Since we determine whether generating SIMD32 based on SIMD16's result, for simplicity, it is better to record whether doing SIMD32 in the context of Stage1 and pass it to Stage2.

Platforms: All

Change-Id: I394ec43c87e9417916020f83ad540bbcfa7843f3
Related-to: SCSP-19136
Resolves:

Author: houjenko

IGC/AdaptorCommon/API/igc.h

@@ -56,13 +56,14 @@ enum {
 };
 
 enum {
-    BIT_CG_SIMD8   = 0b00000001,
-    BIT_CG_SIMD16  = 0b00000010,
-    BIT_CG_SIMD32  = 0b00000100,
-    BIT_CG_SPILL8  = 0b00001000,
-    BIT_CG_SPILL16 = 0b00010000,
-    BIT_CG_SPILL32 = 0b00100000,
-    BIT_CG_RETRY   = 0b01000000,
+    BIT_CG_SIMD8     = 0b00000001,
+    BIT_CG_SIMD16    = 0b00000010,
+    BIT_CG_SIMD32    = 0b00000100,
+    BIT_CG_SPILL8    = 0b00001000,
+    BIT_CG_SPILL16   = 0b00010000,
+    BIT_CG_SPILL32   = 0b00100000,
+    BIT_CG_RETRY     = 0b01000000,
+    BIT_CG_DO_SIMD32 = 0b10000000,
 };
 
 typedef char CG_CTX_STATS_t;
@@ -74,10 +75,12 @@ typedef struct {
 } CG_CTX_t;
 
 #define IsRetry(stats)               (stats & (BIT_CG_RETRY))
+#define DoSimd32(stats)              (stats & (BIT_CG_DO_SIMD32))
 #define HasSimd(MODE, stats)         (stats & (BIT_CG_SIMD##MODE))
 #define HasSimdSpill(MODE, stats)   ((stats & (BIT_CG_SIMD##MODE)) &&  (stats & (BIT_CG_SPILL##MODE)))
 #define HasSimdNoSpill(MODE, stats) ((stats & (BIT_CG_SIMD##MODE)) && !(stats & (BIT_CG_SPILL##MODE)))
 #define SetRetry(stats)              (stats = (stats | (BIT_CG_RETRY)))
+#define SetSimd32(stats)             (stats = (stats | (BIT_CG_DO_SIMD32)))
 #define SetSimdSpill(MODE, stats)    (stats = (stats | (BIT_CG_SIMD##MODE) |  (BIT_CG_SPILL##MODE)))
 #define SetSimdNoSpill(MODE, stats)  (stats = (stats | (BIT_CG_SIMD##MODE) & ~(BIT_CG_SPILL##MODE)))
 
@@ -94,6 +97,8 @@ typedef enum {
 #define IsStage1BestPerf(flag, prev_ctx_ptr)    (!IsStage2RestSIMDs(prev_ctx_ptr) && flag == FLAG_CG_STAGE1_BEST_PERF)
 #define IsAllSIMDs(flag, prev_ctx_ptr)          (!IsStage2RestSIMDs(prev_ctx_ptr) && flag == FLAG_CG_ALL_SIMDS)
 
+#define DoSimd32Stage2(prev_ctx_ptr) (IsStage2RestSIMDs(prev_ctx_ptr) && DoSimd32(prev_ctx_ptr->m_stats))
+
 // We don't need compile continuation if no staged compilation enabled denoted by RegKeys.
 // If the staged compilation enabled, we don't need compile continuation when SIMD8 is spilled.
 #define HasCompileContinuation(flag, prev_ctx_ptr, stats) ( \

IGC/Compiler/CISACodeGen/EmitVISAPass.cpp

@@ -802,6 +802,12 @@ bool EmitPass::runOnFunction(llvm::Function& F)
         }
     }
 
+    if (m_SimdMode == SIMDMode::SIMD16 &&
+        IsStage1BestPerf(m_pCtx->m_CgFlag, m_pCtx->m_StagingCtx))
+    {
+        m_pCtx->m_doSimd32Stage2 = m_currShader->CompileSIMDSize(SIMDMode::SIMD32, *this, F);
+    }
+
     return false;
 }
 

IGC/Compiler/CISACodeGen/PixelShaderCodeGen.cpp

@@ -1140,23 +1140,22 @@ namespace IGC
         }
         if (simdMode == SIMDMode::SIMD32)
         {
+            if (DoSimd32Stage2(ctx->m_StagingCtx))
+            {
+                return true;
+            }
+
             if (forceSIMD32)
             {
                 return true;
             }
 
             CShader* simd16Program = m_parent->GetShader(SIMDMode::SIMD16);
-            // We don't check simd16 if we have it in Stage1
-            // 1. The check of simd16 here is to disable simd32 if simd16 failed to be generated or spilled.
-            // 2. If we have simd16 in stage1, we are sure that simd16 has no spill so we won't need this check.
-            if (ctx->m_StagingCtx && !HasSimd(16, ctx->m_StagingCtx->m_stats))
+            if ((simd16Program == nullptr ||
+                simd16Program->ProgramOutput()->m_programBin == 0 ||
+                simd16Program->ProgramOutput()->m_scratchSpaceUsedBySpills > 0))
             {
-                if ((simd16Program == nullptr ||
-                    simd16Program->ProgramOutput()->m_programBin == 0 ||
-                    simd16Program->ProgramOutput()->m_scratchSpaceUsedBySpills > 0))
-                {
-                    return false;
-                }
+                return false;
             }
 
             const PixelShaderInfo& psInfo = ctx->getModuleMetaData()->psInfo;

IGC/Compiler/CodeGenPublic.h

@@ -768,6 +768,10 @@ namespace IGC
         CG_FLAG_t m_CgFlag = FLAG_CG_ALL_SIMDS;
         // Staging context passing from Stage 1 for compile continuation
         CG_CTX_t* m_StagingCtx = nullptr;
+        // We determine whether generating SIMD32 based on SIMD16's result
+        // For staged compilation, we record if SIMD32 will be generated in Stage1, and
+        // pass it to Stage2.
+        bool m_doSimd32Stage2 = false;
 
     protected:
         // Objects pointed to by these pointers are owned by this class.