Reland "Reland "add op array32 for indirect uniform access""
Forgot to load the pointer to the array in the arm64 case.
This is a reland of fe2506f3cac8edee2fb83b2ba681d8be0d3a20ae
Bug=skia:11822
Original change's description:
> Reland "add op array32 for indirect uniform access"
>
> This is a reland of ac2d053ccfe80775b8144c069bf1f8660a5e8f9a
>
> The original CL was reverted because of a bug in the hash
> function.
>
> Bug=skia:11822
>
> Original change's description:
> > add op array32 for indirect uniform access
> >
> > Change-Id: I6249594a2348c7b24e4f057cce2f4e8a6a2c4409
> > Reviewed-on: https://skia-review.googlesource.com/c/skia/+/431676
> > Reviewed-by: Brian Osman <brianosman@google.com>
> > Commit-Queue: Herb Derby <herb@google.com>
>
> Change-Id: I94604f5589c72d342c39cad44540d810ed7f31a1
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/432797
> Reviewed-by: Brian Osman <brianosman@google.com>
> Commit-Queue: Herb Derby <herb@google.com>
Change-Id: I185fc9554dcb6aa7ce367814ce2c69603074c434
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/433356
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Herb Derby <herb@google.com>
diff --git a/src/core/SkVM.cpp b/src/core/SkVM.cpp
index bc49f16..cf9514b 100644
--- a/src/core/SkVM.cpp
+++ b/src/core/SkVM.cpp
@@ -233,7 +233,8 @@
z = inst.z,
w = inst.w;
int immA = inst.immA,
- immB = inst.immB;
+ immB = inst.immB,
+ immC = inst.immC;
switch (op) {
case Op::assert_true: write(o, op, V{x}, V{y}); break;
@@ -256,6 +257,7 @@
case Op::gather32: write(o, V{id}, "=", op, Ptr{immA}, Hex{immB}, V{x}); break;
case Op::uniform32: write(o, V{id}, "=", op, Ptr{immA}, Hex{immB}); break;
+ case Op::array32: write(o, V{id}, "=", op, Ptr{immA}, Hex{immB}, Hex{immC}); break;
case Op::splat: write(o, V{id}, "=", op, Splat{immA}); break;
@@ -346,7 +348,8 @@
z = inst.z,
w = inst.w;
int immA = inst.immA,
- immB = inst.immB;
+ immB = inst.immB,
+ immC = inst.immC;
switch (op) {
case Op::assert_true: write(o, op, R{x}, R{y}); break;
@@ -369,6 +372,7 @@
case Op::gather32: write(o, R{d}, "=", op, Ptr{immA}, Hex{immB}, R{x}); break;
case Op::uniform32: write(o, R{d}, "=", op, Ptr{immA}, Hex{immB}); break;
+ case Op::array32: write(o, R{d}, "=", op, Ptr{immA}, Hex{immB}, Hex{immC}); break;
case Op::splat: write(o, R{d}, "=", op, Splat{immA}); break;
@@ -467,7 +471,8 @@
std::vector<OptimizedInstruction> optimized(program.size());
for (Val id = 0; id < (Val)program.size(); id++) {
Instruction inst = program[id];
- optimized[id] = {inst.op, inst.x,inst.y,inst.z,inst.w, inst.immA,inst.immB,
+ optimized[id] = {inst.op, inst.x,inst.y,inst.z,inst.w,
+ inst.immA,inst.immB,inst.immC,
/*death=*/id, /*can_hoist=*/true};
}
@@ -540,7 +545,8 @@
&& a.z == b.z
&& a.w == b.w
&& a.immA == b.immA
- && a.immB == b.immB;
+ && a.immB == b.immB
+ && a.immC == b.immC;
}
uint32_t InstructionHash::operator()(const Instruction& inst, uint32_t seed) const {
@@ -618,6 +624,10 @@
return {this, push(Op::uniform32, NA,NA,NA,NA, ptr.ix, offset)};
}
+ I32 Builder::array32 (Ptr ptr, int offset, int index) {
+ return {this, push(Op::array32, NA,NA,NA,NA, ptr.ix, offset, index)};
+ }
+
I32 Builder::splat(int n) { return {this, push(Op::splat, NA,NA,NA,NA, n) }; }
// Be careful peepholing float math! Transformations you might expect to
@@ -2363,7 +2373,6 @@
// Put it all back together, preserving the high 8 bits and low 5.
