Merge develop:Rollback

lib/drivers/include/kpu.h

@@ -680,7 +680,6 @@ typedef struct
             const uint8_t *volatile current_body;
             dmac_channel_number_t dma_ch;
             kpu_done_callback_t done_callback;
-            volatile uint8_t load_first;
             void *userdata;
         };
 

lib/drivers/kpu.c

@@ -1006,8 +1006,12 @@ static void kpu_quantize(const kpu_model_quantize_layer_argument_t *arg, kpu_mod
     size_t count = arg->count;
     const float *src = (const float *)(ctx->main_buffer + arg->main_mem_in_address);
 
-    kpu_model_quant_param_t q = arg->quant_param;
-
+    kpu_model_quant_param_t q;
+#if FIX_CACHE
+    memcpy(&q, &arg->quant_param, sizeof(kpu_model_quant_param_t));
+#else
+    q = arg->quant_param;
+#endif
     float scale = 1.f / q.scale;
 
     uint8_t *dest = (uint8_t *)(ctx->main_buffer + arg->mem_out_address);
@@ -1028,8 +1032,12 @@ static void kpu_kmodel_dequantize(const kpu_model_dequantize_layer_argument_t *a
     const uint8_t *src = (const uint8_t *)(ctx->main_buffer + arg->main_mem_in_address);
     float *dest = (float *)(ctx->main_buffer + arg->main_mem_out_address);
     size_t oc, count = arg->count;
-    kpu_model_quant_param_t q = arg->quant_param;
-
+    kpu_model_quant_param_t q;
+#if FIX_CACHE
+    memcpy(&q, &arg->quant_param, sizeof(kpu_model_quant_param_t));
+#else
+    q = arg->quant_param;
+#endif
     for(oc = 0; oc < count; oc++)
         dest[oc] = *src++ * q.scale + q.bias;
 }
@@ -1134,7 +1142,10 @@ static void kpu_kmodel_fully_connected(const kpu_model_fully_connected_layer_arg
     const float *src = (const float *)(ctx->main_buffer + arg->main_mem_in_address);
     float *dest = (float *)(ctx->main_buffer + arg->main_mem_out_address);
     uint32_t in_channels = arg->in_channels, out_channels = arg->out_channels, ic, oc;
-    const float *weights = arg->weights, *bias = arg->weights + in_channels * out_channels;
+    float *weights = (float *)malloc(in_channels * out_channels * sizeof(float));
+    float *bias = (float *)malloc(out_channels * sizeof(float));
+    memcpy(weights, arg->weights, out_channels * in_channels * sizeof(float));
+    memcpy(bias, arg->weights + in_channels * out_channels, out_channels * sizeof(float));
 
     if(in_channels % 8 == 0)
     {
@@ -1181,7 +1192,8 @@ static void kpu_kmodel_fully_connected(const kpu_model_fully_connected_layer_arg
             dest[oc] = sum + bias[oc];
         }
     }
-
+    free(weights);
+    free(bias);
     kpu_float_activation(dest, out_channels, arg->act);
 }
 
@@ -1265,10 +1277,9 @@ static void kpu_logistic(const kpu_model_logistic_layer_argument_t *arg, kpu_mod
 static void kpu_conv(const kpu_model_conv_layer_argument_t *arg, kpu_model_context_t *ctx)
 {
     volatile kpu_layer_argument_t layer = *(const volatile kpu_layer_argument_t *)(ctx->model_buffer + arg->layer_offset);
-
-    layer.kernel_load_cfg.data.para_start_addr = (uintptr_t)(ctx->model_buffer + arg->weights_offset) - 0x40000000;
-    layer.kernel_pool_type_cfg.data.bwsx_base_addr = (uintptr_t)(ctx->model_buffer + arg->bn_offset) - 0x40000000;
-    layer.kernel_calc_type_cfg.data.active_addr = (uintptr_t)(ctx->model_buffer + arg->act_offset) - 0x40000000;
+    layer.kernel_load_cfg.data.para_start_addr = (uintptr_t)(ctx->model_buffer + arg->weights_offset);
+    layer.kernel_pool_type_cfg.data.bwsx_base_addr = (uintptr_t)(ctx->model_buffer + arg->bn_offset);
+    layer.kernel_calc_type_cfg.data.active_addr = (uintptr_t)(ctx->model_buffer + arg->act_offset);
 
     if(arg->flags & KLF_MAIN_MEM_OUT)
     {
@@ -1360,14 +1371,14 @@ static void kpu_upload(const kpu_model_upload_layer_argument_t *arg, kpu_model_c
 
 int kpu_load_kmodel(kpu_model_context_t *ctx, const uint8_t *buffer)
 {
+#if FIX_CACHE
+    configASSERT(!is_memory_cache((uintptr_t)buffer));
+#endif
     uintptr_t base_addr = (uintptr_t)buffer;
     const kpu_kmodel_header_t *header = (const kpu_kmodel_header_t *)buffer;
 
-    configASSERT(is_memory_cache((uintptr_t)buffer))
-
     if(header->version == 3 && header->arch == 0)
     {
-        ctx->load_first = 1;
         ctx->is_nncase = 0;
         ctx->model_buffer = buffer;
         ctx->output_count = header->output_count;
@@ -1482,7 +1493,6 @@ static int kpu_kmodel_done(kpu_model_context_t *ctx)
         .calc_done_int = 1,
         .layer_cfg_almost_empty_int = 1,
         .layer_cfg_almost_full_int = 1};
-    ctx->load_first = 0;
 #if KPU_DEBUG
     uint32_t cnt_layer_id = ctx->current_layer - 1;
     uint64_t time = sysctl_get_time_us();
@@ -1582,13 +1592,6 @@ static int ai_step(void *userdata)
             kpu_logistic((const kpu_model_logistic_layer_argument_t *)layer_body, ctx);
             break;
         case KL_K210_CONV:
-            if(ctx->load_first)
-            {
-				for(int i=0; i<cnt_layer_header->body_size; i++)
-				{
-					*((uint8_t *)((uintptr_t)layer_body-0x40000000)+i) = *((uint8_t *)layer_body+i);
-				}
-            }
             kpu_conv((const kpu_model_conv_layer_argument_t *)layer_body, ctx);
             return 0;
         case KL_K210_ADD_PADDING:

src/face_detect/main.c

@@ -24,6 +24,7 @@
 #define INCBIN_PREFIX
 #include "incbin.h"
 #include "utils.h"
+#include "iomem.h"
 
 #define PLL0_OUTPUT_FREQ 800000000UL
 #define PLL1_OUTPUT_FREQ 400000000UL
@@ -194,7 +195,7 @@ int main(void)
     w25qxx_init(3, 0);
     w25qxx_enable_quad_mode();
 #if LOAD_KMODEL_FROM_FLASH
-    model_data = (uint8_t *)malloc(KMODEL_SIZE);
+    model_data = (uint8_t *)iomem_malloc(KMODEL_SIZE);
     w25qxx_read_data(0xA00000, model_data, KMODEL_SIZE, W25QXX_QUAD_FAST);
 #endif
     /* LCD init */