diff --git a/src/CodeGen_X86.cpp b/src/CodeGen_X86.cpp
index 7529c31688f4..e208b0d10981 100644
--- a/src/CodeGen_X86.cpp
+++ b/src/CodeGen_X86.cpp
@@ -131,6 +131,11 @@ const x86Intrinsic intrinsic_defs[] = {
     {"llvm.ssub.sat.v16i16", Int(16, 16), "saturating_sub", {Int(16, 16), Int(16, 16)}, Target::AVX2},
     {"llvm.ssub.sat.v8i16", Int(16, 8), "saturating_sub", {Int(16, 8), Int(16, 8)}},
 
+    // Sum of absolute differences
+    {"llvm.x86.sse2.psad.bw", UInt(64, 2), "sum_of_absolute_differences", {UInt(8, 16), UInt(8, 16)}},
+    {"llvm.x86.avx2.psad.bw", UInt(64, 4), "sum_of_absolute_differences", {UInt(8, 32), UInt(8, 32)}, Target::AVX2},
+    {"llvm.x86.avx512.psad.bw.512", UInt(64, 8), "sum_of_absolute_differences", {UInt(8, 64), UInt(8, 64)}, Target::AVX512_Skylake},
+
     // Some of the instructions referred to below only appear with
     // AVX2, but LLVM generates better AVX code if you give it
     // full 256-bit vectors and let it do the slicing up into
@@ -635,10 +640,13 @@ void CodeGen_X86::codegen_vector_reduce(const VectorReduce *op, const Expr &init
         // One could do a horizontal widening addition with
         // other dot_products against a vector of ones. Currently disabled
         // because I haven't found other cases where it's clearly better.
-
         {VectorReduce::Add, 2, u16(wild_u8x_), "horizontal_widening_add", {}, Pattern::SingleArg},
         {VectorReduce::Add, 2, i16(wild_u8x_), "horizontal_widening_add", {}, Pattern::SingleArg},
         {VectorReduce::Add, 2, i16(wild_i8x_), "horizontal_widening_add", {}, Pattern::SingleArg},
+
+        // Sum of absolute differences
+        {VectorReduce::Add, 8, u64(absd(wild_u8x_, wild_u8x_)), "sum_of_absolute_differences", {}},
+
     };
     // clang-format on
 
@@ -708,6 +716,33 @@ void CodeGen_X86::codegen_vector_reduce(const VectorReduce *op, const Expr &init
         }
     }
 
+    // Rewrite non-native sum-of-absolute-difference variants to the native
+    // op. We support reducing to various types. We could consider supporting
+    // multiple reduction factors too, but in general we don't handle non-native
+    // reduction factors for VectorReduce nodes (yet?).
+    if (op->op == VectorReduce::Add &&
+        factor == 8) {
+        const Cast *cast = op->value.as<Cast>();
+        const Call *call = cast ? cast->value.as<Call>() : nullptr;
+        if (call &&
+            call->is_intrinsic(Call::absd) &&
+            cast->type.element_of().can_represent(UInt(8)) &&
+            (cast->type.is_int() || cast->type.is_uint()) &&
+            call->args[0].type().element_of() == UInt(8)) {
+
+            internal_assert(cast->type.element_of() != UInt(64)) << "Should have pattern-matched above\n";
+
+            // Cast to uint64 instead
+            Expr equiv = Cast::make(UInt(64, cast->value.type().lanes()), cast->value);
+            // Reduce on that to hit psadbw
+            equiv = VectorReduce::make(VectorReduce::Add, equiv, op->type.lanes());
+            // Then cast that to the desired type
+            equiv = Cast::make(cast->type.with_lanes(equiv.type().lanes()), equiv);
+            codegen(equiv);
+            return;
+        }
+    }
+
     CodeGen_Posix::codegen_vector_reduce(op, init);
 }
 
diff --git a/test/correctness/simd_op_check.cpp b/test/correctness/simd_op_check.cpp
index 5f2258b65f55..ba3e24d6e80c 100644
--- a/test/correctness/simd_op_check.cpp
+++ b/test/correctness/simd_op_check.cpp
@@ -242,6 +242,18 @@ class SimdOpCheck : public SimdOpCheckTest {
             check(std::string("packssdw") + check_suffix, 4 * w, i16_sat(i32_1));
             check(std::string("packsswb") + check_suffix, 8 * w, i8_sat(i16_1));
             check(std::string("packuswb") + check_suffix, 8 * w, u8_sat(i16_1));
+
+            // Sum-of-absolute-difference ops
+            {
+                const int f = 8;  // reduction factor.
+                RDom r(0, f);
+                check("psadbw", w, sum(u64(absd(in_u8(f * x + r), in_u8(f * x + r + 32)))));
+                check("psadbw", w, sum(u32(absd(in_u8(f * x + r), in_u8(f * x + r + 32)))));
+                check("psadbw", w, sum(u16(absd(in_u8(f * x + r), in_u8(f * x + r + 32)))));
+                check("psadbw", w, sum(i64(absd(in_u8(f * x + r), in_u8(f * x + r + 32)))));
+                check("psadbw", w, sum(i32(absd(in_u8(f * x + r), in_u8(f * x + r + 32)))));
+                check("psadbw", w, sum(i16(absd(in_u8(f * x + r), in_u8(f * x + r + 32)))));
+            }
         }
 
         // SSE 3 / SSSE 3
@@ -518,6 +530,18 @@ class SimdOpCheck : public SimdOpCheckTest {
             check("vpcmpeqq*ymm", 4, select(i64_1 == i64_2, i64(1), i64(2)));
             check("vpackusdw*ymm", 16, u16(clamp(i32_1, 0, max_u16)));
             check("vpcmpgtq*ymm", 4, select(i64_1 > i64_2, i64(1), i64(2)));
+
+            // Sum-of-absolute-difference ops
+            for (int w : {4, 8}) {
+                const int f = 8;  // reduction factor.
+                RDom r(0, f);
+                check("vpsadbw", w, sum(u64(absd(in_u8(f * x + r), in_u8(f * x + r + 32)))));
+                check("vpsadbw", w, sum(u32(absd(in_u8(f * x + r), in_u8(f * x + r + 32)))));
+                check("vpsadbw", w, sum(u16(absd(in_u8(f * x + r), in_u8(f * x + r + 32)))));
+                check("vpsadbw", w, sum(i64(absd(in_u8(f * x + r), in_u8(f * x + r + 32)))));
+                check("vpsadbw", w, sum(i32(absd(in_u8(f * x + r), in_u8(f * x + r + 32)))));
+                check("vpsadbw", w, sum(i16(absd(in_u8(f * x + r), in_u8(f * x + r + 32)))));
+            }
         }
 
         if (use_avx512) {