inst = ((disp << 5) & (19_mask << 5))
| ((inst ) & ~(19_mask << 5));
-
memcpy(fCode + ref, &inst, 4);
}
}
@@ -2481,7 +2490,7 @@
std::vector<llvm::Value*> vals(instructions.size());
auto emit = [&](size_t i, bool scalar, IRBuilder* b) {
- auto [op, x,y,z,w, immA,immB, death,can_hoist] = instructions[i];
+ auto [op, x,y,z,w, immA,immB,immC, death,can_hoist] = instructions[i];
llvm::Type *i1 = llvm::Type::getInt1Ty (*ctx),
*i8 = llvm::Type::getInt8Ty (*ctx),
@@ -2985,6 +2994,7 @@
lookup_register(inst.w),
inst.immA,
inst.immB,
+ inst.immC,
};
fImpl->instructions.push_back(pinst);
};
@@ -3209,7 +3219,8 @@
z = inst.z,
w = inst.w;
const int immA = inst.immA,
- immB = inst.immB;
+ immB = inst.immB,
+ immC = inst.immC;
// alloc_tmp() returns the first of N adjacent temporary registers,
// each freed manually with free_tmp() or noted as our result with mark_tmp_as_dst().
@@ -3613,6 +3624,10 @@
case Op::uniform32: a->vbroadcastss(dst(), A::Mem{arg[immA], immB});
break;
+ case Op::array32: a->mov(GP0, A::Mem{arg[immA], immB});
+ a->vbroadcastss(dst(), A::Mem{GP0, immC});
+ break;
+
case Op::index: a->vmovd((A::Xmm)dst(), N);
a->vbroadcastss(dst(), dst());
a->vpsubd(dst(), dst(), &iota);
@@ -3885,6 +3900,12 @@
a->ld1r4s(dst(), GP0);
break;
+ case Op::array32: a->add(GP0, arg[immA], immB);
+ a->ldrd(GP0, GP0);
+ a->add(GP0, GP0, immC);
+ a->ld1r4s(dst(), GP0);
+ break;
+
case Op::gather8: {
// As usual, the gather base pointer is immB bytes off of uniform immA.
a->add (GP0, arg[immA], immB); // GP0 = &(gather base pointer)
diff --git a/src/core/SkVM.h b/src/core/SkVM.h
index e594a0c..ebe3546 100644
--- a/src/core/SkVM.h
+++ b/src/core/SkVM.h
@@ -439,6 +439,7 @@
M(index) \
M(gather8) M(gather16) M(gather32) \
M(uniform32) \
+ M(array32) \
M(splat) \
M(add_f32) M(add_i32) \
M(sub_f32) M(sub_i32) \
@@ -554,9 +555,9 @@
SK_BEGIN_REQUIRE_DENSE
struct Instruction {
- Op op; // v* = op(x,y,z,w,immA,immB), where * == index of this Instruction.
- Val x,y,z,w; // Enough arguments for Op::store128.
- int immA,immB; // Immediate bit pattern, shift count, pointer index, byte offset, etc.
+ Op op; // v* = op(x,y,z,w,immA,immB), where * == index of this Instruction.
+ Val x,y,z,w; // Enough arguments for Op::store128.
+ int immA,immB,immC; // Immediate bit pattern, shift count, pointer index, byte offset, etc.
