Implement Packed Dot Product intrinsics (#6068)

* implement dot acc intrinsics

* fix sm version

* fix test

* improve comment

---------

Co-authored-by: Yong He <yonghe@outlook.com>

2 files changed, 59 insertions, 0 deletions

diff --git a/tests/hlsl-intrinsic/dot-accumulate.slang b/tests/hlsl-intrinsic/dot-accumulate.slang
new file mode 100644
index 000000000..113ae40e3
--- /dev/null
+++ b/tests/hlsl-intrinsic/dot-accumulate.slang
@@ -0,0 +1,55 @@
+//TEST(compute, vulkan):COMPARE_COMPUTE_EX:-vk -compute -shaderobj -output-using-type
+// Does not run on DX11 as SM 6.4 is required.
+//DISABLE_TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -dx11
+//TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -dx12 -profile cs_6_4 -use-dxil -shaderobj -output-using-type
+//TEST(compute):COMPARE_COMPUTE_EX:-metal -compute -shaderobj -output-using-type
+//TEST(compute):COMPARE_COMPUTE_EX:-wgsl -compute -shaderobj -output-using-type
+//TEST(compute):COMPARE_COMPUTE_EX:-cuda -compute -shaderobj -g0 -output-using-type
+//TEST(compute):COMPARE_COMPUTE_EX:-cpu -compute -shaderobj -output-using-type
+
+//TEST_INPUT:ubuffer(data=[0 0 0], stride=4):out,name outputBuffer
+RWStructuredBuffer<int> outputBuffer;
+
+[numthreads(1, 1, 1)]
+void computeMain(uint3 dispatchThreadID : SV_DispatchThreadID)
+{
+    uint outputIndex = 0;
+
+    //
+    // dot4add_u8packed()
+    // [4 3 2 1]  dot [1 2 4 2] + 5
+    // (4 * 1) + (3 * 2) + (2 * 4) + (1 * 2) + 5 = 25
+    //
+    uint unsignedX = 0x01020304U;
+    uint unsignedY = 0x02040201U;
+    uint unsignedAcc = 5U; 
+    uint unsignedResult = dot4add_u8packed(unsignedX, unsignedY, unsignedAcc);
+    outputBuffer[outputIndex++] = unsignedResult;
+
+    //
+    // dot4add_i8packed()
+    // [6 2 3 -1] dot [-2 -6 2 6] - 100
+    // (6 * -2) + (2 * -6) + (3 * 2) + (-1 * 6) - 100 = -124
+    //
+    int signedX = 0xFF030206;
+    int signedY = 0x0602FAFE;
+    int signedAcc = -100;
+    int signedResult = dot4add_i8packed(signedX, signedY, signedAcc);
+    outputBuffer[outputIndex++] = signedResult;
+
+    //
+    // dot2add()
+    // [10.8 -3.3] dot [1.4 -20.3] - 2.11 
+    // (10.8 * 1.4) + (-3.3 * -20.3) - 2.0 = 80.11
+    //
+    half2 half2X = half2(half(10.8), half(-3.3));
+    half2 half2Y = half2(half(1.4), half(-20.3));
+
+    // `half2Acc` is assigned -2.0 here.
+    // Thread index is used so that `half2Acc` will not be implicitly emitted as literal `-2.0` which
+    // may be treated as a double by DXC and cause it to fail to compile because no overload exists for `dot2add` that
+    // accepts double.
+    float half2Acc = float(dispatchThreadID.x + 1) * -2.0f;
+    float half2Result = dot2add(half2X, half2Y, half2Acc);
+    outputBuffer[outputIndex++] = int(half2Result);
+}
diff --git a/tests/hlsl-intrinsic/dot-accumulate.slang.expected.txt b/tests/hlsl-intrinsic/dot-accumulate.slang.expected.txt
new file mode 100644
index 000000000..184864973
--- /dev/null
+++ b/tests/hlsl-intrinsic/dot-accumulate.slang.expected.txt
@@ -0,0 +1,4 @@
+type: int32_t
+25
+-124
+80