radeonsi: implement AMD_performance_monitor for CIK+

Expose most of the performance counter groups that are exposed by Catalyst.
Ideally, the driver will work with GPUPerfStudio at some point, but we are not
quite there yet. In any case, this is the reason for grouping multiple
instances of hardware blocks in the way it is implemented.

The counters can also be shown using the Gallium HUD. If one is interested to
see how work is distributed across multiple shader engines, one can set the
environment variable RADEON_PC_SEPARATE_SE=1 to obtain finer-grained performance
counter groups.

Part of the implementation is in radeon because an implementation for
older hardware would largely follow along the same lines, but exposing
a different set of blocks which are programmed slightly differently.

Reviewed-by: Marek Olšák <[email protected]>

4 files changed, 702 insertions, 0 deletions

diff --git a/src/gallium/drivers/radeonsi/Makefile.sources b/src/gallium/drivers/radeonsi/Makefile.sources
index 7e997c6d526..53404ab0e58 100644
--- a/src/gallium/drivers/radeonsi/Makefile.sources
+++ b/src/gallium/drivers/radeonsi/Makefile.sources
@@ -12,6 +12,7 @@ C_SOURCES := \
 	si_pipe.h \
 	si_pm4.c \
 	si_pm4.h \
+	si_perfcounter.c \
 	si_public.h \
 	si_shader.c \
 	si_shader.h \
diff --git a/src/gallium/drivers/radeonsi/si_perfcounter.c b/src/gallium/drivers/radeonsi/si_perfcounter.c
new file mode 100644
index 00000000000..f5e479bc549
--- /dev/null
+++ b/src/gallium/drivers/radeonsi/si_perfcounter.c
@@ -0,0 +1,695 @@
+/*
+ * Copyright 2015 Advanced Micro Devices, Inc.
+ *
+ * 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 (including the next
+ * paragraph) 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:
+ *  Nicolai Hähnle <[email protected]>
+ *
+ */
+
+#include "radeon/r600_cs.h"
+#include "radeon/r600_query.h"
+#include "radeon/r600_pipe_common.h"
+#include "util/u_memory.h"
+
+#include "si_pipe.h"
+#include "sid.h"
+
+enum si_pc_reg_layout {
+	/* All secondary selector dwords follow as one block after the primary
+	 * selector dwords for the counters that have secondary selectors.
+	 */
+	SI_PC_MULTI_BLOCK = 0,
+
+	/* Each secondary selector dword follows immediately afters the
+	 * corresponding primary.
+	 */
+	SI_PC_MULTI_ALTERNATE = 1,
+
+	/* All secondary selector dwords follow as one block after all primary
+	 * selector dwords.
+	 */
+	SI_PC_MULTI_TAIL = 2,
+
+	/* Free-form arrangement of selector registers. */
+	SI_PC_MULTI_CUSTOM = 3,
+
+	SI_PC_MULTI_MASK = 3,
+
+	/* Registers are laid out in decreasing rather than increasing order. */
+	SI_PC_REG_REVERSE = 4,
+};
+
+struct si_pc_block_base {
+	const char *name;
+	unsigned num_counters;
+	unsigned flags;
+
+	unsigned select_or;
+	unsigned select0;
+	unsigned counter0_lo;
+	unsigned *select;
+	unsigned *counters;
+	unsigned num_multi;
+	unsigned num_prelude;
+	unsigned layout;
+};
+
+struct si_pc_block {
+	struct si_pc_block_base *b;
+	unsigned selectors;
+	unsigned instances;
+};
+
+
+static struct si_pc_block_base cik_CB = {
