Add test for full reciprocal calculation

buildps3.bat

@@ -9,6 +9,7 @@ spu-lv2-gcc -O3 -c task_putlluc.c -o task_putlluc.o
 spu-lv2-gcc -O3 -c task_putllc.c -o task_putllc.o
 spu-lv2-gcc -O3 -c task_put.c -o task_put.o
 spu-lv2-gcc -O3 -c task_largeput.c -o task_largeput.o
+spu-lv2-gcc -O3 -c task_reacc.c -o task_reacc.o
 
 spu-lv2-gcc -O3 -mspurs-task task.o -o task.elf
 spu-lv2-gcc -O3 -mspurs-task task_pingpong.o -o task_pingpong.elf
@@ -20,6 +21,7 @@ spu-lv2-gcc -O3 -mspurs-task task_putlluc.o -ldma -o task_putlluc.elf
 spu-lv2-gcc -O3 -mspurs-task task_putllc.o -ldma -o task_putllc.elf
 spu-lv2-gcc -O3 -mspurs-task task_put.o -ldma -o task_put.elf
 spu-lv2-gcc -O3 -mspurs-task task_largeput.o -ldma -o task_largeput.elf
+spu-lv2-gcc -O3 -mspurs-task task_reacc.o -o task_reacc.elf
 
 spu_elf-to-ppu_obj task.elf task.ppu.o
 spu_elf-to-ppu_obj task_pingpong.elf task_pingpong.ppu.o
@@ -31,6 +33,7 @@ spu_elf-to-ppu_obj task_putlluc.elf task_putlluc.ppu.o
 spu_elf-to-ppu_obj task_putllc.elf task_putllc.ppu.o
 spu_elf-to-ppu_obj task_put.elf task_put.ppu.o
 spu_elf-to-ppu_obj task_largeput.elf task_largeput.ppu.o
+spu_elf-to-ppu_obj task_reacc.elf task_reacc.ppu.o
 
 ppu-lv2-gcc -O3 -c spurs_helpers.c -o spurs_helpers.o
 ppu-lv2-gcc -O3 -c test_avalanche.c -o test_avalanche.o
@@ -43,5 +46,5 @@ ppu-lv2-gcc -O3 -c test_largeblock.c -o test_largeblock.o
 ppu-lv2-gcc -O3 -c main.c -o main.o
 
 
-ppu-lv2-gcc -O3 spurs_helpers.o test_avalanche.o test_pingpong.o test_mfc64.o test_spu_inst.o test_spinlock.o test_block.o test_largeblock.o main.o task.ppu.o task_pingpong.ppu.o task_mfc64.ppu.o task_spuint.ppu.o task_spufloat.ppu.o task_spuspinlock.ppu.o task_putlluc.ppu.o task_putllc.ppu.o task_put.ppu.o task_largeput.ppu.o -lsysmodule_stub -lspurs_stub -o spurs_test.elf
+ppu-lv2-gcc -O3 spurs_helpers.o test_avalanche.o test_pingpong.o test_mfc64.o test_spu_inst.o test_spinlock.o test_block.o test_largeblock.o main.o task.ppu.o task_pingpong.ppu.o task_mfc64.ppu.o task_spuint.ppu.o task_spufloat.ppu.o task_spuspinlock.ppu.o task_putlluc.ppu.o task_putllc.ppu.o task_put.ppu.o task_largeput.ppu.o task_reacc.ppu.o -lsysmodule_stub -lspurs_stub -o spurs_test.elf
 make_fself spurs_test.elf spurs_test.self

main.c

@@ -28,7 +28,7 @@
 		printf("%s completed in %llu ms (PS3: %u ms)\n", name, (time2 - time1), reference); \
 	}
 
-#define NUM_TESTS 11
+#define NUM_TESTS 12
 
 enum tests
 {
@@ -43,6 +43,7 @@ enum tests
 	TEST_SPUPUTLLC,
 	TEST_SPUPUT,
 	TEST_SPULARGEPUT,
+	TEST_SPUREACC,
 };
 
