!15611 [MS][LITE][Develop] add arm32 fp32 optimize

From: @lzkcode
Reviewed-by: @zhang_xue_tong,@zhanghaibo5
Signed-off-by: @zhang_xue_tong

mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/assembly/arm82_aarch32_fp16/Float16Tofloat32.S

@@ -0,0 +1,55 @@
+#ifdef ENABLE_ARM32
+#include "nnacl/assembly_global.h"
+    .text
+    .align 5
+    .global Float16ToFloat32
+#ifndef __APPLE__
+    .type Float16ToFloat32, %function
+#endif
+
+// void Float16ToFloat32(const float16_t *input, float *output, int number);
+// r0: input, r1: output, r2: number
+Float16ToFloat32:
+    cmp r2, #0
+    beq LoopEnd
+    cmp r2, #16
+    bge Loop16
+    cmp r2, #8
+    bge Loop8
+    b Loop
+    Loop16:
+        vld1.16 {q0, q1}, [r0]!
+        vcvt.f32.f16 q3, d0
+        vcvt.f32.f16 q4, d1
+        vcvt.f32.f16 q5, d2
+        vst1.32 {q3, q4}, [r1]!
+        vcvt.f32.f16 q6, d3
+        subs r2, r2, #16
+        vst1.32 {q5, q6}, [r1]!
+        beq LoopEnd
+        cmp r2, #16
+        bge Loop16
+        cmp r2, #8
+        bge Loop8
+        b Loop
+    Loop8:
+        vld1.16 {q0}, [r0]!
+        vcvt.f32.f16 q1, d0
+        vcvt.f32.f16 q2, d1
+        vst1.32 {q1, q2}, [r1]!
+        subs r2, r2, #8
+        beq LoopEnd
+        cmp r2, #8
+        bge Loop8
+        b Loop
+    Loop:
+        vldr.16 s0, [r0]
+        vcvtb.f32.f16 s0, s0
+        vstr.32 s0, [r1]
+        add r0, r0, #2
+        add r1, r1, #4
+        subs r2, r2, #1
+        bgt Loop
+    LoopEnd:
+        mov pc, lr
+#endif

mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/assembly/arm82_aarch32_fp16/Float32ToFloat16.S

@@ -0,0 +1,55 @@
+#ifdef ENABLE_ARM32
+#include "nnacl/assembly_global.h"
+    .text
+    .align 5
+    .global Float32ToFloat16
+#ifndef __APPLE__
+    .type Float32ToFloat16, %function
+#endif
+
+// void Float32ToFloat16(const float *input, float16_t *output, int number);
+// r0: input, r1: output, r2: number
+Float32ToFloat16:
+    cmp r2, #0
+    beq LoopEnd
+    cmp r2, #16
+    bge Loop16
+    cmp r2, #8
+    bge Loop8
+    b Loop
+    Loop16:
+        vld1.32 {q0, q1}, [r0]!
+        vcvt.f16.f32 d0, q0
+        vcvt.f16.f32 d1, q1
+        vld1.32 {q2, q3}, [r0]!
+        vcvt.f16.f32 d2, q2
+        vcvt.f16.f32 d3, q3
+        vst1.16 {q0, q1}, [r1]!
+        subs r2, r2, #16
+        beq LoopEnd
+        cmp r2, #16
+        bge Loop16
+        cmp r2, #8
+        bge Loop8
+        b Loop
+    Loop8:
+        vld1.32 {q0, q1}, [r0]!
+        vcvt.f16.f32 d0, q0
+        vcvt.f16.f32 d1, q1
+        vst1.16 {q0}, [r1]!
+        subs r2, r2, #8
+        beq LoopEnd
+        cmp r2, #8
+        bge Loop8
+        b Loop
+    Loop:
+        vldr s0, [r0]
+        vcvtb.f16.f32 s0, s0
+        vstr.16 s0, [r1]
+        add r0, r0, #4
+        add r1, r1, #2
+        subs r2, r2, #1
+        bgt Loop
+    LoopEnd:
+        mov pc, lr
+#endif

mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/fp16/cast_fp16.h

@@ -46,6 +46,7 @@ inline void Float16ToInt64(const float16_t *input, int64_t *output, int number)
   }
 }
 
+#ifdef ENABLE_ARM64
 inline void Float32ToFloat16(const float *input, float16_t *output, int number) {
   for (int i = 0; i < number; ++i) {
     output[i] = (float16_t)input[i];
@@ -57,6 +58,11 @@ inline void Float16ToFloat32(const float16_t *input, float *output, int number)
     output[i] = (float)input[i];
   }
 }
+#else
+void Float32ToFloat16(const float *input, float16_t *output, int number);
+
+void Float16ToFloat32(const float16_t *input, float *output, int number);
+#endif
 
 #ifdef __cplusplus
 }

mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/fp16/deconv_winograd_fp16.c

