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 | 2,181 | intel | whitneywhtsang | @@ -714,56 +804,69 @@ struct LoadOpConversion
auto typeConverter = getTypeConverter();
auto *ctx = rewriter.getContext();
- // original values
- Value ptr = op.getPtr();
- Value mask = op.getMask();
- Value other = op.getOther();
-
- // adaptor values
- if (isTensorPointerType(ptr.getType(... | ```suggestion
// contiguous on the fast changing dim when fallback to gather load.
``` |
intel-xpu-backend-for-triton | github_2023 | others | 2,181 | intel | whitneywhtsang | @@ -255,3 +255,44 @@ module attributes {"triton_gpu.num-warps" = 1 : i32, "triton_gpu.threads-per-war
tt.return
}
}
+
+// -----
+
+#dpas = #triton_intel_gpu.dpas<{repeatCount = 8, systolicDepth = 8, executionSize = 16, opsPerChan = 2, threadsPerWarp = 16, warpsPerCTA = [1, 1], repCluster = [1, 1], A = [8, 16... | `dot_a` and `dot_b` are the same, is it a typo? |
intel-xpu-backend-for-triton | github_2023 | others | 2,181 | intel | whitneywhtsang | @@ -255,3 +255,44 @@ module attributes {"triton_gpu.num-warps" = 1 : i32, "triton_gpu.threads-per-war
tt.return
}
}
+
+// -----
+
+#dpas = #triton_intel_gpu.dpas<{repeatCount = 8, systolicDepth = 8, executionSize = 16, opsPerChan = 2, threadsPerWarp = 16, warpsPerCTA = [1, 1], repCluster = [1, 1], A = [8, 16... | can we remove variables not used? |
intel-xpu-backend-for-triton | github_2023 | others | 2,181 | intel | whitneywhtsang | @@ -352,3 +352,22 @@ module attributes {"triton_gpu.num-warps" = 8 : i32, "triton_gpu.threads-per-war
tt.return
}
}
+
+// -----
+
+#blocked = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [1, 16], warpsPerCTA = [2, 4], order = [1, 0]}>
+module attributes {"triton_gpu.num-warps" = 8 : i32, "trit... | please remove unused variables. |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,181 | intel | whitneywhtsang | @@ -187,6 +187,106 @@ struct LoadStoreConversionBase {
return axisAnalysisPass.getMaskAlignment(mask);
}
+ std::tuple<SmallVector<Value>, SmallVector<Value>, SmallVector<Value>>
+ convertBlockPtrToTensorOfPtr(
+ Location loc, Value blockPointerStruct, RankedTensorType tensorType,
+ Type valueElemT... | ```suggestion
unsigned blockOffset = 0, blockShape = 1 * rank,
``` |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,181 | intel | whitneywhtsang | @@ -187,6 +187,106 @@ struct LoadStoreConversionBase {
return axisAnalysisPass.getMaskAlignment(mask);
}
+ std::tuple<SmallVector<Value>, SmallVector<Value>, SmallVector<Value>>
+ convertBlockPtrToTensorOfPtr(
+ Location loc, Value blockPointerStruct, RankedTensorType tensorType,
+ Type valueElemT... | ```suggestion
for (unsigned j = 0; j < rank; ++j) {
``` |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,181 | intel | whitneywhtsang | @@ -187,6 +187,106 @@ struct LoadStoreConversionBase {
return axisAnalysisPass.getMaskAlignment(mask);
}
+ std::tuple<SmallVector<Value>, SmallVector<Value>, SmallVector<Value>>
+ convertBlockPtrToTensorOfPtr(
+ Location loc, Value blockPointerStruct, RankedTensorType tensorType,
+ Type valueElemT... | ```suggestion
for (unsigned i = 0; i < numElems; ++i) {
``` |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,349 | intel | chengjunlu | @@ -41,26 +48,47 @@ struct TritonIntelGPUMaterializeBlockPointerPass
"Expected 'loadOp' to load a tensor value.");
tt::MakeTensorPtrOp makeTensorPtrOp = getMakeTensorPtrOp(ptr);
+ LDBG("Found make tensor ptr op: " << makeTensorPtrOp);
auto ptrType = cast<tt::PointerType>(makeTensorPtr... | Clean the comment out code. |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,349 | intel | whitneywhtsang | @@ -51,17 +51,27 @@ struct TritonIntelGPUMaterializeBlockPointerPass
LDBG("Found make tensor ptr op: " << makeTensorPtrOp);
auto ptrType = cast<tt::PointerType>(makeTensorPtrOp.getType());
auto tensorType = cast<RankedTensorType>(ptrType.getPointeeType());
- ArrayRef<int32_t> order = makeTenso... | ```suggestion
for (size_t i = 0; i < strides.size(); ++i) {
``` |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,349 | intel | whitneywhtsang | @@ -51,17 +51,27 @@ struct TritonIntelGPUMaterializeBlockPointerPass
LDBG("Found make tensor ptr op: " << makeTensorPtrOp);
auto ptrType = cast<tt::PointerType>(makeTensorPtrOp.getType());
auto tensorType = cast<RankedTensorType>(ptrType.getPointeeType());
- ArrayRef<int32_t> order = makeTenso... | is `order` being used? |
intel-xpu-backend-for-triton | github_2023 | python | 2,148 | intel | etiotto | @@ -51,6 +53,7 @@ class XPUOptions:
max_num_imprecise_acc_default: int = 0 # `max_num_imprecise_acc` only applies to fp8 -> fp32 dot on sm_90 for cuda
extern_libs: dict = None
debug: bool = False
+ generate_native_code: bool = True | I would default this to False |
intel-xpu-backend-for-triton | github_2023 | python | 2,148 | intel | etiotto | @@ -296,7 +300,60 @@ def make_spv(src, metadata, options):
else:
metadata["build_flags"] = ""
- return ret
+ if options.generate_native_code is True:
+ with tempfile.NamedTemporaryFile(delete=False, mode='wb', suffix='.spv') as fsrc, \