};
SK_END_REQUIRE_DENSE
@@ -568,7 +569,7 @@
struct OptimizedInstruction {
Op op;
Val x,y,z,w;
- int immA,immB;
+ int immA,immB,immC;
Val death;
bool can_hoist;
@@ -632,6 +633,9 @@
I32 uniform32(Ptr ptr, int offset);
F32 uniformF (Ptr ptr, int offset) { return pun_to_F32(uniform32(ptr,offset)); }
+ // Load i32/f32 uniform with byte-count offset and index.
+ I32 array32 (Ptr ptr, int offset, int index);
+
// Push and load this color as a uniform.
Color uniformColor(SkColor4f, Uniforms*);
@@ -936,8 +940,9 @@
}
private:
- Val push(Op op, Val x=NA, Val y=NA, Val z=NA, Val w=NA, int immA=0, int immB=0) {
- return this->push(Instruction{op, x,y,z,w, immA,immB});
+ Val push(
+ Op op, Val x=NA, Val y=NA, Val z=NA, Val w=NA, int immA=0, int immB=0, int immC=0) {
+ return this->push(Instruction{op, x,y,z,w, immA,immB,immC});
}
template <typename T>
@@ -963,7 +968,7 @@
struct InterpreterInstruction {
Op op;
Reg d,x,y,z,w;
- int immA,immB;
+ int immA,immB,immC;
};
class Program {
diff --git a/src/opts/SkVM_opts.h b/src/opts/SkVM_opts.h
index 3f9132a..cb92fdb 100644
--- a/src/opts/SkVM_opts.h
+++ b/src/opts/SkVM_opts.h
@@ -93,7 +93,8 @@
z = inst.z,
w = inst.w;
int immA = inst.immA,
- immB = inst.immB;
+ immB = inst.immB,
+ immC = inst.immC;
// Ops that interact with memory need to know whether we're stride=1 or K,
// but all non-memory ops can run the same code no matter the stride.
@@ -216,6 +217,12 @@
r[d].i32 = *(const int*)( (const char*)args[immA] + immB );
break;
+ CASE(Op::array32):
+ const int* ptr;
+ memcpy(&ptr, (const uint8_t*)args[immA] + immB, sizeof(ptr));
+ r[d].i32 = ptr[immC/sizeof(int)];
+ break;
+
CASE(Op::splat): r[d].i32 = immA; break;
CASE(Op::add_f32): r[d].f32 = r[x].f32 + r[y].f32; break;
diff --git a/tests/SkVMTest.cpp b/tests/SkVMTest.cpp
index 3d7ef42..0432fc4 100644
--- a/tests/SkVMTest.cpp
+++ b/tests/SkVMTest.cpp
@@ -769,6 +769,46 @@
});
}
+DEF_TEST(SKVM_array32, r) {
+ skvm::Builder b;
+ {
+ skvm::Ptr buf0 = b.varying<int32_t>(),
+ buf1 = b.varying<int32_t>(),
+ uniforms = b.uniform();
+
+ skvm::I32 x = b.array32(uniforms, 0, 0);
+ b.store32(buf0, x);
+ skvm::I32 y = b.array32(uniforms, 0, 4);
+ b.store32(buf1, y);
+ }
+
+ test_jit_and_interpreter(b, [&](const skvm::Program& program) {
+ const int K = 20;
+ int i[2] = {3, 7};
+ struct {
+ int* g;
+ } uniforms{i};
+ int32_t buf0[K];
+ int32_t buf1[K];
+
+ program.eval(K, buf0, buf1, &uniforms);
+ for (auto v : buf0) {
+ REPORTER_ASSERT(r, v == 3);
+ }
+ for (auto v : buf1) {
+ REPORTER_ASSERT(r, v == 7);
+ }
+ i[0] = 4;
+ program.eval(K, buf0, buf1, &uniforms);
+ for (auto v : buf0) {
+ REPORTER_ASSERT(r, v == 4);
+ }
+ for (auto v : buf1) {
+ REPORTER_ASSERT(r, v == 7);
+ }
+ });
+}
+
DEF_TEST(SkVM_sqrt, r) {
skvm::Builder b;
auto buf = b.varying<int>();