[Mesa-dev] [PATCH] nv50: fix PIPE_QUERY_TIMESTAMP_DISJOINT, based on nvc0

Samuel Pitoiset samuel.pitoiset at gmail.com
Sat May 23 09:46:55 PDT 2015


PIPE_QUERY_TIMESTAMP_DISJOINT could not work because q->ready was always
set to FALSE. To fix this issue, add more different states for queries
according to nvc0.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
---
 src/gallium/drivers/nouveau/nv50/nv50_query.c      |  39 +-
 .../drivers/nouveau/nvc0/nvc0_query_hw_sm.c        | 798 +++++++++++++++++++++
 2 files changed, 820 insertions(+), 17 deletions(-)
 create mode 100644 src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_sm.c

diff --git a/src/gallium/drivers/nouveau/nv50/nv50_query.c b/src/gallium/drivers/nouveau/nv50/nv50_query.c
index da41209..55fcac8 100644
--- a/src/gallium/drivers/nouveau/nv50/nv50_query.c
+++ b/src/gallium/drivers/nouveau/nv50/nv50_query.c
@@ -27,6 +27,11 @@
 #include "nv50/nv50_context.h"
 #include "nv_object.xml.h"
 
+#define NV50_QUERY_STATE_READY   0
+#define NV50_QUERY_STATE_ACTIVE  1
+#define NV50_QUERY_STATE_ENDED   2
+#define NV50_QUERY_STATE_FLUSHED 3
+
 /* XXX: Nested queries, and simultaneous queries on multiple gallium contexts
  * (since we use only a single GPU channel per screen) will not work properly.
  *
@@ -42,8 +47,7 @@ struct nv50_query {
    struct nouveau_bo *bo;
    uint32_t base;
    uint32_t offset; /* base + i * 32 */
-   boolean ready;
-   boolean flushed;
+   uint8_t state;
    boolean is64bit;
    struct nouveau_mm_allocation *mm;
    struct nouveau_fence *fence;
@@ -66,7 +70,7 @@ nv50_query_allocate(struct nv50_context *nv50, struct nv50_query *q, int size)
    if (q->bo) {
       nouveau_bo_ref(NULL, &q->bo);
       if (q->mm) {
-         if (q->ready)
+         if (q->state == NV50_QUERY_STATE_READY)
             nouveau_mm_free(q->mm);
          else
             nouveau_fence_work(screen->base.fence.current, nouveau_mm_free_work,
@@ -203,7 +207,7 @@ nv50_query_begin(struct pipe_context *pipe, struct pipe_query *pq)
    default:
       break;
    }
-   q->ready = FALSE;
+   q->state = NV50_QUERY_STATE_ACTIVE;
    return true;
 }
 
@@ -214,6 +218,8 @@ nv50_query_end(struct pipe_context *pipe, struct pipe_query *pq)
    struct nouveau_pushbuf *push = nv50->base.pushbuf;
    struct nv50_query *q = nv50_query(pq);
 
+   q->state = NV50_QUERY_STATE_ENDED;
+
    switch (q->type) {
    case PIPE_QUERY_OCCLUSION_COUNTER:
       nv50_query_get(push, q, 0, 0x0100f002);
@@ -256,29 +262,27 @@ nv50_query_end(struct pipe_context *pipe, struct pipe_query *pq)
       break;
    case PIPE_QUERY_TIMESTAMP_DISJOINT:
       /* This query is not issued on GPU because disjoint is forced to FALSE */
-      q->ready = TRUE;
+      q->state = NV50_QUERY_STATE_READY;
       break;
    default:
       assert(0);
       break;
    }
-   q->ready = q->flushed = FALSE;
 
    if (q->is64bit)
       nouveau_fence_ref(nv50->screen->base.fence.current, &q->fence);
 }
 
-static INLINE boolean
-nv50_query_ready(struct nv50_query *q)
+static INLINE void
+nv50_query_update(struct nv50_query *q)
 {
    if (q->is64bit) {
       if (nouveau_fence_signalled(q->fence))
-         return TRUE;
+         q->state = NV50_QUERY_STATE_READY;
    } else {
       if (q->data[0] == q->sequence)
-         return TRUE;
+         q->state = NV50_QUERY_STATE_READY;
    }
-   return FALSE;
 }
 
 static boolean
@@ -293,13 +297,14 @@ nv50_query_result(struct pipe_context *pipe, struct pipe_query *pq,
    uint64_t *data64 = (uint64_t *)q->data;
    int i;
 
-   if (!q->ready) /* update ? */
-      q->ready = nv50_query_ready(q);
-   if (!q->ready) {
+   if (q->state != NV50_QUERY_STATE_READY)
+      nv50_query_update(q);
+
+   if (q->state != NV50_QUERY_STATE_READY) {
       if (!wait) {
          /* for broken apps that spin on GL_QUERY_RESULT_AVAILABLE */
-         if (!q->flushed) {
-            q->flushed = TRUE;
+         if (q->state != NV50_QUERY_STATE_FLUSHED) {
+            q->state = NV50_QUERY_STATE_FLUSHED;
             PUSH_KICK(nv50->base.pushbuf);
          }
          return FALSE;
@@ -307,7 +312,7 @@ nv50_query_result(struct pipe_context *pipe, struct pipe_query *pq,
       if (nouveau_bo_wait(q->bo, NOUVEAU_BO_RD, nv50->screen->base.client))
          return FALSE;
    }
-   q->ready = TRUE;
+   q->state = NV50_QUERY_STATE_READY;
 
    switch (q->type) {
    case PIPE_QUERY_GPU_FINISHED:
diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_sm.c b/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_sm.c
new file mode 100644
index 0000000..668f0fb
--- /dev/null
+++ b/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_sm.c
@@ -0,0 +1,798 @@
+/*
+ * Copyright 2011 Nouveau Project
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * Authors: Christoph Bumiller
+ */
+
+/*
+ * This implements hardware SM performance counters.
