if (lhsState.hasModulo() && rhsState.hasModulo()) { + op->emitRemark( + "TritonRaiseBlockPointer: do not support adding two point...
Please update the comment if it is no longer the case.
intel-xpu-backend-for-triton
github_2023
cpp
1,746
intel
chengjunlu
@@ -397,31 +497,68 @@ std::optional<LinearLayout> DPAStoLinearLayout(ArrayRef<int64_t> shape, const SmallVector<unsigned> warpsPerCTA = dpas.getWarpsPerCTA(); int threadsPerWarp = triton::gpu::getWarpSize(dpas); + unsigned opsPerChannel = dpas.getOpsPerChannel(); auto repCluster = dpas.getRepCluster(); - S...
I think the logic should be different for A, B and C on the warp dimension. The A operands are duplicated on row dimension and B operands are duplicated on column dimension. In another word, some warp hop should be 0 stride for A and B.
intel-xpu-backend-for-triton
github_2023
cpp
1,746
intel
etiotto
@@ -379,11 +479,11 @@ DPASLaneBasesC(int repeatCount, int executionSize, int threadsPerWarp) { return laneBases; } -std::optional<LinearLayout> DPAStoLinearLayout(ArrayRef<int64_t> shape, - Attribute layout) { +std::optional<LinearLayout> +DPAStoLinearLayout(ArrayRef<...
assert opIdx <3 on function entry
intel-xpu-backend-for-triton
github_2023
cpp
1,746
intel
etiotto
@@ -397,33 +497,82 @@ std::optional<LinearLayout> DPAStoLinearLayout(ArrayRef<int64_t> shape, const SmallVector<unsigned> warpsPerCTA = dpas.getWarpsPerCTA(); int threadsPerWarp = triton::gpu::getWarpSize(dpas); + unsigned opsPerChannel = dpas.getOpsPerChannel(); auto repCluster = dpas.getRepCluster(); - S...
For simplicity I would place these 2 lines before the `if` (is true that both aren't needed unless opIdx == 2 but I think is fine).
intel-xpu-backend-for-triton
github_2023
cpp
1,746
intel
etiotto
@@ -397,33 +497,82 @@ std::optional<LinearLayout> DPAStoLinearLayout(ArrayRef<int64_t> shape, const SmallVector<unsigned> warpsPerCTA = dpas.getWarpsPerCTA(); int threadsPerWarp = triton::gpu::getWarpSize(dpas); + unsigned opsPerChannel = dpas.getOpsPerChannel(); auto repCluster = dpas.getRepCluster(); - S...
[nit]: Initialize KDim to zero (its value is overwritten in all possible code path the if stmt can take)
intel-xpu-backend-for-triton
github_2023
cpp
1,746
intel
etiotto
@@ -336,7 +336,109 @@ LinearLayout combineCtaCgaWithShape(LinearLayout ctaLayout, // The layout example repeat_count=8, systolic_depth=8, // execution_size=16 and operands_per_chan=2 for warp size 32. -// DPASInst layout of C operand: +// For A operand: +// systolic depth = 8 +//<-----------------...
[nit]: opc *= 2 for the step
intel-xpu-backend-for-triton
github_2023
cpp
1,746
intel
etiotto
@@ -336,7 +336,109 @@ LinearLayout combineCtaCgaWithShape(LinearLayout ctaLayout, // The layout example repeat_count=8, systolic_depth=8, // execution_size=16 and operands_per_chan=2 for warp size 32. -// DPASInst layout of C operand: +// For A operand: +// systolic depth = 8 +//<-----------------...
[nit]: warp *= 2 for the step
intel-xpu-backend-for-triton
github_2023
cpp
1,746
intel
etiotto
@@ -336,7 +336,109 @@ LinearLayout combineCtaCgaWithShape(LinearLayout ctaLayout, // The layout example repeat_count=8, systolic_depth=8, // execution_size=16 and operands_per_chan=2 for warp size 32. -// DPASInst layout of C operand: +// For A operand: +// systolic depth = 8 +//<-----------------...
[nit]: remove empty line
intel-xpu-backend-for-triton
github_2023
cpp
1,746
intel
etiotto
@@ -336,7 +336,109 @@ LinearLayout combineCtaCgaWithShape(LinearLayout ctaLayout, // The layout example repeat_count=8, systolic_depth=8, // execution_size=16 and operands_per_chan=2 for warp size 32. -// DPASInst layout of C operand: +// For A operand: +// systolic depth = 8 +//<-----------------...