@@ -44,49 +44,49 @@ void DeConvWgInputPackFp16(float16_t *src_ptr, float16_t *dst_ptr, int channel,
 #ifdef ENABLE_ARM82_A32
 void DeconvWgMergeFp16A32Fun(const float16_t *src_ptr, float16_t *dst_ptr, size_t src_step, size_t dst_step) {
   asm volatile(
-    "mov r7, %[src_ptr]\n"
-    "mov r8, %[dst_ptr]\n"
-    "mov r10, r8\n"
-
-    "vld1.16 {d0}, [r7], %[src_step]\n"
-    "vld1.16 {d2}, [r8], %[dst_step]\n"
-    "vld1.16 {d4}, [r7], %[src_step]\n"
-    "vld1.16 {d6}, [r8], %[dst_step]\n"
+    "mov r0, %[src_ptr]\n"
+    "mov r1, %[dst_ptr]\n"
+    "mov r2, r1\n"
+
+    "vld1.16 {d0}, [r0], %[src_step]\n"
+    "vld1.16 {d2}, [r1], %[dst_step]\n"
+    "vld1.16 {d4}, [r0], %[src_step]\n"
+    "vld1.16 {d6}, [r1], %[dst_step]\n"
     "vadd.f16 d0, d0, d2\n"
-    "vld1.16 {d8}, [r7], %[src_step]\n"
+    "vld1.16 {d8}, [r0], %[src_step]\n"
     "vadd.f16 d4, d4, d6\n"
-    "vst1.16 {d0}, [r10], %[dst_step]\n"
-    "vst1.16 {d4}, [r10], %[dst_step]\n"
+    "vst1.16 {d0}, [r2], %[dst_step]\n"
+    "vst1.16 {d4}, [r2], %[dst_step]\n"
 
-    "vld1.16 {d10}, [r8], %[dst_step]\n"
-    "vld1.16 {d12}, [r7], %[src_step]\n"
+    "vld1.16 {d10}, [r1], %[dst_step]\n"
+    "vld1.16 {d12}, [r0], %[src_step]\n"
     "vadd.f16 d8, d8, d10\n"
-    "vld1.16 {d14}, [r8], %[dst_step]\n"
+    "vld1.16 {d14}, [r1], %[dst_step]\n"
     "vadd.f16 d12, d12, d14\n"
-    "vld1.16 {d0}, [r7], %[src_step]\n"
-    "vst1.16 {d8}, [r10], %[dst_step]\n"
-    "vst1.16 {d12}, [r10], %[dst_step]\n"
+    "vld1.16 {d0}, [r0], %[src_step]\n"
+    "vst1.16 {d8}, [r2], %[dst_step]\n"
+    "vst1.16 {d12}, [r2], %[dst_step]\n"
 
-    "vld1.16 {d2}, [r8], %[dst_step]\n"
-    "vld1.16 {d4}, [r7], %[src_step]\n"
-    "vld1.16 {d6}, [r8], %[dst_step]\n"
+    "vld1.16 {d2}, [r1], %[dst_step]\n"
+    "vld1.16 {d4}, [r0], %[src_step]\n"
+    "vld1.16 {d6}, [r1], %[dst_step]\n"
     "vadd.f16 d0, d0, d2\n"
     "vadd.f16 d4, d4, d6\n"
-    "vst1.16 {d0}, [r10], %[dst_step]\n"
-    "vst1.16 {d4}, [r10], %[dst_step]\n"
+    "vst1.16 {d0}, [r2], %[dst_step]\n"
+    "vst1.16 {d4}, [r2], %[dst_step]\n"
 
-    "vld1.16 {d8}, [r7], %[src_step]\n"
-    "vld1.16 {d10}, [r8], %[dst_step]\n"
-    "vld1.16 {d12}, [r7], %[src_step]\n"
-    "vld1.16 {d14}, [r8], %[dst_step]\n"
+    "vld1.16 {d8}, [r0], %[src_step]\n"
+    "vld1.16 {d10}, [r1], %[dst_step]\n"
+    "vld1.16 {d12}, [r0], %[src_step]\n"
+    "vld1.16 {d14}, [r1], %[dst_step]\n"
     "vadd.f16 d8, d8, d10\n"
     "vadd.f16 d12, d12, d14\n"
-    "vst1.16 {d8}, [r10], %[dst_step]\n"
-    "vst1.16 {d12}, [r10], %[dst_step]\n"
+    "vst1.16 {d8}, [r2], %[dst_step]\n"
+    "vst1.16 {d12}, [r2], %[dst_step]\n"
 
     :
     : [ src_ptr ] "r"(src_ptr), [ dst_ptr ] "r"(dst_ptr), [ src_step ] "r"(src_step), [ dst_step ] "r"(dst_step)
-    : "r7", "r8", "r10", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7");
+    : "r0", "r1", "r2", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7");
 }
 #endif