diff --git a/src/n64/core/RSP.hpp b/src/n64/core/RSP.hpp
index 9490160f..26aa1c04 100644
--- a/src/n64/core/RSP.hpp
+++ b/src/n64/core/RSP.hpp
@@ -142,6 +142,22 @@ struct RSP {
     nextPC = val += 4;
   }
 
+  inline s64 GetACC(int e) {
+    s64 val = s64(acc.h.element[e]) << 32;
+    val |= s64(acc.m.element[e]) << 16;
+    val |= s64(acc.l.element[e]);
+    if((val & 0x800000000000) != 0) {
+      val |= 0xFFFF000000000000;
+    }
+    return val;
+  }
+
+  inline void SetACC(int e, s64 val) {
+    acc.h.element[e] = val >> 32;
+    acc.m.element[e] = val >> 16;
+    acc.l.element[e] = val;
+  }
+
   inline u16 VCOasU16() {
     u16 val = 0;
     for(int i = 0; i < 8; i++) {
@@ -296,6 +312,7 @@ struct RSP {
   void lbu(u32 instr);
   void lhu(u32 instr);
   void lui(u32 instr);
+  void ldv(u32 instr);
   void lqv(u32 instr);
   void j(u32 instr);
   void jal(u32 instr);
@@ -311,6 +328,8 @@ struct RSP {
   void sw(u32 instr);
   void sub(u32 instr);
   void sqv(u32 instr);
+  void sdv(u32 instr);
+  void ssv(u32 instr);
   void sllv(u32 instr);
   void srlv(u32 instr);
   void srav(u32 instr);
@@ -323,9 +342,11 @@ struct RSP {
   void sltiu(u32 instr);
   void vabs(u32 instr);
   void vmov(u32 instr);
+  void vmacf(u32 instr);
   void veq(u32 instr);
   void vne(u32 instr);
   void vsar(u32 instr);
+  void vzero(u32 instr);
   void mfc0(RDP& rdp, u32 instr);
   void mtc0(Registers& regs, Mem& mem, u32 instr);
   void mfc2(u32 instr);
diff --git a/src/n64/core/rsp/decode.cpp b/src/n64/core/rsp/decode.cpp
index 4d9d3512..d70aa9a7 100644
--- a/src/n64/core/rsp/decode.cpp
+++ b/src/n64/core/rsp/decode.cpp
@@ -58,6 +58,7 @@ inline void regimm(RSP& rsp, u32 instr) {
 inline void lwc2(RSP& rsp, u32 instr) {
   u8 mask = (instr >> 11) & 0x1F;
   switch(mask) {
+    case 0x03: rsp.ldv(instr); break;
     case 0x04: rsp.lqv(instr); break;
     default: util::panic("Unhandled RSP LWC2 {:06b}\n", mask);
   }
@@ -66,6 +67,8 @@ inline void lwc2(RSP& rsp, u32 instr) {
 inline void swc2(RSP& rsp, u32 instr) {
   u8 mask = (instr >> 11) & 0x1F;
   switch(mask) {
+    case 0x01: rsp.ssv(instr); break;
+    case 0x03: rsp.sdv(instr); break;
     case 0x04: rsp.sqv(instr); break;
     default: util::panic("Unhandled RSP SWC2 {:06b}\n", mask);
   }
@@ -81,9 +84,11 @@ inline void cop2(RSP& rsp, u32 instr) {
         case 0x02: rsp.cfc2(instr); break;
         case 0x04: rsp.mtc2(instr); break;
         case 0x06: rsp.ctc2(instr); break;
-        default: util::panic("Unhandled RSP COP2 sub ({:06b})\n", mask_sub);
+        case 0x1E: rsp.vzero(instr); break;
+        default: util::panic("Unhandled RSP COP2 sub ({:05b})\n", mask_sub);
       }
       break;
+    case 0x08: rsp.vmacf(instr); break;
     //case 0x13: rsp.vabs(instr); break;
     //case 0x1D: rsp.vsar(instr); break;
     //case 0x21: rsp.veq(instr); break;
diff --git a/src/n64/core/rsp/instructions.cpp b/src/n64/core/rsp/instructions.cpp
index d2a66f18..ea56de1d 100644
--- a/src/n64/core/rsp/instructions.cpp
+++ b/src/n64/core/rsp/instructions.cpp
@@ -205,9 +205,11 @@ void RSP::lui(u32 instr) {
   gpr[RT(instr)] = imm;
 }
 