Can we use SmallVector ?
intel-xpu-backend-for-triton
github_2023
cpp
1,669
intel
FMarno
@@ -774,6 +775,11 @@ void MatchTargetSizePass::transformDotOp(tt::DotOp dot) { subDotC = b.create<tt::DotOp>(loc, subDotA, subDotB, subDotC, dot.getInputPrecisionAttr(), dot.getMaxNumImpreciseAccAttr()); + // hack for attentio...
Why can this not be enabled by default? I believe at this time we are trying to remove the hacks and do things properly
intel-xpu-backend-for-triton
github_2023
others
1,669
intel
FMarno
@@ -243,4 +243,19 @@ def TritonIntelGPUMatchTargetSize : Pass<"tritonintelgpu-match-target-size", "ml "mlir::triton::gpu::intel::TritonIntelGPUDialect"]; } +def TritonIntelGPUScheduleLoad : Pass<"tritonintelgpu-schedule-load", "mlir::ModuleOp"> { + let summary = "naive ra-aware instr sch...
what is RA?
intel-xpu-backend-for-triton
github_2023
others
1,669
intel
FMarno
@@ -243,4 +243,19 @@ def TritonIntelGPUMatchTargetSize : Pass<"tritonintelgpu-match-target-size", "ml "mlir::triton::gpu::intel::TritonIntelGPUDialect"]; } +def TritonIntelGPUScheduleLoad : Pass<"tritonintelgpu-schedule-load", "mlir::ModuleOp"> { + let summary = "naive ra-aware instr sch...
Can we try to be more general that just FlashAttention. Maybe mention the features of the workloads that this optimizes?
intel-xpu-backend-for-triton
github_2023
cpp
1,669
intel
FMarno
@@ -0,0 +1,133 @@ +//===- ScheduleLoad.cpp ----------------------------------------------*-===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===-----...
The use of the function is confusing since it has side effects (modifying `visited`) and returns another value.
intel-xpu-backend-for-triton
github_2023
cpp
1,669
intel
FMarno
@@ -0,0 +1,133 @@ +//===- ScheduleLoad.cpp ----------------------------------------------*-===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===-----...
Please detail the goals of the scheduler
intel-xpu-backend-for-triton
github_2023
cpp
1,669
intel
FMarno
@@ -0,0 +1,133 @@ +//===- ScheduleLoad.cpp ----------------------------------------------*-===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===-----...
I don't think this change fits in here with the scheduling changes
intel-xpu-backend-for-triton
github_2023
cpp
1,669
intel
FMarno
@@ -0,0 +1,133 @@ +//===- ScheduleLoad.cpp ----------------------------------------------*-===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===-----...
you're assigning a negative to an unsigned, maybe use `int` or `int64_t` instead? I believe the return type of `IntegerAttr::getInt` is `int64_t`.
intel-xpu-backend-for-triton
github_2023
cpp
1,669
intel
FMarno
@@ -0,0 +1,133 @@ +//===- ScheduleLoad.cpp ----------------------------------------------*-===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===-----...
why 4?
intel-xpu-backend-for-triton
github_2023
cpp
1,669
intel
FMarno
@@ -0,0 +1,133 @@ +//===- ScheduleLoad.cpp ----------------------------------------------*-===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===-----...
```suggestion if (visited.contains(val)) ```
intel-xpu-backend-for-triton
github_2023
cpp
1,669
intel
FMarno
@@ -0,0 +1,133 @@ +//===- ScheduleLoad.cpp ----------------------------------------------*-===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===-----...
```suggestion if (!visited.contains(base)) { ```
intel-xpu-backend-for-triton
github_2023
cpp
1,669
intel
FMarno
@@ -0,0 +1,133 @@ +//===- ScheduleLoad.cpp ----------------------------------------------*-===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===-----...
```suggestion unsigned opIdx) { ``` This default makes the code less readable by adding extra information to remember without saving much typing
intel-xpu-backend-for-triton
github_2023
cpp
1,669
intel
FMarno
@@ -0,0 +1,133 @@ +//===- ScheduleLoad.cpp ----------------------------------------------*-===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===-----...