+ */
+
+#define NVC0_PUSH_EXPLICIT_SPACE_CHECKING
+
+#include "nvc0/nvc0_context.h"
+#include "nvc0/nvc0_query.h"
+
+#include "nv_object.xml.h"
+#include "nvc0/nve4_compute.xml.h"
+#include "nvc0/nvc0_compute.xml.h"
+
+/* === PERFORMANCE MONITORING COUNTERS for NVE4+ === */
+
+/* Code to read out MP counters: They are accessible via mmio, too, but let's
+ * just avoid mapping registers in userspace. We'd have to know which MPs are
+ * enabled/present, too, and that information is not presently exposed.
+ * We could add a kernel interface for it, but reading the counters like this
+ * has the advantage of being async (if get_result isn't called immediately).
+ */
+static const uint64_t nve4_read_hw_sm_counters_code[] =
+{
+   /* sched 0x20 0x20 0x20 0x20 0x20 0x20 0x20
+    * mov b32 $r8 $tidx
+    * mov b32 $r12 $physid
+    * mov b32 $r0 $pm0
+    * mov b32 $r1 $pm1
+    * mov b32 $r2 $pm2
+    * mov b32 $r3 $pm3
+    * mov b32 $r4 $pm4
+    * sched 0x20 0x20 0x23 0x04 0x20 0x04 0x2b
+    * mov b32 $r5 $pm5
+    * mov b32 $r6 $pm6
+    * mov b32 $r7 $pm7
+    * set $p0 0x1 eq u32 $r8 0x0
+    * mov b32 $r10 c0[0x0]
+    * ext u32 $r8 $r12 0x414
+    * mov b32 $r11 c0[0x4]
+    * sched 0x04 0x2e 0x04 0x20 0x20 0x28 0x04
+    * ext u32 $r9 $r12 0x208
+    * (not $p0) exit
+    * set $p1 0x1 eq u32 $r9 0x0
+    * mul $r8 u32 $r8 u32 96
+    * mul $r12 u32 $r9 u32 16
+    * mul $r13 u32 $r9 u32 4
+    * add b32 $r9 $r8 $r13
+    * sched 0x28 0x04 0x2c 0x04 0x2c 0x04 0x2c
+    * add b32 $r8 $r8 $r12
+    * mov b32 $r12 $r10
+    * add b32 $r10 $c $r10 $r8
+    * mov b32 $r13 $r11
+    * add b32 $r11 $r11 0x0 $c
+    * add b32 $r12 $c $r12 $r9
+    * st b128 wt g[$r10d] $r0q
+    * sched 0x4 0x2c 0x20 0x04 0x2e 0x00 0x00
+    * mov b32 $r0 c0[0x8]
+    * add b32 $r13 $r13 0x0 $c
+    * $p1 st b128 wt g[$r12d+0x40] $r4q
+    * st b32 wt g[$r12d+0x50] $r0
+    * exit */
+   0x2202020202020207ULL,
+   0x2c00000084021c04ULL,
+   0x2c0000000c031c04ULL,
+   0x2c00000010001c04ULL,
+   0x2c00000014005c04ULL,
+   0x2c00000018009c04ULL,
+   0x2c0000001c00dc04ULL,
+   0x2c00000020011c04ULL,
+   0x22b0420042320207ULL,
+   0x2c00000024015c04ULL,
+   0x2c00000028019c04ULL,
+   0x2c0000002c01dc04ULL,
+   0x190e0000fc81dc03ULL,
+   0x2800400000029de4ULL,
+   0x7000c01050c21c03ULL,
+   0x280040001002dde4ULL,
+   0x204282020042e047ULL,
+   0x7000c00820c25c03ULL,
+   0x80000000000021e7ULL,
+   0x190e0000fc93dc03ULL,
+   0x1000000180821c02ULL,
+   0x1000000040931c02ULL,
+   0x1000000010935c02ULL,
+   0x4800000034825c03ULL,
+   0x22c042c042c04287ULL,
+   0x4800000030821c03ULL,
+   0x2800000028031de4ULL,
+   0x4801000020a29c03ULL,
+   0x280000002c035de4ULL,
+   0x0800000000b2dc42ULL,
+   0x4801000024c31c03ULL,
+   0x9400000000a01fc5ULL,
+   0x200002e04202c047ULL,
+   0x2800400020001de4ULL,
+   0x0800000000d35c42ULL,
+   0x9400000100c107c5ULL,
+   0x9400000140c01f85ULL,
+   0x8000000000001de7ULL
+};
+
+/* NOTE: intentionally using the same names as NV */
+static const char *nve4_hw_sm_query_names[] =
+{
+   /* MP counters */
+   "prof_trigger_00",
+   "prof_trigger_01",
+   "prof_trigger_02",
+   "prof_trigger_03",
+   "prof_trigger_04",
+   "prof_trigger_05",
+   "prof_trigger_06",
+   "prof_trigger_07",
+   "warps_launched",
+   "threads_launched",
+   "sm_cta_launched",
+   "inst_issued1",
+   "inst_issued2",
+   "inst_executed",
+   "local_load",
+   "local_store",
+   "shared_load",
+   "shared_store",
+   "l1_local_load_hit",
+   "l1_local_load_miss",
+   "l1_local_store_hit",
+   "l1_local_store_miss",
+   "gld_request",
+   "gst_request",
+   "l1_global_load_hit",
+   "l1_global_load_miss",
+   "uncached_global_load_transaction",
+   "global_store_transaction",
+   "branch",
+   "divergent_branch",
+   "active_warps",
+   "active_cycles",
+   "inst_issued",
+   "atom_count",
+   "gred_count",
+   "shared_load_replay",
+   "shared_store_replay",
+   "local_load_transactions",
+   "local_store_transactions",
+   "l1_shared_load_transactions",
+   "l1_shared_store_transactions",
+   "global_ld_mem_divergence_replays",
+   "global_st_mem_divergence_replays",
+   /* metrics, i.e. functions of the MP counters */
+   "metric-ipc",                   /* inst_executed, clock */
+   "metric-ipac",                  /* inst_executed, active_cycles */
+   "metric-ipec",                  /* inst_executed, (bool)inst_executed */
+   "metric-achieved_occupancy",    /* active_warps, active_cycles */
+   "metric-sm_efficiency",         /* active_cycles, clock */
+   "metric-inst_replay_overhead"   /* inst_issued, inst_executed */
+};
+
+/* For simplicity, we will allocate as many group slots as we allocate counter
+ * slots. This means that a single counter which wants to source from 2 groups
+ * will have to be declared as using 2 counter slots. This shouldn't really be
+ * a problem because such queries don't make much sense ... (unless someone is
+ * really creative).