+	.name = "CB",
+	.num_counters = 4,
+	.flags = R600_PC_BLOCK_SE | R600_PC_BLOCK_INSTANCE_GROUPS,
+
+	.select0 = R_037000_CB_PERFCOUNTER_FILTER,
+	.counter0_lo = R_035018_CB_PERFCOUNTER0_LO,
+	.num_multi = 1,
+	.num_prelude = 1,
+	.layout = SI_PC_MULTI_ALTERNATE,
+};
+
+static unsigned cik_CPC_select[] = {
+	R_036024_CPC_PERFCOUNTER0_SELECT,
+	R_036010_CPC_PERFCOUNTER0_SELECT1,
+	R_03600C_CPC_PERFCOUNTER1_SELECT,
+};
+static struct si_pc_block_base cik_CPC = {
+	.name = "CPC",
+	.num_counters = 2,
+
+	.select = cik_CPC_select,
+	.counter0_lo = R_034018_CPC_PERFCOUNTER0_LO,
+	.num_multi = 1,
+	.layout = SI_PC_MULTI_CUSTOM | SI_PC_REG_REVERSE,
+};
+
+static struct si_pc_block_base cik_CPF = {
+	.name = "CPF",
+	.num_counters = 2,
+
+	.select0 = R_03601C_CPF_PERFCOUNTER0_SELECT,
+	.counter0_lo = R_034028_CPF_PERFCOUNTER0_LO,
+	.num_multi = 1,
+	.layout = SI_PC_MULTI_ALTERNATE | SI_PC_REG_REVERSE,
+};
+
+static struct si_pc_block_base cik_CPG = {
+	.name = "CPG",
+	.num_counters = 2,
+
+	.select0 = R_036008_CPG_PERFCOUNTER0_SELECT,
+	.counter0_lo = R_034008_CPG_PERFCOUNTER0_LO,
+	.num_multi = 1,
+	.layout = SI_PC_MULTI_ALTERNATE | SI_PC_REG_REVERSE,
+};
+
+static struct si_pc_block_base cik_DB = {
+	.name = "DB",
+	.num_counters = 4,
+	.flags = R600_PC_BLOCK_SE | R600_PC_BLOCK_INSTANCE_GROUPS,
+
+	.select0 = R_037100_DB_PERFCOUNTER0_SELECT,
+	.counter0_lo = R_035100_DB_PERFCOUNTER0_LO,
+	.num_multi = 3, // really only 2, but there's a gap between registers
+	.layout = SI_PC_MULTI_ALTERNATE,
+};
+
+static struct si_pc_block_base cik_GDS = {
+	.name = "GDS",
+	.num_counters = 4,
+
+	.select0 = R_036A00_GDS_PERFCOUNTER0_SELECT,
+	.counter0_lo = R_034A00_GDS_PERFCOUNTER0_LO,
+	.num_multi = 1,
+	.layout = SI_PC_MULTI_TAIL,
+};
+
+static unsigned cik_GRBM_counters[] = {
+	R_034100_GRBM_PERFCOUNTER0_LO,
+	R_03410C_GRBM_PERFCOUNTER1_LO,
+};
+static struct si_pc_block_base cik_GRBM = {
+	.name = "GRBM",
+	.num_counters = 2,
+
+	.select0 = R_036100_GRBM_PERFCOUNTER0_SELECT,
+	.counters = cik_GRBM_counters,
+};
+
+static struct si_pc_block_base cik_GRBMSE = {
+	.name = "GRBMSE",
+	.num_counters = 4,
+
+	.select0 = R_036108_GRBM_SE0_PERFCOUNTER_SELECT,
+	.counter0_lo = R_034114_GRBM_SE0_PERFCOUNTER_LO,
+};
+
+static struct si_pc_block_base cik_IA = {
+	.name = "IA",
+	.num_counters = 4,
+
+	.select0 = R_036210_IA_PERFCOUNTER0_SELECT,
+	.counter0_lo = R_034220_IA_PERFCOUNTER0_LO,
+	.num_multi = 1,
+	.layout = SI_PC_MULTI_TAIL,
+};
+
+static struct si_pc_block_base cik_PA_SC = {
+	.name = "PA_SC",
+	.num_counters = 8,
+	.flags = R600_PC_BLOCK_SE,
+
+	.select0 = R_036500_PA_SC_PERFCOUNTER0_SELECT,
+	.counter0_lo = R_034500_PA_SC_PERFCOUNTER0_LO,
+	.num_multi = 1,
+	.layout = SI_PC_MULTI_ALTERNATE,
+};
+
+static struct si_pc_block_base cik_PA_SU = {
+	.name = "PA_SU",
+	.num_counters = 4,
+	.flags = R600_PC_BLOCK_SE,
+
+	.select0 = R_036400_PA_SU_PERFCOUNTER0_SELECT,
+	.counter0_lo = R_034400_PA_SU_PERFCOUNTER0_LO,
+	.num_multi = 2,
+	.layout = SI_PC_MULTI_ALTERNATE,
+};
+
