radeonsi: move r600_query.c/h files to si_query.c/h

Acked-by: Timothy Arceri <[email protected]>

6 files changed, 2395 insertions, 2 deletions

diff --git a/src/gallium/drivers/radeonsi/Makefile.sources b/src/gallium/drivers/radeonsi/Makefile.sources
index be9c171f419..4182b8af489 100644
--- a/src/gallium/drivers/radeonsi/Makefile.sources
+++ b/src/gallium/drivers/radeonsi/Makefile.sources
@@ -23,6 +23,8 @@ C_SOURCES := \
 	si_pm4.h \
 	si_perfcounter.c \
 	si_public.h \
+	si_query.c \
+	si_query.h \
 	si_shader.c \
 	si_shader.h \
 	si_shader_internal.h \
diff --git a/src/gallium/drivers/radeonsi/si_perfcounter.c b/src/gallium/drivers/radeonsi/si_perfcounter.c
index 0593d5523b8..7ab995cacc9 100644
--- a/src/gallium/drivers/radeonsi/si_perfcounter.c
+++ b/src/gallium/drivers/radeonsi/si_perfcounter.c
@@ -23,7 +23,7 @@
  */
 
 #include "si_build_pm4.h"
-#include "radeon/r600_query.h"
+#include "si_query.h"
 #include "util/u_memory.h"
 
 
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index ab58ec5a484..8e00861ae3a 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -840,6 +840,12 @@ void si_init_perfcounters(struct si_screen *screen);
 /* si_pipe.c */
 bool si_check_device_reset(struct si_context *sctx);
 
+/* si_query.c */
+void si_init_screen_query_functions(struct si_screen *sscreen);
+void si_init_query_functions(struct si_context *sctx);
+void si_suspend_queries(struct si_context *sctx);
+void si_resume_queries(struct si_context *sctx);
+
 /* si_test_dma.c */
 void si_test_dma(struct si_screen *sscreen);
 
diff --git a/src/gallium/drivers/radeonsi/si_query.c b/src/gallium/drivers/radeonsi/si_query.c
new file mode 100644
index 00000000000..67c323aa959
--- /dev/null
+++ b/src/gallium/drivers/radeonsi/si_query.c
@@ -0,0 +1,2067 @@
+/*
+ * Copyright 2010 Jerome Glisse <[email protected]>
+ * Copyright 2014 Marek Olšák <[email protected]>
+ * Copyright 2018 Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * 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
+ * on the rights to use, copy, modify, merge, publish, distribute, sub
+ * license, 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 NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHOR(S) AND/OR THEIR SUPPLIERS 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.
+ */
+
+#include "si_pipe.h"
+#include "si_query.h"
+#include "util/u_memory.h"
+#include "util/u_upload_mgr.h"
+#include "util/os_time.h"
+#include "tgsi/tgsi_text.h"
+#include "amd/common/sid.h"
+
+#define SI_MAX_STREAMS 4
+
+struct si_hw_query_params {
+	unsigned start_offset;
+	unsigned end_offset;
+	unsigned fence_offset;
+	unsigned pair_stride;
+	unsigned pair_count;
+};
+
+/* Queries without buffer handling or suspend/resume. */
+struct si_query_sw {
+	struct si_query b;
+
+	uint64_t begin_result;
+	uint64_t end_result;
+
+	uint64_t begin_time;
+	uint64_t end_time;
+
+	/* Fence for GPU_FINISHED. */
+	struct pipe_fence_handle *fence;
+};
+
+static void si_query_sw_destroy(struct si_screen *sscreen,
+				struct si_query *rquery)
+{
+	struct si_query_sw *query = (struct si_query_sw *)rquery;
+
+	sscreen->b.fence_reference(&sscreen->b, &query->fence, NULL);
+	FREE(query);
+}
+
+static enum radeon_value_id winsys_id_from_type(unsigned type)
+{
+	switch (type) {
+	case SI_QUERY_REQUESTED_VRAM: return RADEON_REQUESTED_VRAM_MEMORY;
+	case SI_QUERY_REQUESTED_GTT: return RADEON_REQUESTED_GTT_MEMORY;
+	case SI_QUERY_MAPPED_VRAM: return RADEON_MAPPED_VRAM;
+	case SI_QUERY_MAPPED_GTT: return RADEON_MAPPED_GTT;
+	case SI_QUERY_BUFFER_WAIT_TIME: return RADEON_BUFFER_WAIT_TIME_NS;
+	case SI_QUERY_NUM_MAPPED_BUFFERS: return RADEON_NUM_MAPPED_BUFFERS;
+	case SI_QUERY_NUM_GFX_IBS: return RADEON_NUM_GFX_IBS;
+	case SI_QUERY_NUM_SDMA_IBS: return RADEON_NUM_SDMA_IBS;
+	case SI_QUERY_GFX_BO_LIST_SIZE: return RADEON_GFX_BO_LIST_COUNTER;
+	case SI_QUERY_GFX_IB_SIZE: return RADEON_GFX_IB_SIZE_COUNTER;
+	case SI_QUERY_NUM_BYTES_MOVED: return RADEON_NUM_BYTES_MOVED;
+	case SI_QUERY_NUM_EVICTIONS: return RADEON_NUM_EVICTIONS;
+	case SI_QUERY_NUM_VRAM_CPU_PAGE_FAULTS: return RADEON_NUM_VRAM_CPU_PAGE_FAULTS;
+	case SI_QUERY_VRAM_USAGE: return RADEON_VRAM_USAGE;
+	case SI_QUERY_VRAM_VIS_USAGE: return RADEON_VRAM_VIS_USAGE;
+	case SI_QUERY_GTT_USAGE: return RADEON_GTT_USAGE;
+	case SI_QUERY_GPU_TEMPERATURE: return RADEON_GPU_TEMPERATURE;
+	case SI_QUERY_CURRENT_GPU_SCLK: return RADEON_CURRENT_SCLK;
+	case SI_QUERY_CURRENT_GPU_MCLK: return RADEON_CURRENT_MCLK;
+	case SI_QUERY_CS_THREAD_BUSY: return RADEON_CS_THREAD_TIME;
+	default: unreachable("query type does not correspond to winsys id");
+	}
+}
+
+static bool si_query_sw_begin(struct si_context *sctx,
+			      struct si_query *rquery)
+{
+	struct si_query_sw *query = (struct si_query_sw *)rquery;
+	enum radeon_value_id ws_id;
+
+	switch(query->b.type) {
+	case PIPE_QUERY_TIMESTAMP_DISJOINT:
+	case PIPE_QUERY_GPU_FINISHED:
+		break;
+	case SI_QUERY_DRAW_CALLS:
+		query->begin_result = sctx->b.num_draw_calls;
+		break;
+	case SI_QUERY_DECOMPRESS_CALLS:
+		query->begin_result = sctx->b.num_decompress_calls;
+		break;
+	case SI_QUERY_MRT_DRAW_CALLS:
+		query->begin_result = sctx->b.num_mrt_draw_calls;
+		break;
+	case SI_QUERY_PRIM_RESTART_CALLS:
+		query->begin_result = sctx->b.num_prim_restart_calls;
+		break;
+	case SI_QUERY_SPILL_DRAW_CALLS:
+		query->begin_result = sctx->b.num_spill_draw_calls;
+		break;
+	case SI_QUERY_COMPUTE_CALLS:
+		query->begin_result = sctx->b.num_compute_calls;
+		break;
+	case SI_QUERY_SPILL_COMPUTE_CALLS:
+		query->begin_result = sctx->b.num_spill_compute_calls;
+		break;
+	case SI_QUERY_DMA_CALLS:
+		query->begin_result = sctx->b.num_dma_calls;
+		break;
+	case SI_QUERY_CP_DMA_CALLS:
+		query->begin_result = sctx->b.num_cp_dma_calls;
+		break;
+	case SI_QUERY_NUM_VS_FLUSHES:
+		query->begin_result = sctx->b.num_vs_flushes;
+		break;
+	case SI_QUERY_NUM_PS_FLUSHES:
+		query->begin_result = sctx->b.num_ps_flushes;
+		break;
+	case SI_QUERY_NUM_CS_FLUSHES:
+		query->begin_result = sctx->b.num_cs_flushes;
+		break;
+	case SI_QUERY_NUM_CB_CACHE_FLUSHES:
+		query->begin_result = sctx->b.num_cb_cache_flushes;
+		break;
+	case SI_QUERY_NUM_DB_CACHE_FLUSHES:
+		query->begin_result = sctx->b.num_db_cache_flushes;
+		break;
+	case SI_QUERY_NUM_L2_INVALIDATES:
+		query->begin_result = sctx->b.num_L2_invalidates;
+		break;
+	case SI_QUERY_NUM_L2_WRITEBACKS:
+		query->begin_result = sctx->b.num_L2_writebacks;
+		break;
+	case SI_QUERY_NUM_RESIDENT_HANDLES:
+		query->begin_result = sctx->b.num_resident_handles;
+		break;
+	case SI_QUERY_TC_OFFLOADED_SLOTS:
+		query->begin_result = sctx->b.tc ? sctx->b.tc->num_offloaded_slots : 0;
+		break;
+	case SI_QUERY_TC_DIRECT_SLOTS:
+		query->begin_result = sctx->b.tc ? sctx->b.tc->num_direct_slots : 0;
+		break;
+	case SI_QUERY_TC_NUM_SYNCS:
+		query->begin_result = sctx->b.tc ? sctx->b.tc->num_syncs : 0;
+		break;
+	case SI_QUERY_REQUESTED_VRAM:
+	case SI_QUERY_REQUESTED_GTT:
+	case SI_QUERY_MAPPED_VRAM:
+	case SI_QUERY_MAPPED_GTT:
+	case SI_QUERY_VRAM_USAGE:
+	case SI_QUERY_VRAM_VIS_USAGE:
+	case SI_QUERY_GTT_USAGE:
+	case SI_QUERY_GPU_TEMPERATURE:
+	case SI_QUERY_CURRENT_GPU_SCLK:
+	case SI_QUERY_CURRENT_GPU_MCLK:
+	case SI_QUERY_BACK_BUFFER_PS_DRAW_RATIO:
+	case SI_QUERY_NUM_MAPPED_BUFFERS:
+		query->begin_result = 0;
+		break;
+	case SI_QUERY_BUFFER_WAIT_TIME:
+	case SI_QUERY_GFX_IB_SIZE:
+	case SI_QUERY_NUM_GFX_IBS:
+	case SI_QUERY_NUM_SDMA_IBS:
+	case SI_QUERY_NUM_BYTES_MOVED:
+	case SI_QUERY_NUM_EVICTIONS:
+	case SI_QUERY_NUM_VRAM_CPU_PAGE_FAULTS: {
+		enum radeon_value_id ws_id = winsys_id_from_type(query->b.type);
+		query->begin_result = sctx->b.ws->query_value(sctx->b.ws, ws_id);
+		break;
+	}
+	case SI_QUERY_GFX_BO_LIST_SIZE:
+		ws_id = winsys_id_from_type(query->b.type);
+		query->begin_result = sctx->b.ws->query_value(sctx->b.ws, ws_id);
+		query->begin_time = sctx->b.ws->query_value(sctx->b.ws,
+							  RADEON_NUM_GFX_IBS);
+		break;
+	case SI_QUERY_CS_THREAD_BUSY:
+		ws_id = winsys_id_from_type(query->b.type);
+		query->begin_result = sctx->b.ws->query_value(sctx->b.ws, ws_id);
+		query->begin_time = os_time_get_nano();
+		break;
+	case SI_QUERY_GALLIUM_THREAD_BUSY:
+		query->begin_result =
+			sctx->b.tc ? util_queue_get_thread_time_nano(&sctx->b.tc->queue, 0) : 0;
+		query->begin_time = os_time_get_nano();
+		break;
+	case SI_QUERY_GPU_LOAD:
+	case SI_QUERY_GPU_SHADERS_BUSY:
+	case SI_QUERY_GPU_TA_BUSY:
+	case SI_QUERY_GPU_GDS_BUSY:
+	case SI_QUERY_GPU_VGT_BUSY:
+	case SI_QUERY_GPU_IA_BUSY:
+	case SI_QUERY_GPU_SX_BUSY:
+	case SI_QUERY_GPU_WD_BUSY:
+	case SI_QUERY_GPU_BCI_BUSY:
+	case SI_QUERY_GPU_SC_BUSY:
+	case SI_QUERY_GPU_PA_BUSY:
+	case SI_QUERY_GPU_DB_BUSY:
+	case SI_QUERY_GPU_CP_BUSY:
+	case SI_QUERY_GPU_CB_BUSY:
+	case SI_QUERY_GPU_SDMA_BUSY:
+	case SI_QUERY_GPU_PFP_BUSY:
+	case SI_QUERY_GPU_MEQ_BUSY:
+	case SI_QUERY_GPU_ME_BUSY:
+	case SI_QUERY_GPU_SURF_SYNC_BUSY:
+	case SI_QUERY_GPU_CP_DMA_BUSY:
+	case SI_QUERY_GPU_SCRATCH_RAM_BUSY:
+		query->begin_result = si_begin_counter(sctx->screen,
+							 query->b.type);
+		break;
+	case SI_QUERY_NUM_COMPILATIONS:
+		query->begin_result = p_atomic_read(&sctx->screen->num_compilations);
+		break;
+	case SI_QUERY_NUM_SHADERS_CREATED:
+		query->begin_result = p_atomic_read(&sctx->screen->num_shaders_created);
+		break;
+	case SI_QUERY_NUM_SHADER_CACHE_HITS:
+		query->begin_result =
+			p_atomic_read(&sctx->screen->num_shader_cache_hits);
+		break;
+	case SI_QUERY_GPIN_ASIC_ID:
+	case SI_QUERY_GPIN_NUM_SIMD:
+	case SI_QUERY_GPIN_NUM_RB:
+	case SI_QUERY_GPIN_NUM_SPI:
+	case SI_QUERY_GPIN_NUM_SE:
+		break;
+	default:
+		unreachable("si_query_sw_begin: bad query type");
+	}
+
+	return true;
+}
+
+static bool si_query_sw_end(struct si_context *sctx,
+			    struct si_query *rquery)
+{
+	struct si_query_sw *query = (struct si_query_sw *)rquery;
+	enum radeon_value_id ws_id;
+
+	switch(query->b.type) {
+	case PIPE_QUERY_TIMESTAMP_DISJOINT:
+		break;
+	case PIPE_QUERY_GPU_FINISHED:
+		sctx->b.b.flush(&sctx->b.b, &query->fence, PIPE_FLUSH_DEFERRED);
+		break;
+	case SI_QUERY_DRAW_CALLS:
+		query->end_result = sctx->b.num_draw_calls;
+		break;
+	case SI_QUERY_DECOMPRESS_CALLS:
+		query->end_result = sctx->b.num_decompress_calls;
+		break;
+	case SI_QUERY_MRT_DRAW_CALLS:
+		query->end_result = sctx->b.num_mrt_draw_calls;
+		break;
+	case SI_QUERY_PRIM_RESTART_CALLS:
+		query->end_result = sctx->b.num_prim_restart_calls;
+		break;
+	case SI_QUERY_SPILL_DRAW_CALLS:
+		query->end_result = sctx->b.num_spill_draw_calls;
+		break;
+	case SI_QUERY_COMPUTE_CALLS:
+		query->end_result = sctx->b.num_compute_calls;
+		break;
+	case SI_QUERY_SPILL_COMPUTE_CALLS:
+		query->end_result = sctx->b.num_spill_compute_calls;
+		break;
+	case SI_QUERY_DMA_CALLS:
+		query->end_result = sctx->b.num_dma_calls;
+		break;
+	case SI_QUERY_CP_DMA_CALLS:
+		query->end_result = sctx->b.num_cp_dma_calls;
+		break;
+	case SI_QUERY_NUM_VS_FLUSHES:
+		query->end_result = sctx->b.num_vs_flushes;
+		break;
+	case SI_QUERY_NUM_PS_FLUSHES:
+		query->end_result = sctx->b.num_ps_flushes;
+		break;
+	case SI_QUERY_NUM_CS_FLUSHES:
+		query->end_result = sctx->b.num_cs_flushes;
+		break;
+	case SI_QUERY_NUM_CB_CACHE_FLUSHES:
+		query->end_result = sctx->b.num_cb_cache_flushes;
+		break;
+	case SI_QUERY_NUM_DB_CACHE_FLUSHES:
+		query->end_result = sctx->b.num_db_cache_flushes;
+		break;
+	case SI_QUERY_NUM_L2_INVALIDATES:
+		query->end_result = sctx->b.num_L2_invalidates;
+		break;
+	case SI_QUERY_NUM_L2_WRITEBACKS:
+		query->end_result = sctx->b.num_L2_writebacks;
+		break;
+	case SI_QUERY_NUM_RESIDENT_HANDLES:
+		query->end_result = sctx->b.num_resident_handles;
+		break;
+	case SI_QUERY_TC_OFFLOADED_SLOTS:
+		query->end_result = sctx->b.tc ? sctx->b.tc->num_offloaded_slots : 0;
+		break;
+	case SI_QUERY_TC_DIRECT_SLOTS:
+		query->end_result = sctx->b.tc ? sctx->b.tc->num_direct_slots : 0;
+		break;
+	case SI_QUERY_TC_NUM_SYNCS:
+		query->end_result = sctx->b.tc ? sctx->b.tc->num_syncs : 0;
+		break;
+	case SI_QUERY_REQUESTED_VRAM:
+	case SI_QUERY_REQUESTED_GTT:
+	case SI_QUERY_MAPPED_VRAM:
+	case SI_QUERY_MAPPED_GTT:
+	case SI_QUERY_VRAM_USAGE:
+	case SI_QUERY_VRAM_VIS_USAGE:
+	case SI_QUERY_GTT_USAGE:
+	case SI_QUERY_GPU_TEMPERATURE:
+	case SI_QUERY_CURRENT_GPU_SCLK:
+	case SI_QUERY_CURRENT_GPU_MCLK:
+	case SI_QUERY_BUFFER_WAIT_TIME:
+	case SI_QUERY_GFX_IB_SIZE:
+	case SI_QUERY_NUM_MAPPED_BUFFERS:
+	case SI_QUERY_NUM_GFX_IBS:
+	case SI_QUERY_NUM_SDMA_IBS:
+	case SI_QUERY_NUM_BYTES_MOVED:
+	case SI_QUERY_NUM_EVICTIONS:
+	case SI_QUERY_NUM_VRAM_CPU_PAGE_FAULTS: {
+		enum radeon_value_id ws_id = winsys_id_from_type(query->b.type);
+		query->end_result = sctx->b.ws->query_value(sctx->b.ws, ws_id);
+		break;
+	}
+	case SI_QUERY_GFX_BO_LIST_SIZE:
+		ws_id = winsys_id_from_type(query->b.type);
+		query->end_result = sctx->b.ws->query_value(sctx->b.ws, ws_id);
+		query->end_time = sctx->b.ws->query_value(sctx->b.ws,
+							RADEON_NUM_GFX_IBS);
+		break;
+	case SI_QUERY_CS_THREAD_BUSY:
+		ws_id = winsys_id_from_type(query->b.type);
+		query->end_result = sctx->b.ws->query_value(sctx->b.ws, ws_id);
+		query->end_time = os_time_get_nano();
+		break;
+	case SI_QUERY_GALLIUM_THREAD_BUSY:
+		query->end_result =
+			sctx->b.tc ? util_queue_get_thread_time_nano(&sctx->b.tc->queue, 0) : 0;
+		query->end_time = os_time_get_nano();
+		break;
+	case SI_QUERY_GPU_LOAD:
+	case SI_QUERY_GPU_SHADERS_BUSY:
+	case SI_QUERY_GPU_TA_BUSY:
+	case SI_QUERY_GPU_GDS_BUSY:
+	case SI_QUERY_GPU_VGT_BUSY:
+	case SI_QUERY_GPU_IA_BUSY:
+	case SI_QUERY_GPU_SX_BUSY:
+	case SI_QUERY_GPU_WD_BUSY:
+	case SI_QUERY_GPU_BCI_BUSY:
+	case SI_QUERY_GPU_SC_BUSY:
+	case SI_QUERY_GPU_PA_BUSY:
+	case SI_QUERY_GPU_DB_BUSY:
+	case SI_QUERY_GPU_CP_BUSY:
+	case SI_QUERY_GPU_CB_BUSY:
+	case SI_QUERY_GPU_SDMA_BUSY:
+	case SI_QUERY_GPU_PFP_BUSY:
+	case SI_QUERY_GPU_MEQ_BUSY:
+	case SI_QUERY_GPU_ME_BUSY:
+	case SI_QUERY_GPU_SURF_SYNC_BUSY:
+	case SI_QUERY_GPU_CP_DMA_BUSY:
+	case SI_QUERY_GPU_SCRATCH_RAM_BUSY:
+		query->end_result = si_end_counter(sctx->screen,
+						     query->b.type,
+						     query->begin_result);
+		query->begin_result = 0;
+		break;
+	case SI_QUERY_NUM_COMPILATIONS:
+		query->end_result = p_atomic_read(&sctx->screen->num_compilations);
+		break;
+	case SI_QUERY_NUM_SHADERS_CREATED:
+		query->end_result = p_atomic_read(&sctx->screen->num_shaders_created);
+		break;
+	case SI_QUERY_BACK_BUFFER_PS_DRAW_RATIO:
+		query->end_result = sctx->b.last_tex_ps_draw_ratio;
+		break;
+	case SI_QUERY_NUM_SHADER_CACHE_HITS:
+		query->end_result =
+			p_atomic_read(&sctx->screen->num_shader_cache_hits);
+		break;
+	case SI_QUERY_GPIN_ASIC_ID:
+	case SI_QUERY_GPIN_NUM_SIMD:
+	case SI_QUERY_GPIN_NUM_RB:
+	case SI_QUERY_GPIN_NUM_SPI:
+	case SI_QUERY_GPIN_NUM_SE:
+		break;
+	default:
+		unreachable("si_query_sw_end: bad query type");
+	}
+
+	return true;
+}
+
+static bool si_query_sw_get_result(struct si_context *sctx,
+				   struct si_query *rquery,
+				   bool wait,
+				   union pipe_query_result *result)
+{
+	struct si_query_sw *query = (struct si_query_sw *)rquery;
+
+	switch (query->b.type) {
+	case PIPE_QUERY_TIMESTAMP_DISJOINT:
+		/* Convert from cycles per millisecond to cycles per second (Hz). */
+		result->timestamp_disjoint.frequency =
+			(uint64_t)sctx->screen->info.clock_crystal_freq * 1000;
+		result->timestamp_disjoint.disjoint = false;
+		return true;
+	case PIPE_QUERY_GPU_FINISHED: {
+		struct pipe_screen *screen = sctx->b.b.screen;
+		struct pipe_context *ctx = rquery->b.flushed ? NULL : &sctx->b.b;
+
+		result->b = screen->fence_finish(screen, ctx, query->fence,
+						 wait ? PIPE_TIMEOUT_INFINITE : 0);
+		return result->b;
+	}
+
+	case SI_QUERY_GFX_BO_LIST_SIZE:
+		result->u64 = (query->end_result - query->begin_result) /
+			      (query->end_time - query->begin_time);
+		return true;
+	case SI_QUERY_CS_THREAD_BUSY:
+	case SI_QUERY_GALLIUM_THREAD_BUSY:
+		result->u64 = (query->end_result - query->begin_result) * 100 /
+			      (query->end_time - query->begin_time);
+		return true;
+	case SI_QUERY_GPIN_ASIC_ID:
+		result->u32 = 0;
+		return true;
+	case SI_QUERY_GPIN_NUM_SIMD:
+		result->u32 = sctx->screen->info.num_good_compute_units;
+		return true;
+	case SI_QUERY_GPIN_NUM_RB:
+		result->u32 = sctx->screen->info.num_render_backends;
+		return true;
+	case SI_QUERY_GPIN_NUM_SPI:
+		result->u32 = 1; /* all supported chips have one SPI per SE */
+		return true;
+	case SI_QUERY_GPIN_NUM_SE:
+		result->u32 = sctx->screen->info.max_se;
+		return true;
+	}
+
+	result->u64 = query->end_result - query->begin_result;
+
+	switch (query->b.type) {
+	case SI_QUERY_BUFFER_WAIT_TIME:
+	case SI_QUERY_GPU_TEMPERATURE:
+		result->u64 /= 1000;
+		break;
+	case SI_QUERY_CURRENT_GPU_SCLK:
+	case SI_QUERY_CURRENT_GPU_MCLK:
+		result->u64 *= 1000000;
+		break;
+	}
+
+	return true;
+}
+
+
+static struct si_query_ops sw_query_ops = {
+	.destroy = si_query_sw_destroy,
+	.begin = si_query_sw_begin,
+	.end = si_query_sw_end,
+	.get_result = si_query_sw_get_result,
+	.get_result_resource = NULL
+};
+
+static struct pipe_query *si_query_sw_create(unsigned query_type)
+{
+	struct si_query_sw *query;
+
+	query = CALLOC_STRUCT(si_query_sw);
+	if (!query)
+		return NULL;
+
+	query->b.type = query_type;
+	query->b.ops = &sw_query_ops;
+
+	return (struct pipe_query *)query;
+}
+
+void si_query_hw_destroy(struct si_screen *sscreen,
+			 struct si_query *rquery)
+{
+	struct si_query_hw *query = (struct si_query_hw *)rquery;
+	struct si_query_buffer *prev = query->buffer.previous;
+
+	/* Release all query buffers. */
+	while (prev) {
+		struct si_query_buffer *qbuf = prev;
+		prev = prev->previous;
+		r600_resource_reference(&qbuf->buf, NULL);
+		FREE(qbuf);
+	}
+
+	r600_resource_reference(&query->buffer.buf, NULL);
+	r600_resource_reference(&query->workaround_buf, NULL);
+	FREE(rquery);
+}
+
+static struct r600_resource *si_new_query_buffer(struct si_screen *sscreen,
+						 struct si_query_hw *query)
+{
+	unsigned buf_size = MAX2(query->result_size,
+				 sscreen->info.min_alloc_size);
+
+	/* Queries are normally read by the CPU after
+	 * being written by the gpu, hence staging is probably a good
+	 * usage pattern.