 #define AVALANCHE_NAME "SPU Task Avalanche"
@@ -56,6 +57,7 @@ enum tests
 #define SPUPUTLLC_NAME "PUTLLC Perf"
 #define SPUPUT_NAME "PUT Perf"
 #define SPULARGEPUT_NAME "Large PUT Perf"
+#define SPUREACC_NAME "SPU Reciprocal Accurate"
 
 typedef struct
 {
@@ -75,7 +77,8 @@ const arg_test arg_conv[NUM_TESTS] = {
 	{'U', TEST_SPUPUTLLUC, SPUPUTLLUC_NAME},
 	{'T', TEST_SPUPUTLLC, SPUPUTLLC_NAME},
 	{'R', TEST_SPUPUT, SPUPUT_NAME},
-	{'G', TEST_SPULARGEPUT, SPULARGEPUT_NAME}};
+	{'G', TEST_SPULARGEPUT, SPULARGEPUT_NAME},
+	{'E', TEST_SPUREACC, SPUREACC_NAME}};
 
 extern const CellSpursTaskBinInfo _binary_task_task_spuint_elf_taskbininfo;
 extern const CellSpursTaskBinInfo _binary_task_task_spufloat_elf_taskbininfo;
@@ -85,6 +88,8 @@ extern const CellSpursTaskBinInfo _binary_task_task_putllc_elf_taskbininfo;
 extern const CellSpursTaskBinInfo _binary_task_task_put_elf_taskbininfo;
 extern const CellSpursTaskBinInfo _binary_task_task_largeput_elf_taskbininfo;
 
+extern const CellSpursTaskBinInfo _binary_task_task_reacc_elf_taskbininfo;
+
 bool verbose = false;
 
 uint64_t get_time()
@@ -97,7 +102,7 @@ uint64_t get_time()
 
 int main(int argc, char *argv[])
 {
-	printf("SPU Test v1.1.2 by GalCiv\n");
+	printf("SPU Test v1.2.1 by GalCiv\n");
 
 	unsigned int seed = 12345678;
 	unsigned int repeat = 1;
@@ -221,9 +226,11 @@ int main(int argc, char *argv[])
 		if (tests_to_run[TEST_SPUPUTLLC])
 			DO_A_TEST(SPUPUTLLC_NAME, test_block(spurs2, &_binary_task_task_putllc_elf_taskbininfo), 3364);
 		if (tests_to_run[TEST_SPUPUT])
-			DO_A_TEST(SPUPUT_NAME, test_block(spurs2, &_binary_task_task_put_elf_taskbininfo), 3984);
+			DO_A_TEST(SPUPUT_NAME, test_block(spurs2, &_binary_task_task_put_elf_taskbininfo), 4567);
 		if (tests_to_run[TEST_SPULARGEPUT])
-			DO_A_TEST(SPULARGEPUT_NAME, test_largeblock(spurs2, &_binary_task_task_largeput_elf_taskbininfo), 4454);
+			DO_A_TEST(SPULARGEPUT_NAME, test_largeblock(spurs2, &_binary_task_task_largeput_elf_taskbininfo), 4423);
+		if (tests_to_run[TEST_SPUREACC])
+			DO_A_TEST(SPUREACC_NAME, test_spu_inst(spurs2, &_binary_task_task_reacc_elf_taskbininfo), 6125);
 	}
 
 	timeend = get_time();

task_reacc.c

@@ -0,0 +1,27 @@
+#include <stdint.h>
+#include <stdlib.h>
+#include <spu_intrinsics.h>
+#include <cell/spurs.h>
+
+CELL_SPU_LS_PARAM(16 * 1024, 16 * 1024);
+
+#define NUM_REACC 800000000ull
+
+int cellSpursTaskMain(qword argTask, uint64_t argTaskset)
+{
+	(void)argTaskset;
+
+	vec_float4 values = (vec_float4)argTask;
+	vec_float4 the_ones = {1.0f, 1.0f, 1.0f, 1.0f};
+
+	for (uint64_t counter = 0; counter < NUM_REACC; counter++)
+	{
+		vec_float4 frest_result = __builtin_spu_frest((vec_float4)values);
+		qword fi_result = __builtin_si_fi((qword)values, (qword)frest_result);
+		qword fnms_result = __builtin_si_fnms((qword)values, (qword)fi_result, (qword)the_ones);
+		qword fma_result = __builtin_si_fma((qword)fnms_result, (qword)fi_result, (qword)fi_result);
+		values = (vec_float4)fma_result;
+	}
+
+	return si_to_int((qword)values);
+}