todo: use a gemm pool to deal with the big matrix mul

Author: CodingPlatelets

.gitmodules

@@ -7,3 +7,6 @@
 [submodule "dependencies/fudian"]
 	path = dependencies/fudian
 	url = git@github.com:CodingPlatelets/fudian.git
+[submodule "dependencies/fputil-nopipe"]
+	path = dependencies/fputil-nopipe
+	url = git@github.com:CodingPlatelets/fp-division-no-pipeline.git

build.sbt

@@ -24,6 +24,7 @@ lazy val commonChiselSettings = Seq(
 
 lazy val root = (project in file("."))
   .dependsOn(fputil)
+  .dependsOn(fputilNopipe)
   .dependsOn(hardfloat)
   .settings(
     name := "transformer_MM",
@@ -41,7 +42,15 @@ lazy val fputil = Project("fputil", file("dependencies/fputil/src"))
     Compile / scalaSource := baseDirectory.value / "main" / "scala",
     Compile / resourceDirectory := baseDirectory.value / "main" / "resources"
   )
-
+lazy val fputilNopipe = Project("fputilNopipe", file("dependencies/fputil-nopipe/src"))
+  .settings(
+    name := "fputilNopipe",
+    commonChiselSettings
+  )
+  .settings(
+    Compile / scalaSource := baseDirectory.value / "main" / "scala",
+    Compile / resourceDirectory := baseDirectory.value / "main" / "resources"
+  )
 lazy val hardfloat = Project("hardfloat", file("dependencies/hardfloat/hardfloat/src"))
   .settings(
     name := "hardfloat",

dependencies/fputil-nopipe

@@ -373,8 +373,8 @@ class QKMul(
     val resetBuffer = Input(Bool())
   })
 
-  val qGen = new QKGenerationMatrixMulWarper( gemmType, bufferSizeGemm)
-  val kGen = new QKGenerationMatrixMulWarper( gemmType, bufferSizeGemm)
+  val qGen = new QKGenerationMatrixMulWarper(gemmType, bufferSizeGemm)
+  val kGen = new QKGenerationMatrixMulWarper(gemmType, bufferSizeGemm)
 
   qGen.io.in_a <> io.inputToken
   qGen.io.in_b <> io.weightQ
@@ -399,6 +399,8 @@ class QKMul(
 
 }
 
+// a pool of gemm pe for just one line of input
+// try Mux1H or PriorityMux?
 class GemmPool(
   val n:        Int,
   val poolSize: Int,
@@ -409,7 +411,47 @@ class GemmPool(
     with llamaConfig
     with DebugLog {
   val io = IO(new Bundle {
-    val in = Flipped(Decoupled(new currentSystolicGroupIdx))
+    // implicit a context here
+    val in = Flipped(Decoupled(new Bundle {
+      val a = new currentSystolicGroupIdx
+      val b = new currentSystolicGroupIdx
+    }))
+    val reset = Input(new Bundle {
+      // can reset each gemm unit, 0 is for the whole pool
+      val id = UInt(log2Ceil(poolSize).W)
+      val reset = Bool()
+    })
+    val out = Decoupled(new currentSystolicGroupIdx)
   })
 
+  val gemmUnits = VecInit.fill(poolSize)(Module(new GEMM(n, gemmType)).io)
+
+  val outBuffer = Module(
+    new Queue(
+      new currentSystolicGroupIdx,
+      entries = poolSize,
+      pipe = true,
+      flow = false,
+      useSyncReadMem = false,
+      hasFlush = true
+    )
+  )
+  outBuffer.io.flush.get := Mux(io.reset.id === 0.U, io.reset.reset, false.B)
+  outBuffer.io.deq <> io.out
+
+  // placement reg, point to the gemm unit which is free, start from 0
+  val placementReg = RegInit(0.U(log2Ceil(poolSize).W))
+
+  // if every gemm is full, then ready is false
+  val readyGemmUnit =
+    VecInit(gemmUnits.map(gemmUnit => gemmUnit.in_a.ready & gemmUnit.in_b.ready)).reduceTree(_ & _)
+  io.in.ready := readyGemmUnit
+
+
+  // require the a row equals to b col
+  val context = io.in.bits.a.row
+  when(io.in.valid) {
+
+  }
+
 }

src/main/scala/models/llama3/metrixController.scala

@@ -326,13 +326,13 @@ class GEMMTest extends AnyFlatSpec with ChiselScalatestTester with ParallelTestE
         // println(f"acc: $acc%.4f")
         dut.clock.step()
       }
-      // dut.io.in_h.poke(0.U)
-      // dut.io.in_v.poke(0.U)
+      dut.io.in_h.poke(0.U)
+      dut.io.in_v.poke(0.U)
       dut.clock.step(3)
     }.fork {
-      dut.clock.step(4)
+      dut.clock.step(8)
       val out = java.lang.Float.intBitsToFloat(dut.io.out.peekInt().toInt)
-      assert(math.abs(out - result) < precision)
+      // assert(math.abs(out - result) / result < precision)
       println(f"out: ${out}%.4f\t, result: $result%.4f")
       dut.clock.step(1)
     }.join()
@@ -342,9 +342,10 @@ class GEMMTest extends AnyFlatSpec with ChiselScalatestTester with ParallelTestE
   //   test(new PEFxp()).withAnnotations(Seq(VerilatorBackendAnnotation, WriteVcdAnnotation))(testPEFxp)
   // }
 
-  // "PEFp basic test on Verilator" should "pass" in {
-  //   test(new PEFp(32, 4)).withAnnotations(Seq(VerilatorBackendAnnotation, WriteVcdAnnotation))(testPEFp)
-  // }
+  "PEFp basic test on Verilator" should "pass" in {
+    implicit val fxpConfig: DataWidthConfig = Fp32Config
+    test(new PEFp()).withAnnotations(Seq(VerilatorBackendAnnotation))(testPEFp)
+  }
 
   // "SystolicMM basic test on Verilator" should "pass" in {
   //   implicit val fxpConfig: DataWidthConfig = FxpConfig
@@ -364,9 +365,9 @@ class GEMMTest extends AnyFlatSpec with ChiselScalatestTester with ParallelTestE
   //     .withAnnotations(Seq(VerilatorBackendAnnotation, WriteVcdAnnotation))(testGEMM)
   // }
 
-  "GeMMFp basic test on Verilator" should "pass" in {
-    implicit val fxpConfig: DataWidthConfig = Fp32Config
-    test(new GEMM(6, GEMMDataType.Fp32))
-      .withAnnotations(Seq(VerilatorBackendAnnotation, WriteVcdAnnotation))(testGEMMFp)
-  }
+  // "GeMMFp basic test on Verilator" should "pass" in {
+  //   implicit val fxpConfig: DataWidthConfig = Fp32Config
+  //   test(new GEMM(6, GEMMDataType.Fp32))
+  //     .withAnnotations(Seq(VerilatorBackendAnnotation, WriteVcdAnnotation))(testGEMMFp)
+  // }
 }