Merge branch 'dl_image_simd' into 'master'

Dl image simd

See merge request ai/esp-dl!205

Author: BlueSkyB

.gitignore

@@ -5,6 +5,7 @@ sdkconfig
 bin/
 .idea/
 __pycache__/
+.clangd
 
 # unit test
 test_apps/*/*/build_*

.gitlab/ci/build.yml

@@ -189,7 +189,7 @@ build_example_color_detect:
     - .rules:build:example_color_detect
   parallel:
     matrix:
-      - IMAGE: [espressif/idf:release-v5.3, espressif/idf:latest]
+      - IMAGE: [espressif/idf:release-v5.3, espressif/idf:release-v5.5]
         TARGET: [esp32p4, esp32s3]
   variables:
     EXAMPLE_DIR: examples/color_detect

esp-dl/CMakeLists.txt

@@ -26,6 +26,7 @@ set(include_dirs    ./dl
                     ./fbs_loader/include
                     ./vision/detect
                     ./vision/image
+                    ./vision/image/isa
                     ./vision/recognition
                     ./vision/classification
                     ./audio/common
@@ -49,6 +50,7 @@ elseif(CONFIG_IDF_TARGET_ESP32C3)
 elseif(CONFIG_IDF_TARGET_ESP32P4)
     list(APPEND src_dirs        dl/tool/isa/esp32p4)
     list(APPEND src_dirs        dl/base/isa/esp32p4)
+    list(APPEND src_dirs        ./vision/image/isa/esp32p4)
 endif()
 
 set(requires        dl_fft

esp-dl/dl/tool/include/dl_tool.hpp

@@ -43,7 +43,9 @@ typedef enum {
 } round_mode_t;
 
 void dl_esp32p4_cfg_round(round_mode_t round_mode);
-int dl_esp32p4_round_half_even(float value);
+typedef enum { FORCE_ALIGN, HW_MISALIGN } misalign_mode_t;
+void dl_esp32p4_cfg_misalign(misalign_mode_t store, misalign_mode_t load);
+uint32_t dl_esp32p4_get_cfg();
 #endif
 }
 

esp-dl/dl/tool/isa/esp32p4/dl_esp32p4_cfg.S

@@ -0,0 +1,63 @@
+    .text
+    .global dl_esp32p4_cfg_round
+    .type   dl_esp32p4_cfg_round, @function
+    .balign 4
+    .option norvc
+dl_esp32p4_cfg_round:
+
+    # a0: int8_t round_mode
+    # a1:
+    # a2:
+
+    # a3:
+    # a4:
+    # a5:
+    # t3:
+    # t4:
+    # t5:
+    # t6:
+
+    # s0:
+    # s1:
+    # s8:
+    # s9:
+    # s10:
+    # s11:
+
+    esp.movx.r.cfg a1
+    li a2, 0xffffff0f
+    and a1, a1, a2
+    slli a0, a0, 4
+    or a1, a1, a0
+    esp.movx.w.cfg a1
+    ret
+
+
+    .align 2
+    .text
+    .global dl_esp32p4_cfg_misalign
+    .type   dl_esp32p4_cfg_misalign, @function
+dl_esp32p4_cfg_misalign:
+
+    # a0: bool store
+    # a1: bool load
+
+    slli a1, a1, 1
+    add a0, a0, a1
+    esp.movx.r.cfg a1
+    li a2, 0xfffffffc
+    and a1, a1, a2
+    or a1, a1, a0
+    esp.movx.w.cfg a1
+    ret
+
+    .align 2
+    .text
+    .global dl_esp32p4_get_cfg
+    .type   dl_esp32p4_get_cfg, @function
+dl_esp32p4_get_cfg:
+
+    # a0: uint32_t
+
+    esp.movx.r.cfg a0
+    ret

esp-dl/dl/tool/isa/esp32p4/dl_esp32p4_round.S

@@ -23,7 +23,9 @@ namespace tool {
 int round_half_even(float value)
 {
 #if CONFIG_ESP32P4_BOOST
-    return dl_esp32p4_round_half_even(value);
+    int ret;
+    __asm__ volatile("fcvt.w.s %0, %1, rne" : "=r"(ret) : "f"(value));
+    return ret;
 #else
     float rounded;
     if (value < 0) {

esp-dl/dl/tool/src/dl_tool.cpp

@@ -74,7 +74,7 @@ append_line(".data")
 append_line("#if !defined (__APPLE__) && !defined (__linux__)")
 append_line(".section .rodata.embedded")
 append_line("#endif")
-append_line(".align 16")
+append_line(".balign 16")
 make_and_append_identifier("${varname}")
 make_and_append_identifier("_binary_${varname}_start" "for objcopy compatibility")
 append("${data}")

esp-dl/fbs_loader/cmake/data_file_embed_asm_aligned.cmake

@@ -257,6 +257,7 @@ FbsModel *create_fbs_model(const char *fbs_buf,
             fseek(f, 4, SEEK_CUR);
         }
         fread(model_buf, size, 1, f);
+        fclose(f);
     }
 
     assert(mode == 0 || mode == 1);