Texture copy measurements originally done for Corgi3DS

texcopy.c

static u32 counter = 1;
stopProfiling();
GFX_deinit(false);
TIMER_sleepMs(1000);
*((vu32*)0x10140140) = 0;       // REG_CFG11_GPUPROT
*((vu32*)0x10141200) = 0x1007F; // REG_CFG11_GPU_CNT
IRQ_registerHandler(IRQ_PPF, 14, 0, true, NULL);
__cpsid(i);

static void *const dsts[4] = {(void*)VRAM_BASE, (void*)0x1F000000, (void*)DSP_MEM_BASE, (void*)FCRAM_BASE};
struct
{
	u32 vram[5];
	u32 qtm[5];
	u32 dsp[5];
	u32 fcram[5];
	u32 avg[4];
} data[4] = {0};
for(u32 i = 0; i < 4; i++)
{
	for(u32 n = 0; n < 5; n++)
	{
		u32 tmp = measureTexcopy((void*)VRAM_BASE, dsts[i], 131072);
		data[i].vram[n] = tmp;
		data[i].avg[0] += tmp;

		tmp = measureTexcopy((void*)0x1F000000, dsts[i], 131072);
		data[i].qtm[n] = tmp;
		data[i].avg[1] += tmp;

		tmp = measureTexcopy((void*)DSP_MEM_BASE, dsts[i], 131072);
		data[i].dsp[n] = tmp;
		data[i].avg[2] += tmp;

		tmp = measureTexcopy((void*)FCRAM_BASE, dsts[i], 131072);
		data[i].fcram[n] = tmp;
		data[i].avg[3] += tmp;
	}

	for(u32 n = 0; n < 4; n++) data[i].avg[n] /= 5;
}

__cpsie(i);
IRQ_disable(IRQ_PPF);
*((vu32*)0x10141200) = 0x10001; // REG_CFG11_GPU_CNT
TIMER_sleepMs(1000);
GFX_init(true);
if(counter > 2)
{
	const s32 h = fOpen("texcopy.csv", FS_CREATE_ALWAYS | FS_OPEN_WRITE);
	fWrite(h, "name, run1, run2, run3, run4, run5, average\n", 44);

	char buf[256];
	static const char *const names[4] = {"vram", "qtm", "dsp", "fcram"};
	for(u32 i = 0; i < 4; i++)
	{
		memset(buf, 0, 256);
		ee_sprintf(buf, "%s->%s, %lu, %lu, %lu, %lu, %lu, %lu\n", names[0], names[i], data[i].vram[0], data[i].vram[1], data[i].vram[2], data[i].vram[3], data[i].vram[4], data[i].avg[0]);
		fWrite(h, buf, strlen(buf));

		memset(buf, 0, 256);
		ee_sprintf(buf, "%s->%s, %lu, %lu, %lu, %lu, %lu, %lu\n", names[1], names[i], data[i].qtm[0], data[i].qtm[1], data[i].qtm[2], data[i].qtm[3], data[i].qtm[4], data[i].avg[1]);
		fWrite(h, buf, strlen(buf));

		memset(buf, 0, 256);
		ee_sprintf(buf, "%s->%s, %lu, %lu, %lu, %lu, %lu, %lu\n", names[2], names[i], data[i].dsp[0], data[i].dsp[1], data[i].dsp[2], data[i].dsp[3], data[i].dsp[4], data[i].avg[2]);
		fWrite(h, buf, strlen(buf));

		memset(buf, 0, 256);
		ee_sprintf(buf, "%s->%s, %lu, %lu, %lu, %lu, %lu, %lu\n", names[3], names[i], data[i].fcram[0], data[i].fcram[1], data[i].fcram[2], data[i].fcram[3], data[i].fcram[4], data[i].avg[3]);
		fWrite(h, buf, strlen(buf));
	}

	fClose(h);
}
counter++;

texcopy.s

.align 2
.global measureTexcopy
.type measureTexcopy %function
@ u32 measureTexcopy(const void *src, void *dest, u32 size);
measureTexcopy:
	ldr r3, =0x10400C00
	lsr r0, r0, #3
	lsr r1, r1, #3
	str r0, [r3]       // REGs_TRANS_ENGINE[0]
	mov r0, #(1u<<3)
	str r1, [r3, #4]   // REGs_TRANS_ENGINE[1]
	str r0, [r3, #16]  // REGs_TRANS_ENGINE[4]
	mov r1, #0
	str r2, [r3, #32]  // REGs_TRANS_ENGINE[8]
	str r1, [r3, #36]  // REGs_TRANS_ENGINE[9]
	str r1, [r3, #40]  // REGs_TRANS_ENGINE[10]
	ldr r0, =0x17E00100
	mov r2, #-1
	mov r12, #1
	str r2, [r0, #0x500]  // REG_TIMER_LOAD
	mcr p15, 0, r1, c7, c10, 0      @ "Clean Entire Data Cache"
	mcr p15, 0, r1, c7, c10, 4      @ Data Synchronization Barrier
	str r12, [r0, #0x500 + 8]  // REG_TIMER_CNT
	str r12, [r3, #24] // REGs_TRANS_ENGINE[6]
	wfi
	str r1, [r0, #0x500 + 8]   // REG_TIMER_CNT
	ldr r3, [r0, #12]          // REG_CPU_II_AKN
	str r12, [r0, #0x500 + 12] // REG_TIMER_INT_STAT
	str r3, [r0, #16]          // REG_CPU_II_EOI
	ldr r3, [r0, #0x500 + 4]   // REG_TIMER_COUNTER
	sub r0, r2, r3
	bx lr

.pool

texcopy_n3ds.csv

texcopy_o3ds.csv