* IN THE SOFTWARE.
*/
+#include "common/gen_l3_config.h"
+
#include "brw_context.h"
#include "brw_defines.h"
#include "brw_state.h"
#include "intel_batchbuffer.h"
/**
- * Chunk of L3 cache reserved for some specific purpose.
- */
-enum brw_l3_partition {
- /** Shared local memory. */
- L3P_SLM = 0,
- /** Unified return buffer. */
- L3P_URB,
- /** Union of DC and RO. */
- L3P_ALL,
- /** Data cluster RW partition. */
- L3P_DC,
- /** Union of IS, C and T. */
- L3P_RO,
- /** Instruction and state cache. */
- L3P_IS,
- /** Constant cache. */
- L3P_C,
- /** Texture cache. */
- L3P_T,
- /** Number of supported L3 partitions. */
- NUM_L3P
-};
-
-/**
- * L3 configuration represented as the number of ways allocated for each
- * partition. \sa get_l3_way_size().
+ * Calculate the desired L3 partitioning based on the current state of the
+ * pipeline. For now this simply returns the conservative defaults calculated
+ * by get_default_l3_weights(), but we could probably do better by gathering
+ * more statistics from the pipeline state (e.g. guess of expected URB usage
+ * and bound surfaces), or by using feed-back from performance counters.
*/
-struct brw_l3_config {
- unsigned n[NUM_L3P];
-};
-
-/**
- * IVB/HSW validated L3 configurations.
- */
-static const struct brw_l3_config ivb_l3_configs[] = {
- /* SLM URB ALL DC RO IS C T */
- {{ 0, 32, 0, 0, 32, 0, 0, 0 }},
- {{ 0, 32, 0, 16, 16, 0, 0, 0 }},
- {{ 0, 32, 0, 4, 0, 8, 4, 16 }},
- {{ 0, 28, 0, 8, 0, 8, 4, 16 }},
- {{ 0, 28, 0, 16, 0, 8, 4, 8 }},
- {{ 0, 28, 0, 8, 0, 16, 4, 8 }},
- {{ 0, 28, 0, 0, 0, 16, 4, 16 }},
- {{ 0, 32, 0, 0, 0, 16, 0, 16 }},
- {{ 0, 28, 0, 4, 32, 0, 0, 0 }},
- {{ 16, 16, 0, 16, 16, 0, 0, 0 }},
- {{ 16, 16, 0, 8, 0, 8, 8, 8 }},
- {{ 16, 16, 0, 4, 0, 8, 4, 16 }},
- {{ 16, 16, 0, 4, 0, 16, 4, 8 }},
- {{ 16, 16, 0, 0, 32, 0, 0, 0 }},
- {{ 0 }}
-};
-
-/**
- * VLV validated L3 configurations.
- */
-static const struct brw_l3_config vlv_l3_configs[] = {
- /* SLM URB ALL DC RO IS C T */
- {{ 0, 64, 0, 0, 32, 0, 0, 0 }},
- {{ 0, 80, 0, 0, 16, 0, 0, 0 }},
- {{ 0, 80, 0, 8, 8, 0, 0, 0 }},
- {{ 0, 64, 0, 16, 16, 0, 0, 0 }},
- {{ 0, 60, 0, 4, 32, 0, 0, 0 }},
- {{ 32, 32, 0, 16, 16, 0, 0, 0 }},
- {{ 32, 40, 0, 8, 16, 0, 0, 0 }},
- {{ 32, 40, 0, 16, 8, 0, 0, 0 }},
- {{ 0 }}
-};
-
-/**
- * BDW validated L3 configurations.