+ */
+struct nvc0_mp_counter_cfg
+{
+   uint32_t func    : 16; /* mask or 4-bit logic op (depending on mode) */
+   uint32_t mode    : 4;  /* LOGOP,B6,LOGOP_B6(_PULSE) */
+   uint32_t num_src : 3;  /* number of sources (1 - 6, only for NVC0:NVE4) */
+   uint32_t sig_dom : 1;  /* if 0, MP_PM_A (per warp-sched), if 1, MP_PM_B */
+   uint32_t sig_sel : 8;  /* signal group */
+   uint64_t src_sel;      /* signal selection for up to 6 sources (48 bit) */
+};
+
+#define NVC0_COUNTER_OPn_SUM            0
+#define NVC0_COUNTER_OPn_OR             1
+#define NVC0_COUNTER_OPn_AND            2
+#define NVC0_COUNTER_OP2_REL_SUM_MM     3 /* (sum(ctr0) - sum(ctr1)) / sum(ctr0) */
+#define NVC0_COUNTER_OP2_DIV_SUM_M0     4 /* sum(ctr0) / ctr1 of MP[0]) */
+#define NVC0_COUNTER_OP2_AVG_DIV_MM     5 /* avg(ctr0 / ctr1) */
+#define NVC0_COUNTER_OP2_AVG_DIV_M0     6 /* avg(ctr0) / ctr1 of MP[0]) */
+
+struct nvc0_hw_sm_query_cfg
+{
+   struct nvc0_mp_counter_cfg ctr[4];
+   uint8_t num_counters;
+   uint8_t op;
+   uint8_t norm[2]; /* normalization num,denom */
+};
+
+#define _Q1A(n, f, m, g, s, nu, dn) [NVE4_PM_QUERY_##n] = { { { f, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m, 0, 0, NVE4_COMPUTE_MP_PM_A_SIGSEL_##g, s }, {}, {}, {} }, 1, NVC0_COUNTER_OPn_SUM, { nu, dn } }
+#define _Q1B(n, f, m, g, s, nu, dn) [NVE4_PM_QUERY_##n] = { { { f, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m, 0, 1, NVE4_COMPUTE_MP_PM_B_SIGSEL_##g, s }, {}, {}, {} }, 1, NVC0_COUNTER_OPn_SUM, { nu, dn } }
+#define _M2A(n, f0, m0, g0, s0, f1, m1, g1, s1, o, nu, dn) [NVE4_PM_QUERY_METRIC_##n] = { { \
+   { f0, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m0, 0, 0, NVE4_COMPUTE_MP_PM_A_SIGSEL_##g0, s0 }, \
+   { f1, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m1, 0, 0, NVE4_COMPUTE_MP_PM_A_SIGSEL_##g1, s1 }, \
+   {}, {}, }, 2, NVC0_COUNTER_OP2_##o, { nu, dn } }
+#define _M2B(n, f0, m0, g0, s0, f1, m1, g1, s1, o, nu, dn) [NVE4_PM_QUERY_METRIC_##n] = { { \
+   { f0, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m0, 0, 1, NVE4_COMPUTE_MP_PM_B_SIGSEL_##g0, s0 }, \
+   { f1, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m1, 0, 1, NVE4_COMPUTE_MP_PM_B_SIGSEL_##g1, s1 }, \
+   {}, {}, }, 2, NVC0_COUNTER_OP2_##o, { nu, dn } }
+#define _M2AB(n, f0, m0, g0, s0, f1, m1, g1, s1, o, nu, dn) [NVE4_PM_QUERY_METRIC_##n] = { { \
+   { f0, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m0, 0, 0, NVE4_COMPUTE_MP_PM_A_SIGSEL_##g0, s0 }, \
+   { f1, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m1, 0, 1, NVE4_COMPUTE_MP_PM_B_SIGSEL_##g1, s1 }, \
+   {}, {}, }, 2, NVC0_COUNTER_OP2_##o, { nu, dn } }
+
+/* NOTES:
+ * active_warps: bit 0 alternates btw 0 and 1 for odd nr of warps
+ * inst_executed etc.: we only count a single warp scheduler
+ * metric-ipXc: we simply multiply by 4 to account for the 4 warp schedulers;
+ *  this is inaccurate !
