nv50: add compute-related MP perf counters on G84+
authorSamuel Pitoiset <samuel.pitoiset@gmail.com>
Tue, 10 Nov 2015 00:27:15 +0000 (01:27 +0100)
committerSamuel Pitoiset <samuel.pitoiset@gmail.com>
Sat, 14 Nov 2015 22:42:42 +0000 (23:42 +0100)
These compute-related MP performance counters have been reverse
engineered using CUPTI which is part of NVIDIA CUDA.

As for nvc0, we use a compute kernel to read out those performance
counters, and the command stream to configure them. Note that Tesla
only exposes 4 MP performance counters, while Fermi has 8.

Only G84+ is supported because G80 is an old and weird card.

Tested on G84, G96, G200, MCP79 and GT218 with glxgears, glxspheres64,
xonotic-glx, heaven and valley.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Tested-by: Pierre Moreau <pierre.morrow@free.fr>
Acked-by: Ilia Mirkin <imirkin@alum.mit.edu>
src/gallium/drivers/nouveau/Makefile.sources
src/gallium/drivers/nouveau/nv50/nv50_context.h
src/gallium/drivers/nouveau/nv50/nv50_query.c
src/gallium/drivers/nouveau/nv50/nv50_query_hw.c
src/gallium/drivers/nouveau/nv50/nv50_query_hw.h
src/gallium/drivers/nouveau/nv50/nv50_query_hw_sm.c [new file with mode: 0644]
src/gallium/drivers/nouveau/nv50/nv50_query_hw_sm.h [new file with mode: 0644]
src/gallium/drivers/nouveau/nv50/nv50_screen.c
src/gallium/drivers/nouveau/nv50/nv50_screen.h

index c2ff8e9b46eacb6e392d94f4340b89e63a2cf3e9..a1aa13587a16949069238bf6646856a406c0dae8 100644 (file)
@@ -77,6 +77,8 @@ NV50_C_SOURCES := \
        nv50/nv50_query.h \
        nv50/nv50_query_hw.c \
        nv50/nv50_query_hw.h \
+       nv50/nv50_query_hw_sm.c \
+       nv50/nv50_query_hw_sm.h \
        nv50/nv50_resource.c \
        nv50/nv50_resource.h \
        nv50/nv50_screen.c \
index fbafe029948c39d923a2f7aa889f746a40dc6512..2cebcd994238f7361dd766ca266f015fdf8e0775 100644 (file)
@@ -67,7 +67,8 @@
 /* compute bufctx (during launch_grid) */
 #define NV50_BIND_CP_GLOBAL   0
 #define NV50_BIND_CP_SCREEN   1
-#define NV50_BIND_CP_COUNT    2
+#define NV50_BIND_CP_QUERY    2
+#define NV50_BIND_CP_COUNT    3
 
 /* bufctx for other operations */
 #define NV50_BIND_2D          0
index dd9b85b7208441e4a74c91aeca387d3b6c178e2c..c31bf728c6124b082d0844304b82b75ef8d4c0dc 100644 (file)
@@ -153,3 +153,26 @@ nv50_init_query_functions(struct nv50_context *nv50)
    pipe->get_query_result = nv50_get_query_result;
    pipe->render_condition = nv50_render_condition;
 }
+
+int
+nv50_screen_get_driver_query_info(struct pipe_screen *pscreen,
+                                  unsigned id,
+                                  struct pipe_driver_query_info *info)
+{
+   struct nv50_screen *screen = nv50_screen(pscreen);
+   int num_hw_queries = 0;
+
+   num_hw_queries = nv50_hw_get_driver_query_info(screen, 0, NULL);
+
+   if (!info)
+      return num_hw_queries;
+
+   /* Init default values. */
+   info->name = "this_is_not_the_query_you_are_looking_for";
+   info->query_type = 0xdeadd01d;
+   info->max_value.u64 = 0;
+   info->type = PIPE_DRIVER_QUERY_TYPE_UINT64;
+   info->group_id = -1;
+
+   return nv50_hw_get_driver_query_info(screen, id, info);
+}
index 945ce7abe50c103b7921c87ffc73ceafcdb51865..23108acbef59067bee7b633e96015457b37bab38 100644 (file)
@@ -25,6 +25,7 @@
 
