changed hebi_packet size to 128-bits on 64-bit platforms

Makefile

@@ -226,6 +226,7 @@ BENCH_DEPS := $(DEPS) bench/bench.h
 CHECK_P := \
 	pcmp \
 	pcopy \
+	pmove \
 	pnorm \
 	pzero
 

README

@@ -32,9 +32,9 @@ Hebimath has the following design features:
 - Can use fixed-size buffers instead of an allocator when initializing
   a number primitive, useful for memory allocations on the stack.
 - Can specify separate memory allocators for result and temporary buffers.
-- Smallest unit size for natural numbers and integers is 16 bytes on 32-bit
-  platforms and 32-bytes on 64-bit platforms. This larger unit or packet size
-  aids in shrinking the amount of code bloat for unrolled kernel functions.
+- Smallest unit size for natural numbers and integers is 16 bytes. This
+  larger unit or packet size aids in shrinking the amount of code bloat for
+  unrolled kernel functions.
 - Memory buffers are correctly aligned for use with SIMD instructions.
 
 Hebimath adheres to these principles in its implementation:

bench/p/padd.c

@@ -10,7 +10,7 @@ static volatile uint64_t result_sink;
 int
 main(int argc, char *argv[])
 {
-	bench_args p = { .iter = 5000, .an = 32768, .bn = 16384 };
+	bench_args p = { .iter = 5000, .an = 65536, .bn = 32768 };
 	hebi_packet *a, *b, *r;
 	int i;
 

bench/p/paddu.c

@@ -11,7 +11,7 @@ static const uint64_t b = 42571;
 int
 main(int argc, char *argv[])
 {
-	bench_args p = { .iter = 5000, .an = 32768 };
+	bench_args p = { .iter = 5000, .an = 65536 };
 	hebi_packet *a, *r;
 	int i;
 

bench/p/pclz.c

@@ -10,7 +10,7 @@ static volatile size_t result_sink;
 int
 main(int argc, char *argv[])
 {
-	bench_args p = { .iter = 5000, .an = 32768 };
+	bench_args p = { .iter = 5000, .an = 65536 };
 	hebi_packet *a;
 	int i;
 

bench/p/pcopy.c

@@ -10,7 +10,7 @@ static volatile uint32_t result_sink;
 int
 main(int argc, char *argv[])
 {
-	bench_args p = { .iter = 5000, .an = 32768, .bn = 32768 };
+	bench_args p = { .iter = 5000, .an = 65536, .bn = 65536 };
 	hebi_packet *a, *b;
 	int i;
 

bench/p/pdivrem.c

@@ -10,7 +10,7 @@ static volatile size_t result_sink;
 int
 main(int argc, char *argv[])
 {
-	bench_args p = { .iter = 5000, .an = 512, .bn = 256 };
+	bench_args p = { .iter = 5000, .an = 1024, .bn = 512 };
 	hebi_packet *a, *b, *q, *r, *w;
 	size_t rn;
 	int i;

bench/p/pdivremu.c

@@ -11,7 +11,7 @@ static const uint64_t b = 42571;
 int
 main(int argc, char *argv[])
 {
-	bench_args p = { .iter = 5000, .an = 32768 };
+	bench_args p = { .iter = 5000, .an = 65536 };
 	hebi_packet *a, *r;
 	int i;
 

bench/p/pmove.c

@@ -10,7 +10,7 @@ static volatile uint32_t result_sink;
 int
 main(int argc, char *argv[])
 {
-	bench_args p = { .iter = 5000, .an = 32768, .bn = 32768 };
+	bench_args p = { .iter = 5000, .an = 65536, .bn = 65536 };
 	hebi_packet *a, *b;
 	int i;
 

bench/p/pmul.c

@@ -10,7 +10,7 @@ static volatile uint64_t result_sink;
 int
 main(int argc, char *argv[])
 {
-	bench_args p = { .iter = 5000, .an = 256, .bn = 256 };
+	bench_args p = { .iter = 5000, .an = 512, .bn = 512 };
 	hebi_packet *a, *b, *r;
 	int i;
 

