Halfvec avx512 fp16

.github/workflows/build.yml

@@ -41,6 +41,38 @@ jobs:
           sudo apt-get update
           sudo apt-get install libipc-run-perl
       - run: make prove_installcheck
+  ubuntu_spr:
+    runs-on: ubuntu-24.04
+    if: ${{ !startsWith(github.ref_name, 'mac') && !startsWith(github.ref_name, 'windows') }}
+    steps:
+      - uses: actions/checkout@v4
+      - name: Install Intel SDE
+        run: |
+          curl -o /tmp/sde.tar.xz https://downloadmirror.intel.com/831748/sde-external-9.44.0-2024-08-22-lin.tar.xz
+          mkdir /tmp/sde && tar -xvf /tmp/sde.tar.xz -C /tmp/sde/
+          sudo mv /tmp/sde/* /opt/sde && sudo ln -s /opt/sde/sde64 /usr/bin/sde
+      - name: Install Postgres
+        run: |
+          sudo apt-get install postgresql-16
+          sudo apt-get install postgresql-server-dev-16
+          sudo systemctl stop postgresql
+          pgdir=$(pg_config --bindir)
+          pgdata=/tmp/postgres_data
+          mkdir $pgdata
+          $pgdir/initdb -D $pgdata --username=$USER
+          sudo chmod 777 /var/run/postgresql/
+          sde -spr -mix -- $pgdir/pg_ctl -D $pgdata start
+      - run: make
+        env:
+          PG_CFLAGS: -DUSE_ASSERT_CHECKING -Wall -Wextra -Werror -Wno-unused-parameter -Wno-sign-compare
+      - run: |
+          export PG_CONFIG=`which pg_config`
+          sudo --preserve-env=PG_CONFIG make install
+      - run: sde -spr -mix -- make installcheck
+      - if: ${{ failure() }}
+        run: cat regression.diffs
+      - name: Report AVX512 FP16 FMA instruction count
+        run: cat sde*.txt | grep -E "FMA.*PH"
   mac:
     runs-on: ${{ matrix.os }}
     if: ${{ !startsWith(github.ref_name, 'windows') }}

Makefile

@@ -5,6 +5,9 @@ MODULE_big = vector
 DATA = $(wildcard sql/*--*--*.sql)
 DATA_built = sql/$(EXTENSION)--$(EXTVERSION).sql
 OBJS = src/bitutils.o src/bitvec.o src/halfutils.o src/halfvec.o src/hnsw.o src/hnswbuild.o src/hnswinsert.o src/hnswscan.o src/hnswutils.o src/hnswvacuum.o src/ivfbuild.o src/ivfflat.o src/ivfinsert.o src/ivfkmeans.o src/ivfscan.o src/ivfutils.o src/ivfvacuum.o src/sparsevec.o src/vector.o
+ifneq ($(USE_AVX512), 0)
+	OBJS += src/halfutils_avx512.o
+endif
 HEADERS = src/halfvec.h src/sparsevec.h src/vector.h
 
 TESTS = $(wildcard test/sql/*.sql)
@@ -31,6 +34,9 @@ endif
 # - GCC (needs -ftree-vectorize OR -O3) - https://gcc.gnu.org/projects/tree-ssa/vectorization.html
 # - Clang (could use pragma instead) - https://llvm.org/docs/Vectorizers.html
 PG_CFLAGS += $(OPTFLAGS) -ftree-vectorize -fassociative-math -fno-signed-zeros -fno-trapping-math
+ifneq ($(USE_AVX512), 0)
+	PG_CFLAGS += -DUSE_AVX512
+endif
 
 # Debug GCC auto-vectorization
 # PG_CFLAGS += -fopt-info-vec

src/halfutils.c

@@ -5,6 +5,7 @@
 
 #ifdef HALFVEC_DISPATCH
 #include <immintrin.h>
+#include "halfutils_avx512.h"
 
 #if defined(USE__GET_CPUID)
 #include <cpuid.h>
@@ -24,6 +25,8 @@ float		(*HalfvecInnerProduct) (int dim, half * ax, half * bx);
 double		(*HalfvecCosineSimilarity) (int dim, half * ax, half * bx);
 float		(*HalfvecL1Distance) (int dim, half * ax, half * bx);
 
+void 		(*Float4ToHalfVector) (Vector * vec, HalfVector * result);
+
 static float
 HalfvecL2SquaredDistanceDefault(int dim, half * ax, half * bx)
 {
@@ -237,6 +240,13 @@ HalfvecL1DistanceF16c(int dim, half * ax, half * bx)
 }
 #endif
 
+static void
+Float4ToHalfVectorDefault(Vector * vec, HalfVector * result) {
+	for (int i = 0; i < vec->dim; i++)
+		result->x[i] = Float4ToHalf(vec->x[i]);
+}
+
+
 #ifdef HALFVEC_DISPATCH
 #define CPU_FEATURE_FMA     (1 << 12)
 #define CPU_FEATURE_OSXSAVE (1 << 27)
@@ -284,6 +294,7 @@ HalfvecInit(void)
 	HalfvecInnerProduct = HalfvecInnerProductDefault;
 	HalfvecCosineSimilarity = HalfvecCosineSimilarityDefault;
 	HalfvecL1Distance = HalfvecL1DistanceDefault;
+	Float4ToHalfVector = Float4ToHalfVectorDefault;
 
 #ifdef HALFVEC_DISPATCH
 	if (SupportsCpuFeature(CPU_FEATURE_AVX | CPU_FEATURE_F16C | CPU_FEATURE_FMA))
@@ -294,5 +305,16 @@ HalfvecInit(void)
 		/* Does not require FMA, but keep logic simple */
 		HalfvecL1Distance = HalfvecL1DistanceF16c;
 	}
+
+#if defined(USE_AVX512) && defined(HAVE_AVX512FP16)
+	if (SupportsAvx512Fp16())
+	{
+		HalfvecL2SquaredDistance = HalfvecL2SquaredDistanceAvx512;
+		HalfvecInnerProduct = HalfvecInnerProductAvx512;
+		HalfvecCosineSimilarity = HalfvecCosineSimilarityAvx512;
+		HalfvecL1Distance = HalfvecL1DistanceAvx512;
+		Float4ToHalfVector = Float4ToHalfVectorAvx512;
+	}
+#endif
 #endif
 }

src/halfutils.h

@@ -5,6 +5,7 @@
 
 #include "common/shortest_dec.h"
 #include "halfvec.h"
+#include "vector.h"
 
 #ifdef F16C_SUPPORT
 #include <immintrin.h>
@@ -15,6 +16,8 @@ extern float (*HalfvecInnerProduct) (int dim, half * ax, half * bx);
 extern double (*HalfvecCosineSimilarity) (int dim, half * ax, half * bx);
 extern float (*HalfvecL1Distance) (int dim, half * ax, half * bx);
 
+extern void (*Float4ToHalfVector) (Vector * vec, HalfVector * result);
+
 void		HalfvecInit(void);
 
 /*