mirror of https://github.com/PCSX2/pcsx2.git
GregMiscellaneous: zzogl-pg: Nice comment to decipher asm code.
git-svn-id: http://pcsx2.googlecode.com/svn/branches/GregMiscellaneous@3851 96395faa-99c1-11dd-bbfe-3dabce05a288
This commit is contained in:
gregory.hainaut@gmail.com
parent
c48a820cd5
commit
8d7edf440d
|
|
@ -20,6 +20,11 @@
|
|||
|
||||
#ifdef ZEROGS_SSE2
|
||||
// SSE2 extensions
|
||||
|
||||
// Note: pshufd 0xea <=> movdqa !!!
|
||||
// What the function does is
|
||||
// Interleave s1 and sd0 -> d1 (high) & sd0 (low)
|
||||
// Interleave s3 and sd2 -> d3 (high) & sd2 (low)
|
||||
#define punpck(op, sd0, sd2, s1, s3, d1, d3) \
|
||||
movdqa %xmm##d1, %xmm##sd0; \
|
||||
pshufd %xmm##d3, %xmm##sd2, 0xe4; \
|
||||
|
|
@ -29,6 +34,15 @@
|
|||
punpckh##op %xmm##d3, %xmm##s3; \
|
||||
|
||||
|
||||
// Input xmm7 == 0x0F0F0F0F 0x0F0F0F0F 0x0F0F0F0F 0x0F0F0F0F
|
||||
// DATA xmm[0-3]
|
||||
// This function does a 4-bits interleaving of 4 xmm registers
|
||||
//
|
||||
// ARG Can not put comment in the middle of the define...
|
||||
// After the first por
|
||||
// low 32bits (4bits packed) == 1.6 0.6 1.4 0.4 1.2 0.2 1.0 0.0
|
||||
// After the second one
|
||||
// low 32bits (4bits packed) == 1.7 0.7 1.5 0.5 1.3 0.3 1.1 0.1
|
||||
#define punpcknb \
|
||||
movdqa %xmm4, %xmm0; \
|
||||
pshufd %xmm5, %xmm1, 0xe4; \
|
||||
|
|
@ -48,6 +62,7 @@
|
|||
\
|
||||
movdqa %xmm1, %xmm4; \
|
||||
\
|
||||
\
|
||||
movdqa %xmm4, %xmm2; \
|
||||
pshufd %xmm5, %xmm3, 0xe4; \
|
||||
\
|
||||
|
|
@ -68,6 +83,12 @@
|
|||
\
|
||||
punpck(bw, 0, 2, 1, 3, 4, 6);\
|
||||
|
||||
// output
|
||||
// low 32 bits 0 (4 bits packed) == 1.3 0.3 1.2 0.2 1.1 0.1 1.0 0.0
|
||||
// low 32 bits 4 (4 bits packed) == 1.19 0.19 1.18 0.18 1.17 0.17 1.16 0.16
|
||||
// low 32 bits 2 (4 bits packed) == 3.3 2.3 3.2 2.2 3.1 2.1 3.0 2.0
|
||||
// low 32 bits 6 (4 bits packed) == 3.19 2.19 3.18 2.18 3.17 2.17 3.16 2.16
|
||||
|
||||
|
||||
//
|
||||
// swizzling
|
||||
|
|
@ -84,11 +105,15 @@ SwizzleBlock32_sse2:
|
|||
push %esi
|
||||
push %edi
|
||||
|
||||
// save dst
|
||||
mov %edi, %ecx
|
||||
// save src
|
||||
mov %esi, %edx
|
||||
// get pitch
|
||||
mov %edx, [%esp+4+8]
|
||||
mov %ecx, 4
|
||||
|
||||
// get WriteMask
|
||||
mov %eax, [%esp+8+8]
|
||||
cmp %eax, 0xffffffff
|
||||
jne SwizzleBlock32_sse2_2
|
||||
|
|
@ -100,6 +125,8 @@ SwizzleBlock32_sse2_1:
|
|||
movdqa %xmm1, [%esi+%edx]
|
||||
movdqa %xmm5, [%esi+%edx+16]
|
||||
|
||||
// 64bits interleave 1&0 -> 2&0
|
||||
// 64bits interleave 5&4 -> 6&4
|
||||
punpck(qdq, 0, 4, 1, 5, 2, 6)
|
||||
|
||||
movntps [%edi+16*0], %xmm0
|
||||
|
|
@ -107,6 +134,7 @@ SwizzleBlock32_sse2_1:
|
|||
movntps [%edi+16*2], %xmm4
|
||||
movntps [%edi+16*3], %xmm6
|
||||
|
||||
// update ptr
|
||||
lea %esi, [%esi+%edx*2]
|
||||
add %edi, 64
