Using a blit + sleep does not guarantee generating a high enough load to keep the GPU busy, a busyspin does. Replace the blit at high load with a continuous series spins and low load with a mix of spin and sleep.
References: https://bugs.freedesktop.org/show_bug.cgi?id=104060 Signed-off-by: Chris Wilson <[email protected]> --- tests/pm_rps.c | 160 ++++++++++----------------------------------------------- 1 file changed, 27 insertions(+), 133 deletions(-) diff --git a/tests/pm_rps.c b/tests/pm_rps.c index cc94813e..88aca400 100644 --- a/tests/pm_rps.c +++ b/tests/pm_rps.c @@ -26,7 +26,6 @@ * */ -#include "igt.h" #include <stdio.h> #include <stdlib.h> #include <string.h> @@ -37,7 +36,8 @@ #include <time.h> #include <sys/wait.h> -#include "intel_bufmgr.h" +#include "igt.h" +#include "igt_dummyload.h" IGT_TEST_DESCRIPTION("Render P-States tests - verify GPU frequency changes"); @@ -181,15 +181,9 @@ enum load { }; static struct load_helper { - int devid; - int has_ppgtt; - drm_intel_bufmgr *bufmgr; - struct intel_batchbuffer *batch; - drm_intel_bo *target_buffer; enum load load; bool exit; struct igt_helper_process igt_proc; - drm_intel_bo *src, *dst; } lh; static void load_helper_signal_handler(int sig) @@ -201,29 +195,6 @@ static void load_helper_signal_handler(int sig) lh.exit = true; } -static void emit_store_dword_imm(uint32_t val) -{ - int cmd; - struct intel_batchbuffer *batch = lh.batch; - - cmd = MI_STORE_DWORD_IMM; - if (!lh.has_ppgtt) - cmd |= MI_MEM_VIRTUAL; - - BEGIN_BATCH(4, 0); /* just ignore the reloc we emit and count dwords */ - OUT_BATCH(cmd); - if (batch->gen >= 8) { - OUT_RELOC(lh.target_buffer, I915_GEM_DOMAIN_INSTRUCTION, - I915_GEM_DOMAIN_INSTRUCTION, 0); - } else { - OUT_BATCH(0); /* reserved */ - OUT_RELOC(lh.target_buffer, I915_GEM_DOMAIN_INSTRUCTION, - I915_GEM_DOMAIN_INSTRUCTION, 0); - } - OUT_BATCH(val); - ADVANCE_BATCH(); -} - #define LOAD_HELPER_PAUSE_USEC 500 #define LOAD_HELPER_BO_SIZE (16*1024*1024) static void load_helper_set_load(enum load load) @@ -252,69 +223,45 @@ static void load_helper_run(enum load load) lh.load = load; igt_fork_helper(&lh.igt_proc) { - const uint32_t bbe = MI_BATCH_BUFFER_END; - struct drm_i915_gem_exec_object2 object; - struct drm_i915_gem_execbuffer2 execbuf; - uint32_t fences[3]; - uint32_t val = 0; + igt_spin_t *spin[2] = {}; + uint32_t handle; signal(SIGUSR1, load_helper_signal_handler); signal(SIGUSR2, load_helper_signal_handler); - fences[0] = gem_create(drm_fd, 4096); - gem_write(drm_fd, fences[0], 0, &bbe, sizeof(bbe)); - fences[1] = gem_create(drm_fd, 4096); - gem_write(drm_fd, fences[1], 0, &bbe, sizeof(bbe)); - fences[2] = gem_create(drm_fd, 4096); - gem_write(drm_fd, fences[2], 0, &bbe, sizeof(bbe)); - - memset(&execbuf, 0, sizeof(execbuf)); - execbuf.buffers_ptr = (uintptr_t)&object; - execbuf.buffer_count = 1; - if (intel_gen(lh.devid) >= 6) - execbuf.flags = I915_EXEC_BLT; - igt_debug("Applying %s load...