Optimize dot4{I, U}8Packed for all spv versions

Emit optimized code for `dot4{I, U}8Packed` regardless of SPIR-V version
as long as the required capabilities are available. On SPIR-V < 1.6,
require the extension "SPV_KHR_integer_dot_product" for this. On
SPIR-V >= 1.6, don't require the extension because the corresponding
capabilities are part of SPIR-V >= 1.6 proper.

Author: robamler

naga/src/back/spv/block.rs

@@ -1143,17 +1143,21 @@ impl BlockContext<'_> {
                         ),
                     },
                     fun @ (Mf::Dot4I8Packed | Mf::Dot4U8Packed) => {
-                        if self.writer.lang_version() >= (1, 6)
-                            && self
-                                .writer
-                                .require_all(&[
-                                    spirv::Capability::DotProduct,
-                                    spirv::Capability::DotProductInput4x8BitPacked,
-                                ])
-                                .is_ok()
+                        if self
+                            .writer
+                            .require_all(&[
+                                spirv::Capability::DotProduct,
+                                spirv::Capability::DotProductInput4x8BitPacked,
+                            ])
+                            .is_ok()
                         {
                             // Write optimized code using `PackedVectorFormat4x8Bit`.
-                            self.writer.use_extension("SPV_KHR_integer_dot_product");
+                            if self.writer.lang_version() < (1, 6) {
+                                // SPIR-V 1.6 supports the required capabilities natively, so the extension
+                                // is only required for earlier versions. See right column of
+                                // <https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#OpSDot>.
+                                self.writer.use_extension("SPV_KHR_integer_dot_product");
+                            }
 
                             let op = match fun {
                                 Mf::Dot4I8Packed => spirv::Op::SDot,

naga/tests/in/wgsl/functions-optimized-by-capability.toml

@@ -0,0 +1,9 @@
+# Turn on optimizations for `dot4I8Packed` and `dot4U8Packed` by enabling the
+# required capabilities on a SPIR-V version where these capabilities are only
+# available via the extension "SPV_KHR_integer_dot_product".
+
+targets = "SPIRV"
+
+[spv]
+capabilities = ["DotProduct", "DotProductInput4x8BitPacked"]
+version = [1, 0]

naga/tests/in/wgsl/functions-optimized.wgsl renamed to naga/tests/in/wgsl/functions-optimized-by-capability.wgsl

@@ -0,0 +1,12 @@
+# Turn on optimizations for `dot4I8Packed` and `dot4U8Packed` on SPIR-V and HLSL by
+# using a version of SPIR-V / shader model that supports these without any extensions.
+
+targets = "SPIRV | HLSL"
+
+[spv]
+# We also need to provide the corresponding capabilities (which are part of SPIR-V >= 1.6).
+capabilities = ["DotProduct", "DotProductInput4x8BitPacked"]
+version = [1, 6]
+
+[hlsl]
+shader_model = "V6_4"

naga/tests/in/wgsl/functions-optimized-by-version.toml

@@ -0,0 +1,19 @@
+fn test_packed_integer_dot_product() -> u32 {
+    let a_5 = 1u;
+    let b_5 = 2u;
+    let c_5: i32 = dot4I8Packed(a_5, b_5);
+
+    let a_6 = 3u;
+    let b_6 = 4u;
+    let c_6: u32 = dot4U8Packed(a_6, b_6);
+
+    // test baking of arguments
+    let c_7: i32 = dot4I8Packed(5u + c_6, 6u + c_6);
+    let c_8: u32 = dot4U8Packed(7u + c_6, 8u + c_6);
+    return c_8;
+}
+
+@compute @workgroup_size(1)
+fn main() {
+    let c = test_packed_integer_dot_product();
+}

naga/tests/in/wgsl/functions-optimized-by-version.wgsl

@@ -1,5 +1,5 @@
 ; SPIR-V
-; Version: 1.6
+; Version: 1.0
 ; Generator: rspirv
 ; Bound: 30
 OpCapability Shader

naga/tests/in/wgsl/functions-optimized.toml

@@ -0,0 +1,45 @@
+; SPIR-V
+; Version: 1.6
+; Generator: rspirv
+; Bound: 30
+OpCapability Shader
+OpCapability DotProductKHR
+OpCapability DotProductInput4x8BitPackedKHR
+%1 = OpExtInstImport "GLSL.std.450"
+OpMemoryModel Logical GLSL450
+OpEntryPoint GLCompute %26 "main"
+OpExecutionMode %26 LocalSize 1 1 1
+%2 = OpTypeVoid
+%3 = OpTypeInt 32 0
+%6 = OpTypeFunction %3
+%7 = OpConstant  %3  1
+%8 = OpConstant  %3  2
+%9 = OpConstant  %3  3
+%10 = OpConstant  %3  4
+%11 = OpConstant  %3  5
+%12 = OpConstant  %3  6
+%13 = OpConstant  %3  7
+%14 = OpConstant  %3  8
+%16 = OpTypeInt 32 1
+%27 = OpTypeFunction %2
+%5 = OpFunction  %3  None %6
+%4 = OpLabel
+OpBranch %15
+%15 = OpLabel
+%17 = OpSDotKHR  %16  %7 %8 PackedVectorFormat4x8BitKHR
+%18 = OpUDotKHR  %3  %9 %10 PackedVectorFormat4x8BitKHR
+%19 = OpIAdd  %3  %11 %18
+%20 = OpIAdd  %3  %12 %18
+%21 = OpSDotKHR  %16  %19 %20 PackedVectorFormat4x8BitKHR
+%22 = OpIAdd  %3  %13 %18
+%23 = OpIAdd  %3  %14 %18
+%24 = OpUDotKHR  %3  %22 %23 PackedVectorFormat4x8BitKHR
+OpReturnValue %24
+OpFunctionEnd
+%26 = OpFunction  %2  None %27
+%25 = OpLabel
+OpBranch %28
+%28 = OpLabel
+%29 = OpFunctionCall  %3  %5
+OpReturn
+OpFunctionEnd