- */
-static const struct brw_l3_config bdw_l3_configs[] = {
- /* SLM URB ALL DC RO IS C T */
- {{ 0, 48, 48, 0, 0, 0, 0, 0 }},
- {{ 0, 48, 0, 16, 32, 0, 0, 0 }},
- {{ 0, 32, 0, 16, 48, 0, 0, 0 }},
- {{ 0, 32, 0, 0, 64, 0, 0, 0 }},
- {{ 0, 32, 64, 0, 0, 0, 0, 0 }},
- {{ 24, 16, 48, 0, 0, 0, 0, 0 }},
- {{ 24, 16, 0, 16, 32, 0, 0, 0 }},
- {{ 24, 16, 0, 32, 16, 0, 0, 0 }},
- {{ 0 }}
-};
-
-/**
- * CHV/SKL validated L3 configurations.
- */
-static const struct brw_l3_config chv_l3_configs[] = {
- /* SLM URB ALL DC RO IS C T */
- {{ 0, 48, 48, 0, 0, 0, 0, 0 }},
- {{ 0, 48, 0, 16, 32, 0, 0, 0 }},
- {{ 0, 32, 0, 16, 48, 0, 0, 0 }},
- {{ 0, 32, 0, 0, 64, 0, 0, 0 }},
- {{ 0, 32, 64, 0, 0, 0, 0, 0 }},
- {{ 32, 16, 48, 0, 0, 0, 0, 0 }},
- {{ 32, 16, 0, 16, 32, 0, 0, 0 }},
- {{ 32, 16, 0, 32, 16, 0, 0, 0 }},
- {{ 0 }}
-};
-
-/**
- * Return a zero-terminated array of validated L3 configurations for the
- * specified device.
- */
-static const struct brw_l3_config *
-get_l3_configs(const struct brw_device_info *devinfo)
+static struct gen_l3_weights
+get_pipeline_state_l3_weights(const struct brw_context *brw)
{
- switch (devinfo->gen) {
- case 7:
- return (devinfo->is_baytrail ? vlv_l3_configs : ivb_l3_configs);
-
- case 8:
- return (devinfo->is_cherryview ? chv_l3_configs : bdw_l3_configs);
+ const struct brw_stage_state *stage_states[] = {
+ [MESA_SHADER_VERTEX] = &brw->vs.base,
+ [MESA_SHADER_TESS_CTRL] = &brw->tcs.base,
+ [MESA_SHADER_TESS_EVAL] = &brw->tes.base,
+ [MESA_SHADER_GEOMETRY] = &brw->gs.base,
+ [MESA_SHADER_FRAGMENT] = &brw->wm.base,
+ [MESA_SHADER_COMPUTE] = &brw->cs.base
+ };
+ bool needs_dc = false, needs_slm = false;
- case 9:
- return chv_l3_configs;
+ for (unsigned i = 0; i < ARRAY_SIZE(stage_states); i++) {
+ const struct gl_program *prog =
+ brw->ctx._Shader->CurrentProgram[stage_states[i]->stage];
+ const struct brw_stage_prog_data *prog_data = stage_states[i]->prog_data;
- default:
- unreachable("Not implemented");
+ needs_dc |= (prog && (prog->sh.data->NumAtomicBuffers ||
+ prog->sh.data->NumShaderStorageBlocks ||
+ prog->info.num_images)) ||
+ (prog_data && prog_data->total_scratch);
+ needs_slm |= prog_data && prog_data->total_shared;
}
-}
-
-/**
- * Return the size of an L3 way in KB.
- */
-static unsigned
-get_l3_way_size(const struct brw_device_info *devinfo)
-{
- if (devinfo->is_baytrail)
- return 2;
- else if (devinfo->is_cherryview || devinfo->gt == 1)
- return 4;
-
- else
- return 8 * devinfo->num_slices;
+ return gen_get_default_l3_weights(&brw->screen->devinfo,
+ needs_dc, needs_slm);
}
/**
* Program the hardware to use the specified L3 configuration.