bench/p/pmul_karatsuba.c

@@ -10,7 +10,7 @@ static volatile uint64_t result_sink;
 int
 main(int argc, char *argv[])
 {
-	bench_args p = { .iter = 5000, .an = 256, .bn = 256 };
+	bench_args p = { .iter = 5000, .an = 512, .bn = 512 };
 	hebi_packet *a, *b, *r, *w;
 	size_t rn, wn;
 	int i;

bench/p/pmulu.c

@@ -11,7 +11,7 @@ static const uint64_t b = 42571;
 int
 main(int argc, char *argv[])
 {
-	bench_args p = { .iter = 5000, .an = 32768 };
+	bench_args p = { .iter = 5000, .an = 65536 };
 	hebi_packet *a, *r;
 	int i;
 

bench/p/pnorm.c

@@ -10,7 +10,7 @@ static volatile size_t result_sink;
 int
 main(int argc, char *argv[])
 {
-	bench_args p = { .iter = 5000, .an = 32768 };
+	bench_args p = { .iter = 5000, .an = 65536 };
 	hebi_packet *a;
 	int i;
 

bench/p/pshl.c

@@ -6,12 +6,12 @@
 #include "../bench.h"
 
 static volatile size_t result_sink;
-static const size_t b = 433;
+static const size_t b = 443;
 
 int
 main(int argc, char *argv[])
 {
-	bench_args p = { .iter = 5000, .an = 32768 };
+	bench_args p = { .iter = 5000, .an = 65536 };
 	hebi_packet *a, *r;
 	size_t rn;
 	int i;

bench/p/pshr.c

@@ -11,7 +11,7 @@ static const size_t b = 433;
 int
 main(int argc, char *argv[])
 {
-	bench_args p = { .iter = 5000, .an = 32768 };
+	bench_args p = { .iter = 5000, .an = 65536 };
 	hebi_packet *a, *r;
 	size_t rn;
 	int i;

bench/p/psqr.c

@@ -10,7 +10,7 @@ static volatile uint64_t result_sink;
 int
 main(int argc, char *argv[])
 {
-	bench_args p = { .iter = 5000, .an = 256 };
+	bench_args p = { .iter = 5000, .an = 512 };
 	hebi_packet *a, *r;
 	int i;
 

bench/p/psqr_karatsuba.c

@@ -10,7 +10,7 @@ static volatile uint64_t result_sink;
 int
 main(int argc, char *argv[])
 {
-	bench_args p = { .iter = 5000, .an = 256 };
+	bench_args p = { .iter = 5000, .an = 512 };
 	hebi_packet *a, *r, *w;
 	size_t rn, wn;
 	int i;

bench/p/psub.c

@@ -10,7 +10,7 @@ static volatile uint64_t result_sink;
 int
 main(int argc, char *argv[])
 {
-	bench_args p = { .iter = 5000, .an = 32768, .bn = 16384 };
+	bench_args p = { .iter = 5000, .an = 65536, .bn = 32768 };
 	hebi_packet *a, *b, *r;
 	int i;
 

bench/p/psubu.c

@@ -11,7 +11,7 @@ static const uint64_t b = 42571;
 int
 main(int argc, char *argv[])
 {
-	bench_args p = { .iter = 5000, .an = 32768 };
+	bench_args p = { .iter = 5000, .an = 65536 };
 	hebi_packet *a, *r;
 	int i;
 

bench/p/pzero.c

@@ -10,7 +10,7 @@ static volatile uint32_t result_sink;
 int
 main(int argc, char *argv[])
 {
-	bench_args p = { .iter = 5000, .an = 32768 };
+	bench_args p = { .iter = 5000, .an = 65536 };
 	hebi_packet *a;
 	int i;
 

check/p/pcopy.c

@@ -1,31 +1,45 @@
 #include "../check.h"
 #include <string.h>
 
-#define NUM_PACKETS 501
-#define NUM_DIGITS (NUM_PACKETS * HEBI_PACKET_LIMBS64)
-#define NUM_BYTES (NUM_PACKETS * sizeof(hebi_packet))
+#define COUNT 1024
+#define GUARD 8
+#define TOTAL (COUNT+GUARD*2)
+
+static const uint32_t a[4] = { 0xECECECEC, 0x38383838, 0x92929292, 0xA7A7A7A7 };
+static const uint32_t b[4] = { 0x3D3D3D3D, 0x6B6B6B6B, 0x71717171, 0xFEFEFEFE };
 
 int main(int argc, char *argv[])
 {
 	hebi_packet *x, *y;
-	size_t i, j;
+	size_t i, j, k;
 
