pa_x86_plain_converters.c

一个开源SIP协议栈

        mov     edx, ditherPrevious
        neg     edx
        lea     edx, [eax+edx]          // highpass = current - previous
        mov     highpassedDither, edx
        mov     ditherPrevious, eax     // previous = current
        mov     eax, sourceByteStride   // restore eax
        fild    highpassedDither
        fmul    const_float_dither_scale_
    // end generate dither, dither signal in st(0)
    
        faddp   st(1), st(0)            // stack: dither + value*(int scaler), int scaler
        fistp   dword ptr [edi]         // pop st(0) into dest, stack:  int scaler
        jmp     Float32_To_Int32_DitherClip_stored
    
    Float32_To_Int32_DitherClip_clamp:
        mov     edx, dword ptr [esi]    // load floating point value into integer register
        shr     edx, 31                 // move sign bit into bit 0
        add     esi, eax                // increment source ptr
        //lea     esi, [esi+eax]
        add     edx, 0x7FFFFFFF         // convert to maximum range integers
        mov     dword ptr [edi], edx

    Float32_To_Int32_DitherClip_stored:

        //add     edi, ebx              // increment destination ptr
        lea     edi, [edi+ebx]

        cmp     esi, ecx                // has src ptr reached end?
        jne     Float32_To_Int32_DitherClip_loop

        ffree   st(0)
        fincstp

        fwait
        fnclex
        fldcw   savedFpuControlWord
    }

    ditherGenerator->previous = ditherPrevious;
    ditherGenerator->randSeed1 = ditherRandSeed1;
    ditherGenerator->randSeed2 = ditherRandSeed2;
}

/* -------------------------------------------------------------------------- */

static void Float32_To_Int24(
    void *destinationBuffer, signed int destinationStride,
    void *sourceBuffer, signed int sourceStride,
    unsigned int count, PaUtilTriangularDitherGenerator *ditherGenerator )
{
/*
    float *src = (float*)sourceBuffer;
    unsigned char *dest = (unsigned char*)destinationBuffer;
    signed long temp;

    (void) ditherGenerator; // unused parameter
    
    while( count-- )
    {
        // convert to 32 bit and drop the low 8 bits
        double scaled = *src * 0x7FFFFFFF;
        temp = (signed long) scaled;

        dest[0] = (unsigned char)(temp >> 8);
        dest[1] = (unsigned char)(temp >> 16);
        dest[2] = (unsigned char)(temp >> 24);

        src += sourceStride;
        dest += destinationStride * 3;
    }
*/

    short savedFpuControlWord;
    
    signed long tempInt32;

    (void) ditherGenerator; /* unused parameter */
                 
    __asm{
        // esi -> source ptr
        // eax -> source byte stride
        // edi -> destination ptr
        // ebx -> destination byte stride
        // ecx -> source end ptr
        // edx -> temp

        mov     esi, sourceBuffer

        mov     edx, 4                  // sizeof float32
        mov     eax, sourceStride
        imul    eax, edx

        mov     ecx, count
        imul    ecx, eax
        add     ecx, esi

        mov     edi, destinationBuffer

        mov     edx, 3                  // sizeof int24
        mov     ebx, destinationStride
        imul    ebx, edx

        fwait
        fstcw   savedFpuControlWord
        fldcw   fpuControlWord_

        fld     int24Scaler_             // stack:  (int)0x7FFFFF

    Float32_To_Int24_loop:

        // load unscaled value into st(0)
        fld     dword ptr [esi]         // stack:  value, (int)0x7FFFFF
        add     esi, eax                // increment source ptr
        //lea     esi, [esi+eax]
        fmul    st(0), st(1)            // st(0) *= st(1), stack:  value*0x7FFFFF, (int)0x7FFFFF
        fistp   tempInt32               // pop st(0) into tempInt32, stack:  (int)0x7FFFFF
        mov     edx, tempInt32

        mov     byte ptr [edi], DL
        shr     edx, 8
        //mov     byte ptr [edi+1], DL
        //mov     byte ptr [edi+2], DH
        mov     word ptr [edi+1], DX