*/
static void
-setup_l3_config(struct brw_context *brw, const struct brw_l3_config *cfg)
+setup_l3_config(struct brw_context *brw, const struct gen_l3_config *cfg)
{
- const bool has_dc = cfg->n[L3P_DC] || cfg->n[L3P_ALL];
- const bool has_is = cfg->n[L3P_IS] || cfg->n[L3P_RO] || cfg->n[L3P_ALL];
- const bool has_c = cfg->n[L3P_C] || cfg->n[L3P_RO] || cfg->n[L3P_ALL];
- const bool has_t = cfg->n[L3P_T] || cfg->n[L3P_RO] || cfg->n[L3P_ALL];
- const bool has_slm = cfg->n[L3P_SLM];
+ const struct gen_device_info *devinfo = &brw->screen->devinfo;
+ const bool has_dc = cfg->n[GEN_L3P_DC] || cfg->n[GEN_L3P_ALL];
+ const bool has_is = cfg->n[GEN_L3P_IS] || cfg->n[GEN_L3P_RO] ||
+ cfg->n[GEN_L3P_ALL];
+ const bool has_c = cfg->n[GEN_L3P_C] || cfg->n[GEN_L3P_RO] ||
+ cfg->n[GEN_L3P_ALL];
+ const bool has_t = cfg->n[GEN_L3P_T] || cfg->n[GEN_L3P_RO] ||
+ cfg->n[GEN_L3P_ALL];
+ const bool has_slm = cfg->n[GEN_L3P_SLM];
/* According to the hardware docs, the L3 partitioning can only be changed
* while the pipeline is completely drained and the caches are flushed,
- * which involves a first PIPE_CONTROL flush which stalls the pipeline and
- * initiates invalidation of the relevant caches...
+ * which involves a first PIPE_CONTROL flush which stalls the pipeline...
+ */
+ brw_emit_pipe_control_flush(brw,
+ PIPE_CONTROL_DATA_CACHE_FLUSH |
+ PIPE_CONTROL_NO_WRITE |
+ PIPE_CONTROL_CS_STALL);
+
+ /* ...followed by a second pipelined PIPE_CONTROL that initiates
+ * invalidation of the relevant caches. Note that because RO invalidation
+ * happens at the top of the pipeline (i.e. right away as the PIPE_CONTROL
+ * command is processed by the CS) we cannot combine it with the previous
+ * stalling flush as the hardware documentation suggests, because that
+ * would cause the CS to stall on previous rendering *after* RO
+ * invalidation and wouldn't prevent the RO caches from being polluted by
+ * concurrent rendering before the stall completes. This intentionally
+ * doesn't implement the SKL+ hardware workaround suggesting to enable CS
+ * stall on PIPE_CONTROLs with the texture cache invalidation bit set for
+ * GPGPU workloads because the previous and subsequent PIPE_CONTROLs
+ * already guarantee that there is no concurrent GPGPU kernel execution
+ * (see SKL HSD 2132585).
*/
brw_emit_pipe_control_flush(brw,
PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
PIPE_CONTROL_CONST_CACHE_INVALIDATE |
PIPE_CONTROL_INSTRUCTION_INVALIDATE |
- PIPE_CONTROL_DATA_CACHE_INVALIDATE |
- PIPE_CONTROL_NO_WRITE |
- PIPE_CONTROL_CS_STALL);
+ PIPE_CONTROL_STATE_CACHE_INVALIDATE |
+ PIPE_CONTROL_NO_WRITE);
- /* ...followed by a second stalling flush which guarantees that
- * invalidation is complete when the L3 configuration registers are
- * modified.
+ /* Now send a third stalling flush to make sure that invalidation is
+ * complete when the L3 configuration registers are modified.