+static struct si_pc_block_base cik_SPI = {
+	.name = "SPI",
+	.num_counters = 6,
+	.flags = R600_PC_BLOCK_SE,
+
+	.select0 = R_036600_SPI_PERFCOUNTER0_SELECT,
+	.counter0_lo = R_034604_SPI_PERFCOUNTER0_LO,
+	.num_multi = 4,
+	.layout = SI_PC_MULTI_BLOCK,
+};
+
+static struct si_pc_block_base cik_SQ = {
+	.name = "SQ",
+	.num_counters = 16,
+	.flags = R600_PC_BLOCK_SE | R600_PC_BLOCK_SHADER,
+
+	.select0 = R_036700_SQ_PERFCOUNTER0_SELECT,
+	.select_or = S_036700_SQC_BANK_MASK(15) |
+			S_036700_SQC_CLIENT_MASK(15) |
+			S_036700_SIMD_MASK(15),
+	.counter0_lo = R_034700_SQ_PERFCOUNTER0_LO,
+};
+
+static struct si_pc_block_base cik_SX = {
+	.name = "SX",
+	.num_counters = 4,
+	.flags = R600_PC_BLOCK_SE,
+
+	.select0 = R_036900_SX_PERFCOUNTER0_SELECT,
+	.counter0_lo = R_034900_SX_PERFCOUNTER0_LO,
+	.num_multi = 2,
+	.layout = SI_PC_MULTI_TAIL,
+};
+
+static struct si_pc_block_base cik_TA = {
+	.name = "TA",
+	.num_counters = 2,
+	.flags = R600_PC_BLOCK_SE | R600_PC_BLOCK_INSTANCE_GROUPS | R600_PC_BLOCK_SHADER_WINDOWED,
+
+	.select0 = R_036B00_TA_PERFCOUNTER0_SELECT,
+	.counter0_lo = R_034B00_TA_PERFCOUNTER0_LO,
+	.num_multi = 1,
+	.layout = SI_PC_MULTI_ALTERNATE,
+};
+
+static struct si_pc_block_base cik_TD = {
+	.name = "TD",
+	.num_counters = 2,
+	.flags = R600_PC_BLOCK_SE | R600_PC_BLOCK_INSTANCE_GROUPS | R600_PC_BLOCK_SHADER_WINDOWED,
+
+	.select0 = R_036C00_TD_PERFCOUNTER0_SELECT,
+	.counter0_lo = R_034C00_TD_PERFCOUNTER0_LO,
+	.num_multi = 1,
+	.layout = SI_PC_MULTI_ALTERNATE,
+};
+
+static struct si_pc_block_base cik_TCA = {
+	.name = "TCA",
+	.num_counters = 4,
+	.flags = R600_PC_BLOCK_INSTANCE_GROUPS,
+
+	.select0 = R_036E40_TCA_PERFCOUNTER0_SELECT,
+	.counter0_lo = R_034E40_TCA_PERFCOUNTER0_LO,
+	.num_multi = 2,
+	.layout = SI_PC_MULTI_ALTERNATE,
+};
+
+static struct si_pc_block_base cik_TCC = {
+	.name = "TCC",
+	.num_counters = 4,
+	.flags = R600_PC_BLOCK_INSTANCE_GROUPS,
+
+	.select0 = R_036E00_TCC_PERFCOUNTER0_SELECT,
+	.counter0_lo = R_034E00_TCC_PERFCOUNTER0_LO,
+	.num_multi = 2,
+	.layout = SI_PC_MULTI_ALTERNATE,
+};
+
+static struct si_pc_block_base cik_TCP = {
+	.name = "TCP",
+	.num_counters = 4,
+	.flags = R600_PC_BLOCK_SE | R600_PC_BLOCK_INSTANCE_GROUPS | R600_PC_BLOCK_SHADER_WINDOWED,
+
+	.select0 = R_036D00_TCP_PERFCOUNTER0_SELECT,
+	.counter0_lo = R_034D00_TCP_PERFCOUNTER0_LO,
+	.num_multi = 2,
+	.layout = SI_PC_MULTI_ALTERNATE,
+};
+
+static struct si_pc_block_base cik_VGT = {
+	.name = "VGT",
+	.num_counters = 4,
+	.flags = R600_PC_BLOCK_SE,
+
+	.select0 = R_036230_VGT_PERFCOUNTER0_SELECT,
+	.counter0_lo = R_034240_VGT_PERFCOUNTER0_LO,
+	.num_multi = 1,
+	.layout = SI_PC_MULTI_TAIL,
+};
+
+static struct si_pc_block_base cik_WD = {
+	.name = "WD",
+	.num_counters = 4,
+
+	.select0 = R_036200_WD_PERFCOUNTER0_SELECT,
+	.counter0_lo = R_034200_WD_PERFCOUNTER0_LO,
+};
+
+/* Both the number of instances and selectors varies between chips of the same
+ * class. We only differentiate by class here and simply expose the maximum
+ * number over all chips in a class.