Why are we visiting without doing anything for `currGroup == 0`? it would be good to add a code comment for that
intel-xpu-backend-for-triton
github_2023
cpp
1,669
intel
FMarno
@@ -0,0 +1,133 @@ +//===- ScheduleLoad.cpp ----------------------------------------------*-===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===-----...
If I understand the code correctly, you are grouping instances of `DotOp` with the same value for `schedule-group`. Why not use something like an [`IndexedMap`](https://llvm.org/docs/ProgrammersManual.html#llvm-adt-indexedmap-h) or a [`DenseMap`](https://llvm.org/docs/ProgrammersManual.html#llvm-adt-densemap-h)?
intel-xpu-backend-for-triton
github_2023
cpp
1,669
intel
whitneywhtsang
@@ -0,0 +1,143 @@ +//===- ScheduleLoad.cpp ----------------------------------------------*-===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===-----...
remove
intel-xpu-backend-for-triton
github_2023
cpp
1,669
intel
whitneywhtsang
@@ -0,0 +1,143 @@ +//===- ScheduleLoad.cpp ----------------------------------------------*-===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===-----...
what happen if the attribute is not present?
intel-xpu-backend-for-triton
github_2023
cpp
1,669
intel
FMarno
@@ -0,0 +1,143 @@ +//===- ScheduleLoad.cpp ----------------------------------------------*-===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===-----...
I'm not sure this is working as intended since `numGroups` will be `0` and it was previously `4`.
intel-xpu-backend-for-triton
github_2023
others
1,392
intel
YarShev
@@ -0,0 +1,29 @@ +# Developer Guide + +This short guide covers creating a basic development environment, installing needed dependencies, and best practices for building and running Triton on different Intel platforms.
```suggestion This short guide covers creating a basic development environment, installing needed dependencies, and best practices for building and running Triton on different Intel platforms. ``` To not have long lines I would break these and all below.
intel-xpu-backend-for-triton
github_2023
others
1,392
intel
YarShev
@@ -0,0 +1,29 @@ +# Developer Guide + +This short guide covers creating a basic development environment, installing needed dependencies, and best practices for building and running Triton on different Intel platforms. + +## Dependencies + +The primary software dependency is oneAPI. The best way to get oneAPI for PyTo...
```suggestion The LTS, Production, and Rolling Stable release streams are supported. Some performance optimizations may not be available with the LTS driver package. For development, we recommend rolling stable. ```
intel-xpu-backend-for-triton
github_2023
others
1,392
intel
vlad-penkin
@@ -0,0 +1,29 @@ +# Developer Guide
I suggest to add a section dedicated to the python base envs and offer two alternatives: - virtualenv - conda For conda the recommended setup is `conda create -n xpu -c conda-forge python=3.10 gh jq` and then do pip install if needed for everything else.
intel-xpu-backend-for-triton
github_2023
others
1,392
intel
vlad-penkin
@@ -0,0 +1,29 @@ +# Developer Guide + +This short guide covers creating a basic development environment, installing needed dependencies, and best practices for building and running Triton on different Intel platforms. + +## Dependencies + +The primary software dependency is oneAPI. The best way to get oneAPI for PyTorc...
1) I suggest to provide the recommendation on the packaging option. My preference is offline installer - it's much more easier to maintain, delete from the system. Offline installer will also allow you to have multiple installs on the same machine (PTDB and Regular oneAPI Base Kit). 2) Lets add recommended env activ...
intel-xpu-backend-for-triton
github_2023
others
1,392
intel
vlad-penkin
@@ -0,0 +1,29 @@ +# Developer Guide + +This short guide covers creating a basic development environment, installing needed dependencies, and best practices for building and running Triton on different Intel platforms. + +## Dependencies + +The primary software dependency is oneAPI. The best way to get oneAPI for PyTorc...
conda-forge binaries are usually up-to-date and a viable option to - `conda -c conda-forge install gh`
intel-xpu-backend-for-triton
github_2023
others
1,392
intel
YarShev
@@ -0,0 +1,31 @@ +# Developer Guide + +This short guide covers creating a basic development environment, installing needed dependencies, and best practices for building and running Triton on different Intel platforms. + +## Dependencies + +The primary software dependency is oneAPI. The best way to get oneAPI for PyTorc...