+#define OFFSET(x) ((x) & 0x7F)
+
 void RSP::lqv(u32 instr) {
   int e = E(instr);
-  u32 addr = gpr[BASE(instr)] + SignExt7bit(instr & 0x7F, 4);
+  u32 addr = gpr[BASE(instr)] + SignExt7bit(OFFSET(instr), 4);
   u32 end = ((addr & ~15) + 15);
 
   for(int i = 0; addr + i <= end && i + e < 16; i++) {
@@ -215,6 +217,17 @@ void RSP::lqv(u32 instr) {
   }
 }
 
+void RSP::ldv(u32 instr) {
+  int e = E(instr);
+  u32 addr = gpr[BASE(instr)] + SignExt7bit(OFFSET(instr), 3);
+  u32 end = e + 8 > 16 ? 16 : e + 8;
+
+  for(int i = e; i < end; i++) {
+    vpr[VT(instr)].byte[BYTE_INDEX(i)] = ReadByte(addr);
+    addr++;
+  }
+}
+
 void RSP::j(u32 instr) {
   u32 target = (instr & 0x3ffffff) << 2;
   nextPC = target;
@@ -281,7 +294,7 @@ void RSP::sub(u32 instr) {
 
 void RSP::sqv(u32 instr) {
   int e = E(instr);
-  u32 addr = gpr[BASE(instr)] + SignExt7bit(instr & 0x7F, 4);
+  u32 addr = gpr[BASE(instr)] + SignExt7bit(OFFSET(instr), 4);
   u32 end = ((addr & ~15) + 15);
 
   for(int i = 0; addr + i <= end; i++) {
@@ -289,6 +302,26 @@ void RSP::sqv(u32 instr) {
   }
 }
 
+void RSP::sdv(u32 instr) {
+  int e = E(instr);
+  u32 addr = gpr[BASE(instr)] + SignExt7bit(OFFSET(instr), 3);
+
+  for(int i = 0; i < 8; i++) {
+    WriteByte(addr + i, vpr[VT(instr)].byte[BYTE_INDEX((i + e) & 0xF)]);
+  }
+}
+
+void RSP::ssv(u32 instr) {
+  int e = E(instr);
+  u32 addr = gpr[BASE(instr)] + SignExt7bit(OFFSET(instr), 1);
+
+  u8 hi = vpr[VT(instr)].byte[BYTE_INDEX(e & 15)];
+  u8 lo = vpr[VT(instr)].byte[BYTE_INDEX((e + 1) & 15)];
+  u16 val = (u16)hi << 8 | lo;
+
+  WriteHalf(addr, val);
+}
+
 void RSP::sllv(u32 instr) {
   u8 sa = (gpr[RS(instr)]) & 0x1F;
   u32 rt = gpr[RT(instr)];
@@ -343,6 +376,12 @@ void RSP::sltiu(u32 instr) {
   gpr[RT(instr)] = (u32)gpr[RS(instr)] < imm;
 }
 
+inline s16 clamp_signed(s64 val) {
+  if(val > 32767) return 32767;
+  if(val < -32768) return -32768;
+  return val;
+}
+
 void RSP::vabs(u32 instr) {
 
 }
@@ -351,6 +390,22 @@ void RSP::vmov(u32 instr) {
 
 }
 
+void RSP::vmacf(u32 instr) {
+  VPR& vd = vpr[VD(instr)];
+  VPR& vs = vpr[VS(instr)];
+  VPR vte = GetVTE(vpr[VT(instr)], E(instr));
+
+  for(int i = 0; i < 8; i++) {
+    s32 prod = vs.selement[i] * vte.selement[i];
+    s64 accDelta = prod * 2;
+    s64 accum = GetACC(i) + accDelta;
+    SetACC(i, accum);
+
+    s16 result = clamp_signed(accum >> 16);
+    vd.element[i] = result;
+  }
+}
+
 void RSP::veq(u32 instr) {
 
 }
@@ -363,6 +418,18 @@ void RSP::vsar(u32 instr) {
 
 }
 
+void RSP::vzero(u32 instr) {
+  VPR& vs = vpr[VS(instr)];
+  VPR& vd = vpr[VD(instr)];
+  VPR vte = GetVTE(vpr[VT(instr)], E(instr));
+
+  for(int i = 0; i < 8; i++) {
+    acc.l.element[i] = vte.element[i] + vs.element[i];
+  }
+
+  memset(&vd, 0, sizeof(VPR));
+}
+
 void RSP::mfc0(RDP& rdp, u32 instr) {
   gpr[RT(instr)] = GetCop0Reg(*this, rdp, RD(instr));
 }