drm/i915/selftests: Unroll the CS frequency loop
authorChris Wilson <chris@chris-wilson.co.uk>
Tue, 21 Apr 2020 17:13:51 +0000 (18:13 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Tue, 21 Apr 2020 19:48:45 +0000 (20:48 +0100)
Having noticed that MI_BB_START is incurring a memory stall (see the
correlation with uncore frequency), we have to unroll the loop in order
to diminish the impact of the MI_BB_START on the instruction throughput.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200421171351.19575-1-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/gt/selftest_rps.c

index e2afc20..0d7ed00 100644 (file)
@@ -49,14 +49,17 @@ create_spin_counter(struct intel_engine_cs *engine,
 #define CS_GPR(x) GEN8_RING_CS_GPR(engine->mmio_base, x)
        struct drm_i915_gem_object *obj;
        struct i915_vma *vma;
+       unsigned long end;
        u32 *base, *cs;
        int loop, i;
        int err;
 
-       obj = i915_gem_object_create_internal(vm->i915, 4096);
+       obj = i915_gem_object_create_internal(vm->i915, 64 << 10);
        if (IS_ERR(obj))
                return ERR_CAST(obj);
 
+       end = obj->base.size / sizeof(u32) - 1;
+
        vma = i915_vma_instance(obj, vm, NULL);
        if (IS_ERR(vma)) {
                i915_gem_object_put(obj);
@@ -90,27 +93,31 @@ create_spin_counter(struct intel_engine_cs *engine,
 
        loop = cs - base;
 
-       *cs++ = MI_MATH(4);
-       *cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCA, MI_MATH_REG(COUNT));
-       *cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCB, MI_MATH_REG(INC));
-       *cs++ = MI_MATH_ADD;
-       *cs++ = MI_MATH_STORE(MI_MATH_REG(COUNT), MI_MATH_REG_ACCU);
-
-       if (srm) {
-               *cs++ = MI_STORE_REGISTER_MEM_GEN8;
-               *cs++ = i915_mmio_reg_offset(CS_GPR(COUNT));
-               *cs++ = lower_32_bits(vma->node.start + 1000 * sizeof(*cs));
-               *cs++ = upper_32_bits(vma->node.start + 1000 * sizeof(*cs));
+       /* Unroll the loop to avoid MI_BB_START stalls impacting measurements */
+       for (i = 0; i < 1024; i++) {
+               *cs++ = MI_MATH(4);
+               *cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCA, MI_MATH_REG(COUNT));
+               *cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCB, MI_MATH_REG(INC));
+               *cs++ = MI_MATH_ADD;
+               *cs++ = MI_MATH_STORE(MI_MATH_REG(COUNT), MI_MATH_REG_ACCU);
+
+               if (srm) {
+                       *cs++ = MI_STORE_REGISTER_MEM_GEN8;
+                       *cs++ = i915_mmio_reg_offset(CS_GPR(COUNT));
+                       *cs++ = lower_32_bits(vma->node.start + end * sizeof(*cs));
+                       *cs++ = upper_32_bits(vma->node.start + end * sizeof(*cs));
+               }
        }
 
        *cs++ = MI_BATCH_BUFFER_START_GEN8;
        *cs++ = lower_32_bits(vma->node.start + loop * sizeof(*cs));
        *cs++ = upper_32_bits(vma->node.start + loop * sizeof(*cs));
+       GEM_BUG_ON(cs - base > end);
 
        i915_gem_object_flush_map(obj);
 
        *cancel = base + loop;
-       *counter = srm ? memset32(base + 1000, 0, 1) : NULL;
+       *counter = srm ? memset32(base + end, 0, 1) : NULL;
        return vma;
 }