Fix DequantizeTwoScale kernel (#45632)

paddle/fluid/operators/fake_dequantize_op.cu.h

@@ -88,16 +88,14 @@ __global__ void DequantizeTwoScale(const T* in,
                                    const T* scale_two,
                                    T max_range,
                                    int num,
-                                   int iter_size,
-                                   int channel,
+                                   int n_scales,
+                                   int quant_stride,
                                    T* out) {
-  int tid = threadIdx.x;
-  int channel_size = num / (iter_size * channel);
-  int scale_index = blockIdx.x % channel;
-  const T* in_c = in + blockIdx.x * channel_size;
-  T* out_c = out + blockIdx.x * channel_size;
-  for (int i = tid; i < channel_size; i += blockDim.x) {
-    out_c[i] = in_c[i] * scale_one[scale_index] * scale_two[0] / max_range;
+  int64_t idx = blockDim.x * blockIdx.x + threadIdx.x;
+  for (int64_t i = idx; i < num; i += blockDim.x * gridDim.x) {
+    int scale_index = (i / quant_stride) % n_scales;
+    T s = scale_one[scale_index] * scale_two[0];
+    out[i] = in[i] * s / max_range;
   }
 }
 
@@ -115,6 +113,8 @@ struct ChannelDequantizeFunctor<phi::GPUContext, T> {
     const T* in_data = in->data<T>();
     T* out_data = out->mutable_data<T>(dev_ctx.GetPlace());
     if (scale_num == 1) {
+      // Dequantize inputs or weights before quantizable operators and after
+      // quantization operators. inputs --> quant -- > deqaunt --> conv2d -->
       int64_t num = in->numel();
       const T* scale_factor = scales[0]->data<T>();
       int64_t block_size = std::min(
@@ -140,24 +140,38 @@ struct ChannelDequantizeFunctor<phi::GPUContext, T> {
                                                            quant_stride,
                                                            out_data);
     } else if (scale_num == 2) {
-      // Not need to consider quant_axis
-      int num = in->numel();
-      int iter_size = 1;
-      for (int i = 0; i < x_num_col_dims; i++) {
-        iter_size *= in->dims()[i];
-      }
-      int channel = in->dims()[x_num_col_dims];
+      // Dequantize activations after quantizable operators.
+      // inputs --> quant --> conv2d --> deqaunt -->
+      // Note 1:  Not need to consider 'quant_axis'. Because 'quant_aixs' is the
+      // axis of weights to be quantized on while dequantization is applied on
+      // activations. Note 2: 'x_num_col_dims' is the axis of activations to be
+      // quantized on. `x_num_col_dims` is -1 for operator in ['matmul',
+      // 'matmul_v2', 'mul'] and is 1 for other operators.
+      int64_t num = in->numel();
+      int n_scales = in->dims()[x_num_col_dims];
       const T* scale_one = scales[0]->data<T>();
       const T* scale_two = scales[1]->data<T>();
-      int block = 1024;
-      int grid = iter_size * channel;
-      DequantizeTwoScale<T><<<grid, block, 0, dev_ctx.stream()>>>(in_data,
+
+      int64_t block_size = std::min(
+          num, static_cast<int64_t>(dev_ctx.GetMaxThreadsPerBlock() / 4));
+      int64_t max_threads =
+          dev_ctx.GetMaxPhysicalThreadCount();  // SM * block_per_SM
+      const int64_t max_blocks = std::max(((max_threads - 1) / block_size + 1),
+                                          static_cast<int64_t>(1));
+      const int64_t grid_size =
+          std::min(max_blocks, (num + block_size - 1) / block_size);
+      int quant_stride = 1;
+      for (int i = x_num_col_dims + 1; i < in_dims.size(); i++) {
+        quant_stride *= in_dims[i];
+      }
+      DequantizeTwoScale<T>
+          <<<grid_size, block_size, 0, dev_ctx.stream()>>>(in_data,
                                                            scale_one,
                                                            scale_two,
                                                            max_range,
                                                            num,
-                                                                  iter_size,
-                                                                  channel,
+                                                           n_scales,
+                                                           quant_stride,
                                                            out_data);
     }
   }