+	 */
+	struct r600_resource *buf = (struct r600_resource*)
+		pipe_buffer_create(&sscreen->b, 0,
+				   PIPE_USAGE_STAGING, buf_size);
+	if (!buf)
+		return NULL;
+
+	if (!query->ops->prepare_buffer(sscreen, query, buf)) {
+		r600_resource_reference(&buf, NULL);
+		return NULL;
+	}
+
+	return buf;
+}
+
+static bool si_query_hw_prepare_buffer(struct si_screen *sscreen,
+				       struct si_query_hw *query,
+				       struct r600_resource *buffer)
+{
+	/* Callers ensure that the buffer is currently unused by the GPU. */
+	uint32_t *results = sscreen->ws->buffer_map(buffer->buf, NULL,
+						   PIPE_TRANSFER_WRITE |
+						   PIPE_TRANSFER_UNSYNCHRONIZED);
+	if (!results)
+		return false;
+
+	memset(results, 0, buffer->b.b.width0);
+
+	if (query->b.type == PIPE_QUERY_OCCLUSION_COUNTER ||
+	    query->b.type == PIPE_QUERY_OCCLUSION_PREDICATE ||
+	    query->b.type == PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE) {
+		unsigned max_rbs = sscreen->info.num_render_backends;
+		unsigned enabled_rb_mask = sscreen->info.enabled_rb_mask;
+		unsigned num_results;
+		unsigned i, j;
+
+		/* Set top bits for unused backends. */
+		num_results = buffer->b.b.width0 / query->result_size;
+		for (j = 0; j < num_results; j++) {
+			for (i = 0; i < max_rbs; i++) {
+				if (!(enabled_rb_mask & (1<<i))) {
+					results[(i * 4)+1] = 0x80000000;
+					results[(i * 4)+3] = 0x80000000;
+				}
+			}
+			results += 4 * max_rbs;
+		}
+	}
+
+	return true;
+}
+
+static void si_query_hw_get_result_resource(struct si_context *sctx,
+					    struct si_query *rquery,
+					    bool wait,
+					    enum pipe_query_value_type result_type,
+					    int index,
+					    struct pipe_resource *resource,
+					    unsigned offset);
+
+static struct si_query_ops query_hw_ops = {
+	.destroy = si_query_hw_destroy,
+	.begin = si_query_hw_begin,
+	.end = si_query_hw_end,
+	.get_result = si_query_hw_get_result,
+	.get_result_resource = si_query_hw_get_result_resource,
+};
+
+static void si_query_hw_do_emit_start(struct si_context *sctx,
+				      struct si_query_hw *query,
+				      struct r600_resource *buffer,
+				      uint64_t va);
+static void si_query_hw_do_emit_stop(struct si_context *sctx,
+				     struct si_query_hw *query,
+				     struct r600_resource *buffer,
+				     uint64_t va);
+static void si_query_hw_add_result(struct si_screen *sscreen,
+				   struct si_query_hw *, void *buffer,
+				   union pipe_query_result *result);
+static void si_query_hw_clear_result(struct si_query_hw *,
+				     union pipe_query_result *);
+
+static struct si_query_hw_ops query_hw_default_hw_ops = {
+	.prepare_buffer = si_query_hw_prepare_buffer,
+	.emit_start = si_query_hw_do_emit_start,
+	.emit_stop = si_query_hw_do_emit_stop,
+	.clear_result = si_query_hw_clear_result,
+	.add_result = si_query_hw_add_result,
+};
+
+bool si_query_hw_init(struct si_screen *sscreen,
+		      struct si_query_hw *query)
+{
+	query->buffer.buf = si_new_query_buffer(sscreen, query);
+	if (!query->buffer.buf)
+		return false;
+
+	return true;
+}
+
+static struct pipe_query *si_query_hw_create(struct si_screen *sscreen,
+					     unsigned query_type,
+					     unsigned index)
+{
+	struct si_query_hw *query = CALLOC_STRUCT(si_query_hw);
+	if (!query)
+		return NULL;
+
+	query->b.type = query_type;
+	query->b.ops = &query_hw_ops;
+	query->ops = &query_hw_default_hw_ops;
+
+	switch (query_type) {
+	case PIPE_QUERY_OCCLUSION_COUNTER:
+	case PIPE_QUERY_OCCLUSION_PREDICATE:
+	case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
+		query->result_size = 16 * sscreen->info.num_render_backends;
+		query->result_size += 16; /* for the fence + alignment */
+		query->num_cs_dw_end = 6 + si_gfx_write_fence_dwords(sscreen);
+		break;
+	case PIPE_QUERY_TIME_ELAPSED:
+		query->result_size = 24;
+		query->num_cs_dw_end = 8 + si_gfx_write_fence_dwords(sscreen);
+		break;
+	case PIPE_QUERY_TIMESTAMP:
+		query->result_size = 16;
+		query->num_cs_dw_end = 8 + si_gfx_write_fence_dwords(sscreen);
+		query->flags = SI_QUERY_HW_FLAG_NO_START;
+		break;
+	case PIPE_QUERY_PRIMITIVES_EMITTED:
+	case PIPE_QUERY_PRIMITIVES_GENERATED:
+	case PIPE_QUERY_SO_STATISTICS:
+	case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
+		/* NumPrimitivesWritten, PrimitiveStorageNeeded. */
+		query->result_size = 32;
+		query->num_cs_dw_end = 6;
+		query->stream = index;
+		break;
+	case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
+		/* NumPrimitivesWritten, PrimitiveStorageNeeded. */
+		query->result_size = 32 * SI_MAX_STREAMS;
+		query->num_cs_dw_end = 6 * SI_MAX_STREAMS;
+		break;
+	case PIPE_QUERY_PIPELINE_STATISTICS:
+		/* 11 values on GCN. */
+		query->result_size = 11 * 16;
+		query->result_size += 8; /* for the fence + alignment */
+		query->num_cs_dw_end = 6 + si_gfx_write_fence_dwords(sscreen);
+		break;
+	default:
+		assert(0);
+		FREE(query);
+		return NULL;
+	}
+
+	if (!si_query_hw_init(sscreen, query)) {
+		FREE(query);
+		return NULL;
+	}
+
+	return (struct pipe_query *)query;
+}
+
+static void si_update_occlusion_query_state(struct si_context *sctx,
+					    unsigned type, int diff)
+{
+	if (type == PIPE_QUERY_OCCLUSION_COUNTER ||
+	    type == PIPE_QUERY_OCCLUSION_PREDICATE ||
+	    type == PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE) {
+		bool old_enable = sctx->b.num_occlusion_queries != 0;
+		bool old_perfect_enable =
+			sctx->b.num_perfect_occlusion_queries != 0;
+		bool enable, perfect_enable;
+
+		sctx->b.num_occlusion_queries += diff;
+		assert(sctx->b.num_occlusion_queries >= 0);
+
+		if (type != PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE) {
+			sctx->b.num_perfect_occlusion_queries += diff;
+			assert(sctx->b.num_perfect_occlusion_queries >= 0);
+		}
+
+		enable = sctx->b.num_occlusion_queries != 0;
+		perfect_enable = sctx->b.num_perfect_occlusion_queries != 0;
+
+		if (enable != old_enable || perfect_enable != old_perfect_enable) {
+			si_set_occlusion_query_state(sctx, old_perfect_enable);
+		}
+	}
+}
+
+static unsigned event_type_for_stream(unsigned stream)
+{
+	switch (stream) {
+	default:
+	case 0: return V_028A90_SAMPLE_STREAMOUTSTATS;
+	case 1: return V_028A90_SAMPLE_STREAMOUTSTATS1;
+	case 2: return V_028A90_SAMPLE_STREAMOUTSTATS2;
+	case 3: return V_028A90_SAMPLE_STREAMOUTSTATS3;
+	}
+}
+
+static void emit_sample_streamout(struct radeon_winsys_cs *cs, uint64_t va,
+				  unsigned stream)
+{
+	radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
+	radeon_emit(cs, EVENT_TYPE(event_type_for_stream(stream)) | EVENT_INDEX(3));
+	radeon_emit(cs, va);
+	radeon_emit(cs, va >> 32);
+}
+
+static void si_query_hw_do_emit_start(struct si_context *sctx,
+					struct si_query_hw *query,
+					struct r600_resource *buffer,
+					uint64_t va)
+{
+	struct radeon_winsys_cs *cs = sctx->b.gfx_cs;
+
+	switch (query->b.type) {
+	case PIPE_QUERY_OCCLUSION_COUNTER:
+	case PIPE_QUERY_OCCLUSION_PREDICATE:
+	case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
+		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
+		radeon_emit(cs, EVENT_TYPE(V_028A90_ZPASS_DONE) | EVENT_INDEX(1));
+		radeon_emit(cs, va);
+		radeon_emit(cs, va >> 32);
+		break;
+	case PIPE_QUERY_PRIMITIVES_EMITTED:
+	case PIPE_QUERY_PRIMITIVES_GENERATED:
+	case PIPE_QUERY_SO_STATISTICS:
+	case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
+		emit_sample_streamout(cs, va, query->stream);
+		break;
+	case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
+		for (unsigned stream = 0; stream < SI_MAX_STREAMS; ++stream)
+			emit_sample_streamout(cs, va + 32 * stream, stream);
+		break;
+	case PIPE_QUERY_TIME_ELAPSED:
+		/* Write the timestamp from the CP not waiting for
+		 * outstanding draws (top-of-pipe).
