optimize unique and index_put (#56582)

lj970926 · web-flow · commit d674ea95886d · 2023-08-28T11:51:05.000+08:00
diff --git a/paddle/phi/kernels/funcs/index_put_utils.h b/paddle/phi/kernels/funcs/index_put_utils.h
@@ -15,6 +15,7 @@
 #pragma once
 
 #include <vector>
+#include "paddle/phi/backends/context_pool.h"
 #include "paddle/phi/common/int_array.h"
 #include "paddle/phi/common/memory_utils.h"
 #include "paddle/phi/common/place.h"
@@ -106,7 +107,14 @@ std::vector<const phi::DenseTensor*> DealWithBoolIndices(
       SplitWithNumKernel<int64_t, Context>(
           dev_ctx, nonzero_indices, rank, 1, integer_indices);
 #ifdef PADDLE_WITH_XPU
-      dev_ctx.Wait();
+      auto place = dev_ctx.GetPlace();
+      if (place.GetType() == phi::AllocationType::XPU) {
+        auto& pool = phi::DeviceContextPool::Instance();
+        auto* xpu_ctx = static_cast<phi::XPUContext*>(pool.Get(place));
+        if (xpu_ctx->x_context()->xpu_stream) {
+          dev_ctx.Wait();
+        }
+      }
 #endif
 
     } else if ((indices_v[i]->dtype() == phi::DataType::INT64) ||
diff --git a/paddle/phi/kernels/xpu/index_put_kernel.cc b/paddle/phi/kernels/xpu/index_put_kernel.cc
@@ -65,7 +65,9 @@ void XPUDealWithIndices(const Context& dev_ctx,
   }
 
   StackKernel<int64_t, Context>(dev_ctx, tmp_indices_ptr, -1, out);
-  dev_ctx.Wait();
+  if (dev_ctx.x_context()->xpu_stream) {
+    dev_ctx.Wait();
+  }
 }
 
 template <typename T, typename Context>
@@ -140,7 +142,9 @@ void IndexPutKernel(const Context& dev_ctx,
                                     index_shape,
                                     accumulate);
   PADDLE_ENFORCE_XDNN_SUCCESS(r, "index_put");
-  dev_ctx.Wait();
+  if (dev_ctx.x_context()->xpu_stream) {
+    dev_ctx.Wait();
+  }
 }
 }  // namespace phi
 
diff --git a/paddle/phi/kernels/xpu/unique_kernel.cc b/paddle/phi/kernels/xpu/unique_kernel.cc
@@ -228,26 +228,37 @@ void XPUDimUniqueKernelImpl(const Context& dev_ctx,
   inverse_cpu[ori_idx_cpu[0]] = 0;
   IndexT unique_len = 1;
   IndexT repeat_cnt = 1;
-  for (IndexT i = 1; i < axis_len; ++i) {
-    int cnt_cpu = 0;
-    int* cnt_xpu = RAII_GUARD.alloc_l3_or_gm<int>(1);
-    r = xpu::nonzero_count<bool>(dev_ctx.x_context(),
-                                 compare_results + (i - 1) * slice_size,
-                                 cnt_xpu,
-                                 slice_size);
-    PADDLE_ENFORCE_XDNN_SUCCESS(r, "nonzero_count");
-    memory_utils::Copy(
-        phi::CPUPlace(), &cnt_cpu, dev_ctx.GetPlace(), cnt_xpu, sizeof(int));
-    if (cnt_cpu != slice_size) {
-      unique_axis.push_back(i);
-      indices_cpu.push_back(ori_idx_cpu[i]);
-      counts_cpu.push_back(repeat_cnt);
-      ++unique_len;
-      repeat_cnt = 1;
-    } else {
-      ++repeat_cnt;
+  if (axis_len > 1) {
+    DenseTensor adj_identical_cpu;
+    adj_identical_cpu.Resize({axis_len - 1});
+    bool* adj_identical_cpu_data =
+        dev_ctx.template HostAlloc<bool>(&adj_identical_cpu);
+    auto* adj_identical_xpu = RAII_GUARD.alloc_l3_or_gm<bool>(axis_len - 1);
+    r = xpu::reduce_all<bool>(dev_ctx.x_context(),
+                              compare_results,
+                              adj_identical_xpu,
+                              {axis_len - 1, slice_size},
+                              {1});
+    PADDLE_ENFORCE_XDNN_SUCCESS(r, "reduce_all");
+
+    memory_utils::Copy(phi::CPUPlace(),
+                       adj_identical_cpu_data,
+                       dev_ctx.GetPlace(),
+                       adj_identical_xpu,
+                       (axis_len - 1) * sizeof(bool));
+
+    for (IndexT i = 1; i < axis_len; ++i) {
+      if (!adj_identical_cpu_data[i - 1]) {
+        unique_axis.push_back(i);
+        indices_cpu.push_back(ori_idx_cpu[i]);
+        counts_cpu.push_back(repeat_cnt);
+        ++unique_len;
+        repeat_cnt = 1;
+      } else {
+        ++repeat_cnt;
+      }
+      inverse_cpu[ori_idx_cpu[i]] = unique_len - 1;
     }
-    inverse_cpu[ori_idx_cpu[i]] = unique_len - 1;
   }
   counts_cpu.push_back(repeat_cnt);
   DDim out_dims = x.dims();

Original file line number	Diff line number	Diff line change
`@@ -65,7 +65,9 @@ void XPUDealWithIndices(const Context& dev_ctx,`
`65`	`65`	`}`
`66`	`66`
`67`	`67`	`StackKernel<int64_t, Context>(dev_ctx, tmp_indices_ptr, -1, out);`
`68`		`- dev_ctx.Wait();`
	`68`	`+ if (dev_ctx.x_context()->xpu_stream) {`
	`69`	`+ dev_ctx.Wait();`
	`70`	`+ }`
`69`	`71`	`}`
`70`	`72`
`71`	`73`	`template <typename T, typename Context>`
`@@ -140,7 +142,9 @@ void IndexPutKernel(const Context& dev_ctx,`
`140`	`142`	`index_shape,`
`141`	`143`	`accumulate);`
`142`	`144`	`PADDLE_ENFORCE_XDNN_SUCCESS(r, "index_put");`
`143`		`- dev_ctx.Wait();`
	`145`	`+ if (dev_ctx.x_context()->xpu_stream) {`
	`146`	`+ dev_ctx.Wait();`
	`147`	`+ }`
`144`	`148`	`}`
`145`	`149`	`} // namespace phi`
`146`	`150`