+ */
+static struct si_pc_block groups_CIK[] = {
+	{ &cik_CB, 226, 4 },
+	{ &cik_CPC, 22 },
+	{ &cik_CPF, 17 },
+	{ &cik_CPG, 46 },
+	{ &cik_DB, 257, 4 },
+	{ &cik_GDS, 121 },
+	{ &cik_GRBM, 34 },
+	{ &cik_GRBMSE, 15 },
+	{ &cik_IA, 22 },
+	{ &cik_PA_SC, 395 },
+	{ &cik_PA_SU, 153 },
+	{ &cik_SPI, 186 },
+	{ &cik_SQ, 252 },
+	{ &cik_SX, 32 },
+	{ &cik_TA, 111, 11 },
+	{ &cik_TCA, 39, 2 },
+	{ &cik_TCC, 160, 16 },
+	{ &cik_TCP, 154, 11 },
+	{ &cik_TD, 55, 11 },
+	{ &cik_VGT, 140 },
+	{ &cik_WD, 22 },
+};
+
+static struct si_pc_block groups_VI[] = {
+	{ &cik_CB, 396, 4 },
+	{ &cik_CPC, 24 },
+	{ &cik_CPF, 19 },
+	{ &cik_CPG, 48 },
+	{ &cik_DB, 257, 4 },
+	{ &cik_GDS, 121 },
+	{ &cik_GRBM, 34 },
+	{ &cik_GRBMSE, 15 },
+	{ &cik_IA, 24 },
+	{ &cik_PA_SC, 397 },
+	{ &cik_PA_SU, 153 },
+	{ &cik_SPI, 197 },
+	{ &cik_SQ, 273 },
+	{ &cik_SX, 34 },
+	{ &cik_TA, 119, 16 },
+	{ &cik_TCA, 35, 2 },
+	{ &cik_TCC, 192, 16 },
+	{ &cik_TCP, 180, 16 },
+	{ &cik_TD, 55, 16 },
+	{ &cik_VGT, 147 },
+	{ &cik_WD, 37 },
+};
+
+static void si_pc_get_size(struct r600_perfcounter_block *group,
+			unsigned count, unsigned *selectors,
+			unsigned *num_select_dw, unsigned *num_read_dw)
+{
+	struct si_pc_block *sigroup = (struct si_pc_block *)group->data;
+	struct si_pc_block_base *regs = sigroup->b;
+	unsigned layout_multi = regs->layout & SI_PC_MULTI_MASK;
+
+	if (layout_multi == SI_PC_MULTI_BLOCK) {
+		if (count < regs->num_multi)
+			*num_select_dw = 2 * (count + 2) + regs->num_prelude;
+		else
+			*num_select_dw = 2 + count + regs->num_multi + regs->num_prelude;
+	} else if (layout_multi == SI_PC_MULTI_TAIL) {
+		*num_select_dw = 4 + count + MIN2(count, regs->num_multi) + regs->num_prelude;
+	} else if (layout_multi == SI_PC_MULTI_CUSTOM) {
+		assert(regs->num_prelude == 0);
+		*num_select_dw = 3 * (count + MIN2(count, regs->num_multi));
+	} else {
+		assert(layout_multi == SI_PC_MULTI_ALTERNATE);
+
+		*num_select_dw = 2 + count + MIN2(count, regs->num_multi) + regs->num_prelude;
+	}
+
+	*num_read_dw = 6 * count;
+}
+
+static void si_pc_emit_instance(struct r600_common_context *ctx,
+				int se, int instance)
+{
+	struct radeon_winsys_cs *cs = ctx->gfx.cs;
+	unsigned value = S_030800_SH_BROADCAST_WRITES(1);
+
+	if (se >= 0) {
+		value |= S_030800_SE_INDEX(se);
+	} else {
+		value |= S_030800_SE_BROADCAST_WRITES(1);
+	}
+
+	if (instance >= 0) {
+		value |= S_030800_INSTANCE_INDEX(instance);
+	} else {
+		value |= S_030800_INSTANCE_BROADCAST_WRITES(1);
+	}
+
+	radeon_set_uconfig_reg(cs, R_030800_GRBM_GFX_INDEX, value);
+}
+
+static void si_pc_emit_shaders(struct r600_common_context *ctx,
+			       unsigned shaders)
+{
+	struct radeon_winsys_cs *cs = ctx->gfx.cs;
+
+	radeon_set_uconfig_reg_seq(cs, R_036780_SQ_PERFCOUNTER_CTRL, 2);
+	radeon_emit(cs, shaders & 0x7f);
+	radeon_emit(cs, 0xffffffff);
+}
+
+static void si_pc_emit_select(struct r600_common_context *ctx,
+		        struct r600_perfcounter_block *group,
+		        unsigned count, unsigned *selectors)
+{
+	struct si_pc_block *sigroup = (struct si_pc_block *)group->data;
+	struct si_pc_block_base *regs = sigroup->b;