Why are we missing "Step 1: Install Intel Data Center GPU Drivers"?
intel-xpu-backend-for-triton
github_2023
others
1,711
intel
whitneywhtsang
@@ -324,6 +324,10 @@ jobs: - name: Run triton gemm kernel benchmark run: | cd benchmarks/triton_kernels_benchmark + # Dry run to validate no functionality regression for splitk and streamk in defaul path
```suggestion # Dry run to validate no functionality regression for splitk and streamk in default path ```
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -58,6 +58,39 @@ namespace { /// Import the GPU Ops to TritonGEN Patterns. #include "GPUToTritonGEN.cpp.inc" +struct GPUSubgroupReduceOpLowering + : public ConvertOpToLLVMPattern<mlir::gpu::SubgroupReduceOp> { + using ConvertOpToLLVMPattern< + mlir::gpu::SubgroupReduceOp>::ConvertOpToLLVMPattern; + Logi...
```suggestion LogicalResult matchAndRewrite(mlir::gpu::SubgroupReduceOp op, OpAdaptor adaptor, ```
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -58,6 +58,39 @@ namespace { /// Import the GPU Ops to TritonGEN Patterns. #include "GPUToTritonGEN.cpp.inc" +struct GPUSubgroupReduceOpLowering + : public ConvertOpToLLVMPattern<mlir::gpu::SubgroupReduceOp> { + using ConvertOpToLLVMPattern< + mlir::gpu::SubgroupReduceOp>::ConvertOpToLLVMPattern; + Logi...
Are we expecting non-uniform reductions? SPIR-V supports non-uniform reductions, so we could lower to those if needed here.
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -58,6 +58,39 @@ namespace { /// Import the GPU Ops to TritonGEN Patterns. #include "GPUToTritonGEN.cpp.inc" +struct GPUSubgroupReduceOpLowering + : public ConvertOpToLLVMPattern<mlir::gpu::SubgroupReduceOp> { + using ConvertOpToLLVMPattern< + mlir::gpu::SubgroupReduceOp>::ConvertOpToLLVMPattern; + Logi...
Can we have this in a function on its own and covering all of the possible values? We may run into other reductions.
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -58,6 +58,39 @@ namespace { /// Import the GPU Ops to TritonGEN Patterns. #include "GPUToTritonGEN.cpp.inc" +struct GPUSubgroupReduceOpLowering + : public ConvertOpToLLVMPattern<mlir::gpu::SubgroupReduceOp> { + using ConvertOpToLLVMPattern< + mlir::gpu::SubgroupReduceOp>::ConvertOpToLLVMPattern; + Logi...
Shouldn't we use subgroup size here instead of 32? I understand using a number different than subgroup size will lead to using `WaveAll` and that might give better performance, but that is not what `gpu.subgroup_reduce` represents.
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -92,8 +92,13 @@ struct ConvertTritonGPUToLLVM int numCTAs = triton::gpu::TritonGPUDialect::getNumCTAs(mod); int threadsPerWarp = triton::gpu::TritonGPUDialect::getThreadsPerWarp(mod); - // Allocate shared memory and set barrier - if (!pipelineManager.skipSharedMemoryAllocation()) { + // fixed 0 ...
```suggestion Builder builder(context); mod->setAttr( "triton_gpu.shared", builder.getI32IntegerAttr(0)); ``` Why do we need this anyway?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -189,6 +195,17 @@ class LoadStorePrefetchOpConversion getVectorType(cast<RankedTensorType>(op.getResult().getType()), isDword ? i32_ty : i16_ty); bool vnni = (idx == 1) && dataSize < 32; + // fixed f16 for now
Why?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -170,16 +173,19 @@ class LoadStorePrefetchOpConversion return truncatedShape; }; - Value surfaceW = calculateSurface(ptrOp.getShape()[1], true); - Value surfaceH = calculateSurface(ptrOp.getShape()[0], false); - Value surfaceP = calculateSurface(ptrOp.getStrides()[0], true); + Value surface...
I see at the beginning of the function we're asserting the op is 2D. Should we fail conversion instead?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -198,7 +215,14 @@ class LoadStorePrefetchOpConversion return failure(); } rewriter.replaceOp(op, bitcast(load, resType)); - } else if constexpr (std::is_same_v<OpType, PrefetchOp>) { + } else if constexpr (std::is_same_v<OpType, ttgi::PrefetchOp>) { + if (ptrOp.getOrder()[0] == 0) {...
Use `std::swap`