+		 */
+		radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
+		radeon_emit(cs, COPY_DATA_COUNT_SEL |
+				COPY_DATA_SRC_SEL(COPY_DATA_TIMESTAMP) |
+				COPY_DATA_DST_SEL(COPY_DATA_MEM_ASYNC));
+		radeon_emit(cs, 0);
+		radeon_emit(cs, 0);
+		radeon_emit(cs, va);
+		radeon_emit(cs, va >> 32);
+		break;
+	case PIPE_QUERY_PIPELINE_STATISTICS:
+		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
+		radeon_emit(cs, EVENT_TYPE(V_028A90_SAMPLE_PIPELINESTAT) | EVENT_INDEX(2));
+		radeon_emit(cs, va);
+		radeon_emit(cs, va >> 32);
+		break;
+	default:
+		assert(0);
+	}
+	radeon_add_to_buffer_list(sctx, sctx->b.gfx_cs, query->buffer.buf, RADEON_USAGE_WRITE,
+				  RADEON_PRIO_QUERY);
+}
+
+static void si_query_hw_emit_start(struct si_context *sctx,
+				   struct si_query_hw *query)
+{
+	uint64_t va;
+
+	if (!query->buffer.buf)
+		return; // previous buffer allocation failure
+
+	si_update_occlusion_query_state(sctx, query->b.type, 1);
+	si_update_prims_generated_query_state(sctx, query->b.type, 1);
+
+	si_need_gfx_cs_space(sctx);
+
+	/* Get a new query buffer if needed. */
+	if (query->buffer.results_end + query->result_size > query->buffer.buf->b.b.width0) {
+		struct si_query_buffer *qbuf = MALLOC_STRUCT(si_query_buffer);
+		*qbuf = query->buffer;
+		query->buffer.results_end = 0;
+		query->buffer.previous = qbuf;
+		query->buffer.buf = si_new_query_buffer(sctx->screen, query);
+		if (!query->buffer.buf)
+			return;
+	}
+
+	/* emit begin query */
+	va = query->buffer.buf->gpu_address + query->buffer.results_end;
+
+	query->ops->emit_start(sctx, query, query->buffer.buf, va);
+
+	sctx->b.num_cs_dw_queries_suspend += query->num_cs_dw_end;
+}
+
+static void si_query_hw_do_emit_stop(struct si_context *sctx,
+				       struct si_query_hw *query,
+				       struct r600_resource *buffer,
+				       uint64_t va)
+{
+	struct radeon_winsys_cs *cs = sctx->b.gfx_cs;
+	uint64_t fence_va = 0;
+
+	switch (query->b.type) {
+	case PIPE_QUERY_OCCLUSION_COUNTER:
+	case PIPE_QUERY_OCCLUSION_PREDICATE:
+	case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
+		va += 8;
+		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
+		radeon_emit(cs, EVENT_TYPE(V_028A90_ZPASS_DONE) | EVENT_INDEX(1));
+		radeon_emit(cs, va);
+		radeon_emit(cs, va >> 32);
+
+		fence_va = va + sctx->screen->info.num_render_backends * 16 - 8;
+		break;
+	case PIPE_QUERY_PRIMITIVES_EMITTED:
+	case PIPE_QUERY_PRIMITIVES_GENERATED:
+	case PIPE_QUERY_SO_STATISTICS:
+	case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
+		va += 16;
+		emit_sample_streamout(cs, va, query->stream);
+		break;
+	case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
+		va += 16;
+		for (unsigned stream = 0; stream < SI_MAX_STREAMS; ++stream)
+			emit_sample_streamout(cs, va + 32 * stream, stream);
+		break;
+	case PIPE_QUERY_TIME_ELAPSED:
+		va += 8;
+		/* fall through */
+	case PIPE_QUERY_TIMESTAMP:
+		si_gfx_write_event_eop(sctx, V_028A90_BOTTOM_OF_PIPE_TS,
+				       0, EOP_DATA_SEL_TIMESTAMP, NULL, va,
+				       0, query->b.type);
+		fence_va = va + 8;
+		break;
+	case PIPE_QUERY_PIPELINE_STATISTICS: {
+		unsigned sample_size = (query->result_size - 8) / 2;
+
+		va += sample_size;
+		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
+		radeon_emit(cs, EVENT_TYPE(V_028A90_SAMPLE_PIPELINESTAT) | EVENT_INDEX(2));
+		radeon_emit(cs, va);
+		radeon_emit(cs, va >> 32);
+
+		fence_va = va + sample_size;
+		break;
+	}
+	default:
+		assert(0);
+	}
+	radeon_add_to_buffer_list(sctx, sctx->b.gfx_cs, query->buffer.buf, RADEON_USAGE_WRITE,
+				  RADEON_PRIO_QUERY);
+
+	if (fence_va)
+		si_gfx_write_event_eop(sctx, V_028A90_BOTTOM_OF_PIPE_TS, 0,
+				       EOP_DATA_SEL_VALUE_32BIT,
+				       query->buffer.buf, fence_va, 0x80000000,
+				       query->b.type);
+}
+
+static void si_query_hw_emit_stop(struct si_context *sctx,
+				  struct si_query_hw *query)
+{
+	uint64_t va;
+
+	if (!query->buffer.buf)
+		return; // previous buffer allocation failure
+
+	/* The queries which need begin already called this in begin_query. */
+	if (query->flags & SI_QUERY_HW_FLAG_NO_START)
+		si_need_gfx_cs_space(sctx);
+
+	/* emit end query */
+	va = query->buffer.buf->gpu_address + query->buffer.results_end;
+
+	query->ops->emit_stop(sctx, query, query->buffer.buf, va);
+
+	query->buffer.results_end += query->result_size;
+
+	if (!(query->flags & SI_QUERY_HW_FLAG_NO_START))
+		sctx->b.num_cs_dw_queries_suspend -= query->num_cs_dw_end;
+
+	si_update_occlusion_query_state(sctx, query->b.type, -1);
+	si_update_prims_generated_query_state(sctx, query->b.type, -1);
+}
+
+static void emit_set_predicate(struct si_context *ctx,
+			       struct r600_resource *buf, uint64_t va,
+			       uint32_t op)
+{
+	struct radeon_winsys_cs *cs = ctx->b.gfx_cs;
+
+	if (ctx->b.chip_class >= GFX9) {
+		radeon_emit(cs, PKT3(PKT3_SET_PREDICATION, 2, 0));
+		radeon_emit(cs, op);
+		radeon_emit(cs, va);
+		radeon_emit(cs, va >> 32);
+	} else {
+		radeon_emit(cs, PKT3(PKT3_SET_PREDICATION, 1, 0));
+		radeon_emit(cs, va);
+		radeon_emit(cs, op | ((va >> 32) & 0xFF));
+	}
+	radeon_add_to_buffer_list(ctx, ctx->b.gfx_cs, buf, RADEON_USAGE_READ,
+				  RADEON_PRIO_QUERY);
+}
+
+static void si_emit_query_predication(struct si_context *ctx,
+				      struct r600_atom *atom)
+{
+	struct si_query_hw *query = (struct si_query_hw *)ctx->b.render_cond;
+	struct si_query_buffer *qbuf;
+	uint32_t op;
+	bool flag_wait, invert;
+
+	if (!query)
+		return;
+
+	invert = ctx->b.render_cond_invert;
+	flag_wait = ctx->b.render_cond_mode == PIPE_RENDER_COND_WAIT ||
+		    ctx->b.render_cond_mode == PIPE_RENDER_COND_BY_REGION_WAIT;
+
+	if (query->workaround_buf) {
+		op = PRED_OP(PREDICATION_OP_BOOL64);
+	} else {
+		switch (query->b.type) {
+		case PIPE_QUERY_OCCLUSION_COUNTER:
+		case PIPE_QUERY_OCCLUSION_PREDICATE:
+		case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
+			op = PRED_OP(PREDICATION_OP_ZPASS);
+			break;
+		case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
+		case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
+			op = PRED_OP(PREDICATION_OP_PRIMCOUNT);
+			invert = !invert;
+			break;
+		default:
+			assert(0);
+			return;
+		}
+	}
+
+	/* if true then invert, see GL_ARB_conditional_render_inverted */
+	if (invert)
+		op |= PREDICATION_DRAW_NOT_VISIBLE; /* Draw if not visible or overflow */
+	else
+		op |= PREDICATION_DRAW_VISIBLE; /* Draw if visible or no overflow */
+
+	/* Use the value written by compute shader as a workaround. Note that
+	 * the wait flag does not apply in this predication mode.
+	 *
+	 * The shader outputs the result value to L2. Workarounds only affect VI
+	 * and later, where the CP reads data from L2, so we don't need an
+	 * additional flush.
