Basic RSP LTV/STV support

CMakeLists.txt

@@ -256,6 +256,7 @@ set(ARCH_X86_64_SOURCES
   ${PROJECT_SOURCE_DIR}/arch/x86_64/rsp/vdivh.c
   ${PROJECT_SOURCE_DIR}/arch/x86_64/rsp/rsp.c
   ${PROJECT_SOURCE_DIR}/arch/x86_64/rsp/vrsq.c
+  ${PROJECT_SOURCE_DIR}/arch/x86_64/rsp/transpose.c
 )
 
 set(BUS_SOURCES

arch/x86_64/rsp/rsp.h

@@ -209,6 +209,7 @@ void rsp_vstore_group4(struct rsp *rsp, uint32_t addr, unsigned element,
   uint16_t *regp, rsp_vect_t reg, rsp_vect_t dqm);
 
 #include "arch/x86_64/rsp/clamp.h"
+#include "arch/x86_64/rsp/transpose.h"
 #include "arch/x86_64/rsp/vabs.h"
 #include "arch/x86_64/rsp/vadd.h"
 #include "arch/x86_64/rsp/vaddc.h"

arch/x86_64/rsp/transpose.c

@@ -0,0 +1,36 @@
+//
+// arch/x86_64/rsp/transpose.c
+//
+// This file is subject to the terms and conditions defined in
+// 'LICENSE', which is part of this source code package.
+//
+
+#include "common.h"
+#include "rsp/cpu.h"
+#include "rsp/rsp.h"
+
+void rsp_ltv(struct rsp *rsp, uint32_t addr, unsigned element, unsigned vt) {
+  struct rsp_exdf_latch *exdf_latch = &rsp->pipeline.exdf_latch;
+
+  for(int i = 0; i < 8; i++){
+    uint16_t slice;
+
+    memcpy(&slice, rsp->mem + addr + (i << 1), sizeof(slice));
+    slice = byteswap_16(slice);
+
+    rsp->cp2.regs[vt + i].e[(i -  element) & 7] = slice;
+  }
+}
+
+void rsp_stv(struct rsp *rsp, uint32_t addr, unsigned element, unsigned vt) {
+  struct rsp_exdf_latch *exdf_latch = &rsp->pipeline.exdf_latch;
+
+  for(int i = 0; i < 8; i++){
+    uint16_t slice = rsp->cp2.regs[vt + ((i + element) & 7)].e[i];
+    slice = byteswap_16(slice);
+
+    memcpy(rsp->mem + addr + (i << 1), &slice, sizeof(slice));
+  }
+}
+
+

arch/x86_64/rsp/transpose.h

@@ -0,0 +1,10 @@
+//
+// arch/x86_64/rsp/transpose.h
+//
+// This file is subject to the terms and conditions defined in
+// 'LICENSE', which is part of this source code package.
+//
+
+void rsp_ltv(struct rsp *rsp, uint32_t addr, unsigned vt, unsigned element);
+
+void rsp_stv(struct rsp *rsp, uint32_t addr, unsigned vt, unsigned element);

common.h.in

@@ -162,6 +162,23 @@ static inline uint32_t byteswap_32(uint32_t word) {
 #endif
 }
 
+#ifdef __GNUC__
+__attribute__((pure))
+#endif
+static inline uint16_t byteswap_16(uint16_t hword) {
+#ifdef BIG_ENDIAN_HOST
+  return hword;
+#elif defined(_MSC_VER)
+  return _byteswap_ushort(hword);
+#elif defined(__GNUC__)
+  return __builtin_bswap16(hword);
+#else
+  return
+  ((((hword) >> 8) & 0x00FF) | \
+    (((hword) <<  8) & 0xFF00));
+#endif
+}
+
 // Return from simulation function.
 struct bus_controller;
 

rsp/functions.c

@@ -471,6 +471,27 @@ void RSP_LQRV_SQRV(struct rsp *rsp,
   exdf_latch->request.packet.p_vect.dest = dest;
 }
 
+//
+// LTV
+// STV
+//
+void RSP_LTV_STV(struct rsp *rsp,
+  uint32_t iw, uint32_t rs, uint32_t rt) {
+  struct rsp_exdf_latch *exdf_latch = &rsp->pipeline.exdf_latch;
+  unsigned op = iw >> 29 & 0x1;
+
+  exdf_latch->request.addr = rs + (sign_extend_6(iw) << 4);
+  exdf_latch->request.packet.p_vect.element = GET_EL(iw);
+  exdf_latch->request.type = RSP_MEM_REQUEST_TRANSPOSE;
+
+  exdf_latch->request.packet.p_transpose.vt = GET_VT(iw) & 0x18;
+  exdf_latch->request.packet.p_transpose.element = GET_EL(iw) >> 1;
+
+  exdf_latch->request.packet.p_transpose.transpose_func = op
+    ? rsp_stv
+    : rsp_ltv;
+}
+
 //
 // NOR
 //

rsp/opcodes_priv.h

@@ -129,7 +129,7 @@
 #define LQV RSP_BUILD_OP(LQV, LQRV_SQRV, INFO3(NEEDRS, NEEDVT, LOAD))
 #define LRV RSP_BUILD_OP(LRV, LQRV_SQRV, INFO3(NEEDRS, NEEDVT, LOAD))
 #define LSV RSP_BUILD_OP(LSV, LBDLSV_SBDLSV, INFO3(NEEDRS, NEEDVT, LOAD))
-#define LTV RSP_BUILD_OP(LTV, INVALID, INFO1(NONE))
+#define LTV RSP_BUILD_OP(LTV, LTV_STV, INFO3(NEEDRS, NEEDVT, LOAD))
 #define LUV RSP_BUILD_OP(LUV, LFHPUV_SFHPUV, INFO3(NEEDRS, NEEDVT, LOAD))
 