*/
brw_emit_pipe_control_flush(brw,
- PIPE_CONTROL_DATA_CACHE_INVALIDATE |
+ PIPE_CONTROL_DATA_CACHE_FLUSH |
PIPE_CONTROL_NO_WRITE |
PIPE_CONTROL_CS_STALL);
- if (brw->gen >= 8) {
- assert(!cfg->n[L3P_IS] && !cfg->n[L3P_C] && !cfg->n[L3P_T]);
+ if (devinfo->gen >= 8) {
+ assert(!cfg->n[GEN_L3P_IS] && !cfg->n[GEN_L3P_C] && !cfg->n[GEN_L3P_T]);
BEGIN_BATCH(3);
OUT_BATCH(MI_LOAD_REGISTER_IMM | (3 - 2));
/* Set up the L3 partitioning. */
OUT_BATCH(GEN8_L3CNTLREG);
OUT_BATCH((has_slm ? GEN8_L3CNTLREG_SLM_ENABLE : 0) |
- SET_FIELD(cfg->n[L3P_URB], GEN8_L3CNTLREG_URB_ALLOC) |
- SET_FIELD(cfg->n[L3P_RO], GEN8_L3CNTLREG_RO_ALLOC) |
- SET_FIELD(cfg->n[L3P_DC], GEN8_L3CNTLREG_DC_ALLOC) |
- SET_FIELD(cfg->n[L3P_ALL], GEN8_L3CNTLREG_ALL_ALLOC));
+ SET_FIELD(cfg->n[GEN_L3P_URB], GEN8_L3CNTLREG_URB_ALLOC) |
+ SET_FIELD(cfg->n[GEN_L3P_RO], GEN8_L3CNTLREG_RO_ALLOC) |
+ SET_FIELD(cfg->n[GEN_L3P_DC], GEN8_L3CNTLREG_DC_ALLOC) |
+ SET_FIELD(cfg->n[GEN_L3P_ALL], GEN8_L3CNTLREG_ALL_ALLOC));
ADVANCE_BATCH();
} else {
- assert(!cfg->n[L3P_ALL]);
+ assert(!cfg->n[GEN_L3P_ALL]);
/* When enabled SLM only uses a portion of the L3 on half of the banks,
* the matching space on the remaining banks has to be allocated to a
* client (URB for all validated configurations) set to the
* lower-bandwidth 2-bank address hashing mode.
*/
- const bool urb_low_bw = has_slm && !brw->is_baytrail;
- assert(!urb_low_bw || cfg->n[L3P_URB] == cfg->n[L3P_SLM]);
+ const bool urb_low_bw = has_slm && !devinfo->is_baytrail;
+ assert(!urb_low_bw || cfg->n[GEN_L3P_URB] == cfg->n[GEN_L3P_SLM]);
/* Minimum number of ways that can be allocated to the URB. */
- const unsigned n0_urb = (brw->is_baytrail ? 32 : 0);
- assert(cfg->n[L3P_URB] >= n0_urb);
+ const unsigned n0_urb = (devinfo->is_baytrail ? 32 : 0);
+ assert(cfg->n[GEN_L3P_URB] >= n0_urb);
BEGIN_BATCH(7);
OUT_BATCH(MI_LOAD_REGISTER_IMM | (7 - 2));
/* Demote any clients with no ways assigned to LLC. */
OUT_BATCH(GEN7_L3SQCREG1);
- OUT_BATCH((brw->is_haswell ? HSW_L3SQCREG1_SQGHPCI_DEFAULT :
- brw->is_baytrail ? VLV_L3SQCREG1_SQGHPCI_DEFAULT :
+ OUT_BATCH((devinfo->is_haswell ? HSW_L3SQCREG1_SQGHPCI_DEFAULT :
+ devinfo->is_baytrail ? VLV_L3SQCREG1_SQGHPCI_DEFAULT :
IVB_L3SQCREG1_SQGHPCI_DEFAULT) |
(has_dc ? 0 : GEN7_L3SQCREG1_CONV_DC_UC) |
(has_is ? 0 : GEN7_L3SQCREG1_CONV_IS_UC) |
/* Set up the L3 partitioning. */
OUT_BATCH(GEN7_L3CNTLREG2);
OUT_BATCH((has_slm ? GEN7_L3CNTLREG2_SLM_ENABLE : 0) |