|
||||
|
||||
|
|
@ -120,6 +148,7 @@ SwizzleBlock32_sse2_1:
|
|||
|
||||
SwizzleBlock32_sse2_2:
|
||||
|
||||
// WriteMask: 32bits to 4*32bits
|
||||
movd %xmm7, %eax
|
||||
pshufd %xmm7, %xmm7, 0
|
||||
|
||||
|
|
@ -130,13 +159,19 @@ SwizzleBlock32_sse2_3:
|
|||
movdqa %xmm1, [%esi+%edx]
|
||||
movdqa %xmm5, [%esi+%edx+16]
|
||||
|
||||
// 64bits interleave 1&0 -> 2&0
|
||||
// 64bits interleave 5&4 -> 6&4
|
||||
punpck(qdq, 0, 4, 1, 5, 2, 6)
|
||||
|
||||
// save a mask copy
|
||||
movdqa %xmm3, %xmm7
|
||||
pshufd %xmm5, %xmm7, 0xe4
|
||||
|
||||
// *dst & ~WriteMask
|
||||
pandn %xmm3, [%edi+16*0]
|
||||
// *src & WriteMask
|
||||
pand %xmm0, %xmm7
|
||||
// Final value to save
|
||||
por %xmm0, %xmm3
|
||||
movntps [%edi+16*0], %xmm0
|
||||
|
||||
|
|
@ -158,6 +193,7 @@ SwizzleBlock32_sse2_3:
|
|||
por %xmm6, %xmm5
|
||||
movntps [%edi+16*3], %xmm6
|
||||
|
||||
// update ptr
|
||||
lea %esi, [%esi+%edx*2]
|
||||
add %edi, 64
|
||||
|
||||
|
|
@ -179,6 +215,7 @@ SwizzleBlock16_sse2:
|
|||
|
||||
push %ebx
|
||||
|
||||
// srcpitch
|
||||
mov %ebx, [%esp+4+4]
|
||||
mov %eax, 4
|
||||
|
||||
|
|
@ -189,7 +226,11 @@ SwizzleBlock16_sse2_1:
|
|||
movdqa %xmm2, [%edx+%ebx]
|
||||
movdqa %xmm3, [%edx+%ebx+16]
|
||||
|
||||
// 16bits interleave 1&0 -> 4&0
|
||||
// 16bits interleave 3&2 -> 6&2
|
||||
punpck(wd, 0, 2, 1, 3, 4, 6)
|
||||
// 64bits interleave 2&0 -> 1&0
|
||||
// 64bits interleave 6&4 -> 5&4
|
||||
punpck(qdq, 0, 4, 2, 6, 1, 5)
|
||||
|
||||
movntps [%ecx+16*0], %xmm0
|
||||
|
|
@ -197,6 +238,7 @@ SwizzleBlock16_sse2_1:
|
|||
movntps [%ecx+16*2], %xmm4
|
||||
movntps [%ecx+16*3], %xmm5
|
||||
|
||||
// update ptr
|
||||
lea %edx, [%edx+%ebx*2]
|
||||
add %ecx, 64
|
||||
|
||||
|
|
@ -217,7 +259,9 @@ SwizzleBlock8_sse2:
|
|||
|
||||
push %ebx
|
||||
|
||||
// load srcpitch
|
||||
mov %ebx, [%esp+4+4]
|
||||
// basic counter
|
||||
mov %eax, 2
|
||||
|
||||
.align 16
|
||||
|
|
@ -226,14 +270,23 @@ SwizzleBlock8_sse2_1:
|
|||
|
||||
movdqa %xmm0, [%edx]
|
||||
movdqa %xmm2, [%edx+%ebx]
|
||||
// update src pointer
|
||||
lea %edx, [%edx+%ebx*2]
|
||||
|
||||
// 2 3 0 1
|
||||
pshufd %xmm1, [%edx], 0xb1
|
||||
pshufd %xmm3, [%edx+%ebx], 0xb1
|
||||
// update src pointer
|
||||
lea %edx, [%edx+%ebx*2]
|
||||
|
||||
// 8bits interleave 1&0 -> 4&0
|
||||
// 8bits interleave 3&2 -> 6&2
|
||||
punpck(bw, 0, 2, 1, 3, 4, 6)
|
||||
// 16bits interleave 4&0 -> 1&0
|
||||
// 16bits interleave 6&2 -> 3&2
|
||||
punpck(wd, 0, 2, 4, 6, 1, 3)
|
||||
// 64bits interleave 2&0 -> 4&0
|
||||
// 64bits interleave 3&1 -> 5&1
|
||||
punpck(qdq, 0, 1, 2, 3, 4, 5)
|
||||
|
||||
movntps [%ecx+16*0], %xmm0
|
||||
|
|
@ -241,18 +294,27 @@ SwizzleBlock8_sse2_1:
|
|||
movntps [%ecx+16*2], %xmm1
|
||||
movntps [%ecx+16*3], %xmm5
|
||||
|
||||
// col 1, 3
|
||||
