arch-gcn3: fix bug with DPP support

Instructions that use the DPP field need to use the extra SRC0
register associated with the DPP instruction instead of the
"default" SRC0 register, since the default SRC0 register contains
the DPP information when DPP is being used.  This commit fixes
2735c3bb88 to take this into account.  Additionally, this commit
removes write of the src register from the DPP helper functions,
to avoid overwriting any changes made to the destination register.
Finally, this change modifies the instructions that use DPP to
simplify the flow through the execute() functions.

Change-Id: I80fd0af1f131f287f18ff73b3c1c9122d8c60823
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/29947
Maintainer: Anthony Gutierrez <anthony.gutierrez@amd.com>
Tested-by: kokoro <noreply+kokoro@google.com>
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>

diff --git a/src/arch/gcn3/insts/inst_util.hh b/src/arch/gcn3/insts/inst_util.hh
index 433ccbe8dc9f42024ee9f2e1c05ab22a0d23a5f1..b40e890f6aa411fcd93579d9d9c26a1e7ca8412c 100644 (file)
--- a/src/arch/gcn3/insts/inst_util.hh
+++ b/src/arch/gcn3/insts/inst_util.hh
@@ -505,7 +505,6 @@ namespace Gcn3ISA
                 src0[lane] = 0;
             }
 
-            src0.write();
             // reset for next iteration
             laneDisabled = false;
         }

diff --git a/src/arch/gcn3/insts/instructions.cc b/src/arch/gcn3/insts/instructions.cc
index b852281b13d5aa135c858da0c39fc29f844ae981..79e7ddacfd9495380b05203b33ae6adc221a82af 100644 (file)
--- a/src/arch/gcn3/insts/instructions.cc
+++ b/src/arch/gcn3/insts/instructions.cc
@@ -5296,8 +5296,12 @@ namespace Gcn3ISA
         VecOperandF32 src1(gpuDynInst, instData.VSRC1);
         VecOperandF32 vdst(gpuDynInst, instData.VDST);
 
+        src0.readSrc();
+        src1.read();
+
         if (isDPPInst()) {
             VecOperandF32 src0_dpp(gpuDynInst, extData.iFmt_VOP_DPP.SRC0);
+            src0_dpp.read();
 
             DPRINTF(GCN3, "Handling V_ADD_F32 SRC DPP. SRC0: register v[%d], "
                     "DPP_CTRL: 0x%#x, SRC0_ABS: %d, SRC0_NEG: %d, "
@@ -5313,14 +5317,17 @@ namespace Gcn3ISA
                     extData.iFmt_VOP_DPP.ROW_MASK);
 
             processDPP(gpuDynInst, extData.iFmt_VOP_DPP, src0_dpp, src1);
-        }
-
-        src0.readSrc();
-        src1.read();
 
-        for (int lane = 0; lane < NumVecElemPerVecReg; ++lane) {
-            if (wf->execMask(lane)) {
-                vdst[lane] = src0[lane] + src1[lane];
+            for (int lane = 0; lane < NumVecElemPerVecReg; ++lane) {
+                if (wf->execMask(lane)) {
+                    vdst[lane] = src0_dpp[lane] + src1[lane];
+                }
+            }
+        } else {
+            for (int lane = 0; lane < NumVecElemPerVecReg; ++lane) {
+                if (wf->execMask(lane)) {
+                    vdst[lane] = src0[lane] + src1[lane];
+                }
             }
         }
 
@@ -6164,6 +6171,7 @@ namespace Gcn3ISA
 
         if (isDPPInst()) {
             VecOperandF32 src0_dpp(gpuDynInst, extData.iFmt_VOP_DPP.SRC0);
+            src0_dpp.read();
 
             DPRINTF(GCN3, "Handling V_MAC_F32 SRC DPP. SRC0: register v[%d], "
                     "DPP_CTRL: 0x%#x, SRC0_ABS: %d, SRC0_NEG: %d, "
@@ -6179,11 +6187,18 @@ namespace Gcn3ISA
                     extData.iFmt_VOP_DPP.ROW_MASK);
 
             processDPP(gpuDynInst, extData.iFmt_VOP_DPP, src0_dpp, src1);
-        }
 
-        for (int lane = 0; lane < NumVecElemPerVecReg; ++lane) {
-            if (wf->execMask(lane)) {
-                vdst[lane] = std::fma(src0[lane], src1[lane], vdst[lane]);
+            for (int lane = 0; lane < NumVecElemPerVecReg; ++lane) {
+                if (wf->execMask(lane)) {
+                    vdst[lane] = std::fma(src0_dpp[lane], src1[lane],
+                                          vdst[lane]);
+                }
+            }
+        } else {
+            for (int lane = 0; lane < NumVecElemPerVecReg; ++lane) {
+                if (wf->execMask(lane)) {
+                    vdst[lane] = std::fma(src0[lane], src1[lane], vdst[lane]);
+                }
             }
         }
 
@@ -7117,8 +7132,11 @@ namespace Gcn3ISA
         ConstVecOperandU32 src(gpuDynInst, instData.SRC0);
         VecOperandU32 vdst(gpuDynInst, instData.VDST);
 
+        src.readSrc();
+
         if (isDPPInst()) {
-            VecOperandU32 src0_dpp(gpuDynInst, extData.iFmt_VOP_DPP.SRC0);
+            VecOperandU32 src_dpp(gpuDynInst, extData.iFmt_VOP_DPP.SRC0);
+            src_dpp.read();
 
             DPRINTF(GCN3, "Handling V_MOV_B32 SRC DPP. SRC0: register v[%d], "
                     "DPP_CTRL: 0x%#x, SRC0_ABS: %d, SRC0_NEG: %d, "
@@ -7137,14 +7155,18 @@ namespace Gcn3ISA
             // to negate it or take the absolute value of it
             assert(!extData.iFmt_VOP_DPP.SRC1_ABS);
             assert(!extData.iFmt_VOP_DPP.SRC1_NEG);
-            processDPP(gpuDynInst, extData.iFmt_VOP_DPP, src0_dpp);
-        }
-
-        src.readSrc();
+            processDPP(gpuDynInst, extData.iFmt_VOP_DPP, src_dpp);
 
-        for (int lane = 0; lane < NumVecElemPerVecReg; ++lane) {
-            if (wf->execMask(lane)) {
-                vdst[lane] = src[lane];
+            for (int lane = 0; lane < NumVecElemPerVecReg; ++lane) {
+                if (wf->execMask(lane)) {
+                    vdst[lane] = src_dpp[lane];
+                }
+            }
+        } else {
+            for (int lane = 0; lane < NumVecElemPerVecReg; ++lane) {
+                if (wf->execMask(lane)) {
+                    vdst[lane] = src[lane];
+                }
             }
         }