Merge pull request #59 from InfiniTensor/dev

Dev

Author: YdrMaster

scripts/compare/compare.py

@@ -23,6 +23,7 @@ def parse_args():
         args.actual,
     )
 
+
 def getDiff(base, test):
     absolute_diff = np.subtract(base, test)
     max_absolute_diff = np.max(np.abs(absolute_diff))
@@ -35,16 +36,19 @@ def getDiff(base, test):
 
     return max_absolute_diff, max_relative_diff
 
-def compare_npy(actual_path, expect_path, edge, node):
+
+def compare_npy(node, actual_path, expect_path):
     actual = np.load(actual_path)
     expect = np.load(expect_path)
     if np.isnan(actual).any():
-        print(f"NAN value in node:{node} edge:{edge}")
+        print(f"NAN value in node:{node}\t{actual_path}\t{expect_path}")
         return
-    
+
     max_absolute_diff, max_relative_diff = getDiff(expect, actual)
-    if max_absolute_diff != 0.0: ## No need to print tensor with no diff
-        print(f'{max_absolute_diff}\t{max_relative_diff}\t{node}\t{edge}')
+    if max_absolute_diff != 0.0:  ## No need to print tensor with no diff
+        print(
+            f"{max_absolute_diff}\t{max_relative_diff}\t{node}\t{actual_path}\t{expect_path}"
+        )
 
 
 def main():
@@ -70,9 +74,7 @@ def main():
                     expect_file = expect_file + ".npy"
                     expect_file_path = os.path.join(expect_dir, expect_file)
                     if os.path.exists(expect_file_path):
-                        compare_npy(
-                            actual_file_path, expect_file_path, edge_name, node_name
-                        )
+                        compare_npy(meta_file, actual_file_path, expect_file_path)
 
 
 if __name__ == "__main__":

src/04kernel/include/kernel/attributes/broadcaster.h

@@ -15,6 +15,7 @@ namespace refactor::kernel {
         explicit Broadcaster(std::vector<slice_t<dim_t>>);
         explicit Broadcaster(TensorRefs const &inputs);
         void locate(dim_t k, dim_t ans[]) const noexcept;
+        bool needBroadcast() const noexcept;
     };
 
 }// namespace refactor::kernel

src/04kernel/include/kernel/attributes/matmul_info.h renamed to src/04kernel/include/kernel/attributes/mat_mul_info.h

@@ -1,21 +1,20 @@
-#ifndef KERNEL_MATMUL_INFO_H
-#define KERNEL_MATMUL_INFO_H
+#ifndef KERNEL_MAT_MUL_INFO_H
+#define KERNEL_MAT_MUL_INFO_H
 
 #include "kernel/attributes/broadcaster.h"
 #include "kernel/attributes/expand_info.h"
-#include <variant>
 
 namespace refactor::kernel {
 
     struct MatMulInfo {
         DataType dataType;
         float alpha, beta;
         bool transA, transB;
-        size_t m, k, n;
+        dim_t m, k, n;
         // Expand operation info for biasd
         std::optional<ExpandInfo> biasExpand;
-        // A constant batch or a 2-directional broadcaster that deals with dimensions before the last 2 dimensions
-        std::variant<Broadcaster, size_t> broadcasterOrBatch;
+        // A 2-directional broadcaster that deals with dimensions before the last 2 dimensions
+        Broadcaster broadcaster;
 
         MatMulInfo(Tensor const &, Tensor const &,
                    std::optional<std::reference_wrapper<Tensor const>>,
@@ -24,4 +23,4 @@ namespace refactor::kernel {
 
