thirdparty/ppsspp
2
0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

vertexjit: Use NEON loads for up to 8 weights.

Browse source
This commit is contained in:
Unknown W. Brackets 2014-03-22 16:35:28 -07:00
parent cb02d66d8f
commit 5ee4ba54b7
1 changed files with 49 additions and 35 deletions
Download patch file Download diff file Expand all files Collapse all files

84
GPU/GLES/VertexDecoderArm.cpp
View file

@ -426,21 +426,17 @@ void VertexDecoderJitCache::Jit_ApplyWeights() {
}
void VertexDecoderJitCache::Jit_WeightsU8Skin() {
if (NEONSkinning && dec_->nweights <= 4) {
// Most common cases.
if (NEONSkinning) {
// Weight is first so srcReg is correct.
switch (dec_->nweights) {
case 1: LDRB(scratchReg2, srcReg, 0); break;
case 2: LDRH(scratchReg2, srcReg, 0); break;
case 3:
LDR(scratchReg2, srcReg, 0);
ANDI2R(scratchReg2, scratchReg2, 0xFFFFFF, scratchReg);
break;
case 4:
VLD1_lane(I_32, neonScratchReg, srcReg, 0, false);
break;
}
if (dec_->nweights != 4) {
if (dec_->nweights == 1 || dec_->nweights == 2) {
VMOV_neon(I_32, neonScratchReg, scratchReg2, 0);
}
// This can be represented as a constant.
@ -449,8 +445,26 @@ void VertexDecoderJitCache::Jit_WeightsU8Skin() {
VMOVL(I_16 | I_UNSIGNED, neonScratchRegQ, neonScratchReg);
VCVT(F_32 | I_UNSIGNED, neonScratchRegQ, neonScratchRegQ);
VMUL(F_32, neonWeightRegsQ[0], neonScratchRegQ, Q3);
if (dec_->nweights > 4) {
ADD(tempReg1, srcReg, 4 * sizeof(u8));
switch (dec_->nweights) {
case 5: LDRB(scratchReg2, tempReg1, 0); break;
case 6: LDRH(scratchReg2, tempReg1, 0); break;
case 7:
case 8:
VLD1_lane(I_32, neonScratchReg, tempReg1, 0, false);
break;
}
if (dec_->nweights == 5 || dec_->nweights == 6) {
VMOV_neon(I_32, neonScratchReg, scratchReg2, 0);
}
VMOVL(I_8 | I_UNSIGNED, neonScratchRegQ, neonScratchReg);
VMOVL(I_16 | I_UNSIGNED, neonScratchRegQ, neonScratchReg);
VCVT(F_32 | I_UNSIGNED, neonScratchRegQ, neonScratchRegQ);
VMUL(F_32, neonWeightRegsQ[1], neonScratchRegQ, Q3);
}
} else {
// Fallback and non-neon
for (int j = 0; j < dec_->nweights; j++) {
LDRB(tempReg1, srcReg, dec_->weightoff + j);
VMOV(fpScratchReg, tempReg1);
@ -463,27 +477,41 @@ void VertexDecoderJitCache::Jit_WeightsU8Skin() {
}
void VertexDecoderJitCache::Jit_WeightsU16Skin() {
if (NEONSkinning && dec_->nweights <= 4) {
// Most common cases.
if (NEONSkinning) {
switch (dec_->nweights) {
case 1: LDRH(scratchReg, srcReg, 0); break;
case 1:
LDRH(scratchReg, srcReg, 0);
VMOV_neon(I_32, neonScratchReg, scratchReg, 0);
break;
case 2: VLD1_lane(I_32, neonScratchReg, srcReg, 0, false); break;
case 3:
LDR(scratchReg, srcReg, 0);
LDRH(scratchReg2, srcReg, 4);
break;
case 4:
VLD1(I_32, neonScratchReg, srcReg, 1, ALIGN_NONE);
break;
}
if (dec_->nweights != 2 && dec_->nweights != 4) {
VMOV(neonScratchReg, scratchReg, scratchReg2);
}
// This can be represented as a constant.
VMOV_neon(F_32, Q3, by32768);
VMOVL(I_16 | I_UNSIGNED, neonScratchRegQ, neonScratchReg);
VCVT(F_32 | I_UNSIGNED, neonScratchRegQ, neonScratchRegQ);
VMUL(F_32, neonWeightRegsQ[0], neonScratchRegQ, Q3);
if (dec_->nweights > 4) {
ADD(tempReg1, srcReg, 4 * sizeof(u16));
switch (dec_->nweights) {
case 5:
LDRH(scratchReg, tempReg1, 0);
VMOV_neon(I_32, neonScratchReg, scratchReg, 0);
break;
case 6: VLD1_lane(I_32, neonScratchReg, tempReg1, 0, false); break;
case 7:
case 8:
VLD1(I_32, neonScratchReg, srcReg, 1, ALIGN_NONE);
break;
}
VMOVL(I_16 | I_UNSIGNED, neonScratchRegQ, neonScratchReg);
VCVT(F_32 | I_UNSIGNED, neonScratchRegQ, neonScratchRegQ);
VMUL(F_32, neonWeightRegsQ[1], neonScratchRegQ, Q3);
}
} else {
// Fallback and non-neon
for (int j = 0; j < dec_->nweights; j++) {
@ -505,25 +533,11 @@ void VertexDecoderJitCache::Jit_WeightsFloatSkin() {
// Weights are always first, so we can use srcReg directly.
if (NEONSkinning) {
switch (dec_->nweights) {
case 1: VLD1_lane(F_32, neonWeightRegsD[0], srcReg, 0, true); break;
case 2: VLD1(F_32, neonWeightRegsD[0], srcReg, 1, ALIGN_NONE); break;
case 3:
// TODO: Do we really need to skip the 4th lanes? 8/16 seem so careful to do that...
VLD1(F_32, neonWeightRegsD[0], srcReg, 1, ALIGN_NONE);
VLD1_lane(F_32, neonWeightRegsD[1], srcReg, 0, true);
break;
case 4: VLD1(F_32, neonWeightRegsD[0], srcReg, 2, ALIGN_NONE); break;
case 5:
VLD1(F_32, neonWeightRegsD[0], srcReg, 2, ALIGN_NONE);
VLD1_lane(F_32, neonWeightRegsD[2], srcReg, 0, true);
break;
case 6: VLD1(F_32, neonWeightRegsD[0], srcReg, 3, ALIGN_NONE); break;
case 7:
VLD1(F_32, neonWeightRegsD[0], srcReg, 3, ALIGN_NONE);
VLD1_lane(F_32, neonWeightRegsD[3], srcReg, 0, true);
break;
case 8: VLD1(F_32, neonWeightRegsD[0], srcReg, 4, ALIGN_NONE); break;
if (dec_->nweights == 1) {
VLD1_lane(F_32, neonWeightRegsD[0], srcReg, 0, true);
} else {
// We over-read by one float but this is not a tragedy.
VLDMIA(srcReg, false, neonWeightRegsD[0], (dec_->nweights + 1) / 2);
}
} else {
VLDMIA(srcReg, false, weightRegs[0], dec_->nweights);