+ */
+static const struct nvc0_hw_sm_query_cfg nve4_hw_sm_queries[] =
+{
+   _Q1A(PROF_TRIGGER_0, 0x0001, B6, USER, 0x00000000, 1, 1),
+   _Q1A(PROF_TRIGGER_1, 0x0001, B6, USER, 0x00000004, 1, 1),
+   _Q1A(PROF_TRIGGER_2, 0x0001, B6, USER, 0x00000008, 1, 1),
+   _Q1A(PROF_TRIGGER_3, 0x0001, B6, USER, 0x0000000c, 1, 1),
+   _Q1A(PROF_TRIGGER_4, 0x0001, B6, USER, 0x00000010, 1, 1),
+   _Q1A(PROF_TRIGGER_5, 0x0001, B6, USER, 0x00000014, 1, 1),
+   _Q1A(PROF_TRIGGER_6, 0x0001, B6, USER, 0x00000018, 1, 1),
+   _Q1A(PROF_TRIGGER_7, 0x0001, B6, USER, 0x0000001c, 1, 1),
+   _Q1A(LAUNCHED_WARPS,    0x0001, B6, LAUNCH, 0x00000004, 1, 1),
+   _Q1A(LAUNCHED_THREADS,  0x003f, B6, LAUNCH, 0x398a4188, 1, 1),
+   _Q1B(LAUNCHED_CTA,      0x0001, B6, WARP, 0x0000001c, 1, 1),
+   _Q1A(INST_ISSUED1,  0x0001, B6, ISSUE, 0x00000004, 1, 1),
+   _Q1A(INST_ISSUED2,  0x0001, B6, ISSUE, 0x00000008, 1, 1),
+   _Q1A(INST_ISSUED,   0x0003, B6, ISSUE, 0x00000104, 1, 1),
+   _Q1A(INST_EXECUTED, 0x0003, B6, EXEC,  0x00000398, 1, 1),
+   _Q1A(LD_SHARED,   0x0001, B6, LDST, 0x00000000, 1, 1),
+   _Q1A(ST_SHARED,   0x0001, B6, LDST, 0x00000004, 1, 1),
+   _Q1A(LD_LOCAL,    0x0001, B6, LDST, 0x00000008, 1, 1),
+   _Q1A(ST_LOCAL,    0x0001, B6, LDST, 0x0000000c, 1, 1),
+   _Q1A(GLD_REQUEST, 0x0001, B6, LDST, 0x00000010, 1, 1),
+   _Q1A(GST_REQUEST, 0x0001, B6, LDST, 0x00000014, 1, 1),
+   _Q1B(L1_LOCAL_LOAD_HIT,   0x0001, B6, L1, 0x00000000, 1, 1),
+   _Q1B(L1_LOCAL_LOAD_MISS,  0x0001, B6, L1, 0x00000004, 1, 1),
+   _Q1B(L1_LOCAL_STORE_HIT,  0x0001, B6, L1, 0x00000008, 1, 1),
+   _Q1B(L1_LOCAL_STORE_MISS, 0x0001, B6, L1, 0x0000000c, 1, 1),
+   _Q1B(L1_GLOBAL_LOAD_HIT,  0x0001, B6, L1, 0x00000010, 1, 1),
+   _Q1B(L1_GLOBAL_LOAD_MISS, 0x0001, B6, L1, 0x00000014, 1, 1),
+   _Q1B(GLD_TRANSACTIONS_UNCACHED, 0x0001, B6, MEM, 0x00000000, 1, 1),
+   _Q1B(GST_TRANSACTIONS,          0x0001, B6, MEM, 0x00000004, 1, 1),
+   _Q1A(BRANCH,           0x0001, B6, BRANCH, 0x0000000c, 1, 1),
+   _Q1A(BRANCH_DIVERGENT, 0x0001, B6, BRANCH, 0x00000010, 1, 1),
+   _Q1B(ACTIVE_WARPS,  0x003f, B6, WARP, 0x31483104, 2, 1),
+   _Q1B(ACTIVE_CYCLES, 0x0001, B6, WARP, 0x00000000, 1, 1),
+   _Q1A(ATOM_COUNT, 0x0001, B6, BRANCH, 0x00000000, 1, 1),
+   _Q1A(GRED_COUNT, 0x0001, B6, BRANCH, 0x00000008, 1, 1),
+   _Q1B(LD_SHARED_REPLAY, 0x0001, B6, REPLAY, 0x00000008, 1, 1),
+   _Q1B(ST_SHARED_REPLAY, 0x0001, B6, REPLAY, 0x0000000c, 1, 1),
+   _Q1B(LD_LOCAL_TRANSACTIONS, 0x0001, B6, TRANSACTION, 0x00000000, 1, 1),
+   _Q1B(ST_LOCAL_TRANSACTIONS, 0x0001, B6, TRANSACTION, 0x00000004, 1, 1),
+   _Q1B(L1_LD_SHARED_TRANSACTIONS, 0x0001, B6, TRANSACTION, 0x00000008, 1, 1),
+   _Q1B(L1_ST_SHARED_TRANSACTIONS, 0x0001, B6, TRANSACTION, 0x0000000c, 1, 1),
+   _Q1B(GLD_MEM_DIV_REPLAY, 0x0001, B6, REPLAY, 0x00000010, 1, 1),
+   _Q1B(GST_MEM_DIV_REPLAY, 0x0001, B6, REPLAY, 0x00000014, 1, 1),
+   _M2AB(IPC, 0x3, B6, EXEC, 0x398, 0xffff, LOGOP, WARP, 0x0, DIV_SUM_M0, 10, 1),
+   _M2AB(IPAC, 0x3, B6, EXEC, 0x398, 0x1, B6, WARP, 0x0, AVG_DIV_MM, 10, 1),
+   _M2A(IPEC, 0x3, B6, EXEC, 0x398, 0xe, LOGOP, EXEC, 0x398, AVG_DIV_MM, 10, 1),
+   _M2A(INST_REPLAY_OHEAD, 0x3, B6, ISSUE, 0x104, 0x3, B6, EXEC, 0x398, REL_SUM_MM, 100, 1),
+   _M2B(MP_OCCUPANCY, 0x3f, B6, WARP, 0x31483104, 0x01, B6, WARP, 0x0, AVG_DIV_MM, 200, 64),
+   _M2B(MP_EFFICIENCY, 0x01, B6, WARP, 0x0, 0xffff, LOGOP, WARP, 0x0, AVG_DIV_M0, 100, 1),
+};
+
+#undef _Q1A
+#undef _Q1B
+#undef _M2A
+#undef _M2B
+
+/* === PERFORMANCE MONITORING COUNTERS for NVC0:NVE4 === */
+static const uint64_t nvc0_read_hw_sm_counters_code[] =
+{
+   /* mov b32 $r8 $tidx
+    * mov b32 $r9 $physid
+    * mov b32 $r0 $pm0
+    * mov b32 $r1 $pm1
+    * mov b32 $r2 $pm2
+    * mov b32 $r3 $pm3
+    * mov b32 $r4 $pm4
+    * mov b32 $r5 $pm5
+    * mov b32 $r6 $pm6
+    * mov b32 $r7 $pm7
+    * set $p0 0x1 eq u32 $r8 0x0
+    * mov b32 $r10 c0[0x0]
+    * mov b32 $r11 c0[0x4]
+    * ext u32 $r8 $r9 0x414
+    * (not $p0) exit
+    * mul $r8 u32 $r8 u32 36
+    * add b32 $r10 $c $r10 $r8
+    * add b32 $r11 $r11 0x0 $c
+    * mov b32 $r8 c0[0x8]
+    * st b128 wt g[$r10d+0x00] $r0q
+    * st b128 wt g[$r10d+0x10] $r4q
+    * st b32 wt g[$r10d+0x20] $r8
+    * exit */
+   0x2c00000084021c04ULL,
+   0x2c0000000c025c04ULL,
+   0x2c00000010001c04ULL,
+   0x2c00000014005c04ULL,