+ tempfile.NamedTemp... | The device is hardcoded to PVC. I do not think that would work for older devices. Can we choose the device based on the information PyTorch passes to the compiler ? |
intel-xpu-backend-for-triton | github_2023 | c | 2,148 | intel | etiotto | @@ -169,35 +173,39 @@ static PyObject *loadBinary(PyObject *self, PyObject *args) {
int32_t n_spills = props.spillMemSize;
const int32_t n_regs = 0;
- constexpr int32_t max_reg_spill = 1000;
- std::string build_flags_str(build_flags);
- bool is_GRF_mode_specified = false;
-
- // Check whether the GRF mode i... | [nit]: early exit if `is_spv` is false |
intel-xpu-backend-for-triton | github_2023 | python | 2,148 | intel | victor-eds | @@ -296,7 +300,60 @@ def make_spv(src, metadata, options):
else:
metadata["build_flags"] = ""
- return ret
+ if options.generate_native_code is True: | ```suggestion
if options.generate_native_code:
``` |
intel-xpu-backend-for-triton | github_2023 | python | 2,148 | intel | victor-eds | @@ -296,7 +300,60 @@ def make_spv(src, metadata, options):
else:
metadata["build_flags"] = ""
- return ret
+ if options.generate_native_code is True:
+ with tempfile.NamedTemporaryFile(delete=False, mode='wb', suffix='.spv') as fsrc, \
+ tempfile.NamedTemp... | ```suggestion
metadata["build_flags"]
```
Why format string? |
intel-xpu-backend-for-triton | github_2023 | python | 2,148 | intel | victor-eds | @@ -296,7 +300,60 @@ def make_spv(src, metadata, options):
else:
metadata["build_flags"] = ""
- return ret
+ if options.generate_native_code is True:
+ with tempfile.NamedTemporaryFile(delete=False, mode='wb', suffix='.spv') as fsrc, \
+ tempfile.NamedTemp... | ```suggestion
subprocess.run(ocloc_cmd, check=True, close_fds=False, stdout=flog, stderr=subprocess.STDOUT)
```
Why do we redirect `stderr` to `stdout`? |
intel-xpu-backend-for-triton | github_2023 | python | 2,148 | intel | victor-eds | @@ -296,7 +300,60 @@ def make_spv(src, metadata, options):
else:
metadata["build_flags"] = ""
- return ret
+ if options.generate_native_code is True:
+ with tempfile.NamedTemporaryFile(delete=False, mode='wb', suffix='.spv') as fsrc, \
+ tempfile.NamedTemp... | Why stringdoc? Can we `# comment` instead? |
intel-xpu-backend-for-triton | github_2023 | python | 2,148 | intel | victor-eds | @@ -296,7 +300,60 @@ def make_spv(src, metadata, options):
else:
metadata["build_flags"] = ""
- return ret
+ if options.generate_native_code is True:
+ with tempfile.NamedTemporaryFile(delete=False, mode='wb', suffix='.spv') as fsrc, \
+ tempfile.NamedTemp... | ```suggestion
ocloc_cmd[-1] = metadata["build_flags"]
```
Again, why string format? |
intel-xpu-backend-for-triton | github_2023 | python | 2,148 | intel | victor-eds | @@ -296,7 +300,60 @@ def make_spv(src, metadata, options):
else:
metadata["build_flags"] = ""
- return ret
+ if options.generate_native_code is True:
+ with tempfile.NamedTemporaryFile(delete=False, mode='wb', suffix='.spv') as fsrc, \
+ tempfile.NamedTemp... | Why `stderr > stdout`? |
intel-xpu-backend-for-triton | github_2023 | python | 2,148 | intel | victor-eds | @@ -296,7 +300,60 @@ def make_spv(src, metadata, options):
else:
metadata["build_flags"] = ""
- return ret
+ if options.generate_native_code is True:
+ with tempfile.NamedTemporaryFile(delete=False, mode='wb', suffix='.spv') as fsrc, \
+ tempfile.NamedTemp... | Why this instead of relying on `delete=True` from `NamedTemporaryFile`? |
intel-xpu-backend-for-triton | github_2023 | python | 2,148 | intel | victor-eds | @@ -296,7 +300,60 @@ def make_spv(src, metadata, options):
else:
metadata["build_flags"] = ""
- return ret
+ if options.generate_native_code is True:
+ with tempfile.NamedTemporaryFile(delete=False, mode='wb', suffix='.spv') as fsrc, \
+ tempfile.NamedTemp... | Wouldn't it be open already as `flog`? |
intel-xpu-backend-for-triton | github_2023 | python | 2,148 | intel | victor-eds | @@ -296,7 +300,60 @@ def make_spv(src, metadata, options):
else:
metadata["build_flags"] = ""
- return ret
+ if options.generate_native_code is True:
+ with tempfile.NamedTemporaryFile(delete=False, mode='wb', suffix='.spv') as fsrc, \
+ tempfile.NamedTemp... | ```suggestion
error = f'`ocloc` failed with error code {e.returncode}'
```
Isn't this + printing log enough? |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,148 | intel | victor-eds | @@ -37,6 +52,15 @@ inline std::string parseZeResultCode(const ze_result_t code) {
} \
}
+// TODO: share Triton GetEnv.hpp impl
+inline std::string getStrEnv(const std::string &env) {
+ const char *cstr = std::getenv(env.c_str());
+ if ... | Why not do it now? |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,148 | intel | victor-eds | @@ -23,6 +24,20 @@ typedef struct l0_resc_handles {
using SyclQueueMap = std::unordered_map<sycl::queue, l0_resc_handles>;
+// Create an exception handler for asynchronous SYCL exceptions
+auto exception_handler = [](sycl::exception_list e_list) {
+ for (std::exception_ptr const &e : e_list) {
+ try {
+ s... | 1. Why only report message on debug?
