[Bug Fix] add broadcast mechanism before calculating PReLu when the i…

…nput layout is NCHW [BUG FIX] avoid braodcast in certain cases [BUG FIX] reformat based on comments
openvinotoolkit · Aug 6, 2024 · aed3348 · aed3348
1 parent 3a7d230
commit aed3348
Showing 1 changed file with 45 additions and 13 deletions.
diff --git a/source/thead_rvv/fp32/prelu.c b/source/thead_rvv/fp32/prelu.c
@@ -17,6 +17,7 @@
  */
 
 #include "rvv/rvv.h"
+#include "reference/ref.h"
 
 int shl_rvv_prelu_fp32(struct csinn_tensor *input, struct csinn_tensor *alpha,
                        struct csinn_tensor *output, struct csinn_prelu_params *params)
@@ -53,22 +54,53 @@ int shl_rvv_prelu_fp32(struct csinn_tensor *input, struct csinn_tensor *alpha,
             output->layout = CSINN_LAYOUT_NC1HWC0;
         }
     } else if (input->layout == CSINN_LAYOUT_NCHW) {
-        for (int n = 0; n < input->dim[0]; ++n) {
-            for (int c = 0; c < input->dim[1]; ++c) {
-                float a = alpha_data[c];
-                int inner_size = input->dim[2] * input->dim[3];
-                while (inner_size > 0) {
-                    int vl = vsetvl_e32m2(inner_size);
-                    vfloat32m2_t _input = vle32_v_f32m2(input_data, vl);
-                    vbool16_t _mask = vmflt_vf_f32m2_b16(_input, 0.0f, vl);
-                    vfloat32m2_t _res = vfmul_vf_f32m2_m(_mask, _input, _input, a, vl);
-                    vse32_v_f32m2(output_data, _res, vl);
-                    input_data += vl;
-                    output_data += vl;
-                    inner_size -= vl;
+        if (alpha->dim[1] == csinn_tensor_size(alpha)) {
+            // simplify the calculation by avoiding broadcast
+            for (int n = 0; n < input->dim[0]; ++n) {
+                for (int c = 0; c < input->dim[1]; ++c) {
+                    float a = alpha_data[c];
+                    int inner_size = input->dim[2] * input->dim[3];
+                    while (inner_size > 0) {
+                        int vl = vsetvl_e32m2(inner_size);
+                        vfloat32m2_t _input = vle32_v_f32m2(input_data, vl);
+                        vbool16_t _mask = vmflt_vf_f32m2_b16(_input, 0.0f, vl);
+                        vfloat32m2_t _res = vfmul_vf_f32m2_m(_mask, _input, _input, a, vl);
+                        vse32_v_f32m2(output_data, _res, vl);
+                        input_data += vl;
+                        output_data += vl;
+                        inner_size -= vl;
+                    }
                 }
             }
+        } else {
+            // broadcast alpha
+            int input_size = csinn_tensor_size(input);
+            float *alpha_data_b = shl_mem_alloc(input_size * sizeof(float));
+            struct csinn_tensor *alpha_ = csinn_alloc_tensor(NULL);
+            csinn_tensor_copy(alpha_, input);
+            alpha_->data = alpha_data_b;
+            shl_ref_broadcast_to_shape_f32(alpha, alpha_, alpha_->dim, alpha_->dim_count);
+            alpha_data = (float *)alpha_->data;
+
+            // calculation
+            while (input_size > 0) {
+                int vl = vsetvl_e32m2(input_size);
+                vfloat32m2_t _input = vle32_v_f32m2(input_data, vl);
+                vfloat32m2_t _a = vle32_v_f32m2(alpha_data, vl);
+                vbool16_t _mask = vmflt_vf_f32m2_b16(_input, 0.0f, vl);
+                vfloat32m2_t _res = vfmul_vv_f32m2_m(_mask, _input, _input, _a, vl);
+                vse32_v_f32m2(output_data, _res, vl);
+                input_data += vl;
+                alpha_data += vl;
+                output_data += vl;
+                input_size -= vl;
+            }
+
+            // free memory and tensor
+            shl_mem_free(alpha_data_b);
+            csinn_free_tensor(alpha_);
         }
+
         if (output->layout == CSINN_LAYOUT_NC1HWC0) {
             const int packn = csrr_vlenb() / sizeof(float);
             output->dim[1] *= packn;