+	 */
+	if (query->workaround_buf) {
+		uint64_t va = query->workaround_buf->gpu_address + query->workaround_offset;
+		emit_set_predicate(ctx, query->workaround_buf, va, op);
+		return;
+	}
+
+	op |= flag_wait ? PREDICATION_HINT_WAIT : PREDICATION_HINT_NOWAIT_DRAW;
+
+	/* emit predicate packets for all data blocks */
+	for (qbuf = &query->buffer; qbuf; qbuf = qbuf->previous) {
+		unsigned results_base = 0;
+		uint64_t va_base = qbuf->buf->gpu_address;
+
+		while (results_base < qbuf->results_end) {
+			uint64_t va = va_base + results_base;
+
+			if (query->b.type == PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE) {
+				for (unsigned stream = 0; stream < SI_MAX_STREAMS; ++stream) {
+					emit_set_predicate(ctx, qbuf->buf, va + 32 * stream, op);
+
+					/* set CONTINUE bit for all packets except the first */
+					op |= PREDICATION_CONTINUE;
+				}
+			} else {
+				emit_set_predicate(ctx, qbuf->buf, va, op);
+				op |= PREDICATION_CONTINUE;
+			}
+
+			results_base += query->result_size;
+		}
+	}
+}
+
+static struct pipe_query *si_create_query(struct pipe_context *ctx, unsigned query_type, unsigned index)
+{
+	struct si_screen *sscreen =
+		(struct si_screen *)ctx->screen;
+
+	if (query_type == PIPE_QUERY_TIMESTAMP_DISJOINT ||
+	    query_type == PIPE_QUERY_GPU_FINISHED ||
+	    query_type >= PIPE_QUERY_DRIVER_SPECIFIC)
+		return si_query_sw_create(query_type);
+
+	return si_query_hw_create(sscreen, query_type, index);
+}
+
+static void si_destroy_query(struct pipe_context *ctx, struct pipe_query *query)
+{
+	struct si_context *sctx = (struct si_context *)ctx;
+	struct si_query *rquery = (struct si_query *)query;
+
+	rquery->ops->destroy(sctx->screen, rquery);
+}
+
+static boolean si_begin_query(struct pipe_context *ctx,
+                                struct pipe_query *query)
+{
+	struct si_context *sctx = (struct si_context *)ctx;
+	struct si_query *rquery = (struct si_query *)query;
+
+	return rquery->ops->begin(sctx, rquery);
+}
+
+void si_query_hw_reset_buffers(struct si_context *sctx,
+			       struct si_query_hw *query)
+{
+	struct si_query_buffer *prev = query->buffer.previous;
+
+	/* Discard the old query buffers. */
+	while (prev) {
+		struct si_query_buffer *qbuf = prev;
+		prev = prev->previous;
+		r600_resource_reference(&qbuf->buf, NULL);
+		FREE(qbuf);
+	}
+
+	query->buffer.results_end = 0;
+	query->buffer.previous = NULL;
+
+	/* Obtain a new buffer if the current one can't be mapped without a stall. */
+	if (si_rings_is_buffer_referenced(sctx, query->buffer.buf->buf, RADEON_USAGE_READWRITE) ||
+	    !sctx->b.ws->buffer_wait(query->buffer.buf->buf, 0, RADEON_USAGE_READWRITE)) {
+		r600_resource_reference(&query->buffer.buf, NULL);
+		query->buffer.buf = si_new_query_buffer(sctx->screen, query);
+	} else {
+		if (!query->ops->prepare_buffer(sctx->screen, query, query->buffer.buf))
+			r600_resource_reference(&query->buffer.buf, NULL);
+	}
+}
+
+bool si_query_hw_begin(struct si_context *sctx,
+		       struct si_query *rquery)
+{
+	struct si_query_hw *query = (struct si_query_hw *)rquery;
+
+	if (query->flags & SI_QUERY_HW_FLAG_NO_START) {
+		assert(0);
+		return false;
+	}
+
+	if (!(query->flags & SI_QUERY_HW_FLAG_BEGIN_RESUMES))
+		si_query_hw_reset_buffers(sctx, query);
+
+	r600_resource_reference(&query->workaround_buf, NULL);
+
+	si_query_hw_emit_start(sctx, query);
+	if (!query->buffer.buf)
+		return false;
+
+	LIST_ADDTAIL(&query->list, &sctx->b.active_queries);
+	return true;
+}
+
+static bool si_end_query(struct pipe_context *ctx, struct pipe_query *query)
+{
+	struct si_context *sctx = (struct si_context *)ctx;
+	struct si_query *rquery = (struct si_query *)query;
+
+	return rquery->ops->end(sctx, rquery);
+}
+
+bool si_query_hw_end(struct si_context *sctx,
+		     struct si_query *rquery)
+{
+	struct si_query_hw *query = (struct si_query_hw *)rquery;
+
+	if (query->flags & SI_QUERY_HW_FLAG_NO_START)
+		si_query_hw_reset_buffers(sctx, query);
+
+	si_query_hw_emit_stop(sctx, query);
+
+	if (!(query->flags & SI_QUERY_HW_FLAG_NO_START))
+		LIST_DELINIT(&query->list);
+
+	if (!query->buffer.buf)
+		return false;
+
+	return true;
+}
+
+static void si_get_hw_query_params(struct si_context *sctx,
+				   struct si_query_hw *rquery, int index,
+				   struct si_hw_query_params *params)
+{
+	unsigned max_rbs = sctx->screen->info.num_render_backends;
+
+	params->pair_stride = 0;
+	params->pair_count = 1;
+
+	switch (rquery->b.type) {
+	case PIPE_QUERY_OCCLUSION_COUNTER:
+	case PIPE_QUERY_OCCLUSION_PREDICATE:
+	case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
+		params->start_offset = 0;
+		params->end_offset = 8;
+		params->fence_offset = max_rbs * 16;
+		params->pair_stride = 16;
+		params->pair_count = max_rbs;
+		break;
+	case PIPE_QUERY_TIME_ELAPSED:
+		params->start_offset = 0;
+		params->end_offset = 8;
+		params->fence_offset = 16;
+		break;
+	case PIPE_QUERY_TIMESTAMP:
+		params->start_offset = 0;
+		params->end_offset = 0;
+		params->fence_offset = 8;
+		break;
+	case PIPE_QUERY_PRIMITIVES_EMITTED:
+		params->start_offset = 8;
+		params->end_offset = 24;
+		params->fence_offset = params->end_offset + 4;
+		break;
+	case PIPE_QUERY_PRIMITIVES_GENERATED:
+		params->start_offset = 0;
+		params->end_offset = 16;
+		params->fence_offset = params->end_offset + 4;
+		break;
+	case PIPE_QUERY_SO_STATISTICS:
+		params->start_offset = 8 - index * 8;
+		params->end_offset = 24 - index * 8;
+		params->fence_offset = params->end_offset + 4;
+		break;
+	case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
+		params->pair_count = SI_MAX_STREAMS;
+		params->pair_stride = 32;
+	case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
+		params->start_offset = 0;
+		params->end_offset = 16;
+
+		/* We can re-use the high dword of the last 64-bit value as a
+		 * fence: it is initialized as 0, and the high bit is set by
+		 * the write of the streamout stats event.
+		 */
+		params->fence_offset = rquery->result_size - 4;
+		break;
+	case PIPE_QUERY_PIPELINE_STATISTICS:
+	{
+		/* Offsets apply to EG+ */
+		static const unsigned offsets[] = {56, 48, 24, 32, 40, 16, 8, 0, 64, 72, 80};
+		params->start_offset = offsets[index];
+		params->end_offset = 88 + offsets[index];
+		params->fence_offset = 2 * 88;
+		break;
+	}
+	default:
+		unreachable("si_get_hw_query_params unsupported");
+	}
+}
+
+static unsigned si_query_read_result(void *map, unsigned start_index, unsigned end_index,
+				     bool test_status_bit)
+{
+	uint32_t *current_result = (uint32_t*)map;
+	uint64_t start, end;
+
+	start = (uint64_t)current_result[start_index] |
+		(uint64_t)current_result[start_index+1] << 32;
+	end = (uint64_t)current_result[end_index] |
+	      (uint64_t)current_result[end_index+1] << 32;
+
+	if (!test_status_bit ||
+	    ((start & 0x8000000000000000UL) && (end & 0x8000000000000000UL))) {
+		return end - start;
+	}
+	return 0;
+}
+
+static void si_query_hw_add_result(struct si_screen *sscreen,
+				     struct si_query_hw *query,
+				     void *buffer,
+				     union pipe_query_result *result)
+{
+	unsigned max_rbs = sscreen->info.num_render_backends;
+
+	switch (query->b.type) {
+	case PIPE_QUERY_OCCLUSION_COUNTER: {
+		for (unsigned i = 0; i < max_rbs; ++i) {
+			unsigned results_base = i * 16;
+			result->u64 +=
+				si_query_read_result(buffer + results_base, 0, 2, true);
+		}
+		break;
+	}
+	case PIPE_QUERY_OCCLUSION_PREDICATE:
+	case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE: {
+		for (unsigned i = 0; i < max_rbs; ++i) {