2. We are reporting only first as we call `std::terminate`
3. Reporting and `std::terminate`-ing is exactly what the default handler does. Why override? |
intel-xpu-backend-for-triton | github_2023 | python | 2,331 | intel | jopperm | @@ -179,6 +215,78 @@ def sqrt(arg0, _builder=None):
}, is_pure=True, _builder=_builder)
+@core.extern
+def add_rn(arg0, arg1, _builder=None):
+ return core.extern_elementwise(
+ "", "", [arg0, arg1], {
+ (core.dtype("fp64"), core.dtype("fp64")): ("__imf_dadd_rn", core.dtype("fp64")),
+... | Is it intentional that the fp64 version is specified first? |
intel-xpu-backend-for-triton | github_2023 | python | 2,331 | intel | FMarno | @@ -1158,6 +1316,15 @@ def erfcx(arg0, _builder=None):
}, is_pure=True, _builder=_builder)
+@core.extern
+def erfcinv(arg0, _builder=None):
+ return core.extern_elementwise(
+ "", "", [arg0], {
+ (core.dtype("fp32"), ): ("__imf_erfcinvf", core.dtype("fp32")),
+ (core.dtype("... | ```suggestion
(core.dtype("fp64"), ): ("__imf_erfcinvd", core.dtype("fp64")),
```
Is this missing the return? I'm just guessing from usually mangling rules, but I guess these may not be mangled. |
intel-xpu-backend-for-triton | github_2023 | others | 2,318 | intel | victor-eds | @@ -284,4 +284,15 @@ def TritonIntelGPUMaterializeBlockPointer : Pass<"tritonintelgpu-materialize-blo
"mlir::arith::ArithDialect"];
}
+def TritonIntelGPUScheduleOps : Pass<"tritonintelgpu-schedule-ops", "mlir::ModuleOp"> {
+ let summary = "";
+
+ let description = [{
+ }]; | Can we add both summary and description? |
intel-xpu-backend-for-triton | github_2023 | others | 2,304 | intel | whitneywhtsang | @@ -210,6 +210,9 @@ run_interpreter_tests() {
TRITON_INTERPRET=1 TRITON_TEST_SUITE=interpreter \
pytest -vvv -n 16 -m interpreter language/test_core.py language/test_standard.py \
language/test_random.py --device cpu
+
+ TRITON_DISABLE_LINE_INFO=0 TRITON_INTERPRET=1 TRITON_TEST_SUITE=interpreter_line_info... | Is there a reason why not run it as the same command line as above like https://github.com/intel/intel-xpu-backend-for-triton/blob/main/.github/workflows/integration-tests.yml#L255? |
intel-xpu-backend-for-triton | github_2023 | others | 2,272 | intel | pbchekin | @@ -126,6 +126,6 @@ jobs:
mode: ${{ matrix.mode }}
test_mode: accuracy
dtype: ${{ matrix.dtype }}
- models: "all"
+ models: "subset" | Nit: for consistency
```suggestion
models: subset
``` |
intel-xpu-backend-for-triton | github_2023 | others | 2,293 | intel | anmyachev | @@ -121,9 +121,10 @@ jobs:
else
skiplist=
fi
+
{
echo SKIPLIST="$skiplist"
- echo TRITON_TEST_CMD="bash -v -x scripts/test-triton.sh --warning-reports --skip-pytorch-install --reports-dir $GITHUB_WORKSPACE/reports --ignore-errors $skiplist"
+ ... | ?
```suggestion
echo TRITON_TEST_CMD="bash -v -x scripts/test-triton.sh --warning-reports --skip-pytorch-install --reports-dir $GITHUB_WORKSPACE/reports ${{ inputs.ignore_errors && '--ignore-errors' || '' }} $skiplist"
``` |
intel-xpu-backend-for-triton | github_2023 | python | 2,060 | intel | sommerlukas | @@ -230,19 +275,26 @@ def main():
print("Failed to obtain raw data")
sys.exit(1)
- df = parse_directory(num_cfg, None, num_dir)
- df = parse_directory(denom_cfg, df, denom_dir)
+ df = parse_directory(args, num_cfg, None, num_dir)
+ df = parse_directory(args, denom... | Nit: I'd prefer a regular if/else for a multi-line list construction. |
intel-xpu-backend-for-triton | github_2023 | python | 2,060 | intel | sommerlukas | @@ -188,12 +203,40 @@ def eval_data(df: pd.DataFrame, numerator: str, denominator: str, plot: bool):
ax = sns.boxplot(df, x="xlabel", y="relative difference", order=order)
- ax.set(xlabel=None, ylabel="Relative difference in speedup")
+ ax.set(xlabel=None, ylabel="Relative differe... | Why do we need to pass around the entire `args` structure? AFAICT, only two boolean value (no-plot, triton-benchmark) are used from it. |
intel-xpu-backend-for-triton | github_2023 | python | 2,060 | intel | sommerlukas | @@ -74,48 +74,63 @@ def get_raw_data(args: argparse.Namespace) -> tuple[Optional[Path], Optional[Pat
return num_dir, denom_dir
-def parse_data(config: str, df: pd.DataFrame, file: Path) -> pd.DataFrame:
+def parse_data(args: argparse.Namespace, config: str, df: pd.DataFrame, file: Path) -> pd.DataFrame:
"... | The reuse of code between the two different paths in this function seems negligible. Why not have two different functions and call them depending on whether we parse Triton benchmarks or Pytorch benchmarks? |
intel-xpu-backend-for-triton | github_2023 | python | 2,060 | intel | sommerlukas | @@ -142,13 +157,13 @@ def eval_data(df: pd.DataFrame, numerator: str, denominator: str, plot: bool):
df["relative difference"] = (df[num_col] - df[denom_col]) / df[denom_col]
- print("Overview of relative difference in speedup.\n"
+ print("Overview of relative difference in perf.\n"
"Relative ... | What number is printed in the Triton benchmark case? Is that also speedup? |
intel-xpu-backend-for-triton | github_2023 | python | 2,060 | intel | victor-eds | @@ -97,25 +96,65 @@ def parse_data(config: str, df: pd.DataFrame, file: Path) -> pd.DataFrame:
return pd.concat([df, raw_data], ignore_index=True)
-def parse_directory(config: str, previous: pd.DataFrame, directory: Path) -> pd.DataFrame:
+def merge_triton_xetla_reports_data(config: str, triton_file: Path, xet... | ```suggestion
def build_triton_benchmark_reports_path(directory: Path, report_name: str) -> str:
"""Construct the full file path for a given report name."""