 #include "nv50/nv50_context.h"
 #include "nv50/nv50_query_hw.h"
+#include "nv50/nv50_query_hw_sm.h"
 #include "nv_object.xml.h"
 
 #define NV50_HW_QUERY_STATE_READY   0
@@ -41,7 +42,7 @@
 
 #define NV50_HW_QUERY_ALLOC_SPACE 256
 
-static bool
+bool
 nv50_hw_query_allocate(struct nv50_context *nv50, struct nv50_query *q,
                        int size)
 {
@@ -122,6 +123,9 @@ nv50_hw_begin_query(struct nv50_context *nv50, struct nv50_query *q)
    struct nouveau_pushbuf *push = nv50->base.pushbuf;
    struct nv50_hw_query *hq = nv50_hw_query(q);
 
+   if (hq->funcs && hq->funcs->begin_query)
+      return hq->funcs->begin_query(nv50, hq);
+
    /* For occlusion queries we have to change the storage, because a previous
     * query might set the initial render condition to false even *after* we re-
     * initialized it to true.
@@ -193,6 +197,11 @@ nv50_hw_end_query(struct nv50_context *nv50, struct nv50_query *q)
    struct nouveau_pushbuf *push = nv50->base.pushbuf;
    struct nv50_hw_query *hq = nv50_hw_query(q);
 
+   if (hq->funcs && hq->funcs->end_query) {
+      hq->funcs->end_query(nv50, hq);
+      return;
+   }
+
    hq->state = NV50_HW_QUERY_STATE_ENDED;
 
    switch (q->type) {
@@ -261,6 +270,9 @@ nv50_hw_get_query_result(struct nv50_context *nv50, struct nv50_query *q,
    uint64_t *data64 = (uint64_t *)hq->data;
    int i;
 
+   if (hq->funcs && hq->funcs->get_query_result)
+      return hq->funcs->get_query_result(nv50, hq, wait, result);
+
    if (hq->state != NV50_HW_QUERY_STATE_READY)
       nv50_hw_query_update(q);
 
@@ -331,6 +343,12 @@ nv50_hw_create_query(struct nv50_context *nv50, unsigned type, unsigned index)
    struct nv50_hw_query *hq;
    struct nv50_query *q;
 
+   hq = nv50_hw_sm_create_query(nv50, type);
+   if (hq) {
+      hq->base.funcs = &hw_query_funcs;
+      return (struct nv50_query *)hq;
+   }
+
    hq = CALLOC_STRUCT(nv50_hw_query);
    if (!hq)
       return NULL;
@@ -375,6 +393,20 @@ nv50_hw_create_query(struct nv50_context *nv50, unsigned type, unsigned index)
    return q;
 }
 
+int
+nv50_hw_get_driver_query_info(struct nv50_screen *screen, unsigned id,
+                              struct pipe_driver_query_info *info)
+{
+   int num_hw_sm_queries = 0;
+
+   num_hw_sm_queries = nv50_hw_sm_get_driver_query_info(screen, 0, NULL);
+
+   if (!info)
+      return num_hw_sm_queries;
+
+   return nv50_hw_sm_get_driver_query_info(screen, id, info);
+}
+
 void
 nv50_hw_query_pushbuf_submit(struct nouveau_pushbuf *push, uint16_t method,
                              struct nv50_query *q, unsigned result_offset)
index 294c67de9a4084cca0b6fe323c7447b6043f1b9d..82ec6bd2d96c0cb6d7a57d994074a9c104973005 100644 (file)
@@ -8,8 +8,19 @@
 
 #define NVA0_HW_QUERY_STREAM_OUTPUT_BUFFER_OFFSET (PIPE_QUERY_TYPES + 0)
 
+struct nv50_hw_query;
+
+struct nv50_hw_query_funcs {
+   void (*destroy_query)(struct nv50_context *, struct nv50_hw_query *);
+   boolean (*begin_query)(struct nv50_context *, struct nv50_hw_query *);
+   void (*end_query)(struct nv50_context *, struct nv50_hw_query *);
+   boolean (*get_query_result)(struct nv50_context *, struct nv50_hw_query *,
+                               boolean, union pipe_query_result *);
+};
+
 struct nv50_hw_query {
    struct nv50_query base;
+   const struct nv50_hw_query_funcs *funcs;
    uint32_t *data;
    uint32_t sequence;
    struct nouveau_bo *bo;
@@ -31,6 +42,11 @@ nv50_hw_query(struct nv50_query *q)
 
 struct nv50_query *
 nv50_hw_create_query(struct nv50_context *, unsigned, unsigned);
+int
+nv50_hw_get_driver_query_info(struct nv50_screen *, unsigned,
+                              struct pipe_driver_query_info *);
+bool
+nv50_hw_query_allocate(struct nv50_context *, struct nv50_query *, int);
 void
 nv50_hw_query_pushbuf_submit(struct nouveau_pushbuf *, uint16_t,
                              struct nv50_query *, unsigned);
diff --git a/src/gallium/drivers/nouveau/nv50/nv50_query_hw_sm.c b/src/gallium/drivers/nouveau/nv50/nv50_query_hw_sm.c
new file mode 100644 (file)
index 0000000..e75b428
--- /dev/null
@@ -0,0 +1,417 @@
+/*
+ * Copyright 2015 Samuel Pitoiset
+ *
+ * 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.
+ */
+
+#define NV50_PUSH_EXPLICIT_SPACE_CHECKING
+
+#include "nv50/nv50_context.h"
+#include "nv50/nv50_query_hw_sm.h"
+
+#include "nv_object.xml.h"
+#include "nv50/nv50_compute.xml.h"
+
+/* === PERFORMANCE MONITORING COUNTERS for NV84+ === */
+
+/* NOTE: intentionally using the same names as NV */
+static const char *nv50_hw_sm_query_names[] =
+{
+   "branch",
+   "divergent_branch",
+   "instructions",
+   "prof_trigger_00",
+   "prof_trigger_01",
+   "prof_trigger_02",
+   "prof_trigger_03",
+   "prof_trigger_04",
+   "prof_trigger_05",
+   "prof_trigger_06",
+   "prof_trigger_07",
+   "sm_cta_launched",
+   "warp_serialize",
+};
+
+static const uint64_t nv50_read_hw_sm_counters_code[] =
+{
+   /* and b32 $r0 $r0 0x0000ffff
+    * add b32 $c0 $r0 $r0 $r0
+    * (lg $c0) ret
+    * mov $r0 $pm0
+    * mov $r1 $pm1
+    * mov $r2 $pm2
+    * mov $r3 $pm3
+    * mov $r4 $physid
+    * ld $r5 b32 s[0x10]
+    * ld $r6 b32 s[0x14]
+    * and b32 $r4 $r4 0x000f0000
+    * shr u32 $r4 $r4 0x10
+    * mul $r4 u24 $r4 0x14
+    * add b32 $r5 $r5 $r4
+    * st b32 g15[$r5] $r0
+    * add b32 $r5 $r5 0x04
+    * st b32 g15[$r5] $r1
+    * add b32 $r5 $r5 0x04
+    * st b32 g15[$r5] $r2
+    * add b32 $r5 $r5 0x04
+    * st b32 g15[$r5] $r3
+    * add b32 $r5 $r5 0x04
+    * exit st b32 g15[$r5] $r6 */
+   0x00000fffd03f0001ULL,
+   0x040007c020000001ULL,
+   0x0000028030000003ULL,
+   0x6001078000000001ULL,
+   0x6001478000000005ULL,
+   0x6001878000000009ULL,
+   0x6001c7800000000dULL,
+   0x6000078000000011ULL,
+   0x4400c78010000815ULL,
+   0x4400c78010000a19ULL,
+   0x0000f003d0000811ULL,
+   0xe410078030100811ULL,
+   0x0000000340540811ULL,
+   0x0401078020000a15ULL,
+   0xa0c00780d00f0a01ULL,
+   0x0000000320048a15ULL,
+   0xa0c00780d00f0a05ULL,
+   0x0000000320048a15ULL,
+   0xa0c00780d00f0a09ULL,
+   0x0000000320048a15ULL,
+   0xa0c00780d00f0a0dULL,
+   0x0000000320048a15ULL,
+   0xa0c00781d00f0a19ULL,
+};
+
+struct nv50_hw_sm_counter_cfg
+{
+   uint32_t mode : 4;    /* LOGOP, LOGOP_PULSE */
+   uint32_t unit : 8;    /* UNK[0-5] */
+   uint32_t sig  : 8;    /* signal selection */
+};
+
+struct nv50_hw_sm_query_cfg
+{
+   struct nv50_hw_sm_counter_cfg ctr[4];
+   uint8_t num_counters;
+};
+
+#define _Q(n, m, u, s) [NV50_HW_SM_QUERY_##n] = { { { NV50_COMPUTE_MP_PM_CONTROL_MODE_##m, NV50_COMPUTE_MP_PM_CONTROL_UNIT_##u, s, }, {}, {}, {} }, 1 }
+
+/* ==== Compute capability 1.1 (G84+) ==== */
+static const struct nv50_hw_sm_query_cfg sm11_hw_sm_queries[] =
+{
+   _Q(BRANCH,           LOGOP, UNK4, 0x02),
+   _Q(DIVERGENT_BRANCH, LOGOP, UNK4, 0x09),
+   _Q(INSTRUCTIONS,     LOGOP, UNK4, 0x04),
+   _Q(PROF_TRIGGER_0,   LOGOP, UNK1, 0x26),
+   _Q(PROF_TRIGGER_1,   LOGOP, UNK1, 0x27),
+   _Q(PROF_TRIGGER_2,   LOGOP, UNK1, 0x28),
+   _Q(PROF_TRIGGER_3,   LOGOP, UNK1, 0x29),
+   _Q(PROF_TRIGGER_4,   LOGOP, UNK1, 0x2a),
+   _Q(PROF_TRIGGER_5,   LOGOP, UNK1, 0x2b),
+   _Q(PROF_TRIGGER_6,   LOGOP, UNK1, 0x2c),
+   _Q(PROF_TRIGGER_7,   LOGOP, UNK1, 0x2d),
+   _Q(SM_CTA_LAUNCHED,  LOGOP, UNK1, 0x33),
+   _Q(WARP_SERIALIZE,   LOGOP, UNK0, 0x0b),
+};
+
+static inline uint16_t nv50_hw_sm_get_func(uint8_t slot)
+{
+   switch (slot) {
+   case 0: return 0xaaaa;
+   case 1: return 0xcccc;
+   case 2: return 0xf0f0;
+   case 3: return 0xff00;
+   }
+   return 0;
+}
+
+static const struct nv50_hw_sm_query_cfg *
+nv50_hw_sm_query_get_cfg(struct nv50_context *nv50, struct nv50_hw_query *hq)
+{
+   struct nv50_query *q = &hq->base;
+   return &sm11_hw_sm_queries[q->type - NV50_HW_SM_QUERY(0)];
+}
+
+static void
+nv50_hw_sm_destroy_query(struct nv50_context *nv50, struct nv50_hw_query *hq)
+{
+   struct nv50_query *q = &hq->base;
+   q->funcs->destroy_query(nv50, q);
+}
+
+static boolean
+nv50_hw_sm_begin_query(struct nv50_context *nv50, struct nv50_hw_query *hq)
+{
+   struct nv50_screen *screen = nv50->screen;
+   struct nouveau_pushbuf *push = nv50->base.pushbuf;
+   struct nv50_hw_sm_query *hsq = nv50_hw_sm_query(hq);
+   const struct nv50_hw_sm_query_cfg *cfg;
+   uint16_t func;
+   int i, c;
+
+   cfg = nv50_hw_sm_query_get_cfg(nv50, hq);
+
+   /* check if we have enough free counter slots */
+   if (screen->pm.num_hw_sm_active + cfg->num_counters > 4) {
+      NOUVEAU_ERR("Not enough free MP counter slots !\n");
+      return false;
+   }
+
+   assert(cfg->num_counters <= 4);
+   PUSH_SPACE(push, 4 * 4);
+
+   /* set sequence field to 0 (used to check if result is available) */
+   for (i = 0; i < screen->MPsInTP; ++i) {
+      const unsigned b = (0x14 / 4) * i;
+      hq->data[b + 16] = 0;
+   }
+   hq->sequence++;
+
+   for (i = 0; i < cfg->num_counters; i++) {
+      screen->pm.num_hw_sm_active++;
+
+      /* find free counter slots */
+      for (c = 0; c < 4; ++c) {
+         if (!screen->pm.mp_counter[c]) {
+            hsq->ctr[i] = c;
+            screen->pm.mp_counter[c] = hsq;
+            break;
+         }
+      }
+
+      /* select func to aggregate counters */
+      func = nv50_hw_sm_get_func(c);
+
+      /* configure and reset the counter(s) */
+      BEGIN_NV04(push, NV50_COMPUTE(MP_PM_CONTROL(c)), 1);
+      PUSH_DATA (push, (cfg->ctr[i].sig << 24) | (func << 8)
+                        | cfg->ctr[i].unit | cfg->ctr[i].mode);
+      BEGIN_NV04(push, NV50_COMPUTE(MP_PM_SET(c)), 1);
+      PUSH_DATA (push, 0);
+   }
+   return true;
+}
+
+static void
+nv50_hw_sm_end_query(struct nv50_context *nv50, struct nv50_hw_query *hq)
+{
+   struct nv50_screen *screen = nv50->screen;
+   struct pipe_context *pipe = &nv50->base.pipe;
+   struct nouveau_pushbuf *push = nv50->base.pushbuf;
+   struct nv50_hw_sm_query *hsq = nv50_hw_sm_query(hq);
+   uint32_t mask;
+   uint32_t input[3];
+   const uint block[3] = { 32, 1, 1 };
+   const uint grid[3] = { screen->MPsInTP, screen->TPs, 1 };
+   int c;
+
+   if (unlikely(!screen->pm.prog)) {
+      struct nv50_program *prog = CALLOC_STRUCT(nv50_program);
+      prog->type = PIPE_SHADER_COMPUTE;
+      prog->translated = true;
+      prog->max_gpr = 7;
+      prog->parm_size = 8;
+      prog->code = (uint32_t *)nv50_read_hw_sm_counters_code;
+      prog->code_size = sizeof(nv50_read_hw_sm_counters_code);
+      screen->pm.prog = prog;
+   }
+
+   /* disable all counting */
+   PUSH_SPACE(push, 8);
+   for (c = 0; c < 4; c++) {
+      if (screen->pm.mp_counter[c]) {
+         BEGIN_NV04(push, NV50_COMPUTE(MP_PM_CONTROL(c)), 1);
+         PUSH_DATA (push, 0);
+      }
+   }
+
+   /* release counters for this query */
+   for (c = 0; c < 4; c++) {
+      if (screen->pm.mp_counter[c] == hsq) {
+         screen->pm.num_hw_sm_active--;
+         screen->pm.mp_counter[c] = NULL;
+      }
+   }
+
+   BCTX_REFN_bo(nv50->bufctx_cp, CP_QUERY, NOUVEAU_BO_GART | NOUVEAU_BO_WR,
+                hq->bo);
+
+   PUSH_SPACE(push, 2);
+   BEGIN_NV04(push, SUBC_COMPUTE(NV50_GRAPH_SERIALIZE), 1);
+   PUSH_DATA (push, 0);
+
+   pipe->bind_compute_state(pipe, screen->pm.prog);
+   input[0] = hq->bo->offset + hq->base_offset;
+   input[1] = hq->sequence;
+   pipe->launch_grid(pipe, block, grid, 0, input);
+
+   nouveau_bufctx_reset(nv50->bufctx_cp, NV50_BIND_CP_QUERY);
+
+   /* re-active other counters */
+   PUSH_SPACE(push, 8);
+   mask = 0;
+   for (c = 0; c < 4; c++) {
+      const struct nv50_hw_sm_query_cfg *cfg;
+      unsigned i;
+
+      hsq = screen->pm.mp_counter[c];
+      if (!hsq)
+         continue;
+
+      cfg = nv50_hw_sm_query_get_cfg(nv50, &hsq->base);
+      for (i = 0; i < cfg->num_counters; i++) {
+         uint16_t func;
+
+         if (mask & (1 << hsq->ctr[i]))
+            break;
+
+         mask |= 1 << hsq->ctr[i];
+         func  = nv50_hw_sm_get_func(hsq->ctr[i]);
+
+         BEGIN_NV04(push, NV50_COMPUTE(MP_PM_CONTROL(hsq->ctr[i])), 1);
+         PUSH_DATA (push, (cfg->ctr[i].sig << 24) | (func << 8)
+                    | cfg->ctr[i].unit | cfg->ctr[i].mode);
+      }
+   }
+}
+
+static inline bool
+nv50_hw_sm_query_read_data(uint32_t count[32][4],
+                           struct nv50_context *nv50, bool wait,
+                           struct nv50_hw_query *hq,
+                           const struct nv50_hw_sm_query_cfg *cfg,
+                           unsigned mp_count)
+{
+   struct nv50_hw_sm_query *hsq = nv50_hw_sm_query(hq);
+   unsigned p, c;
+
+   for (p = 0; p < mp_count; ++p) {
+      const unsigned b = (0x14 / 4) * p;
+
+      for (c = 0; c < cfg->num_counters; ++c) {
+         if (hq->data[b + 4] != hq->sequence) {
+            if (!wait)
+               return false;
+            if (nouveau_bo_wait(hq->bo, NOUVEAU_BO_RD, nv50->base.client))
+               return false;
+         }
+         count[p][c] = hq->data[b + hsq->ctr[c]];
+      }
+   }
+   return true;
+}
+
+static boolean
+nv50_hw_sm_get_query_result(struct nv50_context *nv50, struct nv50_hw_query *hq,
+                            boolean wait, union pipe_query_result *result)
+{
+   uint32_t count[32][4];
+   uint64_t value = 0;
+   unsigned mp_count = MIN2(nv50->screen->MPsInTP, 32);
+   unsigned p, c;
+   const struct nv50_hw_sm_query_cfg *cfg;
+   bool ret;
+
+   cfg = nv50_hw_sm_query_get_cfg(nv50, hq);
+
+   ret = nv50_hw_sm_query_read_data(count, nv50, wait, hq, cfg, mp_count);
+   if (!ret)
+      return false;
+
+   for (c = 0; c < cfg->num_counters; ++c)
+      for (p = 0; p < mp_count; ++p)
+         value += count[p][c];
+
+   /* We only count a single TP, and simply multiply by the total number of
+    * TPs to compute result over all TPs. This is inaccurate, but enough! */
+   value *= nv50->screen->TPs;
+
+   *(uint64_t *)result = value;
+   return true;
+}
+
+static const struct nv50_hw_query_funcs hw_sm_query_funcs = {
+   .destroy_query = nv50_hw_sm_destroy_query,
+   .begin_query = nv50_hw_sm_begin_query,
+   .end_query = nv50_hw_sm_end_query,
+   .get_query_result = nv50_hw_sm_get_query_result,
+};
+
+struct nv50_hw_query *
+nv50_hw_sm_create_query(struct nv50_context *nv50, unsigned type)
+{
+   struct nv50_hw_sm_query *hsq;
+   struct nv50_hw_query *hq;
+   unsigned space;
+
+   if (type < NV50_HW_SM_QUERY(0) || type > NV50_HW_SM_QUERY_LAST)
+      return NULL;
+
+   hsq = CALLOC_STRUCT(nv50_hw_sm_query);
+   if (!hsq)
+      return NULL;
+
+   hq = &hsq->base;
+   hq->funcs = &hw_sm_query_funcs;
+   hq->base.type = type;
+
+   /*
+    * for each MP:
+    * [00] = MP.C0
+    * [04] = MP.C1
+    * [08] = MP.C2
+    * [0c] = MP.C3
+    * [10] = MP.sequence
+    */
+   space = (4 + 1) * nv50->screen->MPsInTP * sizeof(uint32_t);
+
+   if (!nv50_hw_query_allocate(nv50, &hq->base, space)) {
+      FREE(hq);
+      return NULL;
+   }
+
+   return hq;
+}
+
+int
+nv50_hw_sm_get_driver_query_info(struct nv50_screen *screen, unsigned id,
+                                 struct pipe_driver_query_info *info)
+{
+   int count = 0;
+
+   if (screen->compute)
+      if (screen->base.class_3d >= NV84_3D_CLASS)
+         count += NV50_HW_SM_QUERY_COUNT;
+
+   if (!info)
+      return count;
+
