Rearranged the ppu_state to be even more cacheline aware, gained another 1% performance

4 files changed, 17 insertions, 16 deletions

diff --git a/build.sh b/build.sh
index a5e2f9d..d2cb7b2 100755
--- a/build.sh
+++ b/build.sh
@@ -16,6 +16,8 @@ WIN_CC=x86_64-w64-mingw32-gcc
 # Base configuration common to all builds
 CFLAGS="-std=gnu11 -mtune=generic "
 # -fdump-tree-alias "
+CFLAGS+="-falign-functions=32 -falign-loops=32 "
+CFLAGS+="-finline-limit=800 "
 CFLAGS+="-mbmi -fno-argument-alias "
 CFLAGS+="-mfunction-return=keep -mindirect-branch=keep "
 CFLAGS+="-fwrapv -ffast-math -fno-trapping-math -fvisibility=hidden "
@@ -93,7 +95,7 @@ set -e
 
 # Build Linux version
 (
-	$CC $CFLAGS ${PROJECT_NAME}_sdl.c -o ${PROJECT_NAME} $INCLUDE_PATHS $LINUX_INCLUDE $LDFLAGS $LINUX_LIBS
+	$CC $CFLAGS ${PROJECT_NAME}.c -o ${PROJECT_NAME} $INCLUDE_PATHS $LINUX_INCLUDE $LDFLAGS $LINUX_LIBS
 	objdump -d -Mintel mknes > mknes.s
 ) &
 
diff --git a/mknes.h b/mknes.h
index b6739d3..944f762 100644
--- a/mknes.h
+++ b/mknes.h
@@ -56,7 +56,6 @@ struct ppu_state {
 	uint8_t open_bus;						// 29
 
 	uint8_t sprite_count;				// 30 - Number of sprites in secondary OAM (0-8)
-	uint8_t palette[32];					// 31
 
 	// NOTE(peter): CACHELINE 2
 	uint8_t secondary_oam[32] __attribute__((aligned(64)));
@@ -68,16 +67,18 @@ struct ppu_state {
 		uint8_t position;
 		uint8_t priority;
 		uint8_t palette;
-	} __attribute__((packed)) sprites[8];
+	} __attribute__((packed)) sprites[8];	// 32 -> +40
 
-	uint8_t input[2];						// 40 - Controller 1 & 2
-	uint8_t input_latch[2];				// 42 - Latched inputs after strobe
-	uint8_t input_bit[2];				// 44 - Current bit position being shifted out
-	uint8_t input_strobe;				// 46 - Control bit (0 or 1)
-	uint8_t frame_ready;					// 47
-	uint8_t sprite_zero_in_range;		// 48 - Boolean: is sprite 0 in range (will always be slot 0 if true)
+	uint8_t input_strobe;				// 76 - Control bit (0 or 1)
+	uint8_t input[2];						// 73 - Controller 1 & 2
+	uint8_t input_latch[2];				// 74 - Latched inputs after strobe
+	uint8_t input_bit[2];				// 75 - Current bit position being shifted out
+	uint8_t frame_ready;					// 77
+	uint8_t sprite_zero_in_range;		// 78 - Boolean: is sprite 0 in range (will always be slot 0 if true)
 
-	// NOTE(peter): CACHELINE 4
+	uint8_t palette[32];					// 79
+
+	// NOTE(peter): CACHELINE 5
 	uint8_t oam[256] __attribute__((aligned(64)));
 } __attribute__((packed));
 
diff --git a/mknes_bench.c b/mknes_bench.c
index b064266..e7d53fa 100644
--- a/mknes_bench.c
+++ b/mknes_bench.c
@@ -333,17 +333,17 @@ static void run_benchmark(struct nes_state *nstate, uint32_t num_runs, uint32_t
 
 	// Calculate per-frame statistics using static storage
 	double *cycles_per_frame = &per_frame_storage[0];
-	double *insn_per_frame = &per_frame_storage[num_runs];
+	double *ipc_per_run = &per_frame_storage[num_runs];
 	double *time_ms = &per_frame_storage[num_runs * 2];
 
 	for(uint32_t i = 0; i < num_runs; i++) {
 		cycles_per_frame[i] = (double)runs[i].cycles / frames_per_run;
-		insn_per_frame[i] = (double)runs[i].instructions / frames_per_run;
+		ipc_per_run[i] = (double)runs[i].instructions / (double)runs[i].cycles;
 		time_ms[i] = (double)runs[i].time_ns / 1000000.0;
 	}
 
 	struct bench_stats cycles_stats = calculate_stats(cycles_per_frame, num_runs);
-	struct bench_stats insn_stats = calculate_stats(insn_per_frame, num_runs);
+	struct bench_stats ipc_stats = calculate_stats(ipc_per_run, num_runs);
 	struct bench_stats time_stats = calculate_stats(time_ms, num_runs);
 
 	// Print results
@@ -368,7 +368,7 @@ static void run_benchmark(struct nes_state *nstate, uint32_t num_runs, uint32_t
 	printf("Throughput: %.2f MIPS, %.2f Mcycles/sec\n", mips, mcps);
 	printf("\n");
 	printf("cycles/frame  mean=%9.0f  sd=%6.0f  relSD=%.3f%%  n=%d\n", cycles_stats.mean, cycles_stats.sd, cycles_stats.rel_sd, cycles_stats.n);
-	printf("insn/frame    mean=%9.0f  sd=%6.0f  relSD=%.3f%%  n=%d\n", insn_stats.mean, insn_stats.sd, insn_stats.rel_sd, insn_stats.n);
+	printf("IPC           mean=%9.3f  sd=%6.3f  relSD=%.3f%%  n=%d\n", ipc_stats.mean, ipc_stats.sd, ipc_stats.rel_sd, ipc_stats.n);
 	printf("time (ms)     mean=%9.3f  sd=%6.3f  relSD=%.3f%%  n=%d\n", time_stats.mean, time_stats.sd, time_stats.rel_sd, time_stats.n);
 
 	double fps = (double)frames_per_run / (time_stats.mean / 1000.0);
diff --git a/mknes_ppu.c b/mknes_ppu.c
index b0f6479..a4f7a2c 100644
--- a/mknes_ppu.c
+++ b/mknes_ppu.c
@@ -277,8 +277,6 @@ shift_and_fetch:
 
 					switch(dot % 8) {
 						case 1: {
-							// uint32_t nt_addr = 0x2000 | (ppu->vram_addr & 0xfff);
-							// ppu->bg_next_tile_id = state->mapper_function.ciram_read(state, nt_addr);
 							ppu->bg_next_tile_id = state->mapper_function.ciram_read(state, ppu->vram_addr);
 						} break;