return os.path.join(directory, "benchmark-reports", f"{report_name}-report.csv")
```
NIT: use `os.path.join` |
intel-xpu-backend-for-triton | github_2023 | python | 2,060 | intel | victor-eds | @@ -97,25 +96,65 @@ def parse_data(config: str, df: pd.DataFrame, file: Path) -> pd.DataFrame:
return pd.concat([df, raw_data], ignore_index=True)
-def parse_directory(config: str, previous: pd.DataFrame, directory: Path) -> pd.DataFrame:
+def merge_triton_xetla_reports_data(config: str, triton_file: Path, xet... | ```suggestion
reports = ["softmax", "gemm", "attn"]
reports_list = [df]
for report in reports:
triton_file = f"{report}-triton"
xetla_file = f"{report}-xetla"
triton_path = build_triton_benchmark_reports_path(Path(directory), triton_file)
xetla_path = build_triton_be... |
intel-xpu-backend-for-triton | github_2023 | python | 2,060 | intel | victor-eds | @@ -97,25 +96,65 @@ def parse_data(config: str, df: pd.DataFrame, file: Path) -> pd.DataFrame:
return pd.concat([df, raw_data], ignore_index=True)
-def parse_directory(config: str, previous: pd.DataFrame, directory: Path) -> pd.DataFrame:
+def merge_triton_xetla_reports_data(config: str, triton_file: Path, xet... | ```suggestion
on = ["params", "benchmark"] if triton_benchmark else ["suite", "datatype", "mode", "name", "dev"]
df = df.merge(previous, how="outer", on=on)
```
More readable IMO |
intel-xpu-backend-for-triton | github_2023 | python | 2,060 | intel | victor-eds | @@ -169,12 +208,13 @@ def eval_data(df: pd.DataFrame, numerator: str, denominator: str, plot: bool):
import matplotlib.pyplot as plt
from matplotlib.backends.backend_pdf import PdfPages
- df["xlabel"] = df[["suite", "mode", "datatype"]].agg(", ".join, axis=1)
+ df["xlabel"] = df[["para... | ```suggestion
keys = ["params", "benchmark"] if triton_benchmark else ["suite", "mode", "datatype"]
df["xlabel"] = df[keys].agg(", ".join, axis=1)
```
Readability |
intel-xpu-backend-for-triton | github_2023 | python | 2,060 | intel | anmyachev | @@ -97,25 +96,64 @@ def parse_data(config: str, df: pd.DataFrame, file: Path) -> pd.DataFrame:
return pd.concat([df, raw_data], ignore_index=True)
-def parse_directory(config: str, previous: pd.DataFrame, directory: Path) -> pd.DataFrame:
+def merge_triton_xetla_reports_data(config: str, triton_file: Path, xet... | ```suggestion
def parse_triton_benchmark_data(config: str, directory: Path) -> pd.DataFrame:
``` |
intel-xpu-backend-for-triton | github_2023 | python | 2,060 | intel | anmyachev | @@ -97,25 +96,64 @@ def parse_data(config: str, df: pd.DataFrame, file: Path) -> pd.DataFrame:
return pd.concat([df, raw_data], ignore_index=True)
-def parse_directory(config: str, previous: pd.DataFrame, directory: Path) -> pd.DataFrame:
+def merge_triton_xetla_reports_data(config: str, triton_file: Path, xet... | ```suggestion
reports_list = []
``` |
intel-xpu-backend-for-triton | github_2023 | python | 2,060 | intel | anmyachev | @@ -97,25 +96,64 @@ def parse_data(config: str, df: pd.DataFrame, file: Path) -> pd.DataFrame:
return pd.concat([df, raw_data], ignore_index=True)
-def parse_directory(config: str, previous: pd.DataFrame, directory: Path) -> pd.DataFrame:
+def merge_triton_xetla_reports_data(config: str, triton_file: Path, xet... | ```suggestion
triton_path = build_triton_benchmark_reports_path(directory, triton_file)
xetla_path = build_triton_benchmark_reports_path(directory, xetla_file)
``` |
intel-xpu-backend-for-triton | github_2023 | python | 2,060 | intel | anmyachev | @@ -97,25 +96,64 @@ def parse_data(config: str, df: pd.DataFrame, file: Path) -> pd.DataFrame:
return pd.concat([df, raw_data], ignore_index=True)
-def parse_directory(config: str, previous: pd.DataFrame, directory: Path) -> pd.DataFrame:
+def merge_triton_xetla_reports_data(config: str, triton_file: Path, xet... | ```suggestion
df = parse_triton_benchmark_data(config, directory)
``` |
intel-xpu-backend-for-triton | github_2023 | python | 2,060 | intel | anmyachev | @@ -97,25 +96,64 @@ def parse_data(config: str, df: pd.DataFrame, file: Path) -> pd.DataFrame:
return pd.concat([df, raw_data], ignore_index=True)
-def parse_directory(config: str, previous: pd.DataFrame, directory: Path) -> pd.DataFrame:
+def merge_triton_xetla_reports_data(config: str, triton_file: Path, xet... | ```suggestion
def parse_directory(triton_benchmark: bool, config: str, directory: Path) -> pd.DataFrame:
``` |
intel-xpu-backend-for-triton | github_2023 | python | 2,060 | intel | anmyachev | @@ -230,19 +288,29 @@ def main():
print("Failed to obtain raw data")
sys.exit(1)
- df = parse_directory(num_cfg, None, num_dir)
- df = parse_directory(denom_cfg, df, denom_dir)
+ df = parse_directory(args.triton_benchmark, num_cfg, None, num_dir)
+ df = parse_dire... | ```suggestion
num_df = parse_directory(args.triton_benchmark, num_cfg, num_dir)
denom_df = parse_directory(args.triton_benchmark, denom_cfg, denom_dir)
on = ["params", "benchmark"] if args.triton_benchmark else ["suite", "datatype", "mode", "name", "dev"]
df = denom_df.merge(num_df... |
intel-xpu-backend-for-triton | github_2023 | python | 2,300 | intel | whitneywhtsang | @@ -229,8 +229,11 @@ def benchmark(Z, H, N_CTX, D_HEAD, provider):
# FIXME: use torch sdpa for result check after https://github.com/intel/intel-xpu-backend-for-triton/issues/2042 fixed
torch_fn = lambda: torch.nn.functional.scaled_dot_product_attention(
q, k, v, attn_mask=Non... | Can we move the above FIXME to the else clause? |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,283 | intel | victor-eds | @@ -249,8 +206,9 @@ static bool isOCLBuiltinAvailable(TritonGEN::Matrix2DBlockLoadOp op) {
return false;
}
-static Value createGenISA2DBlockRead(TritonGEN::Matrix2DBlockLoadOp op,
- ConversionPatternRewriter &rewriter) {
+[[maybe_unused]] static Value
+createGenISA2DBlockRead(T... | This is used, right? |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,283 | intel | victor-eds | @@ -1013,9 +919,8 @@ struct TritonMatrix2DBlockLoadLowering
return success();
}
- // TODO: Remove GenISA lowering after PoC productization is completed.
