config: reduce Radiance shared memory to 32 KiB

generators/chipyard/src/main/scala/config/RadianceConfigs.scala

@@ -25,6 +25,8 @@ class WithRadBootROM(address: BigInt = 0x10000, size: Int = 0x10000, hang: BigIn
 // aliases for virgo
 class VirgoConfig extends RadianceClusterConfig
 class VirgoFP16Config extends RadianceFP16ClusterConfig
+class VirgoVoltaConfig extends RadianceFP16ClusterConfig
+class VirgoAmpereConfig extends RadianceFP16ClusterConfig
 class VirgoHopperConfig extends Radiance4CFP16ClusterConfig
 class VirgoBlackwellConfig extends RadianceBlackwellClusterConfig
 class VirgoFlashConfig extends RadianceClusterConfig
@@ -35,7 +37,7 @@ class VirgoHopperSynConfig extends Radiance4CFP16ClusterSynConfig
 class RadianceBaseConfig extends Config(
   // NOTE: when changing these, remember to change NUM_CORES/THREADS/WARPS in
   // the verilog source as well!
-  new radiance.subsystem.WithSimtConfig(nWarps = 8, nCoreLanes = 8, nMemLanes = 8, nSrcIds = 32) ++
+  new radiance.subsystem.WithSimtConfig(nWarps = 4, nCoreLanes = 4, nMemLanes = 4, nSrcIds = 32) ++
   new chipyard.config.WithSystemBusWidth(bitWidth = 256) ++
   new WithExtMemSize(BigInt("80000000", 16)) ++
   new WithRadBootROM() ++
@@ -56,8 +58,8 @@ class RadianceBaseConfig extends Config(
 
 class RadianceFP16ClusterConfig extends Config(
   new radiance.subsystem.WithRadianceGemmini(location = InCluster(0), dim = 16, accSizeInKB = 32, tileSize = (8, 4, 8), dataType = RadianceGemminiDataType.FP16) ++
-  new radiance.subsystem.WithRadianceCores(8, location = InCluster(0), tensorCoreFP16 = true, tensorCoreDecoupled = false, useVxCache = false) ++
+  new radiance.subsystem.WithRadianceCores(1, location = InCluster(0), tensorCoreFP16 = true, tensorCoreDecoupled = false, useVxCache = false) ++
-  new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 128 << 10, numBanks = 4, numWords = 16) ++
+  new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 32 << 10, numBanks = 4, numWords = 16) ++
   new radiance.subsystem.WithCoalescer(nNewSrcIds = 16) ++
   new radiance.subsystem.WithVortexL1Banks(nBanks = 8) ++
   new radiance.subsystem.WithRadianceCluster(0) ++
@@ -66,7 +68,7 @@ class RadianceFP16ClusterConfig extends Config(
 class Radiance8B8WFP16ClusterConfig extends Config(
   new radiance.subsystem.WithRadianceGemmini(location = InCluster(0), dim = 16, accSizeInKB = 32, tileSize = (8, 4, 8), dataType = RadianceGemminiDataType.FP16) ++
   new radiance.subsystem.WithRadianceCores(8, location = InCluster(0), tensorCoreFP16 = true, tensorCoreDecoupled = false, useVxCache = false) ++
-  new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 128 << 10, numBanks = 8, numWords = 8) ++
+  new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 32 << 10, numBanks = 8, numWords = 8) ++
   new radiance.subsystem.WithCoalescer(nNewSrcIds = 16) ++
   new radiance.subsystem.WithVortexL1Banks(nBanks = 8) ++
   new radiance.subsystem.WithRadianceCluster(0) ++
@@ -74,13 +76,13 @@ class Radiance8B8WFP16ClusterConfig extends Config(
 