+   0x2c00000018009c04ULL,
+   0x2c0000001c00dc04ULL,
+   0x2c00000020011c04ULL,
+   0x2c00000024015c04ULL,
+   0x2c00000028019c04ULL,
+   0x2c0000002c01dc04ULL,
+   0x190e0000fc81dc03ULL,
+   0x2800400000029de4ULL,
+   0x280040001002dde4ULL,
+   0x7000c01050921c03ULL,
+   0x80000000000021e7ULL,
+   0x1000000090821c02ULL,
+   0x4801000020a29c03ULL,
+   0x0800000000b2dc42ULL,
+   0x2800400020021de4ULL,
+   0x9400000000a01fc5ULL,
+   0x9400000040a11fc5ULL,
+   0x9400000080a21f85ULL,
+   0x8000000000001de7ULL
+};
+
+static const char *nvc0_hw_sm_query_names[] =
+{
+   /* MP counters */
+   "inst_executed",
+   "branch",
+   "divergent_branch",
+   "active_warps",
+   "active_cycles",
+   "warps_launched",
+   "threads_launched",
+   "shared_load",
+   "shared_store",
+   "local_load",
+   "local_store",
+   "gred_count",
+   "atom_count",
+   "gld_request",
+   "gst_request",
+   "inst_issued1_0",
+   "inst_issued1_1",
+   "inst_issued2_0",
+   "inst_issued2_1",
+   "thread_inst_executed_0",
+   "thread_inst_executed_1",
+   "thread_inst_executed_2",
+   "thread_inst_executed_3",
+   "prof_trigger_00",
+   "prof_trigger_01",
+   "prof_trigger_02",
+   "prof_trigger_03",
+   "prof_trigger_04",
+   "prof_trigger_05",
+   "prof_trigger_06",
+   "prof_trigger_07",
+};
+
+#define _Q(n, f, m, g, c, s0, s1, s2, s3, s4, s5) [NVC0_PM_QUERY_##n] = { { { f, NVC0_COMPUTE_MP_PM_OP_MODE_##m, c, 0, g, s0|(s1 << 8)|(s2 << 16)|(s3 << 24)|(s4##ULL << 32)|(s5##ULL << 40) }, {}, {}, {} }, 1, NVC0_COUNTER_OPn_SUM, { 1, 1 } }
+
+static const struct nvc0_hw_sm_query_cfg nvc0_hw_sm_queries[] =
+{
+   _Q(INST_EXECUTED,       0xaaaa, LOGOP, 0x2d, 3, 0x00, 0x11, 0x22, 0x00, 0x00, 0x00),
+   _Q(BRANCH,              0xaaaa, LOGOP, 0x1a, 2, 0x00, 0x11, 0x00, 0x00, 0x00, 0x00),
+   _Q(BRANCH_DIVERGENT,    0xaaaa, LOGOP, 0x19, 2, 0x20, 0x31, 0x00, 0x00, 0x00, 0x00),
+   _Q(ACTIVE_WARPS,        0xaaaa, LOGOP, 0x24, 6, 0x10, 0x21, 0x32, 0x43, 0x54, 0x65),
+   _Q(ACTIVE_CYCLES,       0xaaaa, LOGOP, 0x11, 1, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(LAUNCHED_WARPS,      0xaaaa, LOGOP, 0x26, 1, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(LAUNCHED_THREADS,    0xaaaa, LOGOP, 0x26, 6, 0x10, 0x21, 0x32, 0x43, 0x54, 0x65),
+   _Q(LD_SHARED,           0xaaaa, LOGOP, 0x64, 1, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(ST_SHARED,           0xaaaa, LOGOP, 0x64, 1, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(LD_LOCAL,            0xaaaa, LOGOP, 0x64, 1, 0x20, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(ST_LOCAL,            0xaaaa, LOGOP, 0x64, 1, 0x50, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(GRED_COUNT,          0xaaaa, LOGOP, 0x63, 1, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(ATOM_COUNT,          0xaaaa, LOGOP, 0x63, 1, 0x30, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(GLD_REQUEST,         0xaaaa, LOGOP, 0x64, 1, 0x30, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(GST_REQUEST,         0xaaaa, LOGOP, 0x64, 1, 0x60, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(INST_ISSUED1_0,      0xaaaa, LOGOP, 0x7e, 1, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(INST_ISSUED1_1,      0xaaaa, LOGOP, 0x7e, 1, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(INST_ISSUED2_0,      0xaaaa, LOGOP, 0x7e, 1, 0x20, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(INST_ISSUED2_1,      0xaaaa, LOGOP, 0x7e, 1, 0x50, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(TH_INST_EXECUTED_0,  0xaaaa, LOGOP, 0xa3, 6, 0x00, 0x11, 0x22, 0x33, 0x44, 0x55),
+   _Q(TH_INST_EXECUTED_1,  0xaaaa, LOGOP, 0xa5, 6, 0x00, 0x11, 0x22, 0x33, 0x44, 0x55),
+   _Q(TH_INST_EXECUTED_2,  0xaaaa, LOGOP, 0xa4, 6, 0x00, 0x11, 0x22, 0x33, 0x44, 0x55),
+   _Q(TH_INST_EXECUTED_3,  0xaaaa, LOGOP, 0xa6, 6, 0x00, 0x11, 0x22, 0x33, 0x44, 0x55),
+   _Q(PROF_TRIGGER_0,      0xaaaa, LOGOP, 0x01, 1, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(PROF_TRIGGER_1,      0xaaaa, LOGOP, 0x01, 1, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(PROF_TRIGGER_2,      0xaaaa, LOGOP, 0x01, 1, 0x20, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(PROF_TRIGGER_3,      0xaaaa, LOGOP, 0x01, 1, 0x30, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(PROF_TRIGGER_4,      0xaaaa, LOGOP, 0x01, 1, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(PROF_TRIGGER_5,      0xaaaa, LOGOP, 0x01, 1, 0x50, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(PROF_TRIGGER_6,      0xaaaa, LOGOP, 0x01, 1, 0x60, 0x00, 0x00, 0x00, 0x00, 0x00),