 }// namespace refactor::kernel
 
-#endif// KERNEL_MATMUL_INFO_H
+#endif// KERNEL_MAT_MUL_INFO_H

src/04kernel/include/kernel/attributes/mat_mul_integer_info.h

@@ -0,0 +1,28 @@
+#ifndef KERNEL_MAT_MUL_INTEGER_INFO_H
+#define KERNEL_MAT_MUL_INTEGER_INFO_H
+
+#include "kernel/attributes/broadcaster.h"
+
+namespace refactor::kernel {
+
+    struct MatMulIntegerInfo {
+        struct Input {
+            bool
+                withZeroPoint,
+                signed_,
+                scalar;
+
+            Input(TensorRefs const &, size_t i) noexcept;
+        };
+
+        Input a, b;
+        dim_t m, k, n;
+        Broadcaster broadcaster;
+
+        explicit MatMulIntegerInfo(TensorRefs const &inputs) noexcept;
+        dim_t batch() const noexcept;
+    };
+
+}// namespace refactor::kernel
+
+#endif// KERNEL_MAT_MUL_INTEGER_INFO_H

src/04kernel/include/kernel/collectors/dequantize_linear.h

@@ -0,0 +1,18 @@
+#ifndef KERNEL_DEQUANTIZE_LINEAR_H
+#define KERNEL_DEQUANTIZE_LINEAR_H
+
+#include "../collector.h"
+
+namespace refactor::kernel {
+
+    struct DequantizeLinearCollector final : public InfoCollector {
+
+        explicit DequantizeLinearCollector(decltype(_target)) noexcept;
+
+        std::vector<KernelBox>
+        filter(TensorRefs inputs, TensorRefs outputs) const final;
+    };
+
+}// namespace refactor::kernel
+
+#endif// KERNEL_DEQUANTIZE_LINEAR_H

src/04kernel/include/kernel/collectors/dynamic_quantize_linear.h

@@ -0,0 +1,18 @@
+#ifndef KERNEL_DYNAMIC_QUANTIZE_LINEAR_H
+#define KERNEL_DYNAMIC_QUANTIZE_LINEAR_H
+
+#include "../collector.h"
+
+namespace refactor::kernel {
+
+    struct DynamicQuantizeLinearCollector final : public InfoCollector {
+
+        explicit DynamicQuantizeLinearCollector(decltype(_target)) noexcept;
+
+        std::vector<KernelBox>
+        filter(TensorRefs inputs, TensorRefs outputs) const final;
+    };
+
+}// namespace refactor::kernel
+
+#endif// KERNEL_DYNAMIC_QUANTIZE_LINEAR_H

src/04kernel/include/kernel/collectors/mat_mul_integer.h

@@ -0,0 +1,19 @@
+#ifndef KERNEL_MAT_MUL_INTEGER_H
+#define KERNEL_MAT_MUL_INTEGER_H
+
+#include "../collector.h"
+
+namespace refactor::kernel {
+
+    struct MatMulIntegerCollector final : public InfoCollector {
+
+        constexpr MatMulIntegerCollector(decltype(_target) target) noexcept
+            : InfoCollector(target) {}
+
+        std::vector<KernelBox>
+        filter(TensorRefs inputs, TensorRefs outputs) const final;
+    };
+
+}// namespace refactor::kernel
+
+#endif// KERNEL_MAT_MUL_INTEGER_H

src/04kernel/src/attributes/broadcaster.cc

@@ -96,4 +96,8 @@ namespace refactor::kernel {
         }
     }
 
+    bool Broadcaster::needBroadcast() const noexcept {
+        return !strides.empty();
+    }
+
 }// namespace refactor::kernel

src/04kernel/src/attributes/matmul_info.cc renamed to src/04kernel/src/attributes/mat_mul_info.cc