- SET_FIELD(cfg->n[L3P_URB] - n0_urb, GEN7_L3CNTLREG2_URB_ALLOC) |
+ SET_FIELD(cfg->n[GEN_L3P_URB] - n0_urb, GEN7_L3CNTLREG2_URB_ALLOC) |
(urb_low_bw ? GEN7_L3CNTLREG2_URB_LOW_BW : 0) |
- SET_FIELD(cfg->n[L3P_ALL], GEN7_L3CNTLREG2_ALL_ALLOC) |
- SET_FIELD(cfg->n[L3P_RO], GEN7_L3CNTLREG2_RO_ALLOC) |
- SET_FIELD(cfg->n[L3P_DC], GEN7_L3CNTLREG2_DC_ALLOC));
+ SET_FIELD(cfg->n[GEN_L3P_ALL], GEN7_L3CNTLREG2_ALL_ALLOC) |
+ SET_FIELD(cfg->n[GEN_L3P_RO], GEN7_L3CNTLREG2_RO_ALLOC) |
+ SET_FIELD(cfg->n[GEN_L3P_DC], GEN7_L3CNTLREG2_DC_ALLOC));
OUT_BATCH(GEN7_L3CNTLREG3);
- OUT_BATCH(SET_FIELD(cfg->n[L3P_IS], GEN7_L3CNTLREG3_IS_ALLOC) |
- SET_FIELD(cfg->n[L3P_C], GEN7_L3CNTLREG3_C_ALLOC) |
- SET_FIELD(cfg->n[L3P_T], GEN7_L3CNTLREG3_T_ALLOC));
+ OUT_BATCH(SET_FIELD(cfg->n[GEN_L3P_IS], GEN7_L3CNTLREG3_IS_ALLOC) |
+ SET_FIELD(cfg->n[GEN_L3P_C], GEN7_L3CNTLREG3_C_ALLOC) |
+ SET_FIELD(cfg->n[GEN_L3P_T], GEN7_L3CNTLREG3_T_ALLOC));
ADVANCE_BATCH();
- if (brw->is_haswell && brw->intelScreen->cmd_parser_version >= 4) {
+ if (can_do_hsw_l3_atomics(brw->screen)) {
/* Enable L3 atomics on HSW if we have a DC partition, otherwise keep
* them disabled to avoid crashing the system hard.
*/
}
}
}
+
+/**
+ * Update the URB size in the context state for the specified L3
+ * configuration.
+ */
+static void
+update_urb_size(struct brw_context *brw, const struct gen_l3_config *cfg)
+{
+ const struct gen_device_info *devinfo = &brw->screen->devinfo;
+ const unsigned sz = gen_get_l3_config_urb_size(devinfo, cfg);
+
+ if (brw->urb.size != sz) {
+ brw->urb.size = sz;
+ brw->ctx.NewDriverState |= BRW_NEW_URB_SIZE;
+
+ /* If we change the total URB size, reset the individual stage sizes to
+ * zero so that, even if there is no URB size change, gen7_upload_urb
+ * still re-emits 3DSTATE_URB_*.
+ */
+ brw->urb.vsize = 0;
+ brw->urb.gsize = 0;
+ brw->urb.hsize = 0;
+ brw->urb.dsize = 0;
+ }
+}
+
+static void
+emit_l3_state(struct brw_context *brw)
+{
+ const struct gen_l3_weights w = get_pipeline_state_l3_weights(brw);
+ const float dw = gen_diff_l3_weights(w, gen_get_l3_config_weights(brw->l3.config));
+ /* The distance between any two compatible weight vectors cannot exceed two
+ * due to the triangle inequality.
+ */
+ const float large_dw_threshold = 2.0;
+ /* Somewhat arbitrary, simply makes sure that there will be no repeated
+ * transitions to the same L3 configuration, could probably do better here.