+			unsigned results_base = i * 16;
+			result->b = result->b ||
+				si_query_read_result(buffer + results_base, 0, 2, true) != 0;
+		}
+		break;
+	}
+	case PIPE_QUERY_TIME_ELAPSED:
+		result->u64 += si_query_read_result(buffer, 0, 2, false);
+		break;
+	case PIPE_QUERY_TIMESTAMP:
+		result->u64 = *(uint64_t*)buffer;
+		break;
+	case PIPE_QUERY_PRIMITIVES_EMITTED:
+		/* SAMPLE_STREAMOUTSTATS stores this structure:
+		 * {
+		 *    u64 NumPrimitivesWritten;
+		 *    u64 PrimitiveStorageNeeded;
+		 * }
+		 * We only need NumPrimitivesWritten here. */
+		result->u64 += si_query_read_result(buffer, 2, 6, true);
+		break;
+	case PIPE_QUERY_PRIMITIVES_GENERATED:
+		/* Here we read PrimitiveStorageNeeded. */
+		result->u64 += si_query_read_result(buffer, 0, 4, true);
+		break;
+	case PIPE_QUERY_SO_STATISTICS:
+		result->so_statistics.num_primitives_written +=
+			si_query_read_result(buffer, 2, 6, true);
+		result->so_statistics.primitives_storage_needed +=
+			si_query_read_result(buffer, 0, 4, true);
+		break;
+	case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
+		result->b = result->b ||
+			si_query_read_result(buffer, 2, 6, true) !=
+			si_query_read_result(buffer, 0, 4, true);
+		break;
+	case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
+		for (unsigned stream = 0; stream < SI_MAX_STREAMS; ++stream) {
+			result->b = result->b ||
+				si_query_read_result(buffer, 2, 6, true) !=
+				si_query_read_result(buffer, 0, 4, true);
+			buffer = (char *)buffer + 32;
+		}
+		break;
+	case PIPE_QUERY_PIPELINE_STATISTICS:
+		result->pipeline_statistics.ps_invocations +=
+			si_query_read_result(buffer, 0, 22, false);
+		result->pipeline_statistics.c_primitives +=
+			si_query_read_result(buffer, 2, 24, false);
+		result->pipeline_statistics.c_invocations +=
+			si_query_read_result(buffer, 4, 26, false);
+		result->pipeline_statistics.vs_invocations +=
+			si_query_read_result(buffer, 6, 28, false);
+		result->pipeline_statistics.gs_invocations +=
+			si_query_read_result(buffer, 8, 30, false);
+		result->pipeline_statistics.gs_primitives +=
+			si_query_read_result(buffer, 10, 32, false);
+		result->pipeline_statistics.ia_primitives +=
+			si_query_read_result(buffer, 12, 34, false);
+		result->pipeline_statistics.ia_vertices +=
+			si_query_read_result(buffer, 14, 36, false);
+		result->pipeline_statistics.hs_invocations +=
+			si_query_read_result(buffer, 16, 38, false);
+		result->pipeline_statistics.ds_invocations +=
+			si_query_read_result(buffer, 18, 40, false);
+		result->pipeline_statistics.cs_invocations +=
+			si_query_read_result(buffer, 20, 42, false);
+#if 0 /* for testing */
+		printf("Pipeline stats: IA verts=%llu, IA prims=%llu, VS=%llu, HS=%llu, "
+		       "DS=%llu, GS=%llu, GS prims=%llu, Clipper=%llu, "
+		       "Clipper prims=%llu, PS=%llu, CS=%llu\n",
+		       result->pipeline_statistics.ia_vertices,
+		       result->pipeline_statistics.ia_primitives,
+		       result->pipeline_statistics.vs_invocations,
+		       result->pipeline_statistics.hs_invocations,
+		       result->pipeline_statistics.ds_invocations,
+		       result->pipeline_statistics.gs_invocations,
+		       result->pipeline_statistics.gs_primitives,
+		       result->pipeline_statistics.c_invocations,
+		       result->pipeline_statistics.c_primitives,
+		       result->pipeline_statistics.ps_invocations,
+		       result->pipeline_statistics.cs_invocations);
+#endif
+		break;
+	default:
+		assert(0);
+	}
+}
+
+static boolean si_get_query_result(struct pipe_context *ctx,
+				   struct pipe_query *query, boolean wait,
+				   union pipe_query_result *result)
+{
+	struct si_context *sctx = (struct si_context *)ctx;
+	struct si_query *rquery = (struct si_query *)query;
+
+	return rquery->ops->get_result(sctx, rquery, wait, result);
+}
+
+static void si_get_query_result_resource(struct pipe_context *ctx,
+					 struct pipe_query *query,
+					 boolean wait,
+					 enum pipe_query_value_type result_type,
+					 int index,
+					 struct pipe_resource *resource,
+					 unsigned offset)
+{
+	struct si_context *sctx = (struct si_context *)ctx;
+	struct si_query *rquery = (struct si_query *)query;
+
+	rquery->ops->get_result_resource(sctx, rquery, wait, result_type, index,
+	                                 resource, offset);
+}
+
+static void si_query_hw_clear_result(struct si_query_hw *query,
+				       union pipe_query_result *result)
+{
+	util_query_clear_result(result, query->b.type);
+}
+
+bool si_query_hw_get_result(struct si_context *sctx,
+			    struct si_query *rquery,
+			    bool wait, union pipe_query_result *result)
+{
+	struct si_screen *sscreen = sctx->screen;
+	struct si_query_hw *query = (struct si_query_hw *)rquery;
+	struct si_query_buffer *qbuf;
+
+	query->ops->clear_result(query, result);
+
+	for (qbuf = &query->buffer; qbuf; qbuf = qbuf->previous) {
+		unsigned usage = PIPE_TRANSFER_READ |
+				 (wait ? 0 : PIPE_TRANSFER_DONTBLOCK);
+		unsigned results_base = 0;
+		void *map;
+
+		if (rquery->b.flushed)
+			map = sctx->b.ws->buffer_map(qbuf->buf->buf, NULL, usage);
+		else
+			map = si_buffer_map_sync_with_rings(sctx, qbuf->buf, usage);
+
+		if (!map)
+			return false;
+
+		while (results_base != qbuf->results_end) {
+			query->ops->add_result(sscreen, query, map + results_base,
+					       result);
+			results_base += query->result_size;
+		}
+	}
+
+	/* Convert the time to expected units. */
+	if (rquery->type == PIPE_QUERY_TIME_ELAPSED ||
+	    rquery->type == PIPE_QUERY_TIMESTAMP) {
+		result->u64 = (1000000 * result->u64) / sscreen->info.clock_crystal_freq;
+	}
+	return true;
+}
+
+/* Create the compute shader that is used to collect the results.
+ *
+ * One compute grid with a single thread is launched for every query result
+ * buffer. The thread (optionally) reads a previous summary buffer, then
+ * accumulates data from the query result buffer, and writes the result either
+ * to a summary buffer to be consumed by the next grid invocation or to the
+ * user-supplied buffer.
+ *
+ * Data layout:
+ *
+ * CONST
+ *  0.x = end_offset
+ *  0.y = result_stride
+ *  0.z = result_count
+ *  0.w = bit field:
+ *          1: read previously accumulated values
+ *          2: write accumulated values for chaining
+ *          4: write result available
+ *          8: convert result to boolean (0/1)
+ *         16: only read one dword and use that as result
+ *         32: apply timestamp conversion
+ *         64: store full 64 bits result
+ *        128: store signed 32 bits result
+ *        256: SO_OVERFLOW mode: take the difference of two successive half-pairs
+ *  1.x = fence_offset
+ *  1.y = pair_stride
+ *  1.z = pair_count
+ *
+ * BUFFER[0] = query result buffer
+ * BUFFER[1] = previous summary buffer
+ * BUFFER[2] = next summary buffer or user-supplied buffer
+ */
+static void si_create_query_result_shader(struct si_context *sctx)
+{
+	/* TEMP[0].xy = accumulated result so far
+	 * TEMP[0].z = result not available
+	 *
+	 * TEMP[1].x = current result index
+	 * TEMP[1].y = current pair index
+	 */
+	static const char text_tmpl[] =
+		"COMP\n"
+		"PROPERTY CS_FIXED_BLOCK_WIDTH 1\n"
+		"PROPERTY CS_FIXED_BLOCK_HEIGHT 1\n"
+		"PROPERTY CS_FIXED_BLOCK_DEPTH 1\n"
+		"DCL BUFFER[0]\n"
+		"DCL BUFFER[1]\n"
+		"DCL BUFFER[2]\n"
+		"DCL CONST[0][0..1]\n"
+		"DCL TEMP[0..5]\n"
+		"IMM[0] UINT32 {0, 31, 2147483647, 4294967295}\n"
+		"IMM[1] UINT32 {1, 2, 4, 8}\n"
+		"IMM[2] UINT32 {16, 32, 64, 128}\n"
+		"IMM[3] UINT32 {1000000, 0, %u, 0}\n" /* for timestamp conversion */