// col 1, 3 (same as previous column)
|
||||
|
||||
// 2 3 0 1
|
||||
pshufd %xmm0, [%edx], 0xb1
|
||||
pshufd %xmm2, [%edx+%ebx], 0xb1
|
||||
// update src pointer
|
||||
lea %edx, [%edx+%ebx*2]
|
||||
|
||||
movdqa %xmm1, [%edx]
|
||||
movdqa %xmm3, [%edx+%ebx]
|
||||
// update src pointer
|
||||
lea %edx, [%edx+%ebx*2]
|
||||
|
||||
// 8bits interleave 1&0 -> 4&0
|
||||
// 8bits interleave 3&2 -> 6&2
|
||||
punpck(bw, 0, 2, 1, 3, 4, 6)
|
||||
// 16bits interleave 4&0 -> 1&0
|
||||
// 16bits interleave 6&2 -> 3&2
|
||||
punpck(wd, 0, 2, 4, 6, 1, 3)
|
||||
// 64bits interleave 2&0 -> 4&0
|
||||
// 64bits interleave 3&1 -> 5&1
|
||||
punpck(qdq, 0, 1, 2, 3, 4, 5)
|
||||
|
||||
movntps [%ecx+16*4], %xmm0
|
||||
|
|
@ -260,6 +322,7 @@ SwizzleBlock8_sse2_1:
|
|||
movntps [%ecx+16*6], %xmm1
|
||||
movntps [%ecx+16*7], %xmm5
|
||||
|
||||
// update dst pointer
|
||||
add %ecx, 128
|
||||
|
||||
dec %eax
|
||||
|
|
@ -279,10 +342,12 @@ SwizzleBlock4_sse2:
|
|||
|
||||
push %ebx
|
||||
|
||||
// load 4 0x0F0F0F0F
|
||||
mov %eax, 0xf0f0f0f
|
||||
movd %xmm7, %eax
|
||||
pshufd %xmm7, %xmm7, 0
|
||||
|
||||
// load srcpitch
|
||||
mov %ebx, [%esp+4+4]
|
||||
mov %eax, 2
|
||||
|
||||
|
|
@ -292,20 +357,32 @@ SwizzleBlock4_sse2_1:
|
|||
|
||||
movdqa %xmm0, [%edx]
|
||||
movdqa %xmm2, [%edx+%ebx]
|
||||
//update src pointer
|
||||
lea %edx, [%edx+%ebx*2]
|
||||
|
||||
movdqa %xmm1, [%edx]
|
||||
movdqa %xmm3, [%edx+%ebx]
|
||||
// update src pointer
|
||||
lea %edx, [%edx+%ebx*2]
|
||||
|
||||
// - - - - 2 3 0 1
|
||||
pshuflw %xmm1, %xmm1, 0xb1
|
||||
pshuflw %xmm3, %xmm3, 0xb1
|
||||
// 6 7 4 5 - - - -
|
||||
pshufhw %xmm1, %xmm1, 0xb1
|
||||
pshufhw %xmm3, %xmm3, 0xb1
|
||||
|
||||
// 4bits interleave 1&0 -> 4&0
|
||||
// 4bits interleave 3&2 -> 6&2
|
||||
punpcknb
|
||||
// 8bits interleave 4&0 -> 1&0
|
||||
// 8bits interleave 6&2 -> 3&2
|
||||
punpck(bw, 0, 2, 4, 6, 1, 3)
|
||||
// 8bits interleave 1&0 -> 4&0
|
||||
// 8bits interleave 3&2 -> 6&2
|
||||
punpck(bw, 0, 2, 1, 3, 4, 6)
|
||||
// 64bits interleave 2&0 -> 1&0
|
||||
// 64bits interleave 6&4 -> 3&4
|
||||
punpck(qdq, 0, 4, 2, 6, 1, 3)
|
||||
|
||||
movntps [%ecx+16*0], %xmm0
|
||||
|
|
@ -313,7 +390,7 @@ SwizzleBlock4_sse2_1:
|
|||
movntps [%ecx+16*2], %xmm4
|
||||
movntps [%ecx+16*3], %xmm3
|
||||
|
||||
// col 1, 3
|
||||
// col 1, 3 (same as previous column)
|
||||
|
||||
movdqa %xmm0, [%edx]
|
||||
movdqa %xmm2, [%edx+%ebx]
|
||||
|
|
@ -349,6 +426,9 @@ SwizzleBlock4_sse2_1:
|
|||
|
||||
//
|
||||
// swizzling with unaligned reads
|
||||
// Same functions as a above with movdqu instead of movdqa for the reads
|
||||
// Movdqu is as fast as movdqa with aligned address... So do not bother, directly
|
||||
// use movdqu
|
||||
//
|
||||
|
||||
//
|
||||
|
|
|
|||
Loading…
Reference in New Issue