        //add     edi, ebx                // increment destination ptr
        lea     edi, [edi+ebx]

        cmp     esi, ecx                // has src ptr reached end?
        jne     Float32_To_Int24_loop

        ffree   st(0)
        fincstp

        fwait
        fnclex
        fldcw   savedFpuControlWord
    }
}

/* -------------------------------------------------------------------------- */

static void Float32_To_Int24_Clip(
    void *destinationBuffer, signed int destinationStride,
    void *sourceBuffer, signed int sourceStride,
    unsigned int count, PaUtilTriangularDitherGenerator *ditherGenerator )
{
/*
    float *src = (float*)sourceBuffer;
    unsigned char *dest = (unsigned char*)destinationBuffer;
    signed long temp;

    (void) ditherGenerator; // unused parameter
    
    while( count-- )
    {
        // convert to 32 bit and drop the low 8 bits
        double scaled = *src * 0x7FFFFFFF;
        PA_CLIP_( scaled, -2147483648., 2147483647.  );
        temp = (signed long) scaled;

        dest[0] = (unsigned char)(temp >> 8);
        dest[1] = (unsigned char)(temp >> 16);
        dest[2] = (unsigned char)(temp >> 24);

        src += sourceStride;
        dest += destinationStride * 3;
    }
*/

    short savedFpuControlWord;
    
    signed long tempInt32;

    (void) ditherGenerator; /* unused parameter */
                 
    __asm{
        // esi -> source ptr
        // eax -> source byte stride
        // edi -> destination ptr
        // ebx -> destination byte stride
        // ecx -> source end ptr
        // edx -> temp

        mov     esi, sourceBuffer

        mov     edx, 4                  // sizeof float32
        mov     eax, sourceStride
        imul    eax, edx

        mov     ecx, count
        imul    ecx, eax
        add     ecx, esi

        mov     edi, destinationBuffer

        mov     edx, 3                  // sizeof int24
        mov     ebx, destinationStride
        imul    ebx, edx

        fwait
        fstcw   savedFpuControlWord
        fldcw   fpuControlWord_

        fld     int24Scaler_             // stack:  (int)0x7FFFFF

    Float32_To_Int24_Clip_loop:

        mov     edx, dword ptr [esi]    // load floating point value into integer register

        and     edx, 0x7FFFFFFF         // mask off sign
        cmp     edx, 0x3F800000         // greater than 1.0 or less than -1.0

        jg      Float32_To_Int24_Clip_clamp

        // load unscaled value into st(0)
        fld     dword ptr [esi]         // stack:  value, (int)0x7FFFFF
        add     esi, eax                // increment source ptr
        //lea     esi, [esi+eax]
        fmul    st(0), st(1)            // st(0) *= st(1), stack:  value*0x7FFFFF, (int)0x7FFFFF
        fistp   tempInt32               // pop st(0) into tempInt32, stack:  (int)0x7FFFFF
        mov     edx, tempInt32
        jmp     Float32_To_Int24_Clip_store
    
    Float32_To_Int24_Clip_clamp:
        mov     edx, dword ptr [esi]    // load floating point value into integer register
        shr     edx, 31                 // move sign bit into bit 0
        add     esi, eax                // increment source ptr
        //lea     esi, [esi+eax]
        add     edx, 0x7FFFFF           // convert to maximum range integers

    Float32_To_Int24_Clip_store:

        mov     byte ptr [edi], DL
        shr     edx, 8
        //mov     byte ptr [edi+1], DL
        //mov     byte ptr [edi+2], DH
        mov     word ptr [edi+1], DX

        //add     edi, ebx                // increment destination ptr
        lea     edi, [edi+ebx]

        cmp     esi, ecx                // has src ptr reached end?
        jne     Float32_To_Int24_Clip_loop

        ffree   st(0)
        fincstp

        fwait
        fnclex
        fldcw   savedFpuControlWord
    }
}

/* -------------------------------------------------------------------------- */

static void Float32_To_Int24_DitherClip(
    void *destinationBuffer, signed int destinationStride,
    void *sourceBuffer, signed int sourceStride,
    unsigned int count, PaUtilTriangularDitherGenerator *ditherGenerator )
{
/*
    float *src = (float*)sourceBuffer;
    unsigned char *dest = (unsigned char*)destinationBuffer;
    signed long temp;
    
    while( count-- )
    {
        // convert to 32 bit and drop the low 8 bits

        // FIXME: the dither amplitude here appears to be too small by 8 bits
        double dither  = PaUtil_GenerateFloatTriangularDither( ditherGenerator );
        // use smaller scaler to prevent overflow when we add the dither
        double dithered = ((double)*src * (2147483646.0)) + dither;
        PA_CLIP_( dithered, -2147483648., 2147483647.  );
        
        temp = (signed long) dithered;

        dest[0] = (unsigned char)(temp >> 8);
        dest[1] = (unsigned char)(temp >> 16);
        dest[2] = (unsigned char)(temp >> 24);

        src += sourceStride;
        dest += destinationStride * 3;
    }
*/

    short savedFpuControlWord;