@@ -1,19 +1,17 @@
-#include "kernel/attributes/matmul_info.h"
-#include <cstddef>
-#include <numeric>
+#include "kernel/attributes/mat_mul_info.h"
 
 namespace refactor::kernel {
 
-    ExpandInfo buildBias(size_t m, size_t n,
+    ExpandInfo buildBias(dim_t m, dim_t n,
                          Tensor const &a,
                          Tensor const &b,
                          Tensor const &c) {
         std::vector<dim_t> output(std::max(a.rank(), b.rank()));
         auto it = output.rbegin();
         *it++ = n;
         *it++ = m;
-        for (auto da = static_cast<size_t>(a.rank() - 2),
-                  db = static_cast<size_t>(b.rank() - 2);
+        for (auto da = static_cast<dim_t>(a.rank() - 2),
+                  db = static_cast<dim_t>(b.rank() - 2);
              auto i : range0_(output.size() - 2)) {
             auto a_ = i < da ? a.shape[da - i - 1] : 1;
             auto b_ = i < db ? b.shape[db - i - 1] : 1;
@@ -26,13 +24,6 @@ namespace refactor::kernel {
             slice(output.data(), output.size()));
     }
 
-    std::variant<Broadcaster, size_t> buildBroadcasterOrBatch(slice_t<dim_t> dimA, slice_t<dim_t> dimB) {
-        if (std::equal(dimA.begin(), dimA.end(), dimB.begin(), dimB.end())) {
-            return std::accumulate(dimA.begin(), dimA.end(), (size_t) 1, std::multiplies<size_t>());
-        }
-        return Broadcaster({dimA, dimB});
-    }
-
     MatMulInfo::MatMulInfo(
         Tensor const &a, Tensor const &b,
         std::optional<std::reference_wrapper<Tensor const>> c,
@@ -44,7 +35,8 @@ namespace refactor::kernel {
           k(transA ? a.shape.rbegin()[1] : a.shape.rbegin()[0]),
           n(transB ? b.shape.rbegin()[1] : b.shape.rbegin()[0]),
           biasExpand(c ? std::make_optional(buildBias(m, n, a, b, *c)) : std::nullopt),
-          broadcasterOrBatch(buildBroadcasterOrBatch(slice(a.shape.data(), a.shape.size() - 2), slice(b.shape.data(), b.shape.size() - 2))) {
+          broadcaster({slice(a.shape.data(), a.shape.size() - 2),
+                       slice(b.shape.data(), b.shape.size() - 2)}) {
         auto kB = transB ? b.shape.rbegin()[0] : b.shape.rbegin()[1];
         ASSERT(k == kB, "MatMul: input shape not matched.");
     }

src/04kernel/src/attributes/mat_mul_integer_info.cc

@@ -0,0 +1,42 @@
+#include "kernel/attributes/mat_mul_integer_info.h"
+
+namespace refactor::kernel {
+
+    MatMulIntegerInfo::Input::Input(TensorRefs const &inputs, size_t i) noexcept
+        : withZeroPoint(false),
+          signed_(true),
+          scalar(true) {
+        if (inputs.size() > i + 2) {
+            auto const &t = inputs[i + 2].get();
+            auto size = t.elementsSize();
+            if (t.data) {
+                auto data = slice(t.data->get<uint8_t>(), size);
+                if (std::all_of(data.begin(), data.end(), [](auto x) { return x == 0; })) {
+                    return;
+                }
+            }
+            withZeroPoint = true;
+            signed_ = t.dataType == DataType::I8;
+            scalar = size == 1;
+        }
+    }
+
+    MatMulIntegerInfo::MatMulIntegerInfo(TensorRefs const &inputs) noexcept
+        : a(inputs, 0),
+          b(inputs, 1),
+#define A (inputs[0].get().shape)
+#define B (inputs[1].get().shape)
+          m(A.rbegin()[1]),
+          k(A.rbegin()[0]),
+          n(B.rbegin()[0]),
+          broadcaster({slice(A.data(), A.size() - 2),
+                       slice(B.data(), B.size() - 2)}) {
+    }
+#undef A
+#undef B
+
+    dim_t MatMulIntegerInfo::batch() const noexcept {
+        return broadcaster.outputsCount;
+    }
+
+}// namespace refactor::kernel