gpu: major update to arrayn variable creation: actual array size is determined when value buffer is set, and also added ReadOnly flag, so it auto-creates read buffer for all !ReadOnly Storage variables.

Author: rcoreilly

goal/gosl/examples/basic/compute.go

@@ -12,6 +12,7 @@ import "cogentcore.org/core/math32"
 //gosl:vars
 var (
 	// Params are the parameters for the computation.
+	//gosl:read-only
 	Params []ParamStruct
 
 	// Data is the data on which the computation operates.

goal/gosl/examples/basic/gosl.go

@@ -12,6 +12,7 @@ import (
 	"runtime"
 
 	"cogentcore.org/core/base/timer"
+	"cogentcore.org/core/gpu"
 )
 
 //go:generate gosl .
@@ -22,6 +23,9 @@ func init() {
 }
 
 func main() {
+	gpu.Debug = true
+	GPUInit()
+
 	n := 2000000 // note: not necc to spec up-front, but easier if so
 
 	Params = make([]ParamStruct, 1)
@@ -37,8 +41,6 @@ func main() {
 		sd[i].Raw = Data[i].Raw
 	}
 
-	GPUInit()
-
 	cpuTmr := timer.Time{}
 	cpuTmr.Start()
 

goal/gosl/examples/basic/main.go

@@ -70,8 +70,6 @@ func main() {
 	gpu.SetValueFrom(cvl, []uint64{seed})
 	gpu.SetValueFrom(dvl, dataG)
 
-	sgp.CreateReadBuffers()
-
 	gpuTmr := timer.Time{}
 	gpuTmr.Start()
 

goal/gosl/examples/rand/main.go

@@ -42,7 +42,7 @@ import (
 //go:embed %s/*.wgsl
 var shaders embed.FS
 
-// GPU is the compute gpu device
+// ComputeGPU is the compute gpu device
 var ComputeGPU *gpu.GPU
 
 // UseGPU indicates whether to use GPU vs. CPU.
@@ -53,6 +53,7 @@ var UseGPU bool
 	b.WriteString(fmt.Sprintf(header, st.Config.Output))
 
 	for _, sy := range st.Systems {
+		b.WriteString(fmt.Sprintf("// %s is a GPU compute System with kernels operating on the\nsame set of data variables.\n", st.genSysVar(sy)))
 		b.WriteString(fmt.Sprintf("var %s *gpu.ComputeSystem\n", st.genSysVar(sy)))
 	}
 
@@ -77,8 +78,8 @@ const (
 	b.WriteString(")\n")
 
 	initf := `
-// GPUInit initializes the GPU compute system
-// Configuring Systems, variables and kernels.
+// GPUInit initializes the GPU compute system,
+// configuring system(s), variables and kernels.
 func GPUInit() {
 	gp := gpu.NewComputeGPU()
 	ComputeGPU = gp
@@ -92,7 +93,8 @@ func GPUInit() {
 	b.WriteString("}\n\n")
 
 	release := `
-// GPURelease releases the GPU compute system.
+// GPURelease releases the GPU compute system resources.
+// Call this at program exit.
 func GPURelease() {
 `
 
@@ -138,9 +140,6 @@ func (st *State) GenGPUSystemInit(sy *System) string {
 			b.WriteString(fmt.Sprintf("\t\t\tsgp.AddStruct(%q, int(unsafe.Sizeof(%s{})), len(%s), gpu.ComputeShader)\n", vr.Name, vr.Type[2:], vr.Name))
 		}
 		b.WriteString("\t\t\tsgp.SetNValues(1)\n")
-		if !gp.Uniform {
-			b.WriteString("\t\t\tsgp.CreateReadBuffers()\n")
-		}
 		b.WriteString("\t\t}\n")
 	}
 	b.WriteString("\t\tsy.Config()\n")
@@ -157,9 +156,9 @@ func (st *State) GenGPUSystemOps(sy *System) string {
 
 	// 1 = kernel, 2 = system, 3 = sysname
 	run := `
-// Run%[1]s runs the %[1]s kernel with given number of items,
-// on either the CPU or GPU depending on the UseGPU.
-// Pass *Var variable names to sync those variables back from the GPU
+// Run%[1]s runs the %[1]s kernel with given number of elements,
+// on either the CPU or GPU depending on the UseGPU variable.
+// Pass *Var variable enums to sync those variables back from the GPU
 // after running (irrelevant for CPU).
 func Run%[1]s(n int, syncVars ...GPUVars) {
 	if UseGPU {

goal/gosl/gotosl/gengpu.go

@@ -54,6 +54,11 @@ type Var struct {
 
 	// Type of variable: either []Type or tensor.Float32, tensor.Int32
 	Type string
+
+	// ReadOnly indicates that this variable is never read back from GPU,
+	// specified by the gosl:read-only property in the variable comments.
+	// It is important to optimize GPU memory usage to indicate this.
+	ReadOnly bool
 }
 
 // Group represents one variable group.

goal/gosl/gotosl/gotosl.go

@@ -59,8 +59,6 @@ func main() {
 	}
 	gpu.SetValueFrom(dvl, sd)
 
-	sgp.CreateReadBuffers()
-
 	ce, _ := sy.BeginComputePass()
 	pl.Dispatch1D(ce, n, threads)
 	ce.End()

gpu/examples/compute/compute.go

@@ -30,10 +30,17 @@ type Value struct {
 	// index of this value within the Var list of values
 	Index int
 
-	// VarSize is the size of each Var element, which includes any fixed ArrayN
+	// VarSize is the size of each Var element, which includes any fixed Var.ArrayN
 	// array size specified on the Var.
+	// The actual buffer size is VarSize * Value.ArrayN (or DynamicN for dynamic).
 	VarSize int
 