+   _Q(PROF_TRIGGER_7,      0xaaaa, LOGOP, 0x01, 1, 0x70, 0x00, 0x00, 0x00, 0x00, 0x00),
+};
+
+#undef _Q
+
+static const struct nvc0_hw_sm_query_cfg *
+nvc0_hw_sm_query_get_cfg(struct nvc0_context *nvc0, struct nvc0_query *q)
+{
+   struct nvc0_screen *screen = nvc0->screen;
+
+   if (screen->base.class_3d >= NVE4_3D_CLASS)
+      return &nve4_hw_sm_queries[q->type - PIPE_QUERY_DRIVER_SPECIFIC];
+   return &nvc0_hw_sm_queries[q->type - NVC0_PM_QUERY(0)];
+}
+
+void
+nvc0_hw_sm_query_create(struct nvc0_context *nvc0, struct nvc0_query *q)
+{
+   struct nvc0_screen *screen = nvc0->screen;
+
+   if (screen->base.class_3d >= NVE4_3D_CLASS) {
+      /* for each MP:
+       * [00] = WS0.C0
+       * [04] = WS0.C1
+       * [08] = WS0.C2
+       * [0c] = WS0.C3
+       * [10] = WS1.C0
+       * [14] = WS1.C1
+       * [18] = WS1.C2
+       * [1c] = WS1.C3
+       * [20] = WS2.C0
+       * [24] = WS2.C1
+       * [28] = WS2.C2
+       * [2c] = WS2.C3
+       * [30] = WS3.C0
+       * [34] = WS3.C1
+       * [38] = WS3.C2
+       * [3c] = WS3.C3
+       * [40] = MP.C4
+       * [44] = MP.C5
+       * [48] = MP.C6
+       * [4c] = MP.C7
+       * [50] = WS0.sequence
+       * [54] = WS1.sequence
+       * [58] = WS2.sequence
+       * [5c] = WS3.sequence
+       */
+      q->space = (4 * 4 + 4 + 4) * screen->mp_count * sizeof(uint32_t);
+   } else {
+      /* for each MP:
+       * [00] = MP.C0
+       * [04] = MP.C1
+       * [08] = MP.C2
+       * [0c] = MP.C3
+       * [10] = MP.C4
+       * [14] = MP.C5
+       * [18] = MP.C6
+       * [1c] = MP.C7
+       * [20] = MP.sequence
+       */
+      q->space = (8 + 1) * screen->mp_count * sizeof(uint32_t);
+   }
+}
+
+boolean
+nvc0_hw_sm_query_begin(struct nvc0_context *nvc0, struct nvc0_query *q)
+{
+   struct nvc0_screen *screen = nvc0->screen;
+   struct nouveau_pushbuf *push = nvc0->base.pushbuf;
+   const boolean is_nve4 = screen->base.class_3d >= NVE4_3D_CLASS;
+   const struct nvc0_hw_sm_query_cfg *cfg;
+   unsigned i, c;
+   unsigned num_ab[2] = { 0, 0 };
+
+   cfg = nvc0_hw_sm_query_get_cfg(nvc0, q);
+
+   /* check if we have enough free counter slots */
+   for (i = 0; i < cfg->num_counters; ++i)
+      num_ab[cfg->ctr[i].sig_dom]++;
+
+   if (screen->pm.num_hw_sm_active[0] + num_ab[0] > 4 ||
+       screen->pm.num_hw_sm_active[1] + num_ab[1] > 4) {
+      NOUVEAU_ERR("Not enough free MP counter slots !\n");
+      return false;
+   }
+
+   assert(cfg->num_counters <= 4);
+   PUSH_SPACE(push, 4 * 8 * (is_nve4 ? 1 : 6) + 6);
+
+   if (!screen->pm.mp_counters_enabled) {
+      screen->pm.mp_counters_enabled = TRUE;
+      BEGIN_NVC0(push, SUBC_SW(0x06ac), 1);
+      PUSH_DATA (push, 0x1fcb);
+   }
+
+   /* set sequence field to 0 (used to check if result is available) */
+   for (i = 0; i < screen->mp_count; ++i)
+      q->data[i * 10 + 10] = 0;
+
+   for (i = 0; i < cfg->num_counters; ++i) {
+      const unsigned d = cfg->ctr[i].sig_dom;
+
+      if (!screen->pm.num_hw_sm_active[d]) {
+         uint32_t m = (1 << 22) | (1 << (7 + (8 * !d)));
+         if (screen->pm.num_hw_sm_active[!d])
+            m |= 1 << (7 + (8 * d));
+         BEGIN_NVC0(push, SUBC_SW(0x0600), 1);
+         PUSH_DATA (push, m);
+      }
+      screen->pm.num_hw_sm_active[d]++;
+
+      for (c = d * 4; c < (d * 4 + 4); ++c) {
+         if (!screen->pm.mp_counter[c]) {
+            q->ctr[i] = c;
+            screen->pm.mp_counter[c] = (struct pipe_query *)q;
+            break;
+         }
+      }
+      assert(c <= (d * 4 + 3)); /* must succeed, already checked for space */
+
+      /* configure and reset the counter(s) */
+      if (is_nve4) {
+         if (d == 0)
+            BEGIN_NVC0(push, NVE4_COMPUTE(MP_PM_A_SIGSEL(c & 3)), 1);
+         else
+            BEGIN_NVC0(push, NVE4_COMPUTE(MP_PM_B_SIGSEL(c & 3)), 1);
+         PUSH_DATA (push, cfg->ctr[i].sig_sel);
+         BEGIN_NVC0(push, NVE4_COMPUTE(MP_PM_SRCSEL(c)), 1);
+         PUSH_DATA (push, cfg->ctr[i].src_sel + 0x2108421 * (c & 3));
+         BEGIN_NVC0(push, NVE4_COMPUTE(MP_PM_FUNC(c)), 1);