+ */
+ const float small_dw_threshold = 0.5;
+ /* If we're emitting a new batch the caches should already be clean and the
+ * transition should be relatively cheap, so it shouldn't hurt much to use
+ * the smaller threshold. Otherwise use the larger threshold so that we
+ * only reprogram the L3 mid-batch if the most recently programmed
+ * configuration is incompatible with the current pipeline state.
+ */
+ const float dw_threshold = (brw->ctx.NewDriverState & BRW_NEW_BATCH ?
+ small_dw_threshold : large_dw_threshold);
+
+ if (dw > dw_threshold && can_do_pipelined_register_writes(brw->screen)) {
+ const struct gen_l3_config *const cfg =
+ gen_get_l3_config(&brw->screen->devinfo, w);
+
+ setup_l3_config(brw, cfg);
+ update_urb_size(brw, cfg);
+ brw->l3.config = cfg;
+
+ if (unlikely(INTEL_DEBUG & DEBUG_L3)) {
+ fprintf(stderr, "L3 config transition (%f > %f): ", dw, dw_threshold);
+ gen_dump_l3_config(cfg, stderr);
+ }
+ }
+}
+
+const struct brw_tracked_state gen7_l3_state = {
+ .dirty = {
+ .mesa = 0,
+ .brw = BRW_NEW_BATCH |
+ BRW_NEW_BLORP |
+ BRW_NEW_CS_PROG_DATA |
+ BRW_NEW_FS_PROG_DATA |
+ BRW_NEW_GS_PROG_DATA |
+ BRW_NEW_VS_PROG_DATA,
+ },
+ .emit = emit_l3_state
+};
+
+/**
+ * Hack to restore the default L3 configuration.
+ *
+ * This will be called at the end of every batch in order to reset the L3
+ * configuration to the default values for the time being until the kernel is
+ * fixed. Until kernel commit 6702cf16e0ba8b0129f5aa1b6609d4e9c70bc13b
+ * (included in v4.1) we would set the MI_RESTORE_INHIBIT bit when submitting
+ * batch buffers for the default context used by the DDX, which meant that any
+ * context state changed by the GL would leak into the DDX, the assumption
+ * being that the DDX would initialize any state it cares about manually. The
+ * DDX is however not careful enough to program an L3 configuration
+ * explicitly, and it makes assumptions about it (URB size) which won't hold
+ * and cause it to misrender if we let our L3 set-up to leak into the DDX.
+ *
+ * Since v4.1 of the Linux kernel the default context is saved and restored
+ * normally, so it's far less likely for our L3 programming to interfere with
+ * other contexts -- In fact restoring the default L3 configuration at the end
+ * of the batch will be redundant most of the time. A kind of state leak is
+ * still possible though if the context making assumptions about L3 state is
+ * created immediately after our context was active (e.g. without the DDX
+ * default context being scheduled in between) because at present the DRM
+ * doesn't fully initialize the contents of newly created contexts and instead
+ * sets the MI_RESTORE_INHIBIT flag causing it to inherit the state from the
+ * last active context.
+ *
+ * It's possible to realize such a scenario if, say, an X server (or a GL
+ * application using an outdated non-L3-aware Mesa version) is started while
+ * another GL application is running and happens to have modified the L3
+ * configuration, or if no X server is running at all and a GL application
+ * using a non-L3-aware Mesa version is started after another GL application
+ * ran and modified the L3 configuration -- The latter situation can actually
+ * be reproduced easily on IVB in our CI system.
+ */
+void
+gen7_restore_default_l3_config(struct brw_context *brw)
+{
+ const struct gen_device_info *devinfo = &brw->screen->devinfo;
+ const struct gen_l3_config *const cfg = gen_get_default_l3_config(devinfo);
+
+ if (cfg != brw->l3.config &&
+ can_do_pipelined_register_writes(brw->screen)) {
+ setup_l3_config(brw, cfg);
+ update_urb_size(brw, cfg);
+ brw->l3.config = cfg;
+ }
+}