- if (tools::getBoolEnv("TRITONGEN_FORCE_GENISA") ||
- !isOCLBuiltinAvailable(op)) {
+ if (!isOCLBuiltinAvailable(op)) {
+ op.emitWa... | Can we split the read in several calls and concat the obtained vectors to avoid using GENISA? |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,291 | intel | victor-eds | @@ -564,13 +564,26 @@ struct LoadOpConversion
unsigned numRepOuter = numReps[opIdx];
unsigned numRepInner = numReps[!opIdx];
+
+ Value pitch = trunc(i32_ty, rowStride);
+ if (!memoryRowMajor) {
+ // Column major memory. We need to swap the width and height because HW
+ // only support row ma... | ```suggestion
Value pitch;
if (memoryRowMajor) {
pitch = trunc(i32_ty, rowStride);
} else {
// Column major memory. We need to swap the width and height because HW
// only support row major memory layout.
pitch = trunc(i32_ty, colStride);
std::swap(baseWidth, baseHeight... |
intel-xpu-backend-for-triton | github_2023 | others | 2,230 | intel | pbchekin | @@ -146,6 +146,12 @@ jobs:
run: |
${{ env.TRITON_TEST_CMD }} --instrumentation --skip-pip-install
+ - name: Run benchmarks | We do not need benchmarks in build-test, they are running daily in triton-benchmarks. |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,294 | intel | victor-eds | @@ -615,9 +615,9 @@ struct TritonGENSubgroupIdLowering
funcAttrs.memEffectsAttr = memory_zero;
intel::AttributeList passthroughAttrs = createFunctionAttributes(
{{llvm::Attribute::NoSync, std::nullopt}}, ctx);
- LLVM::CallOp callOp =
- createDeviceFunctionCall(rewriter, "_Z16get_sub_group_i... | Why move if these are passed by const-ref? |
intel-xpu-backend-for-triton | github_2023 | others | 2,281 | intel | whitneywhtsang | @@ -109,10 +109,12 @@ jobs:
run: |
cd benchmarks/triton_kernels_benchmark
python gemm_benchmark.py --reports $REPORTS
+ mv $REPORTS/matmul-performance.csv $REPORTS/matmul-performance-base.csv
source ../../scripts/capture-hw-details.sh
+
TAG=${{ inputs.tag ||... | Aren't these csv names defined in the benchmarks? Why would this work? |
intel-xpu-backend-for-triton | github_2023 | others | 2,245 | intel | anmyachev | @@ -8,39 +8,96 @@ This is the development repository of Intel® XPU Backend for Triton\*, a new [T
# Compatibility
- |Category|Requirement|Installation|
- |-|-|-|
- |OS|Ubuntu [22.04](http://releases.ubuntu.com/22.04/)| [Install Ubuntu](https://ubuntu.com/tutorials)|
- |GPU Card | Intel® Data Center GPU Max, Fl... | The prerequisites are the same, can there only be one copy of them? |
intel-xpu-backend-for-triton | github_2023 | others | 2,245 | intel | anmyachev | @@ -8,39 +8,96 @@ This is the development repository of Intel® XPU Backend for Triton\*, a new [T
# Compatibility
- |Category|Requirement|Installation|
- |-|-|-|
- |OS|Ubuntu [22.04](http://releases.ubuntu.com/22.04/)| [Install Ubuntu](https://ubuntu.com/tutorials)|
- |GPU Card | Intel® Data Center GPU Max, Fl... | This is not necessary unless we use `--no-build-isolation`. https://github.com/intel/intel-xpu-backend-for-triton/blob/3fc58dff56b558f3814beb038ba234aeb42e9ea1/python/pyproject.toml#L3 will handle build dependencies automatically. |
intel-xpu-backend-for-triton | github_2023 | others | 2,245 | intel | vlad-penkin | @@ -8,39 +8,90 @@ This is the development repository of Intel® XPU Backend for Triton\*, a new [T
# Compatibility
- |Category|Requirement|Installation|
- |-|-|-|
- |OS|Ubuntu [22.04](http://releases.ubuntu.com/22.04/)| [Install Ubuntu](https://ubuntu.com/tutorials)|
- |GPU Card | Intel® Data Center GPU Max, Fl... | I suggest to add the clarification that PTI / Kineto doesn't work for Triton kernels themselves as of now. |
intel-xpu-backend-for-triton | github_2023 | others | 2,245 | intel | vlad-penkin | @@ -8,39 +8,90 @@ This is the development repository of Intel® XPU Backend for Triton\*, a new [T
# Compatibility
- |Category|Requirement|Installation|
- |-|-|-|
- |OS|Ubuntu [22.04](http://releases.ubuntu.com/22.04/)| [Install Ubuntu](https://ubuntu.com/tutorials)|
- |GPU Card | Intel® Data Center GPU Max, Fl... | Let's add a provision in case you installed system wide. |
intel-xpu-backend-for-triton | github_2023 | others | 2,245 | intel | vlad-penkin | @@ -8,39 +8,90 @@ This is the development repository of Intel® XPU Backend for Triton\*, a new [T
# Compatibility
- |Category|Requirement|Installation|
- |-|-|-|
- |OS|Ubuntu [22.04](http://releases.ubuntu.com/22.04/)| [Install Ubuntu](https://ubuntu.com/tutorials)|
- |GPU Card | Intel® Data Center GPU Max, Fl... | Our preference is for user to install our nightly wheels. |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,100 | intel | victor-eds | @@ -50,48 +49,56 @@ using namespace mlir::triton::gpu;
// Helper Functions
//===----------------------------------------------------------------------===//
-static intel::AttributeList createFunctionAttributes(
- ArrayRef<std::pair<llvm::Attribute::AttrKind, std::optional<uint64_t>>>
- attributes,
- ML... | Make static too |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,100 | intel | victor-eds | @@ -50,48 +49,56 @@ using namespace mlir::triton::gpu;
// Helper Functions
//===----------------------------------------------------------------------===//
-static intel::AttributeList createFunctionAttributes(
- ArrayRef<std::pair<llvm::Attribute::AttrKind, std::optional<uint64_t>>>
- attributes,
- ML... | I don't like this. I think simply doing it inline the MLIR way instead of going via the LLVM names is cleaner. |
intel-xpu-backend-for-triton | github_2023 | others | 2,100 | intel | etiotto | @@ -1,5 +1,4 @@
add_triton_library(TritonGENToLLVM
- Attributes.cpp | I think we should continue to build this file. |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,217 | intel | whitneywhtsang | @@ -644,10 +644,9 @@ sycl::event fmha_forward_impl(sycl::queue &q, void *_q, void *_k, void *_v,
uint32_t size_ml = shape.get_ml_size();
// forward
- // T *query = sycl::malloc_shared<T>(size_query, q);
+ T *query = sycl::malloc_shared<T>(size_query, q); | @anmyachev Thanks for the investigation! For a fair comparison, it makes sense to make the operating conditions as close as possible across implementations.