\n", lh.load ? "high" : "low"); while (!lh.exit) { - memset(&object, 0, sizeof(object)); - object.handle = fences[val%3]; - - while (gem_bo_busy(drm_fd, object.handle)) - usleep(100); - - if (lh.load == HIGH) - intel_copy_bo(lh.batch, lh.dst, lh.src, - LOAD_HELPER_BO_SIZE); - - emit_store_dword_imm(val); - intel_batchbuffer_flush_on_ring(lh.batch, - I915_EXEC_BLT); - val++; - - gem_execbuf(drm_fd, &execbuf); + igt_spin_batch_free(drm_fd, spin[0]); + usleep(100); + spin[0] = spin[1]; + spin[lh.load == HIGH] = + igt_spin_batch_new(drm_fd, 0, 0, 0); + } - /* Lower the load by pausing after every submitted - * write. */ - if (lh.load == LOW) - usleep(LOAD_HELPER_PAUSE_USEC); + if (spin[0]) { + handle = spin[0]->handle; + igt_spin_batch_end(spin[0]); + } + if (spin[1]) { + handle = spin[1]->handle; + igt_spin_batch_end(spin[1]); } /* Wait for completion without boosting */ usleep(1000); - while (gem_bo_busy(drm_fd, lh.target_buffer->handle)) + while (gem_bo_busy(drm_fd, handle)) usleep(1000); - igt_debug("load helper sent %u dword writes\n", val); - gem_close(drm_fd, fences[0]); - gem_close(drm_fd, fences[1]); - gem_close(drm_fd, fences[2]); - - /* Idle/boost logic is tied with request retirement. + /* + * Idle/boost logic is tied with request retirement. * Speed up detection of idle state and ensure deboost * after removing load. */ igt_drop_caches_set(drm_fd, DROP_RETIRE); + + igt_spin_batch_free(drm_fd, spin[1]); + igt_spin_batch_free(drm_fd, spin[0]); } } @@ -324,54 +271,6 @@ static void load_helper_stop(void) igt_assert(igt_wait_helper(&lh.igt_proc) == 0); } -static void load_helper_init(void) -{ - lh.devid = intel_get_drm_devid(drm_fd); - lh.has_ppgtt = gem_uses_ppgtt(drm_fd); - - /* MI_STORE_DATA can only use GTT address on gen4+/g33 and needs - * snoopable mem on pre-gen6. Hence load-helper only works on gen6+, but - * that's also all we care about for the rps testcase*/ - igt_assert(intel_gen(lh.devid) >= 6); - lh.bufmgr = drm_intel_bufmgr_gem_init(drm_fd, 4096); - igt_assert(lh.bufmgr); - - drm_intel_bufmgr_gem_enable_reuse(lh.bufmgr); - - lh.batch = intel_batchbuffer_alloc(lh.bufmgr, lh.devid); - igt_assert(lh.batch); - - lh.target_buffer = drm_intel_bo_alloc(lh.bufmgr, "target bo", - 4096, 4096); - igt_assert(lh.target_buffer); - - lh.dst = drm_intel_bo_alloc(lh.bufmgr, "dst bo", - LOAD_HELPER_BO_SIZE, 4096); - igt_assert(lh.dst); - lh.src = drm_intel_bo_alloc(lh.bufmgr, "src bo", - LOAD_HELPER_BO_SIZE, 4096); - igt_assert(lh.src); -} - -static void load_helper_deinit(void) -{ - if (lh.igt_proc.running) - load_helper_stop(); - - if (lh.target_buffer) - drm_intel_bo_unreference(lh.target_buffer); - if (lh.src) - drm_intel_bo_unreference(lh.src); - if (lh.dst) - drm_intel_bo_unreference(lh.dst); - - if (lh.batch) - intel_batchbuffer_free(lh.batch); - - if (lh.bufmgr) - drm_intel_bufmgr_destroy(lh.bufmgr); -} - static void do_load_gpu(void) { load_helper_run(LOW); @@ -582,13 +481,8 @@ static void boost_freq(int fd, int *boost_freqs) { int64_t timeout = 1; igt_spin_t *load; - unsigned int engine; - /* Put boost on the same engine as low load */ - engine = I915_EXEC_RENDER; - if (intel_gen(lh.devid) >= 6) - engine = I915_EXEC_BLT; - load = igt_spin_batch_new(fd, 0, engine, 0); + load = igt_spin_batch_new(fd, 0, 0, 0); /* Waiting will grant us a boost to maximum */ gem_wait(fd, load->handle, &timeout); @@ -655,7 +549,9 @@ static void pm_rps_exit_handler(int sig) writeval(sysfs_files[MAX].filp, origfreqs[MAX]); } - load_helper_deinit(); + if (lh.igt_proc.running) + load_helper_stop(); + close(drm_fd); } @@ -691,8 +587,6 @@ igt_main read_freqs(origfreqs); igt_install_exit_handler(pm_rps_exit_handler); - - load_helper_init(); } igt_subtest("basic-api") -- 2.15.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