+   if (id < count) {
+      if (screen->compute) {
+         if (screen->base.class_3d >= NV84_3D_CLASS) {
+            info->name = nv50_hw_sm_query_names[id];
+            info->query_type = NV50_HW_SM_QUERY(id);
+            info->group_id = -1;
+            return 1;
+         }
+      }
+   }
+   return 0;
+}
diff --git a/src/gallium/drivers/nouveau/nv50/nv50_query_hw_sm.h b/src/gallium/drivers/nouveau/nv50/nv50_query_hw_sm.h
new file mode 100644 (file)
index 0000000..c1a1cd1
--- /dev/null
@@ -0,0 +1,45 @@
+#ifndef __NV50_QUERY_HW_SM_H__
+#define __NV50_QUERY_HW_SM_H__
+
+#include "nv50_query_hw.h"
+
+struct nv50_hw_sm_query {
+   struct nv50_hw_query base;
+   uint8_t ctr[4];
+};
+
+static inline struct nv50_hw_sm_query *
+nv50_hw_sm_query(struct nv50_hw_query *hq)
+{
+   return (struct nv50_hw_sm_query *)hq;
+}
+
+/*
+ * Performance counter queries:
+ */
+#define NV50_HW_SM_QUERY(i)    (PIPE_QUERY_DRIVER_SPECIFIC + (i))
+#define NV50_HW_SM_QUERY_LAST   NV50_HW_SM_QUERY(NV50_HW_SM_QUERY_COUNT - 1)
+enum nv50_hw_sm_queries
+{
+   NV50_HW_SM_QUERY_BRANCH = 0,
+   NV50_HW_SM_QUERY_DIVERGENT_BRANCH,
+   NV50_HW_SM_QUERY_INSTRUCTIONS,
+   NV50_HW_SM_QUERY_PROF_TRIGGER_0,
+   NV50_HW_SM_QUERY_PROF_TRIGGER_1,
+   NV50_HW_SM_QUERY_PROF_TRIGGER_2,
+   NV50_HW_SM_QUERY_PROF_TRIGGER_3,
+   NV50_HW_SM_QUERY_PROF_TRIGGER_4,
+   NV50_HW_SM_QUERY_PROF_TRIGGER_5,
+   NV50_HW_SM_QUERY_PROF_TRIGGER_6,
+   NV50_HW_SM_QUERY_PROF_TRIGGER_7,
+   NV50_HW_SM_QUERY_SM_CTA_LAUNCHED,
+   NV50_HW_SM_QUERY_WARP_SERIALIZE,
+   NV50_HW_SM_QUERY_COUNT,
+};
+
+struct nv50_hw_query *
+nv50_hw_sm_create_query(struct nv50_context *, unsigned);
+int
+nv50_hw_sm_get_driver_query_info(struct nv50_screen *, unsigned,
+                                 struct pipe_driver_query_info *);
+#endif
index 0142e86ba20e4c5cbe53dc5ce11d41e2f2750ec9..4e7201d7dd94fd38a945002b5dc15415123e216a 100644 (file)
@@ -789,6 +789,7 @@ nv50_screen_create(struct nouveau_device *dev)
    pscreen->get_shader_param = nv50_screen_get_shader_param;
    pscreen->get_paramf = nv50_screen_get_paramf;
    pscreen->get_compute_param = nv50_screen_get_compute_param;
+   pscreen->get_driver_query_info = nv50_screen_get_driver_query_info;
 
    nv50_screen_init_resource_functions(pscreen);
 
index 153ceea7a4fc02da140c0cea042e82268fe16bed..c2a16d8bd1d9723ea6adb299961892b6913d9c90 100644 (file)
@@ -95,6 +95,12 @@ struct nv50_screen {
       struct nouveau_bo *bo;
    } fence;
 
+   struct {
+      struct nv50_program *prog; /* compute state object to read MP counters */
+      struct nv50_hw_sm_query *mp_counter[4]; /* counter to query allocation */
+      uint8_t num_hw_sm_active;
+   } pm;
+
    struct nouveau_object *sync;
 
    struct nouveau_object *tesla;
@@ -109,6 +115,9 @@ nv50_screen(struct pipe_screen *screen)
    return (struct nv50_screen *)screen;
 }
 
+int nv50_screen_get_driver_query_info(struct pipe_screen *, unsigned,
+                                      struct pipe_driver_query_info *);
+
 bool nv50_blitter_create(struct nv50_screen *);
 void nv50_blitter_destroy(struct nv50_screen *);