[MSLITE][Develop] support conv_depthwise arm32 int8 weight perchannel

This commit is contained in:
yangruoqi713 2020-09-21 14:06:04 +08:00
parent d60033c8db
commit 53c6862a6f
3 changed files with 117 additions and 4 deletions

View File

@ -0,0 +1,113 @@
#ifdef __arm__
#ifndef __aarch64__
.text
.align 5
.global ConvDwInt8PostAlign4PerChannel
#ifndef __APPLE__
.type ConvDwInt8PostAlign4PerChannel, %function
#endif
// void ConvDwInt8PostAlign4PerChannel(int8_t *dst, int32_t *buffer, int channel4, int32_t output_zp, int32_t *out_multiplier,
// int32_t *left_shift, int32_t *right_shift, int32_t acc_min, int32_t acc_max);
// r0: dst, r1: buffer, r2: num_pixels, r3: output_zp, r4: out_multiplier,
// r5: left_shift, r6: right_shift, r7: acc_min, r8: acc_max
ConvDwInt8PostAlign4PerChannel:
// at return, clang generates "push {lr}, pop {pc}"" while gcc will generate "bx lr"
// according to https://stackoverflow.com/questions/53625807
// even if we jump to link register instead of saving it, we still have to save it in subroutine calls anyway
// clang's rule seems more simple, though there are no subroutine calls here
// r4-r8 and q4-q7 must be saved according to https://static.docs.arm.com/ihi0042/i/aapcs32.pdf
push {r4-r8, r10}
vpush {q4-q7}
add sp, sp, #88
vdup.32 q15, r3 // output_zp
ldr r4, [sp] // out_multiplier
ldr r5, [sp, #4] // left_shift
ldr r6, [sp, #8] // right_shift
ldr r7, [sp, #12] // acc_min
vdup.32 q11, r7
ldr r8, [sp, #16] // acc_max
vdup.32 q10, r8
mov r10, r0
LoopDepth8:
cmp r2, #8
blt End
vld1.32 {q0}, [r1]!
vld1.32 {q13}, [r5]!
vshl.s32 q0, q0, q13
vld1.32 {q14}, [r4]!
vqrdmulh.s32 q0, q0, q14
vld1.32 {q12}, [r6]!
vand q4, q0, q12
vshr.s32 q4, q4, #31
vqadd.s32 q0, q0, q4
vrshl.s32 q0, q0, q12
vadd.i32 q0, q0, q15
vmax.s32 q0, q0, q11
vmin.s32 q0, q0, q10
vqmovn.s32 d4, q0
vld1.32 {q1}, [r1]!
vld1.32 {q13}, [r5]!
vshl.s32 q1, q1, q13
vld1.32 {q14}, [r4]!
vqrdmulh.s32 q1, q1, q14
vld1.32 {q12}, [r6]!
vand q4, q1, q12
vshr.s32 q4, q4, #31
vqadd.s32 q1, q1, q4
vrshl.s32 q1, q1, q12
vadd.i32 q1, q1, q15
vmax.s32 q1, q1, q11
vmin.s32 q1, q1, q10
vqmovn.s32 d5, q1
vqmovn.s16 d4, q2
vst1.8 {d4}, [r10]!
sub r2, r2, #8
b LoopDepth8
LoopDepth4:
cmp r2, #4
blt End
vld1.32 {q0}, [r1]!
vld1.32 {q13}, [r5]!
vshl.s32 q0, q0, q13
vld1.32 {q14}, [r4]!
vqrdmulh.s32 q0, q0, q14
vld1.32 {q12}, [r6]!
vand q4, q0, q12
vshr.s32 q4, q4, #31
vqadd.s32 q0, q0, q4
vrshl.s32 q0, q0, q12
vadd.i32 q0, q0, q15
vmax.s32 q0, q0, q11
vmin.s32 q0, q0, q10
vqmovn.s32 d0, q0
vqmovn.s16 d0, q0
vst1.8 {d0[0]}, [r10]!
vst1.8 {d0[1]}, [r10]!
vst1.8 {d0[2]}, [r10]!
vst1.8 {d0[3]}, [r10]!
sub r2, r2, #4
b LoopDepth4
End:
sub sp, sp, #88
vpop {q4-q7}
pop {r4-r8, r10}
bx lr
#endif
#endif