 class Radiance4CFP16ClusterConfig extends Config(
   new radiance.subsystem.WithRadianceGemmini(location = InCluster(0), dim = 16, accSizeInKB = 32, tileSize = (8, 4, 8), dataType = RadianceGemminiDataType.FP16) ++
-  new radiance.subsystem.WithRadianceCores(4, location = InCluster(0), tensorCoreFP16 = true, tensorCoreDecoupled = true, useVxCache = false) ++
+  new radiance.subsystem.WithRadianceCores(1, location = InCluster(0), tensorCoreFP16 = true, tensorCoreDecoupled = true, useVxCache = false) ++
-  // new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 128 << 10, numBanks = 4, numWords = 16,
+  // new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 32 << 10, numBanks = 4, numWords = 16,
   //                                              memType = radiance.subsystem.TwoReadOneWrite,
   //                                              serializeUnaligned = radiance.subsystem.CoreSerialized) ++
   // NOTE: Hopper Tensor Core does not work with 16-word config due to the
   // address alignment requirement
-  new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 128 << 10, numBanks = 4, numWords = 8) ++
+  new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 32 << 10, numBanks = 4, numWords = 8) ++
   new radiance.subsystem.WithCoalescer(nNewSrcIds = 16) ++
   new radiance.subsystem.WithVortexL1Banks(nBanks = 8) ++
   new radiance.subsystem.WithRadianceCluster(0) ++
@@ -88,8 +90,8 @@ class Radiance4CFP16ClusterConfig extends Config(
 
 class RadianceBlackwellClusterConfig extends Config(
   new radiance.subsystem.WithRadianceGemmini(location = InCluster(0), dim = 16, accSizeInKB = 32, tileSize = (8, 4, 8), dataType = RadianceGemminiDataType.FP16) ++
-  new radiance.subsystem.WithRadianceCores(4, location = InCluster(0), tensorCoreFP16 = true, tensorCoreDecoupled = false, tensorCoreBlackwell = true, startupAddress = BigInt("80000000", 16), useVxCache = false) ++
+  new radiance.subsystem.WithRadianceCores(1, location = InCluster(0), tensorCoreFP16 = true, tensorCoreDecoupled = false, tensorCoreBlackwell = true, startupAddress = BigInt("80000000", 16), useVxCache = false) ++
-  new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 128 << 10, numBanks = 4, numWords = 8) ++
+  new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 32 << 10, numBanks = 4, numWords = 8) ++
   new radiance.subsystem.WithCoalescer(nNewSrcIds = 16) ++
   new radiance.subsystem.WithVortexL1Banks(nBanks = 8) ++
   new radiance.subsystem.WithRadianceCluster(0) ++
@@ -98,9 +100,9 @@ class RadianceBlackwellClusterConfig extends Config(
 class RadianceClusterConfig extends Config(
   // important to keep gemmini tile before RadianceCores to ensure radiance tile id is 0-indexed
   new radiance.subsystem.WithRadianceGemmini(location = InCluster(0), dim = 8, accSizeInKB = 16, tileSize = 8) ++
-  new radiance.subsystem.WithRadianceCores(4, location = InCluster(0), tensorCoreFP16 = false, tensorCoreDecoupled = false, useVxCache = false) ++
+  new radiance.subsystem.WithRadianceCores(1, location = InCluster(0), tensorCoreFP16 = false, tensorCoreDecoupled = false, useVxCache = false) ++
   // new radiance.subsystem.WithRadianceFrameBuffer(x"ff018000", 16, 0x8000, x"ff011000", "fb0") ++
-  new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 256 << 10/*KBytes*/, numBanks = 8, numWords = 8,
+  new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 32 << 10/*KBytes*/, numBanks = 8, numWords = 8,
                                                // memType = radiance.subsystem.TwoReadOneWrite,
                                                serializeUnaligned = radiance.subsystem.CoreSerialized) ++
   new radiance.subsystem.WithCoalescer(nNewSrcIds = 16) ++
@@ -133,7 +135,7 @@ class RadianceBigLittleClusterConfig extends Config(
   new radiance.subsystem.WithRadianceGemmini(location = InCluster(0), dim = 4, accSizeInKB = 16, tileSize = 16) ++
   new radiance.subsystem.WithRadianceGemmini(location = InCluster(0), dim = 8, accSizeInKB = 16, tileSize = 8) ++
   new radiance.subsystem.WithRadianceCores(2, location = InCluster(0), useVxCache = false) ++
-  new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 64 << 10, numBanks = 4, numWords = 8) ++
+  new radiance.subsystem.WithRadianceSharedMem(address = x"ff000000", size = 32 << 10, numBanks = 4, numWords = 8) ++
   new radiance.subsystem.WithCoalescer(nNewSrcIds = 16) ++
   new radiance.subsystem.WithVortexL1Banks(nBanks = 8)++
   new radiance.subsystem.WithRadianceCluster(0) ++