+	// ArrayN is the actual number of array elements, for Uniform or Storage
+	// variables without a fixed array size (i.e., the Var ArrayN = 1).
+	// This is set when the buffer is actually created, based on the data,
+	// or can be set directly prior to buffer creation.
+	ArrayN int
+
 	// DynamicIndex is the current index into a DynamicOffset variable
 	// to use for the SetBindGroup call.  Note that this is an index,
 	// not an offset, so it indexes the DynamicN Vars in the Value,
@@ -112,6 +119,7 @@ func (vl *Value) init(vr *Var, dev *Device, idx int) {
 	vl.Index = idx
 	vl.Name = fmt.Sprintf("%s_%d", vr.Name, vl.Index)
 	vl.VarSize = vr.MemSize()
+	vl.ArrayN = 1
 	vl.alignBytes = vr.alignBytes
 	vl.AlignVarSize = MemSizeAlign(vl.VarSize, vl.alignBytes)
 	vl.isDynamic = vl.role == Vertex || vl.role == Index || vr.DynamicOffset
@@ -129,11 +137,12 @@ func (vl *Value) MemSize() int {
 	if vl.isDynamic {
 		return vl.AlignVarSize * vl.dynamicN
 	}
-	return vl.VarSize
+	return vl.ArrayN * vl.VarSize
 }
 
 // CreateBuffer creates the GPU buffer for this value if it does not
 // yet exist or is not the right size.
+// For !ReadOnly [Storage] buffers, calls [Value.CreateReadBuffer].
 func (vl *Value) CreateBuffer() error {
 	if vl.role == SampledTexture {
 		return nil
@@ -159,6 +168,9 @@ func (vl *Value) CreateBuffer() error {
 	}
 	vl.AllocSize = sz
 	vl.buffer = buf
+	if vl.role == Storage && !vl.vvar.ReadOnly {
+		vl.CreateReadBuffer()
+	}
 	return nil
 }
 
@@ -214,6 +226,9 @@ func (vl *Value) SetDynamicN(n int) {
 
 // SetValueFrom copies given values into value buffer memory,
 // making the buffer if it has not yet been constructed.
+// The actual ArrayN size of Storage or Uniform variables will
+// be computed based on the size of the from bytes, relative to
+// the variable size.
 // IMPORTANT: do not use this for dynamic offset Uniform or
 // Storage variables, as the alignment will not be correct;
 // See [SetDynamicFromBytes].
@@ -223,6 +238,7 @@ func SetValueFrom[E any](vl *Value, from []E) error {
 
 // SetFromBytes copies given bytes into value buffer memory,
 // making the buffer if it has not yet been constructed.
+// For !ReadOnly [Storage] buffers, calls [Value.CreateReadBuffer].
 // IMPORTANT: do not use this for dynamic offset Uniform or
 // Storage variables, as the alignment will not be correct;
 // See [SetDynamicFromBytes].
@@ -232,12 +248,19 @@ func (vl *Value) SetFromBytes(from []byte) error {
 		return errors.Log(err)
 	}
 	nb := len(from)
+	an := nb / vl.VarSize
+	aover := nb % vl.VarSize
+	if aover != 0 {
+		err := fmt.Errorf("gpu.Value SetFromBytes %s, Size passed: %d is not an even multiple of the variable size: %d", vl.Name, nb, vl.VarSize)
+		return errors.Log(err)
+	}
 	if vl.isDynamic { // Vertex, Index at this point
-		dn := nb / vl.VarSize
-		vl.SetDynamicN(dn)
+		vl.SetDynamicN(an)
+	} else {
+		vl.ArrayN = an
 	}
 	tb := vl.MemSize()
-	if nb != tb {
+	if nb != tb { // this should never happen, but justin case
 		err := fmt.Errorf("gpu.Value SetFromBytes %s, Size passed: %d != Size expected %d", vl.Name, nb, tb)
 		return errors.Log(err)
 	}
@@ -254,6 +277,9 @@ func (vl *Value) SetFromBytes(from []byte) error {
 		}
 		vl.buffer = buf
 		vl.AllocSize = nb
+		if vl.role == Storage && !vl.vvar.ReadOnly {
+			vl.CreateReadBuffer()
+		}
 	} else {
 		err := vl.device.Queue.WriteBuffer(vl.buffer, 0, from)
 		if errors.Log(err) != nil {
@@ -406,9 +432,8 @@ func (vl *Value) SetFromTexture(tx *Texture) *Texture {
 }
 
 // CreateReadBuffer creates a read buffer for this value,
-// if it does not yet exist or is not the right size.
+// for [Storage] values only. Automatically called for !ReadOnly.
 // Read buffer is needed for reading values back from the GPU.
-// Only for Storage role variables.
 func (vl *Value) CreateReadBuffer() error {
 	if !(vl.role == Storage || vl.role == StorageTexture) {
 		return nil

gpu/value.go

@@ -147,18 +147,6 @@ func (vs *Values) MemSize() int {
 	return tsz
 }
 
-// CreateReadBuffers creates read buffers for all values.
-func (vs *Values) CreateReadBuffers() error {
-	var errs []error
-	for _, vl := range vs.Values {
-		err := vl.CreateReadBuffer()
-		if err != nil {
-			errs = append(errs, err)
-		}
-	}
-	return errors.Join(errs...)
-}
-
 // bindGroupEntry returns the BindGroupEntry for Current
 // value for this variable.
 func (vs *Values) bindGroupEntry(vr *Var) []wgpu.BindGroupEntry {