 #define SBV RSP_BUILD_OP(SBV, LBDLSV_SBDLSV, INFO3(NEEDRS, NEEDVT, STORE))
@@ -141,7 +141,7 @@
 #define SQV RSP_BUILD_OP(SQV, LQRV_SQRV, INFO3(NEEDRS, NEEDVT, STORE))
 #define SRV RSP_BUILD_OP(SRV, LQRV_SQRV, INFO3(NEEDRS, NEEDVT, STORE))
 #define SSV RSP_BUILD_OP(SSV, LBDLSV_SBDLSV, INFO3(NEEDRS, NEEDVT, STORE))
-#define STV RSP_BUILD_OP(STV, INVALID, INFO1(NONE))
+#define STV RSP_BUILD_OP(STV, LTV_STV, INFO3(NEEDRS, NEEDVT, LOAD))
 #define SUV RSP_BUILD_OP(SUV, LFHPUV_SFHPUV, INFO3(NEEDRS, NEEDVT, STORE))
 #define SWV RSP_BUILD_OP(SWV, INVALID, INFO1(NONE))
 

rsp/pipeline.c

@@ -161,8 +161,32 @@ cen64_flatten static inline void rsp_df_stage(struct rsp *rsp) {
 
   addr = request->addr & 0xFFF;
 
+  // Scalar unit DMEM access.
+  if (request->type == RSP_MEM_REQUEST_INT_MEM) {
+    uint32_t rdqm = request->packet.p_int.rdqm;
+    uint32_t wdqm = request->packet.p_int.wdqm;
+    uint32_t data = request->packet.p_int.data;
+    unsigned rshift = request->packet.p_int.rshift;
+    uint32_t word;
+
+    memcpy(&word, rsp->mem + addr, sizeof(word));
+
+    word = byteswap_32(word);
+    dfwb_latch->result.result = rdqm & (((int32_t) word) >> rshift);
+    word = byteswap_32((word & ~wdqm) | (data & wdqm));
+
+    memcpy(rsp->mem + addr, &word, sizeof(word));
+  }
+  // Transposed vector unit DMEM access.
+  else if (request->type == RSP_MEM_REQUEST_TRANSPOSE) {
+    unsigned element = request->packet.p_transpose.element;
+    unsigned vt = request->packet.p_transpose.vt;
+
+    exdf_latch->request.packet.p_transpose.transpose_func(
+      rsp, addr, element, vt);
+  }
   // Vector unit DMEM access.
-  if (request->type != RSP_MEM_REQUEST_INT_MEM) {
+  else {
     uint16_t *regp = rsp->cp2.regs[request->packet.p_vect.dest].e;
     unsigned element = request->packet.p_vect.element;
     rsp_vect_t reg, dqm;
@@ -179,22 +203,6 @@ cen64_flatten static inline void rsp_df_stage(struct rsp *rsp) {
       rsp, addr, element, regp, reg, dqm);
   }
 
-  // Scalar unit DMEM access.
-  else {
-    uint32_t rdqm = request->packet.p_int.rdqm;
-    uint32_t wdqm = request->packet.p_int.wdqm;
-    uint32_t data = request->packet.p_int.data;
-    unsigned rshift = request->packet.p_int.rshift;
-    uint32_t word;
-
-    memcpy(&word, rsp->mem + addr, sizeof(word));
-
-    word = byteswap_32(word);
-    dfwb_latch->result.result = rdqm & (((int32_t) word) >> rshift);
-    word = byteswap_32((word & ~wdqm) | (data & wdqm));
-
-    memcpy(rsp->mem + addr, &word, sizeof(word));
-  }
 }
 
 // Writeback stage.

rsp/pipeline.h

@@ -26,7 +26,8 @@ enum rsp_mem_request_type {
   RSP_MEM_REQUEST_PACK,
   RSP_MEM_REQUEST_QUAD,
   RSP_MEM_REQUEST_REST,
-  RSP_MEM_REQUEST_UPACK
+  RSP_MEM_REQUEST_UPACK,
+  RSP_MEM_REQUEST_TRANSPOSE,
 };
 
 struct rsp_int_mem_packet {
@@ -37,6 +38,14 @@ struct rsp_int_mem_packet {
   unsigned rshift;
 };
 
+struct rsp_transpose_mem_packet {
+  void (*transpose_func)(struct rsp *rsp, uint32_t addr, unsigned element,
+    unsigned vt);
+
+  unsigned element;
+  unsigned vt;
+};
+
 struct rsp_vect_mem_packet {
   union aligned_rsp_1vect_t vdqm;
 
@@ -49,6 +58,7 @@ struct rsp_vect_mem_packet {
 
 union rsp_mem_packet {
   struct rsp_int_mem_packet p_int;
+  struct rsp_transpose_mem_packet p_transpose;
   struct rsp_vect_mem_packet p_vect;
 };