+	struct radeon_winsys_cs *cs = ctx->gfx.cs;
+	unsigned idx;
+	unsigned layout_multi = regs->layout & SI_PC_MULTI_MASK;
+	unsigned dw;
+
+	assert(count <= regs->num_counters);
+
+	if (layout_multi == SI_PC_MULTI_BLOCK) {
+		assert(!(regs->layout & SI_PC_REG_REVERSE));
+
+		dw = count + regs->num_prelude;
+		if (count >= regs->num_multi)
+			count += regs->num_multi;
+		radeon_set_uconfig_reg_seq(cs, regs->select0, dw);
+		for (idx = 0; idx < regs->num_prelude; ++idx)
+			radeon_emit(cs, 0);
+		for (idx = 0; idx < MIN2(count, regs->num_multi); ++idx)
+			radeon_emit(cs, selectors[idx] | regs->select_or);
+
+		if (count < regs->num_multi) {
+			unsigned select1 =
+				regs->select0 + 4 * regs->num_multi;
+			radeon_set_uconfig_reg_seq(cs, select1, count);
+		}
+
+		for (idx = 0; idx < MIN2(count, regs->num_multi); ++idx)
+			radeon_emit(cs, 0);
+
+		if (count > regs->num_multi) {
+			for (idx = regs->num_multi; idx < count; ++idx)
+				radeon_emit(cs, selectors[idx] | regs->select_or);
+		}
+	} else if (layout_multi == SI_PC_MULTI_TAIL) {
+		unsigned select1, select1_count;
+
+		assert(!(regs->layout & SI_PC_REG_REVERSE));
+
+		radeon_set_uconfig_reg_seq(cs, regs->select0, count + regs->num_prelude);
+		for (idx = 0; idx < regs->num_prelude; ++idx)
+			radeon_emit(cs, 0);
+		for (idx = 0; idx < count; ++idx)
+			radeon_emit(cs, selectors[idx] | regs->select_or);
+
+		select1 = regs->select0 + 4 * regs->num_counters;
+		select1_count = MIN2(count, regs->num_multi);
+		radeon_set_uconfig_reg_seq(cs, select1, select1_count);
+		for (idx = 0; idx < select1_count; ++idx)
+			radeon_emit(cs, 0);
+	} else if (layout_multi == SI_PC_MULTI_CUSTOM) {
+		unsigned *reg = regs->select;
+		for (idx = 0; idx < count; ++idx) {
+			radeon_set_uconfig_reg(cs, *reg++, selectors[idx] | regs->select_or);
+			if (idx < regs->num_multi)
+				radeon_set_uconfig_reg(cs, *reg++, 0);
+		}
+	} else {
+		assert(layout_multi == SI_PC_MULTI_ALTERNATE);
+
+		unsigned reg_base = regs->select0;
+		unsigned reg_count = count + MIN2(count, regs->num_multi);
+		reg_count += regs->num_prelude;
+
+		if (!(regs->layout & SI_PC_REG_REVERSE)) {
+			radeon_set_uconfig_reg_seq(cs, reg_base, reg_count);
+
+			for (idx = 0; idx < regs->num_prelude; ++idx)
+				radeon_emit(cs, 0);
+			for (idx = 0; idx < count; ++idx) {
+				radeon_emit(cs, selectors[idx] | regs->select_or);
+				if (idx < regs->num_multi)
+					radeon_emit(cs, 0);
+			}
+		} else {
+			reg_base -= (reg_count - 1) * 4;
+			radeon_set_uconfig_reg_seq(cs, reg_base, reg_count);
+
+			for (idx = count; idx > 0; --idx) {
+				if (idx <= regs->num_multi)
+					radeon_emit(cs, 0);
+				radeon_emit(cs, selectors[idx - 1] | regs->select_or);
+			}
+			for (idx = 0; idx < regs->num_prelude; ++idx)
+				radeon_emit(cs, 0);
+		}
+	}
+}
+
+static void si_pc_emit_start(struct r600_common_context *ctx,
+			     struct r600_resource *buffer, uint64_t va)
+{
+	struct radeon_winsys_cs *cs = ctx->gfx.cs;
+
+	radeon_add_to_buffer_list(ctx, &ctx->gfx, buffer,
+				  RADEON_USAGE_WRITE, RADEON_PRIO_QUERY);
+
+	radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