    // spill storage:
    signed long sourceByteStride;
    signed long highpassedDither;

    // dither state:
    unsigned long ditherPrevious = ditherGenerator->previous;
    unsigned long ditherRandSeed1 = ditherGenerator->randSeed1;
    unsigned long ditherRandSeed2 = ditherGenerator->randSeed2;
    
    signed long tempInt32;
                 
    __asm{
        // esi -> source ptr
        // eax -> source byte stride
        // edi -> destination ptr
        // ebx -> destination byte stride
        // ecx -> source end ptr
        // edx -> temp

        mov     esi, sourceBuffer

        mov     edx, 4                  // sizeof float32
        mov     eax, sourceStride
        imul    eax, edx

        mov     ecx, count
        imul    ecx, eax
        add     ecx, esi

        mov     edi, destinationBuffer

        mov     edx, 3                  // sizeof int24
        mov     ebx, destinationStride
        imul    ebx, edx

        fwait
        fstcw   savedFpuControlWord
        fldcw   fpuControlWord_

        fld     ditheredInt24Scaler_    // stack:  int scaler

    Float32_To_Int24_DitherClip_loop:

        mov     edx, dword ptr [esi]    // load floating point value into integer register

        and     edx, 0x7FFFFFFF         // mask off sign
        cmp     edx, 0x3F800000         // greater than 1.0 or less than -1.0

        jg      Float32_To_Int24_DitherClip_clamp

        // load unscaled value into st(0)
        fld     dword ptr [esi]         // stack:  value, int scaler
        add     esi, eax                // increment source ptr
        //lea     esi, [esi+eax]
        fmul    st(0), st(1)            // st(0) *= st(1), stack:  value*(int scaler), int scaler

    /*
        // call PaUtil_GenerateFloatTriangularDither with C calling convention
        mov     sourceByteStride, eax   // save eax
        mov     sourceEnd, ecx          // save ecx
        push    ditherGenerator         // pass ditherGenerator parameter on stack
	    call    PaUtil_GenerateFloatTriangularDither  // stack:  dither, value*(int scaler), int scaler
	    pop     edx                     // clear parameter off stack
        mov     ecx, sourceEnd          // restore ecx
        mov     eax, sourceByteStride   // restore eax
    */
    
    // generate dither
        mov     sourceByteStride, eax   // save eax
        mov     edx, 196314165
        mov     eax, ditherRandSeed1
        mul     edx                     // eax:edx = eax * 196314165
        //add     eax, 907633515
        lea     eax, [eax+907633515]
        mov     ditherRandSeed1, eax
        mov     edx, 196314165
        mov     eax, ditherRandSeed2
        mul     edx                     // eax:edx = eax * 196314165
        //add     eax, 907633515
        lea     eax, [eax+907633515]
        mov     edx, ditherRandSeed1
        shr     edx, PA_DITHER_SHIFT_
        mov     ditherRandSeed2, eax
        shr     eax, PA_DITHER_SHIFT_
        //add     eax, edx                // eax -> current
        lea     eax, [eax+edx]
        mov     edx, ditherPrevious
        neg     edx
        lea     edx, [eax+edx]          // highpass = current - previous
        mov     highpassedDither, edx
        mov     ditherPrevious, eax     // previous = current
        mov     eax, sourceByteStride   // restore eax
        fild    highpassedDither
        fmul    const_float_dither_scale_
    // end generate dither, dither signal in st(0)

        faddp   st(1), st(0)            // stack: dither * value*(int scaler), int scaler
        fistp   tempInt32               // pop st(0) into tempInt32, stack:  int scaler
        mov     edx, tempInt32
        jmp     Float32_To_Int24_DitherClip_store
    
    Float32_To_Int24_DitherClip_clamp: