Compiler/TMS320F28377D: Accessing the I16TOF32 instruction in C

IainRist

Part Number: TMS320F28377D

Tool/software: TI C/C++ Compiler

I want to access the I16TOF32 instruction in C. A int-to-float cast results in a load to ACC to perform a sign extension from 16 to 32 bits, a move to a floating-point register, then a conversion to float using I32TOF32. Below is a snippet of assembly output by the C2000 16.9.2.LTS compiler at full optimisation level (it has unrolled the loop by a factor of 2).

MOVB XAR6,#24

MOVL XAR5,#_b
MOVL XAR4,#_a
SETC SXM
RPTB $C$L2,AR6
; repeat block starts
MOV ACC,*XAR4++
MOV32 R0H,ACC
MOV ACC,*XAR4++
MOV32 R3H,ACC
NOP
NOP
NOP
I32TOF32 R1H,R0H
I32TOF32 R0H,R3H
MOV32 *XAR5++,R1H
MOV32 *XAR5++,R0H
; repeat block ends
$C$L2:

This compiler-generated code is too slow for my use case (11 cycles per 2 outputs, vs 2 cycles per 2 outputs for hand-coded pipelined assembly). Is there a way to force the compiler to emit an I16TOF32 without using inline assembly (as it turns off some/most optimisations) and without resorting to re-writing my algorithms in assembly?

over 8 years ago

0 Archaeologist over 8 years ago

TI__Guru* 84285 points

The compiler is capable of generating I16TOF32, as shown by the snippet below. I don't see why it isn't being generated for your loop, as it is clearly doing 16-bit loads. Could you please show me a compilable C test case that demonstrates the problem? Please include the complete command-line options. Also, what version of the compiler are you using? (This is not the same as the CCS version.)

% cat foo.c
float stof(short *x) { return *x; }
% cl2000 --float_support=fpu32 -O0 -k
% grep F32 foo.asm
I16TOF32 R0H,*+XAR4[0]

0 IainRist over 8 years ago in reply to Archaeologist

Intellectual 711 points

I am using C2000 compiler v16.9.2.LTS. I am primarily interested in vector operations, hence the tight loops.

% cat i16tof32_v_attempt1.c

void i16tof32(float *y, const short *x, short num) { while (num--) *y++ = (float)*x++; }

% "C:/ti/ccsv6/tools/compiler/ti-cgt-c2000_16.9.2.LTS/bin/cl2000" -v28 -ml -mt --cla_support=cla1 --float_support=fpu32 --tmu_support=tmu0 --vcu_support=vcu2 -O3 --opt_for_speed=5 --fp_mode=relaxed --include_path="C:/ti/ccsv6/tools/compiler/ti-cgt-c2000_16.9.2.LTS/include" --symdebug:none --diag_warning=225 --diag_wrap=off --display_error_number -k --src_interlist "../i16tof32_v_attempt1.c"

% cat i16_to_f32_v_attempt1.asm

;***************************************************************

;* TMS320C2000 C/C++ Codegen PC v16.9.2.LTS *

;* Date/Time created: Tue Apr 04 11:34:07 2017 *

;***************************************************************

.compiler_opts --abi=coffabi --cla_support=cla1 --diag_wrap=off --float_support=fpu32 --hll_source=on --mem_model:code=flat --mem_model:data=large --object_format=coff --silicon_version=28 --symdebug:none --tmu_support=tmu0

.asg XAR2, FP

; C:\ti\ccsv6\tools\compiler\ti-cgt-c2000_16.9.2.LTS\bin\opt2000.exe C:\\Users\\rist-1\\AppData\\Local\\Temp\\040042 C:\\Users\\rist-1\\AppData\\Local\\Temp\\040044

; C:\ti\ccsv6\tools\compiler\ti-cgt-c2000_16.9.2.LTS\bin\ac2000.exe -@C:\\Users\\rist-1\\AppData\\Local\\Temp\\0400412

.sect ".text"

.clink

.global _i16tof32_v_attempt1

;***************************************************************

;* FNAME: _i16tof32_v_attempt1 FR SIZE: 0 *

;* *

;* FUNCTION ENVIRONMENT *

;* *

;* FUNCTION PROPERTIES *

;* 0 Parameter, 0 Auto, 0 SOE *

;***************************************************************

_i16tof32_v_attempt1:

;*** 3 ----------------------- if ( !num ) goto g7;

CMPB AL,#0 ; [CPU_] |3|

B $C$L3,EQ ; [CPU_] |3|

; branchcc occurs ; [] |3|

;*** 3 ----------------------- d$1 = num&1;

;*** 3 ----------------------- if ( num < 2 ) goto g5;

AND AH,AL,#0x0001 ; [CPU_] |3|

CMPB AL,#2 ; [CPU_] |3|

B $C$L2,LT ; [CPU_] |3|

; branchcc occurs ; [] |3|

;*** ----------------------- L$1 = (num>>1)-1;

;*** ----------------------- #pragma MUST_ITERATE(1, 16383, 1)

;*** ----------------------- // LOOP BELOW UNROLLED BY FACTOR(2)

;*** ----------------------- #pragma LOOP_FLAGS(4102u)

;*** -----------------------g4:

;*** 3 ----------------------- *y++ = (float)*x++;

;*** 3 ----------------------- if ( (--L$1) != (-1) ) goto g4;

ASR AL,1 ; [CPU_]

ADDB AL,#-1 ; [CPU_]

MOVZ AR6,AL ; [CPU_]

RPTB $C$L2,AR6 ; [CPU_] |3|

; repeat block starts ; []

$C$L1:

MOVX TL,*XAR5++ ; [CPU_] |3|

MOV32 R0H,XT ; [CPU_] |3|

MOVX TL,*XAR5++ ; [CPU_] |3|

MOV32 R3H,XT ; [CPU_] |3|

NOP ; [CPU_]

I32TOF32 R1H,R0H ; [CPU_] |3|

I32TOF32 R0H,R3H ; [CPU_] |3|

MOV32 *XAR4++,R1H ; [CPU_] |3|

MOV32 *XAR4++,R0H ; [CPU_] |3|

; repeat block ends ; []

$C$L2:

;*** -----------------------g5:

;*** ----------------------- if ( d$1 <= 0 ) goto g7;

CMPB AH,#0 ; [CPU_]

B $C$L3,LEQ ; [CPU_]

; branchcc occurs ; []

; Peeled loop iterations for unrolled loop:

;*** 3 ----------------------- *y = (float)*x;

;*** -----------------------g7:

;*** ----------------------- return;

I16TOF32 R0H,*+XAR5[0] ; [CPU_] |3|

NOP ; [CPU_]

MOV32 *+XAR4[0],R0H ; [CPU_] |3|

$C$L3:

LRETR ; [CPU_]

; return occurs ; []

Notice how the compiler emits the correct instruction for the possible trailing input.

I also tested your suggestion of dereferencing the pointer within a function, with startling result

% cat i16tof32_v_attempt2

static inline float stof1(const short *x) { return (float)*x; }

void i16tof32_v_attempt2(float *y, const short *x, short num) { while (num--) *y++ = stof1(x++); }

% cat i16tof32_v_attempt2.asm

;***************************************************************

;* TMS320C2000 C/C++ Codegen PC v16.9.2.LTS *

;* Date/Time created: Tue Apr 04 11:34:17 2017 *

;***************************************************************

.asg XAR2, FP

; C:\ti\ccsv6\tools\compiler\ti-cgt-c2000_16.9.2.LTS\bin\opt2000.exe C:\\Users\\rist-1\\AppData\\Local\\Temp\\093762 C:\\Users\\rist-1\\AppData\\Local\\Temp\\093764

; C:\ti\ccsv6\tools\compiler\ti-cgt-c2000_16.9.2.LTS\bin\ac2000.exe -@C:\\Users\\rist-1\\AppData\\Local\\Temp\\0937612

.sect ".text"

.clink

.global _i16tof32_v_attempt2

;***************************************************************

;* FNAME: _i16tof32_v_attempt2 FR SIZE: 0 *

;* *

;* FUNCTION ENVIRONMENT *

;* *

;* FUNCTION PROPERTIES *

;* 0 Parameter, 0 Auto, 0 SOE *

;***************************************************************

_i16tof32_v_attempt2:

;*** 5 ----------------------- if ( !num ) goto g7;

CMPB AL,#0 ; [CPU_] |5|

B $C$L3,EQ ; [CPU_] |5|

; branchcc occurs ; [] |5|

;*** 5 ----------------------- d$1 = num&1;

;*** 5 ----------------------- if ( num < 2 ) goto g5;

AND AH,AL,#0x0001 ; [CPU_] |5|

CMPB AL,#2 ; [CPU_] |5|

B $C$L2,LT ; [CPU_] |5|

; branchcc occurs ; [] |5|

;*** ----------------------- L$1 = (num>>1)-1;

;*** ----------------------- #pragma MUST_ITERATE(1, 16383, 1)

;*** ----------------------- // LOOP BELOW UNROLLED BY FACTOR(2)

;*** ----------------------- #pragma LOOP_FLAGS(4102u)

;*** -----------------------g4:

;*** 1 ----------------------- *y++ = (float)*x++; // [0]

;*** 1 ----------------------- x = x; // [0]

;*** 1 ----------------------- ++x; // [0]

;*** 1 ----------------------- *y++ = (float)*x; // [0]

;*** 5 ----------------------- if ( (--L$1) != (-1) ) goto g4;

ASR AL,1 ; [CPU_]

ADDB AL,#-1 ; [CPU_]

MOVZ AR6,AL ; [CPU_]

RPTB $C$L2,AR6 ; [CPU_] |5|

; repeat block starts ; []

$C$L1:

MOVX TL,*XAR5++ ; [CPU_] |1|

MOV32 R0H,XT ; [CPU_] |1|

NOP ; [CPU_]

I32TOF32 R1H,R0H ; [CPU_] |1|

NOP ; [CPU_]

MOV32 *XAR4++,R1H ; [CPU_] |1|

MOV32 *XAR4++,R0H ; [CPU_] |1|

; repeat block ends ; []

$C$L2:

;*** -----------------------g5:

;*** ----------------------- if ( d$1 <= 0 ) goto g7;

CMPB AH,#0 ; [CPU_]

B $C$L3,LEQ ; [CPU_]

; branchcc occurs ; []

; Peeled loop iterations for unrolled loop:

;*** 1 ----------------------- *y = (float)*x; // [0]

;*** -----------------------g7:

;*** ----------------------- return;

I16TOF32 R0H,*+XAR5[0] ; [CPU_] |1|

NOP ; [CPU_]

MOV32 *+XAR4[0],R0H ; [CPU_] |1|

$C$L3:

LRETR ; [CPU_]

; return occurs ; []

;* Inlined function references:

;* [0] stof1

Notice how the assembly output of the second attempt is incorrectly pushing the sign-extended integer to the output. This is a bug that I will raise independently.

I was expecting the compiler to generate code more akin to the following (not tested but most of the way there):

.global _i16tof32_v_hand_optimised
.sect text
.clink
_i16tof32_v_hand_optimised:

TBIT @AL, #0

SB skip_1, NTC
I16TOF32 R0H, *XAR5++
LSR AL, #1
MOV32 *XAR4++, R0H

skip_1;
TBIT @AL, #0
SB init_loop, NTC
I16TOF32 R0H, *XAR5++
I16TOF32 R1H, *XAR5++
MOV32 *XAR4++, R0H
MOV32 *XAR4++, R1H

init_loop:
LSR AL, #1
SB end_loop, EQ

ADDB AL, #-1
RPTB end_loop, @AL

I16TOF32 R0H, *XAR5++
I16TOF32 R1H, *XAR5++
MOV32 *XAR4++, R0H
MOV32 *XAR4++, R1H
I16TOF32 R0H, *XAR5++
I16TOF32 R1H, *XAR5++
MOV32 *XAR4++, R0H
MOV32 *XAR4++, R1H
end_loop:

LRETR

0 Archaeologist over 8 years ago in reply to IainRist

TI__Guru* 84285 points

Thank you for the test case. This is a performance bug. I've submitted CODEGEN-2179 to track this issue.

With the problem fixed, the compiler will generate:

_i16tof32:
        CMPB      AL,#0                
        B         $C$L3,EQ             
        ; branchcc occurs
        AND       AH,AL,#0x0001        
        CMPB      AL,#2                
        B         $C$L2,LT             
        ; branchcc occurs
        ASR       AL,1                 
        ADDB      AL,#-1               
        MOVZ      AR6,AL               
        RPTB      $C$L2,AR6            
        ; repeat block starts
$C$L1:    
        I16TOF32  R1H,*XAR5++          
        I16TOF32  R0H,*XAR5++          
        MOV32     *XAR4++,R1H          
        MOV32     *XAR4++,R0H          
        NOP      
        ; repeat block ends
$C$L2:    
        CMPB      AH,#0                
        B         $C$L3,LEQ            
        ; branchcc occurs
; Peeled loop iterations for unrolled loop:
        I16TOF32  R0H,*+XAR5[0]        
        NOP      
        MOV32     *+XAR4[0],R0H        
$C$L3:    
        LRETR    
        ; return occurs

Code Composer Studio™︎

Code Composer Studio forum

Compiler/TMS320F28377D: Accessing the I16TOF32 instruction in C