+         PUSH_DATA (push, (cfg->ctr[i].func << 4) | cfg->ctr[i].mode);
+         BEGIN_NVC0(push, NVE4_COMPUTE(MP_PM_SET(c)), 1);
+         PUSH_DATA (push, 0);
+      } else {
+         unsigned s;
+
+         for (s = 0; s < cfg->ctr[i].num_src; s++) {
+            BEGIN_NVC0(push, NVC0_COMPUTE(MP_PM_SIGSEL(s)), 1);
+            PUSH_DATA (push, cfg->ctr[i].sig_sel);
+            BEGIN_NVC0(push, NVC0_COMPUTE(MP_PM_SRCSEL(s)), 1);
+            PUSH_DATA (push, (cfg->ctr[i].src_sel >> (s * 8)) & 0xff);
+            BEGIN_NVC0(push, NVC0_COMPUTE(MP_PM_OP(s)), 1);
+            PUSH_DATA (push, (cfg->ctr[i].func << 4) | cfg->ctr[i].mode);
+            BEGIN_NVC0(push, NVC0_COMPUTE(MP_PM_SET(s)), 1);
+            PUSH_DATA (push, 0);
+         }
+      }
+   }
+   return true;
+}
+
+void
+nvc0_hw_sm_query_end(struct nvc0_context *nvc0, struct nvc0_query *q)
+{
+   struct nvc0_screen *screen = nvc0->screen;
+   struct pipe_context *pipe = &nvc0->base.pipe;
+   struct nouveau_pushbuf *push = nvc0->base.pushbuf;
+   const boolean is_nve4 = screen->base.class_3d >= NVE4_3D_CLASS;
+   uint32_t mask;
+   uint32_t input[3];
+   const uint block[3] = { 32, is_nve4 ? 4 : 1, 1 };
+   const uint grid[3] = { screen->mp_count, 1, 1 };
+   unsigned c;
+   const struct nvc0_hw_sm_query_cfg *cfg;
+
+   cfg = nvc0_hw_sm_query_get_cfg(nvc0, q);
+
+   if (unlikely(!screen->pm.prog)) {
+      struct nvc0_program *prog = CALLOC_STRUCT(nvc0_program);
+      prog->type = PIPE_SHADER_COMPUTE;
+      prog->translated = TRUE;
+      prog->num_gprs = 14;
+      prog->parm_size = 12;
+      if (is_nve4) {
+         prog->code = (uint32_t *)nve4_read_hw_sm_counters_code;
+         prog->code_size = sizeof(nve4_read_hw_sm_counters_code);
+      } else {
+         prog->code = (uint32_t *)nvc0_read_hw_sm_counters_code;
+         prog->code_size = sizeof(nvc0_read_hw_sm_counters_code);
+      }
+      screen->pm.prog = prog;
+   }
+
+   /* disable all counting */
+   PUSH_SPACE(push, 8);
+   for (c = 0; c < 8; ++c)
+      if (screen->pm.mp_counter[c]) {
+         if (is_nve4) {
+            IMMED_NVC0(push, NVE4_COMPUTE(MP_PM_FUNC(c)), 0);
+         } else {
+            IMMED_NVC0(push, NVC0_COMPUTE(MP_PM_OP(c)), 0);
+         }
+      }
+   /* release counters for this query */
+   for (c = 0; c < 8; ++c) {
+      if (nvc0_query(screen->pm.mp_counter[c]) == q) {
+         screen->pm.num_hw_sm_active[c / 4]--;
+         screen->pm.mp_counter[c] = NULL;
+      }
+   }
+
+   BCTX_REFN_bo(nvc0->bufctx_cp, CP_QUERY, NOUVEAU_BO_GART | NOUVEAU_BO_WR,
+                q->bo);
+
+   PUSH_SPACE(push, 1);
+   IMMED_NVC0(push, SUBC_COMPUTE(NV50_GRAPH_SERIALIZE), 0);
+
+   pipe->bind_compute_state(pipe, screen->pm.prog);
+   input[0] = (q->bo->offset + q->base);
+   input[1] = (q->bo->offset + q->base) >> 32;
+   input[2] = q->sequence;
+   pipe->launch_grid(pipe, block, grid, 0, input);
+
+   nouveau_bufctx_reset(nvc0->bufctx_cp, NVC0_BIND_CP_QUERY);
+
+   /* re-activate other counters */
+   PUSH_SPACE(push, 16);
+   mask = 0;
+   for (c = 0; c < 8; ++c) {
+      unsigned i;
+      q = nvc0_query(screen->pm.mp_counter[c]);
+      if (!q)
+         continue;
+      cfg = nvc0_hw_sm_query_get_cfg(nvc0, q);
+      for (i = 0; i < cfg->num_counters; ++i) {
+         if (mask & (1 << q->ctr[i]))
+            break;
+         mask |= 1 << q->ctr[i];
+         if (is_nve4) {
+            BEGIN_NVC0(push, NVE4_COMPUTE(MP_PM_FUNC(q->ctr[i])), 1);
+         } else {
+            BEGIN_NVC0(push, NVC0_COMPUTE(MP_PM_OP(q->ctr[i])), 1);
+         }
+         PUSH_DATA (push, (cfg->ctr[i].func << 4) | cfg->ctr[i].mode);
+      }
+   }
+}
+
+static INLINE boolean
+nvc0_hw_sm_query_read_data(uint32_t count[32][4],
+                           struct nvc0_context *nvc0, boolean wait,
+                           struct nvc0_query *q,
+                           const struct nvc0_hw_sm_query_cfg *cfg,
+                           unsigned mp_count)
+{
+   unsigned p, c;
+
+   for (p = 0; p < mp_count; ++p) {
+      const unsigned b = (0x24 / 4) * p;
+
+      for (c = 0; c < cfg->num_counters; ++c) {
+         if (q->data[b + 8] != q->sequence) {
+            if (!wait)
+               return FALSE;
+            if (nouveau_bo_wait(q->bo, NOUVEAU_BO_RD, nvc0->base.client))
+               return FALSE;