Let's keep this PR up for a day so others can chime in if they have an opinion. |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,227 | intel | victor-eds | @@ -48,6 +48,19 @@ template <typename Op> static LogicalResult verifyMatrixInput(Op op) {
return success();
}
+template <typename Op>
+static LogicalResult verifySIMDBlockTy(Op op, VectorType vecTy) {
+ unsigned numElems = vecTy.getNumElements();
+ IntegerType elemTy = cast<IntegerType>(vecTy.getElementType());... | ```suggestion
static LogicalResult verifySIMDBlockTy(Operation *op, VectorType vecTy) {
unsigned numElems = vecTy.getNumElements();
IntegerType elemTy = cast<IntegerType>(vecTy.getElementType());
// FIXME: Allow 16xi16 when SPIRV-LLVM translator supports it.
if (numElems != 1 && numElems != 2 && numElems... |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,193 | intel | victor-eds | @@ -282,8 +282,17 @@ class LoadStorePrefetchOpConversion
if constexpr (std::is_same_v<OpType, LoadOp>) {
rewriter.restoreInsertionPoint(insertPoint);
- TritonGEN::SIMDBlockReadOp simdRead =
- rewriter.create<TritonGEN::SIMDBlockReadOp>(loc, v64i16Ty, base);
+ VectorType v8i16Ty = Vector... |
```suggestion
VectorType v8i16Ty = VectorType::get(8, i16_ty);
SmallVector<Value> values;
Value offset = i32_val(/*getSubGroupSize()*/);
for (int i = 0; i < 8; ++i) {
auto simdRead =
rewriter.create<TritonGEN::SIMDBlockReadOp>(loc, v8i16Ty, base);
values.push... |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,193 | intel | victor-eds | @@ -435,3 +435,20 @@ LogicalResult TritonGEN::Matrix2DBlockPrefetchOp::verify() {
return success();
}
+
+//===----------------------------------------------------------------------===//
+// gen.simdblockread
+//===----------------------------------------------------------------------===//
+
+LogicalResult TritonG... | For future: Change operation signature to use scalars instead of vectors of length 1. |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,193 | intel | etiotto | @@ -282,8 +282,20 @@ class LoadStorePrefetchOpConversion
if constexpr (std::is_same_v<OpType, LoadOp>) {
rewriter.restoreInsertionPoint(insertPoint);
- TritonGEN::SIMDBlockReadOp simdRead =
- rewriter.create<TritonGEN::SIMDBlockReadOp>(loc, v64i16Ty, base);
+ const unsigned maxOCLVector... | Use constexpr to declare the variable |
intel-xpu-backend-for-triton | github_2023 | others | 2,219 | intel | pbchekin | @@ -66,11 +66,7 @@ jobs:
- name: Matrix
id: matrix
run: |
- if [[ -n "${{ inputs.runner_label }}" ]]; then
- matrix='{"python": ["3.9", "3.10", "3.11", "3.12"]}'
- else
- matrix='{"python": ["3.9", "3.10", "3.11", "3.12"], "driver": ["rolling", "lts"]}'
-... | This is not what we discussed.
For the normal flow you need 3 matrixes:
* python for inductor tests
* python x driver for integration tests
* e2e matrix
|
intel-xpu-backend-for-triton | github_2023 | cpp | 2,216 | intel | whitneywhtsang | @@ -402,25 +405,55 @@ void PrefetchBlockPass::injectPrefetchOpsInBody(
auto yield = cast<scf::YieldOp>(newLoop.getBody()->getTerminator());
loop.erase();
+ Attribute attr = loop->getAttr(AttrWorkloadName);
+ assert(attr && "Expecting a workload attribute");
+ Workload workload = static_cast<Workload>(cast<In... | please remove |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,233 | intel | victor-eds | @@ -853,7 +853,7 @@ static Value allocateSLMForTransposedReduction(tt::ReduceOp op, unsigned step,
Location loc = op.getLoc();
// Fixed size for num_warps matrices of sg_size^2 shape.