+	radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_IMM) |
+			COPY_DATA_DST_SEL(COPY_DATA_MEM));
+	radeon_emit(cs, 1); /* immediate */
+	radeon_emit(cs, 0); /* unused */
+	radeon_emit(cs, va);
+	radeon_emit(cs, va >> 32);
+
+	radeon_set_uconfig_reg(cs, R_036020_CP_PERFMON_CNTL,
+			       S_036020_PERFMON_STATE(V_036020_DISABLE_AND_RESET));
+	radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+	radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_PERFCOUNTER_START) | EVENT_INDEX(0));
+	radeon_set_uconfig_reg(cs, R_036020_CP_PERFMON_CNTL,
+			       S_036020_PERFMON_STATE(V_036020_START_COUNTING));
+}
+
+/* Note: The buffer was already added in si_pc_emit_start, so we don't have to
+ * do it again in here. */
+static void si_pc_emit_stop(struct r600_common_context *ctx,
+			    struct r600_resource *buffer, uint64_t va)
+{
+	struct radeon_winsys_cs *cs = ctx->gfx.cs;
+
+	if (ctx->screen->chip_class == CIK) {
+		/* Workaround for cache flush problems: send two EOP events. */
+		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0));
+		radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH_AND_INV_TS_EVENT) |
+				EVENT_INDEX(5));
+		radeon_emit(cs, va);
+		radeon_emit(cs, (va >> 32) | EOP_DATA_SEL(1));
+		radeon_emit(cs, 0); /* immediate data */
+		radeon_emit(cs, 0); /* unused */
+	}
+
+	radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0));
+	radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH_AND_INV_TS_EVENT) |
+			EVENT_INDEX(5));
+	radeon_emit(cs, va);
+	radeon_emit(cs, (va >> 32) | EOP_DATA_SEL(1));
+	radeon_emit(cs, 0); /* immediate data */
+	radeon_emit(cs, 0); /* unused */
+
+	radeon_emit(cs, PKT3(PKT3_WAIT_REG_MEM, 5, 0));
+	radeon_emit(cs, WAIT_REG_MEM_EQUAL | WAIT_REG_MEM_MEM_SPACE(1));
+	radeon_emit(cs, va);
+	radeon_emit(cs, va >> 32);
+	radeon_emit(cs, 0); /* reference value */
+	radeon_emit(cs, 0xffffffff); /* mask */
+	radeon_emit(cs, 4); /* poll interval */
+
+	radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+	radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_PERFCOUNTER_SAMPLE) | EVENT_INDEX(0));
+	radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+	radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_PERFCOUNTER_STOP) | EVENT_INDEX(0));
+	radeon_set_uconfig_reg(cs, R_036020_CP_PERFMON_CNTL,
+			       S_036020_PERFMON_STATE(V_036020_STOP_COUNTING) |
+			       S_036020_PERFMON_SAMPLE_ENABLE(1));
+}
+
+static void si_pc_emit_read(struct r600_common_context *ctx,
+			    struct r600_perfcounter_block *group,
+			    unsigned count, unsigned *selectors,
+			    struct r600_resource *buffer, uint64_t va)
+{
+	struct si_pc_block *sigroup = (struct si_pc_block *)group->data;
+	struct si_pc_block_base *regs = sigroup->b;
+	struct radeon_winsys_cs *cs = ctx->gfx.cs;
+	unsigned idx;
+	unsigned reg = regs->counter0_lo;
+	unsigned reg_delta = 8;
+
+	if (regs->layout & SI_PC_REG_REVERSE)
+		reg_delta = -reg_delta;
+
+	for (idx = 0; idx < count; ++idx) {
+		if (regs->counters)
+			reg = regs->counters[idx];
+
+		radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
+		radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_PERF) |
