impl gather8/gather16 with gather32
This is our quick path to JIT small gathers.
The idea is roughly,
const uint32_t* ptr32 = ptr8;
uint32_t abcd = ptr32[ix/4];
switch (ix & 3) {
case 3: return (abcd >> 24) ;
case 2: return (abcd >> 16) & 0xff;
case 1: return (abcd >> 8) & 0xff;
case 0: return (abcd ) & 0xff;
}
With the idea that if we may load a given byte,
we should also be allowed to load the four byte
aligned word that byte falls within.
Change-Id: I7fb1085306050c918ccf505f1d2e1e87db3b8c9a
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/268381
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
diff --git a/src/core/SkVM.cpp b/src/core/SkVM.cpp
index 10bd38e..be91326 100644
--- a/src/core/SkVM.cpp
+++ b/src/core/SkVM.cpp
@@ -143,8 +143,6 @@
case Op::load16: write(o, V{id}, "=", op, Arg{immy}); break;
case Op::load32: write(o, V{id}, "=", op, Arg{immy}); break;
- case Op::gather8: write(o, V{id}, "=", op, Arg{immy}, Hex{immz}, V{x}); break;
- case Op::gather16: write(o, V{id}, "=", op, Arg{immy}, Hex{immz}, V{x}); break;
case Op::gather32: write(o, V{id}, "=", op, Arg{immy}, Hex{immz}, V{x}); break;
case Op::uniform8: write(o, V{id}, "=", op, Arg{immy}, Hex{immz}); break;
@@ -268,8 +266,6 @@
case Op::load16: write(o, R{d}, "=", op, Arg{immy}); break;
case Op::load32: write(o, R{d}, "=", op, Arg{immy}); break;
- case Op::gather8: write(o, R{d}, "=", op, Arg{immy}, Hex{immz}, R{x}); break;
- case Op::gather16: write(o, R{d}, "=", op, Arg{immy}, Hex{immz}, R{x}); break;
case Op::gather32: write(o, R{d}, "=", op, Arg{immy}, Hex{immz}, R{x}); break;
case Op::uniform8: write(o, R{d}, "=", op, Arg{immy}, Hex{immz}); break;
@@ -529,10 +525,24 @@
I32 Builder::load32(Arg ptr) { return {this->push(Op::load32, NA,NA,NA, ptr.ix) }; }
I32 Builder::gather8 (Arg ptr, int offset, I32 index) {
- return {this->push(Op::gather8 , index.id,NA,NA, ptr.ix,offset)};
+ // We'll gather32() the right 4 bytes, then select the correct byte.
+ skvm::I32 val = gather32(ptr, offset, shr(index,2)),
+ rem = bit_and(index, splat(3));
+
+ val = bit_or(bit_and(eq(rem, splat(3)), shr(val, 24)),
+ bit_or(bit_and(eq(rem, splat(2)), shr(val, 16)),
+ bit_or(bit_and(eq(rem, splat(1)), shr(val, 8)),
+ bit_and(eq(rem, splat(0)), val ))));
+ return bit_and(val, splat(0xff));
}
I32 Builder::gather16(Arg ptr, int offset, I32 index) {
- return {this->push(Op::gather16, index.id,NA,NA, ptr.ix,offset)};
+ // We'll gather32() the right 4 bytes, then select the correct two.
+ skvm::I32 val = gather32(ptr, offset, shr(index,1)),
+ rem = bit_and(index, splat(1));
+
+ val = bit_or(bit_and(eq(rem, splat(1)), shr(val, 16)),
+ bit_and(eq(rem, splat(0)), val ));
+ return bit_and(val, splat(0xffff));
}
I32 Builder::gather32(Arg ptr, int offset, I32 index) {
return {this->push(Op::gather32, index.id,NA,NA, ptr.ix,offset)};
@@ -1700,32 +1710,12 @@
// - *(const T**)foo loads the gather base and casts it to the right type.
// After all that we have an ordinary (uniform) pointer `ptr` to load from,
// and we then gather from it using the varying indices in r(x).
- STRIDE_1(Op::gather8):
- for (int i = 0; i < K; i++) {
- auto ptr = *(const uint8_t**)((const uint8_t*)arg(immy) + immz);
- r(d).i32[i] = (i==0) ? ptr[ r(x).i32[i] ] : 0;
- } break;
- STRIDE_1(Op::gather16):
- for (int i = 0; i < K; i++) {
- auto ptr = *(const uint16_t**)((const uint8_t*)arg(immy) + immz);
- r(d).i32[i] = (i==0) ? ptr[ r(x).i32[i] ] : 0;
- } break;
STRIDE_1(Op::gather32):
for (int i = 0; i < K; i++) {
auto ptr = *(const int**)((const uint8_t*)arg(immy) + immz);
r(d).i32[i] = (i==0) ? ptr[ r(x).i32[i] ] : 0;
} break;
- STRIDE_K(Op::gather8):
- for (int i = 0; i < K; i++) {
- auto ptr = *(const uint8_t**)((const uint8_t*)arg(immy) + immz);
- r(d).i32[i] = ptr[ r(x).i32[i] ];
- } break;
- STRIDE_K(Op::gather16):
- for (int i = 0; i < K; i++) {
- auto ptr = *(const uint16_t**)((const uint8_t*)arg(immy) + immz);
- r(d).i32[i] = ptr[ r(x).i32[i] ];
- } break;
STRIDE_K(Op::gather32):
for (int i = 0; i < K; i++) {
auto ptr = *(const int**)((const uint8_t*)arg(immy) + immz);
diff --git a/src/core/SkVM.h b/src/core/SkVM.h
index 47cda38..5f733be 100644
--- a/src/core/SkVM.h
+++ b/src/core/SkVM.h
@@ -286,7 +286,7 @@
M(store8) M(store16) M(store32) \
M(index) \
M(load8) M(load16) M(load32) \
- M(gather8) M(gather16) M(gather32) \
+ M(gather32) \
M(uniform8) M(uniform16) M(uniform32) \
M(splat) \
M(add_f32) M(add_i32) M(add_i16x2) \
diff --git a/tests/SkVMTest.cpp b/tests/SkVMTest.cpp
index 2f69e16..08f9540 100644
--- a/tests/SkVMTest.cpp
+++ b/tests/SkVMTest.cpp
@@ -378,7 +378,7 @@
b.store8 (buf8 , b.gather8 (uniforms,0, b.bit_and(x, b.splat(31))));
}
- test_interpreter_only(r, b.done(), [&](const skvm::Program& program) {
+ test_jit_and_interpreter(r, b.done(), [&](const skvm::Program& program) {
const int img[] = {12,34,56,78, 90,98,76,54};
constexpr int N = 20;