- int64_t size = step * step * srcTy.getElementTypeBitWidth() / 8 *
+ int64_t size = (int64_t)step * step * srcTy.getElementTy... | ```suggestion
int64_t size = static_cast<int64_t>(step) * step * srcTy.getElementTypeBitWidth() / 8 *
```
Can we go with that instead? (No C-cast) |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,180 | intel | jopperm | @@ -428,7 +428,8 @@ class TritonIntelGPUDistributeToWarpsPass
else if (auto makeRange = dyn_cast<tt::MakeRangeOp>(op))
distributeMakeRangeOp(makeRange, warpId);
else if (isa<tt::LoadOp, tt::DotOp, tt::AdvanceOp, tt::ReduceOp,
- tt::SplatOp, tt::BroadcastOp, tt::ExpandDim... | Can both broadcast ops still occur? |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,180 | intel | jopperm | @@ -751,6 +766,8 @@ class ConvertTritonToTritonGPUWarp
} else if (isa<tt::SplatOp, tt::BroadcastOp, tt::ReduceOp,
tt::MakeRangeOp>(def)) {
chainedVals.insert(def->getResult(0));
+ } else if (isa<tt::ExpandDimsOp>(def)) {
+ ; | Maybe add a comment why this is a no-op? |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -58,6 +58,7 @@ permissions: read-all
env:
TRITON_DISABLE_LINE_INFO: 1
TEST_UNSKIP: ${{ inputs.enable_unskip }}
+ GH_TOKEN: ${{ github.token }} | `GH_TOKEN` is not required to run tests. If `test-triton.sh` requires it, then the script needs to be refactored to not use it to just run tests. |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -37,8 +37,8 @@ on:
branches:
- llvm-target
push:
- branches:
- - llvm-target
+ # branches:
+ # - llvm-target | Why this is commented out? If PR is not ready for merging it needs to marked as draft to avoid accidental merging, |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -270,7 +228,10 @@ jobs:
- name: Pass rate
run: |
+ set -x
source ./scripts/capture-hw-details.sh
+ find $TRITON_TEST_REPORTS_DIR | Why is this needed? |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -270,7 +228,10 @@ jobs:
- name: Pass rate
run: |
+ set -x | This step is now overflown with bash output, do we really need `set -x`? |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -311,6 +315,26 @@ run_benchmark_attention() {
python ${BENCHMARK_TEST_DIR}/flash_attention_fwd_benchmark.py
}
+run_instrumentation_tests() {
+ set -x
+ # FIXME: the "instrumentation" test suite currently contains only one test, when all tests
+ # are skipped pytest reports an error. If the only test is the ... | Out of curiosity, why `exit -1`? |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -115,8 +115,6 @@ jobs:
- name: Create directory for tests reports
run: |
mkdir reports
- echo "TRITON_TEST_REPORTS=true" >> $GITHUB_ENV
- echo "TRITON_TEST_WARNING_REPORTS=true" >> $GITHUB_ENV | If you keep these environment variables, then you can skip specifying `--reports --warning-reports` for `test-triton.sh` in the steps below. |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -112,117 +112,46 @@ jobs:
cd python
lit -v build/*/test
- - name: Create directory for tests reports
+ - name: Create test-triton command line
run: |
- mkdir reports
- echo "TRITON_TEST_REPORTS=true" >> $GITHUB_ENV
- echo "TRITON_TEST_WARNING_REPO... | Why there is no `--skip-...` option as in other steps? |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -112,117 +112,46 @@ jobs:
cd python
lit -v build/*/test
- - name: Create directory for tests reports
+ - name: Create test-triton command line
run: |
- mkdir reports
- echo "TRITON_TEST_REPORTS=true" >> $GITHUB_ENV
- echo "TRITON_TEST_WARNING_REPO... | Why only `--skip-pip-install`? PyTorch should be installed at this moment already. |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -13,14 +18,17 @@ TEST_MICRO_BENCHMARKS=false
TEST_BENCHMARK_SOFTMAX=false
TEST_BENCHMARK_GEMM=false
TEST_BENCHMARK_ATTENTION=false
+TEST_INSTRUMENTATION=false
VENV=false
TRITON_TEST_REPORTS=false
TRITON_TEST_WARNING_REPORTS=false
TRITON_TEST_IGNORE_ERRORS=false
-SKIP_DEPS=false
+SKIP_PIP=false
+SKIP_PYTORCH=f... | Not used. |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -77,19 +98,21 @@ for arg in "$@"; do
TRITON_TEST_IGNORE_ERRORS=true
shift
;;
+ --skip-list)
+ TRITON_TEST_SKIPLIST_DIR="$2"
+ shift 2
+ ;;
--help)
- echo "Example usage: ./test-triton.sh [--core | --tutorial | --unit | --microbench | --softmax | --gemm | --attention |... | The list of command line options is not up to date. |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -106,59 +129,58 @@ if [ "$VENV" = true ]; then
source .venv/bin/activate
fi
-SCRIPTS_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
-TRITON_PROJ="$BASE/intel-xpu-backend-for-triton"
-
+export SCRIPTS_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
source ... | `TRITON_PROJ_BUILD` is required only for unit tests. |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -106,59 +129,58 @@ if [ "$VENV" = true ]; then
source .venv/bin/activate
fi
-SCRIPTS_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
-TRITON_PROJ="$BASE/intel-xpu-backend-for-triton"
-
+export SCRIPTS_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
source ... | Regression, this was moved to `run_unit_tests` in #2213. |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -106,59 +129,58 @@ if [ "$VENV" = true ]; then
source .venv/bin/activate
fi
-SCRIPTS_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
-TRITON_PROJ="$BASE/intel-xpu-backend-for-triton"
-
+export SCRIPTS_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
source ... | Regression, the list of dependencies has been moved to `requirements-test.txt` in #2213. |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -106,59 +129,58 @@ if [ "$VENV" = true ]; then
source .venv/bin/activate
fi
-SCRIPTS_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
-TRITON_PROJ="$BASE/intel-xpu-backend-for-triton"
-
+export SCRIPTS_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
source ... | Not needed, it is installed in `install_deps`. |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -106,59 +129,58 @@ if [ "$VENV" = true ]; then
source .venv/bin/activate
fi
-SCRIPTS_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
-TRITON_PROJ="$BASE/intel-xpu-backend-for-triton"
-
+export SCRIPTS_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
source ... | We want to keep the error message. When Triton is installed, for example, from wheels the `python/build` directory does not exist and unit tests cannot be run. A user needs to know that Triton must be built from sources to run unit tests. |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -271,18 +293,35 @@ run_benchmark_attention() {
echo "Default path:"
TRITON_INTEL_ADVANCED_PATH=0 \
- TRITON_INTEL_ENABLE_ADDRESS_PAYLOAD_OPT=1 \
- IGC_VISAOptions=" -enableBCR -nolocalra -printregusage -DPASTokenReduction -enableHalfLSC" \
- IGC_DisableLoopUnroll=1 \
- python $TRITON_PROJ/benchmarks/trit... | The error message does not make sense: if there is an error with `ls` then `SHARED_LIB_DIR` is not set and `err` prints
```
Could not find ''
``` |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -106,59 +128,51 @@ if [ "$VENV" = true ]; then
source .venv/bin/activate
fi
-SCRIPTS_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
-TRITON_PROJ="$BASE/intel-xpu-backend-for-triton"
-
+export SCRIPTS_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )" | ```suggestion
SCRIPTS_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
``` |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -106,59 +128,51 @@ if [ "$VENV" = true ]; then
source .venv/bin/activate
fi
-SCRIPTS_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
-TRITON_PROJ="$BASE/intel-xpu-backend-for-triton"
-
+export SCRIPTS_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
source ... | ```suggestion
TRITON_PROJ=$BASE/intel-xpu-backend-for-triton
``` |
intel-xpu-backend-for-triton | github_2023 | python | 2,183 | intel | pbchekin | @@ -198,7 +202,9 @@ def print_json_stats(stats: ReportStats):
def main():
"""Main."""
args = create_argument_parser().parse_args()
- stats = parse_reports(pathlib.Path(args.reports))
+ skiplist_dir = args.skip_list if args.skip_list else os.getenv('TRITON_TEST_SKIPLIST_DIR',
+ ... | Since `skiplist_dir` is a path, i would recommend to use use `pathlib.Path` for it.