+         }
+         count[p][c] = q->data[b + q->ctr[c]];
+      }
+   }
+   return TRUE;
+}
+
+static INLINE boolean
+nve4_hw_sm_query_read_data(uint32_t count[32][4],
+                           struct nvc0_context *nvc0, boolean wait,
+                           struct nvc0_query *q,
+                           const struct nvc0_hw_sm_query_cfg *cfg,
+                           unsigned mp_count)
+{
+   unsigned p, c, d;
+
+   for (p = 0; p < mp_count; ++p) {
+      const unsigned b = (0x60 / 4) * p;
+
+      for (c = 0; c < cfg->num_counters; ++c) {
+         count[p][c] = 0;
+         for (d = 0; d < ((q->ctr[c] & ~3) ? 1 : 4); ++d) {
+            if (q->data[b + 20 + d] != q->sequence) {
+               if (!wait)
+                  return FALSE;
+               if (nouveau_bo_wait(q->bo, NOUVEAU_BO_RD, nvc0->base.client))
+                  return FALSE;
+            }
+            if (q->ctr[c] & ~0x3)
+               count[p][c] = q->data[b + 16 + (q->ctr[c] & 3)];
+            else
+               count[p][c] += q->data[b + d * 4 + q->ctr[c]];
+         }
+      }
+   }
+   return TRUE;
+}
+
+/* Metric calculations:
+ * sum(x) ... sum of x over all MPs
+ * avg(x) ... average of x over all MPs
+ *
+ * IPC              : sum(inst_executed) / clock
+ * INST_REPLAY_OHEAD: (sum(inst_issued) - sum(inst_executed)) / sum(inst_issued)
+ * MP_OCCUPANCY     : avg((active_warps / 64) / active_cycles)
+ * MP_EFFICIENCY    : avg(active_cycles / clock)
+ *
+ * NOTE: Interpretation of IPC requires knowledge of MP count.
+ */
+boolean
+nvc0_hw_sm_query_result(struct nvc0_context *nvc0, struct nvc0_query *q,
+                        void *result, boolean wait)
+{
+   uint32_t count[32][4];
+   uint64_t value = 0;
+   unsigned mp_count = MIN2(nvc0->screen->mp_count_compute, 32);
+   unsigned p, c;
+   const struct nvc0_hw_sm_query_cfg *cfg;
+   boolean ret;
+
+   cfg = nvc0_hw_sm_query_get_cfg(nvc0, q);
+
+   if (nvc0->screen->base.class_3d >= NVE4_3D_CLASS)
+      ret = nve4_hw_sm_query_read_data(count, nvc0, wait, q, cfg, mp_count);
+   else
+      ret = nvc0_hw_sm_query_read_data(count, nvc0, wait, q, cfg, mp_count);
+   if (!ret)
+      return FALSE;
+
+   if (cfg->op == NVC0_COUNTER_OPn_SUM) {
+      for (c = 0; c < cfg->num_counters; ++c)
+         for (p = 0; p < mp_count; ++p)
+            value += count[p][c];
+      value = (value * cfg->norm[0]) / cfg->norm[1];
+   } else
+   if (cfg->op == NVC0_COUNTER_OPn_OR) {
+      uint32_t v = 0;
+      for (c = 0; c < cfg->num_counters; ++c)
+         for (p = 0; p < mp_count; ++p)
+            v |= count[p][c];
+      value = ((uint64_t)v * cfg->norm[0]) / cfg->norm[1];
+   } else
+   if (cfg->op == NVC0_COUNTER_OPn_AND) {
+      uint32_t v = ~0;
+      for (c = 0; c < cfg->num_counters; ++c)
+         for (p = 0; p < mp_count; ++p)
+            v &= count[p][c];
+      value = ((uint64_t)v * cfg->norm[0]) / cfg->norm[1];
+   } else
+   if (cfg->op == NVC0_COUNTER_OP2_REL_SUM_MM) {
+      uint64_t v[2] = { 0, 0 };
+      for (p = 0; p < mp_count; ++p) {
+         v[0] += count[p][0];
+         v[1] += count[p][1];
+      }
+      if (v[0])
+         value = ((v[0] - v[1]) * cfg->norm[0]) / (v[0] * cfg->norm[1]);
+   } else
+   if (cfg->op == NVC0_COUNTER_OP2_DIV_SUM_M0) {
+      for (p = 0; p < mp_count; ++p)
+         value += count[p][0];
+      if (count[0][1])
+         value = (value * cfg->norm[0]) / (count[0][1] * cfg->norm[1]);
+      else
+         value = 0;
+   } else
+   if (cfg->op == NVC0_COUNTER_OP2_AVG_DIV_MM) {
+      unsigned mp_used = 0;
+      for (p = 0; p < mp_count; ++p, mp_used += !!count[p][0])
+         if (count[p][1])
+            value += (count[p][0] * cfg->norm[0]) / count[p][1];
+      if (mp_used)
+         value /= (uint64_t)mp_used * cfg->norm[1];
+   } else
+   if (cfg->op == NVC0_COUNTER_OP2_AVG_DIV_M0) {
+      unsigned mp_used = 0;
+      for (p = 0; p < mp_count; ++p, mp_used += !!count[p][0])
+         value += count[p][0];
+      if (count[0][1] && mp_used) {
+         value *= cfg->norm[0];
+         value /= (uint64_t)count[0][1] * mp_used * cfg->norm[1];
+      } else {
+         value = 0;
+      }
+   }
+
+   *(uint64_t *)result = value;
+   return TRUE;
+}
-- 
2.4.1



More information about the mesa-dev mailing list