-	x = hebi_palloc(HEBI_ALLOC_DEFAULT, NUM_PACKETS);
-	y = hebi_palloc(HEBI_ALLOC_DEFAULT, NUM_PACKETS);
+	x = hebi_palloc(HEBI_ALLOC_DEFAULT, TOTAL);
+	y = hebi_palloc(HEBI_ALLOC_DEFAULT, TOTAL);
 	assert(x && y);
 
-	for (i = 0; i < NUM_PACKETS; ++i) {
-		for (j = 0; j < HEBI_PACKET_LIMBS64; ++j) {
-			x[i].hp_limbs64[j] = (uint64_t)i;
-			y[i].hp_limbs64[j] = (uint64_t)i + 55;
+	for (k = 1; k <= COUNT; k++) {
+		for (i = 0; i < k+GUARD*2; i++) {
+			for (j = 0; j < HEBI_PACKET_LIMBS32; j++) {
+				x[i].hp_limbs32[j] = a[(k+j)&7];
+				y[i].hp_limbs32[j] = b[(k+j)&7];
+			}
 		}
-	}
 
-	assert(memcmp(x, y, NUM_BYTES) != 0);
-	hebi_pcopy(y, x, NUM_PACKETS);
-	assert(memcmp(x, y, NUM_BYTES) == 0);
+		assert(memcmp(y+GUARD, x+GUARD, k*sizeof(hebi_packet)) != 0);
+
+		hebi_pcopy(y+GUARD, x+GUARD, k);
+
+		assert(memcmp(y+GUARD, x+GUARD, k*sizeof(hebi_packet)) == 0);
+
+		for (i = 0; i < GUARD; i++) {
+			for (j = 0; j < HEBI_PACKET_LIMBS32; j++) {
+				assert(y[i].hp_limbs32[j] == b[(k+j)&7]);
+				assert(y[i+k+GUARD].hp_limbs32[j] == b[(k+j)&7]);
+			}
+		}
+	}
 
-	hebi_pfree(HEBI_ALLOC_DEFAULT, x, NUM_PACKETS);
-	hebi_pfree(HEBI_ALLOC_DEFAULT, y, NUM_PACKETS);
+	hebi_pfree(HEBI_ALLOC_DEFAULT, x, TOTAL);
+	hebi_pfree(HEBI_ALLOC_DEFAULT, y, TOTAL);
 	return 0;
 }

check/p/pmove.c

@@ -0,0 +1,45 @@
+#include "../check.h"
+#include <string.h>
+
+#define COUNT 1024
+#define GUARD 8
+#define TOTAL (COUNT+GUARD*2)
+
+static const uint32_t a[4] = { 0xECECECEC, 0x38383838, 0x92929292, 0xA7A7A7A7 };
+static const uint32_t b[4] = { 0x3D3D3D3D, 0x6B6B6B6B, 0x71717171, 0xFEFEFEFE };
+
+int main(int argc, char *argv[])
+{
+	hebi_packet *x, *y;
+	size_t i, j, k;
+
+	x = hebi_palloc(HEBI_ALLOC_DEFAULT, TOTAL);
+	y = hebi_palloc(HEBI_ALLOC_DEFAULT, TOTAL);
+	assert(x && y);
+
+	for (k = 1; k <= COUNT; k++) {
+		for (i = 0; i < k+GUARD*2; i++) {
+			for (j = 0; j < HEBI_PACKET_LIMBS32; j++) {
+				x[i].hp_limbs32[j] = a[(k+j)&7];
+				y[i].hp_limbs32[j] = b[(k+j)&7];
+			}
+		}
+
+		assert(memcmp(y+GUARD, x+GUARD, k*sizeof(hebi_packet)) != 0);
+
+		hebi_pmove(y+GUARD, x+GUARD, k);
+
+		assert(memcmp(y+GUARD, x+GUARD, k*sizeof(hebi_packet)) == 0);
+
+		for (i = 0; i < GUARD; i++) {
+			for (j = 0; j < HEBI_PACKET_LIMBS32; j++) {
+				assert(y[i].hp_limbs32[j] == b[(k+j)&7]);
+				assert(y[i+k+GUARD].hp_limbs32[j] == b[(k+j)&7]);
+			}
+		}
+	}
+
+	hebi_pfree(HEBI_ALLOC_DEFAULT, x, TOTAL);
+	hebi_pfree(HEBI_ALLOC_DEFAULT, y, TOTAL);
+	return 0;
+}