```suggestion
skiplist_dir = args.skip_list if args.skip_list else os.getenv('TRITON_TEST_SKIPLIST_DIR',
'scripts/skiplist/default')
stats = parse_reports(pa... |
intel-xpu-backend-for-triton | github_2023 | python | 2,183 | intel | pbchekin | @@ -72,7 +76,7 @@ def get_deselected(report_path: pathlib.Path) -> int:
return len([line for line in f.readlines() if line and not line.startswith('#')])
-def parse_report(report_path: pathlib.Path) -> ReportStats:
+def parse_report(report_path: pathlib.Path, skiplist_dir: str) -> ReportStats: | Use `pathlib.Path` instead `str`:
```suggestion
def parse_report(report_path: pathlib.Path, skiplist_dir: pathlib.Path) -> ReportStats:
``` |
intel-xpu-backend-for-triton | github_2023 | python | 2,183 | intel | pbchekin | @@ -58,12 +58,16 @@ def create_argument_parser() -> argparse.ArgumentParser:
default='all',
help='name of the test suite, default: %(default)s',
)
+ argument_parser.add_argument(
+ '--skip-list',
+ type=str,
+ help='an exclude list dir used in pass rate calculation, can be... | ```suggestion
def get_deselected(report_path: pathlib.Path, skiplist_dir: pathlib.Path) -> int:
``` |
intel-xpu-backend-for-triton | github_2023 | python | 2,183 | intel | pbchekin | @@ -58,12 +58,16 @@ def create_argument_parser() -> argparse.ArgumentParser:
default='all',
help='name of the test suite, default: %(default)s',
)
+ argument_parser.add_argument(
+ '--skip-list',
+ type=str,
+ help='an exclude list dir used in pass rate calculation, can be... | ```suggestion
skiplist_path = skiplist_dir / f'{report_path.stem}.txt'
``` |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -106,59 +128,51 @@ if [ "$VENV" = true ]; then
source .venv/bin/activate
fi
-SCRIPTS_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
-TRITON_PROJ="$BASE/intel-xpu-backend-for-triton"
-
+export SCRIPTS_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
source ... | This check must not be here, see #2187. The script should work for all test suites except `unit` and `instrumentation` even if `python/build` does not exist. |
intel-xpu-backend-for-triton | github_2023 | others | 2,183 | intel | pbchekin | @@ -97,11 +97,6 @@ jobs:
repository: pytorch/pytorch
ref: ${{ inputs.pytorch_ref }}
- - name: Install test dependencies
- run: |
- pip install pytest pytest-xdist pytest-rerunfailures pytest-select pytest-timeout expecttest defusedxml | You still need to install `defusedxml`. It is required for `pass-rate.py`. |
intel-xpu-backend-for-triton | github_2023 | others | 2,221 | intel | pbchekin | @@ -184,7 +184,7 @@ build_triton() {
pip install -vvv -e '.[build,tests]'
# Copy compile_commands.json in the build directory (so that cland vscode plugin can find it).
- cp $TRITON_PROJ_BUILD/"$(ls $TRITON_PROJ_BUILD)"/compile_commands.json $TRITON_PROJ/
+ cp $(find $TRITON_PROJ_BUILD -name compile_commands.... | This should also work:
```suggestion
cp $TRITON_PROJ_BUILD/*/compile_commands.json $TRITON_PROJ/
``` |
intel-xpu-backend-for-triton | github_2023 | python | 2,200 | intel | ZzEeKkAa | @@ -112,6 +32,124 @@ def extract_kernels(funcs):
return getattr(torch, return_mode)(times).item()
+if USE_IPEX_OPTION: | Can we have definde two functions unconditionally let's say `do_bench_ipex` and `do_bench_no_ipex`. And define do_bench conditionally. It is to avoid waterfall code. |
intel-xpu-backend-for-triton | github_2023 | others | 1,269 | intel | etiotto | @@ -1,6 +1,5 @@
// RUN: triton-opt %s -split-input-file --convert-triton-intel-gpu-to-llvm | FileCheck %s
-// CHECK-DAG: llvm.func spir_funccc @_Z27__spirv_ConvertFToBF16INTELf(f32) -> i16 | Add the CHECK for the OpenCL function declaration `_Z32intel_convert_bfloat16_as_ushortf`. It would be useful to check that the SPIRV calling convention is on the function decl. |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,109 | intel | Dewei-Wang-sh | @@ -793,14 +830,24 @@ void MatchTargetSizePass::transformReduceOp(tt::ReduceOp op) {
int64_t outer = dims == 2 ? srcTy.getShape()[0] : 1;
SmallVector<Value> glueVals;
- unsigned step = 8; // FIXME: fixed to 8 for now.
+ // Fixed for reduce transpose. | can you put this to a separate function?
the default way is not to use the transposed reduction. |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,109 | intel | Dewei-Wang-sh | @@ -0,0 +1,23 @@
+//===- Utils.h - TritonIntelGPU Utils -----------------------------------*-===//
+//
+// 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
+//
+//===---... | triton has a default method wrapper to get bool env. |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,109 | intel | whitneywhtsang | @@ -551,6 +555,135 @@ class ReduceOpConversion : public ConvertTritonGPUOpToLLVMPattern<ReduceOp> {
}
};
+class TransposedReduceOpConversion
+ : public ConvertTritonGPUOpToLLVMPattern<ReduceOp> {
+public:
+ using ConvertTritonGPUOpToLLVMPattern<
+ ReduceOp>::ConvertTritonGPUOpToLLVMPattern;
+ LogicalRes... | `TritonGEN::SubGroupShuffleOp` is removed. |
intel-xpu-backend-for-triton | github_2023 | cpp | 2,109 | intel | whitneywhtsang | @@ -805,7 +807,141 @@ Value MatchTargetSizePass::getSubVal(Operation *op, Value val,
return dstVal;
}
+static Value hackAlloc(OpBuilder &b, Location loc, Type ptrTy, int64_t size) {
+ auto func = static_cast<FunctionOpInterface>(
+ &*b.getInsertionPoint()
+ ->getParentWithTrait<FunctionOpInterfac... | is it more robust to get subgroup size from module? |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.