View File

@ -35,6 +35,9 @@ void PostFuncInt8C4(const int32_t *in, const int32_t *bias, int8_t *out, size_t
#ifdef ENABLE_ARM #ifdef ENABLE_ARM
void ConvDwInt8Row(int32_t *output_ptr, const int8_t *input_ptr, const int16_t *weight_ptr, int num_pixels, void ConvDwInt8Row(int32_t *output_ptr, const int8_t *input_ptr, const int16_t *weight_ptr, int num_pixels,
int output_channel, int input_step, int8_t input_zp); int output_channel, int input_step, int8_t input_zp);
void ConvDwInt8PostAlign4PerChannel(int8_t *dst, int32_t *buffer, int channel4, int32_t output_zp,
int32_t *out_multiplier, int32_t *left_shift, int32_t *right_shift, int32_t acc_min,
int32_t acc_max);
void ConvDwInt8PostAlign4(int8_t *dst, int32_t *buffer, int num_pixels, int32_t output_zp, int32_t out_multiplier, void ConvDwInt8PostAlign4(int8_t *dst, int32_t *buffer, int num_pixels, int32_t output_zp, int32_t out_multiplier,
int32_t left_shift, int32_t right_shift, int32_t acc_min, int32_t acc_max); int32_t left_shift, int32_t right_shift, int32_t acc_min, int32_t acc_max);
void IndirectGemmInt16to32_8x4(int32_t *dst, const int16_t *src, const int16_t *weight, size_t ksize, size_t ic8, void IndirectGemmInt16to32_8x4(int32_t *dst, const int16_t *src, const int16_t *weight, size_t ksize, size_t ic8,
@ -64,9 +67,6 @@ void IndirectGemmInt8_4x4(int8_t *output, const int8_t *input, const int8_t *wei
void DeconvDwInt8Center(int32_t *dst, const int16_t *src, const int16_t *weight, size_t height, size_t width, void DeconvDwInt8Center(int32_t *dst, const int16_t *src, const int16_t *weight, size_t height, size_t width,
size_t kernel_h, size_t kernel_w, size_t out_h_step, size_t block_channel, size_t in_sh_step, size_t kernel_h, size_t kernel_w, size_t out_h_step, size_t block_channel, size_t in_sh_step,
size_t in_sw_step, size_t in_kh_step, size_t in_kw_step); size_t in_sw_step, size_t in_kh_step, size_t in_kw_step);
void ConvDwInt8PostAlign4PerChannel(int8_t *dst, int32_t *buffer, int channel4, int32_t output_zp,
int32_t *out_multiplier, int32_t *left_shift, int32_t *right_shift, int32_t acc_min,
int32_t acc_max);
#endif #endif
#ifdef __cplusplus #ifdef __cplusplus

View File

@ -39,7 +39,7 @@ void ConvDwInt8Post(int8_t *dst, int32_t *buffer, int output_w, int channel, int
// support perchannel // support perchannel
for (int w = 0; w < output_w; w++) { for (int w = 0; w < output_w; w++) {
int channel4 = 0; int channel4 = 0;
#ifdef ENABLE_ARM64 #ifdef ENABLE_ARM
channel4 = channel / 4 * 4; channel4 = channel / 4 * 4;
ConvDwInt8PostAlign4PerChannel(dst, buffer, channel4, output_zp, out_multiplier, left_shift, right_shift, acc_min, ConvDwInt8PostAlign4PerChannel(dst, buffer, channel4, output_zp, out_multiplier, left_shift, right_shift, acc_min,
acc_max); acc_max);