check/p/pzero.c

@@ -2,27 +2,35 @@
 #include <limits.h>
 #include <string.h>
 
-#define NUM_PACKETS 501
-#define NUM_DIGITS (NUM_PACKETS * HEBI_PACKET_LIMBS64)
-#define NUM_BYTES (NUM_PACKETS * sizeof(hebi_packet))
+#define COUNT 1024
+#define GUARD 8
+#define TOTAL (COUNT+GUARD*2)
 
 int main(int argc, char *argv[])
 {
 	hebi_packet *x;
-	size_t i, j;
+	size_t i, j, k;
 
-	x = hebi_palloc(HEBI_ALLOC_DEFAULT, NUM_PACKETS);
+	x = hebi_palloc(HEBI_ALLOC_DEFAULT, TOTAL);
 	assert(x);
 
-	memset(x, UCHAR_MAX, NUM_BYTES);
+	for (k = 1; k <= COUNT; k++) {
+		memset(x, UCHAR_MAX, (k+GUARD*2)*sizeof(hebi_packet));
 
-	hebi_pzero(x, NUM_PACKETS);
+		hebi_pzero(x+GUARD, k);
 
-	for (i = 0; i < NUM_PACKETS; ++i)
-		for (j = 0; j < HEBI_PACKET_LIMBS64; ++j)
-			assert(x[i].hp_limbs64[j] == 0);
+		for (i = 0; i < k; i++)
+			for (j = 0; j < HEBI_PACKET_LIMBS32; j++)
+				assert(x[i+GUARD].hp_limbs32[j] == 0);
 
-	hebi_pfree(HEBI_ALLOC_DEFAULT, x, NUM_PACKETS);
+		for (i = 0; i < GUARD; i++) {
+			for (j = 0; j < HEBI_PACKET_LIMBS32; j++) {
+				assert(x[i].hp_limbs32[j] == UINT32_MAX);
+				assert(x[i+k+GUARD].hp_limbs32[j] == UINT32_MAX);
+			}
+		}
+	}
 
+	hebi_pfree(HEBI_ALLOC_DEFAULT, x, TOTAL);
 	return 0;
 }

check/z/zgetset.c

@@ -90,7 +90,7 @@ int main(int argc, char *argv[])
 
 	assert(hebi_zsetstr(a, s1, NULL, 10) > 0);
 	assert(hebi_zsign(a) > 0);
-	assert(hebi_zused(a) == 1);
+	assert(hebi_zused(a) == 2);
 	assert(hebi_zeven(a));
 	assert(!hebi_zodd(a));
 	assert(!hebi_zzero(a));
@@ -101,7 +101,7 @@ int main(int argc, char *argv[])
 
 	assert(hebi_zsetstr(a, s2, NULL, 0) > 0);
 	assert(hebi_zsign(a) < 0);
-	assert(hebi_zused(a) == 3);
+	assert(hebi_zused(a) == 6);
 	assert(!hebi_zeven(a));
 	assert(hebi_zodd(a));
 	assert(!hebi_zzero(a));

check/z/zshl.c

@@ -1,4 +1,5 @@
 #include "../check.h"
+#include <stdio.h>
 
 int
 main(int argc, char *argv[])