+				COPY_DATA_DST_SEL(COPY_DATA_MEM));
+		radeon_emit(cs, reg >> 2);
+		radeon_emit(cs, 0); /* unused */
+		radeon_emit(cs, va);
+		radeon_emit(cs, va >> 32);
+		va += 4;
+		reg += reg_delta;
+	}
+}
+
+static void si_pc_cleanup(struct r600_common_screen *rscreen)
+{
+	r600_perfcounters_do_destroy(rscreen->perfcounters);
+	rscreen->perfcounters = NULL;
+}
+
+void si_init_perfcounters(struct si_screen *screen)
+{
+	struct r600_perfcounters *pc;
+	struct si_pc_block *blocks;
+	unsigned num_blocks;
+	unsigned i;
+
+	switch (screen->b.chip_class) {
+	case CIK:
+		blocks = groups_CIK;
+		num_blocks = ARRAY_SIZE(groups_CIK);
+		break;
+	case VI:
+		blocks = groups_VI;
+		num_blocks = ARRAY_SIZE(groups_VI);
+		break;
+	case SI:
+	default:
+		return; /* not implemented */
+	}
+
+	if (screen->b.info.max_sh_per_se != 1) {
+		/* This should not happen on non-SI chips. */
+		fprintf(stderr, "si_init_perfcounters: max_sh_per_se = %d not "
+			"supported (inaccurate performance counters)\n",
+			screen->b.info.max_sh_per_se);
+	}
+
+	pc = CALLOC_STRUCT(r600_perfcounters);
+	if (!pc)
+		return;
+
+	pc->num_start_cs_dwords = 14;
+	pc->num_stop_cs_dwords = 20;
+	pc->num_instance_cs_dwords = 3;
+	pc->num_shaders_cs_dwords = 4;
+
+	if (screen->b.chip_class == CIK) {
+		pc->num_stop_cs_dwords += 6;
+	}
+
+	pc->get_size = si_pc_get_size;
+	pc->emit_instance = si_pc_emit_instance;
+	pc->emit_shaders = si_pc_emit_shaders;
+	pc->emit_select = si_pc_emit_select;
+	pc->emit_start = si_pc_emit_start;
+	pc->emit_stop = si_pc_emit_stop;
+	pc->emit_read = si_pc_emit_read;
+	pc->cleanup = si_pc_cleanup;
+
+	if (!r600_perfcounters_init(pc, num_blocks))
+		goto error;
+
+	for (i = 0; i < num_blocks; ++i) {
+		struct si_pc_block *block = &blocks[i];
+		unsigned instances = block->instances;
+
+		if (!strcmp(block->b->name, "IA")) {
+			if (screen->b.info.max_se > 2)
+				instances = 2;
+		}
+
+		if (!r600_perfcounters_add_block(&screen->b, pc,
+						 block->b->name,
+						 block->b->flags,
+						 block->b->num_counters,
+						 block->selectors,
+						 instances,
+						 block))
+			goto error;
+	}
+
+	screen->b.perfcounters = pc;
+	return;
+
+error:
+	r600_perfcounters_do_destroy(pc);
+}
diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index 9a0fe808ee3..81d809b6b9f 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -630,6 +630,9 @@ struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws)
 		return NULL;
 	}
 
+	if (!debug_get_bool_option("RADEON_DISABLE_PERFCOUNTERS", FALSE))
+		si_init_perfcounters(sscreen);
+
 	sscreen->b.has_cp_dma = true;
 	sscreen->b.has_streamout = true;
 
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 05d52fe19dc..834c3587569 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -346,6 +346,9 @@ void si_need_cs_space(struct si_context *ctx);
 /* si_compute.c */
 void si_init_compute_functions(struct si_context *sctx);
 
+/* si_perfcounters.c */
+void si_init_perfcounters(struct si_screen *screen);
+
 /* si_uvd.c */
 struct pipe_video_codec *si_uvd_create_decoder(struct pipe_context *context,
 					       const struct pipe_video_codec *templ);