+		"IMM[4] UINT32 {256, 0, 0, 0}\n"
+
+		"AND TEMP[5], CONST[0][0].wwww, IMM[2].xxxx\n"
+		"UIF TEMP[5]\n"
+			/* Check result availability. */
+			"LOAD TEMP[1].x, BUFFER[0], CONST[0][1].xxxx\n"
+			"ISHR TEMP[0].z, TEMP[1].xxxx, IMM[0].yyyy\n"
+			"MOV TEMP[1], TEMP[0].zzzz\n"
+			"NOT TEMP[0].z, TEMP[0].zzzz\n"
+
+			/* Load result if available. */
+			"UIF TEMP[1]\n"
+				"LOAD TEMP[0].xy, BUFFER[0], IMM[0].xxxx\n"
+			"ENDIF\n"
+		"ELSE\n"
+			/* Load previously accumulated result if requested. */
+			"MOV TEMP[0], IMM[0].xxxx\n"
+			"AND TEMP[4], CONST[0][0].wwww, IMM[1].xxxx\n"
+			"UIF TEMP[4]\n"
+				"LOAD TEMP[0].xyz, BUFFER[1], IMM[0].xxxx\n"
+			"ENDIF\n"
+
+			"MOV TEMP[1].x, IMM[0].xxxx\n"
+			"BGNLOOP\n"
+				/* Break if accumulated result so far is not available. */
+				"UIF TEMP[0].zzzz\n"
+					"BRK\n"
+				"ENDIF\n"
+
+				/* Break if result_index >= result_count. */
+				"USGE TEMP[5], TEMP[1].xxxx, CONST[0][0].zzzz\n"
+				"UIF TEMP[5]\n"
+					"BRK\n"
+				"ENDIF\n"
+
+				/* Load fence and check result availability */
+				"UMAD TEMP[5].x, TEMP[1].xxxx, CONST[0][0].yyyy, CONST[0][1].xxxx\n"
+				"LOAD TEMP[5].x, BUFFER[0], TEMP[5].xxxx\n"
+				"ISHR TEMP[0].z, TEMP[5].xxxx, IMM[0].yyyy\n"
+				"NOT TEMP[0].z, TEMP[0].zzzz\n"
+				"UIF TEMP[0].zzzz\n"
+					"BRK\n"
+				"ENDIF\n"
+
+				"MOV TEMP[1].y, IMM[0].xxxx\n"
+				"BGNLOOP\n"
+					/* Load start and end. */
+					"UMUL TEMP[5].x, TEMP[1].xxxx, CONST[0][0].yyyy\n"
+					"UMAD TEMP[5].x, TEMP[1].yyyy, CONST[0][1].yyyy, TEMP[5].xxxx\n"
+					"LOAD TEMP[2].xy, BUFFER[0], TEMP[5].xxxx\n"
+
+					"UADD TEMP[5].y, TEMP[5].xxxx, CONST[0][0].xxxx\n"
+					"LOAD TEMP[3].xy, BUFFER[0], TEMP[5].yyyy\n"
+
+					"U64ADD TEMP[4].xy, TEMP[3], -TEMP[2]\n"
+
+					"AND TEMP[5].z, CONST[0][0].wwww, IMM[4].xxxx\n"
+					"UIF TEMP[5].zzzz\n"
+						/* Load second start/end half-pair and
+						 * take the difference
+						 */
+						"UADD TEMP[5].xy, TEMP[5], IMM[1].wwww\n"
+						"LOAD TEMP[2].xy, BUFFER[0], TEMP[5].xxxx\n"
+						"LOAD TEMP[3].xy, BUFFER[0], TEMP[5].yyyy\n"
+
+						"U64ADD TEMP[3].xy, TEMP[3], -TEMP[2]\n"
+						"U64ADD TEMP[4].xy, TEMP[4], -TEMP[3]\n"
+					"ENDIF\n"
+
+					"U64ADD TEMP[0].xy, TEMP[0], TEMP[4]\n"
+
+					/* Increment pair index */
+					"UADD TEMP[1].y, TEMP[1].yyyy, IMM[1].xxxx\n"
+					"USGE TEMP[5], TEMP[1].yyyy, CONST[0][1].zzzz\n"
+					"UIF TEMP[5]\n"
+						"BRK\n"
+					"ENDIF\n"
+				"ENDLOOP\n"
+
+				/* Increment result index */
+				"UADD TEMP[1].x, TEMP[1].xxxx, IMM[1].xxxx\n"
+			"ENDLOOP\n"
+		"ENDIF\n"
+
+		"AND TEMP[4], CONST[0][0].wwww, IMM[1].yyyy\n"
+		"UIF TEMP[4]\n"
+			/* Store accumulated data for chaining. */
+			"STORE BUFFER[2].xyz, IMM[0].xxxx, TEMP[0]\n"
+		"ELSE\n"
+			"AND TEMP[4], CONST[0][0].wwww, IMM[1].zzzz\n"
+			"UIF TEMP[4]\n"
+				/* Store result availability. */
+				"NOT TEMP[0].z, TEMP[0]\n"
+				"AND TEMP[0].z, TEMP[0].zzzz, IMM[1].xxxx\n"
+				"STORE BUFFER[2].x, IMM[0].xxxx, TEMP[0].zzzz\n"
+
+				"AND TEMP[4], CONST[0][0].wwww, IMM[2].zzzz\n"
+				"UIF TEMP[4]\n"
+					"STORE BUFFER[2].y, IMM[0].xxxx, IMM[0].xxxx\n"
+				"ENDIF\n"
+			"ELSE\n"
+				/* Store result if it is available. */
+				"NOT TEMP[4], TEMP[0].zzzz\n"
+				"UIF TEMP[4]\n"
+					/* Apply timestamp conversion */
+					"AND TEMP[4], CONST[0][0].wwww, IMM[2].yyyy\n"
+					"UIF TEMP[4]\n"
+						"U64MUL TEMP[0].xy, TEMP[0], IMM[3].xyxy\n"
+						"U64DIV TEMP[0].xy, TEMP[0], IMM[3].zwzw\n"
+					"ENDIF\n"
+
+					/* Convert to boolean */
+					"AND TEMP[4], CONST[0][0].wwww, IMM[1].wwww\n"
+					"UIF TEMP[4]\n"
+						"U64SNE TEMP[0].x, TEMP[0].xyxy, IMM[4].zwzw\n"
+						"AND TEMP[0].x, TEMP[0].xxxx, IMM[1].xxxx\n"
+						"MOV TEMP[0].y, IMM[0].xxxx\n"
+					"ENDIF\n"
+
+					"AND TEMP[4], CONST[0][0].wwww, IMM[2].zzzz\n"
+					"UIF TEMP[4]\n"
+						"STORE BUFFER[2].xy, IMM[0].xxxx, TEMP[0].xyxy\n"
+					"ELSE\n"
+						/* Clamping */
+						"UIF TEMP[0].yyyy\n"
+							"MOV TEMP[0].x, IMM[0].wwww\n"
+						"ENDIF\n"
+
+						"AND TEMP[4], CONST[0][0].wwww, IMM[2].wwww\n"
+						"UIF TEMP[4]\n"
+							"UMIN TEMP[0].x, TEMP[0].xxxx, IMM[0].zzzz\n"
+						"ENDIF\n"
+
+						"STORE BUFFER[2].x, IMM[0].xxxx, TEMP[0].xxxx\n"
+					"ENDIF\n"
+				"ENDIF\n"
+			"ENDIF\n"
+		"ENDIF\n"
+
+		"END\n";
+
+	char text[sizeof(text_tmpl) + 32];
+	struct tgsi_token tokens[1024];
+	struct pipe_compute_state state = {};
+
+	/* Hard code the frequency into the shader so that the backend can
+	 * use the full range of optimizations for divide-by-constant.
+	 */
+	snprintf(text, sizeof(text), text_tmpl,
+		 sctx->screen->info.clock_crystal_freq);
+
+	if (!tgsi_text_translate(text, tokens, ARRAY_SIZE(tokens))) {
+		assert(false);
+		return;
+	}
+
+	state.ir_type = PIPE_SHADER_IR_TGSI;
+	state.prog = tokens;
+
+	sctx->b.query_result_shader = sctx->b.b.create_compute_state(&sctx->b.b, &state);
+}
+
+static void si_restore_qbo_state(struct si_context *sctx,
+				 struct si_qbo_state *st)
+{
+	sctx->b.b.bind_compute_state(&sctx->b.b, st->saved_compute);
+
+	sctx->b.b.set_constant_buffer(&sctx->b.b, PIPE_SHADER_COMPUTE, 0, &st->saved_const0);
+	pipe_resource_reference(&st->saved_const0.buffer, NULL);
+
+	sctx->b.b.set_shader_buffers(&sctx->b.b, PIPE_SHADER_COMPUTE, 0, 3, st->saved_ssbo);
+	for (unsigned i = 0; i < 3; ++i)
+		pipe_resource_reference(&st->saved_ssbo[i].buffer, NULL);
+}
+
+static void si_query_hw_get_result_resource(struct si_context *sctx,
+                                              struct si_query *rquery,
+                                              bool wait,
+                                              enum pipe_query_value_type result_type,
+                                              int index,
+                                              struct pipe_resource *resource,
+                                              unsigned offset)
+{
+	struct si_query_hw *query = (struct si_query_hw *)rquery;
+	struct si_query_buffer *qbuf;
+	struct si_query_buffer *qbuf_prev;
+	struct pipe_resource *tmp_buffer = NULL;
+	unsigned tmp_buffer_offset = 0;
+	struct si_qbo_state saved_state = {};
+	struct pipe_grid_info grid = {};
+	struct pipe_constant_buffer constant_buffer = {};
+	struct pipe_shader_buffer ssbo[3];
+	struct si_hw_query_params params;
+	struct {
+		uint32_t end_offset;
+		uint32_t result_stride;
+		uint32_t result_count;
+		uint32_t config;
+		uint32_t fence_offset;
+		uint32_t pair_stride;
+		uint32_t pair_count;
+	} consts;
+
+	if (!sctx->b.query_result_shader) {
+		si_create_query_result_shader(sctx);
+		if (!sctx->b.query_result_shader)
+			return;
+	}
+
+	if (query->buffer.previous) {
+		u_suballocator_alloc(sctx->b.allocator_zeroed_memory, 16, 16,
+				     &tmp_buffer_offset, &tmp_buffer);
+		if (!tmp_buffer)
+			return;
+	}
+
+	si_save_qbo_state(sctx, &saved_state);
+
+	si_get_hw_query_params(sctx, query, index >= 0 ? index : 0, &params);
+	consts.end_offset = params.end_offset - params.start_offset;
+	consts.fence_offset = params.fence_offset - params.start_offset;
+	consts.result_stride = query->result_size;
+	consts.pair_stride = params.pair_stride;
+	consts.pair_count = params.pair_count;
+
+	constant_buffer.buffer_size = sizeof(consts);
+	constant_buffer.user_buffer = &consts;
+
+	ssbo[1].buffer = tmp_buffer;
+	ssbo[1].buffer_offset = tmp_buffer_offset;
+	ssbo[1].buffer_size = 16;
+
+	ssbo[2] = ssbo[1];
+
+	sctx->b.b.bind_compute_state(&sctx->b.b, sctx->b.query_result_shader);
+
+	grid.block[0] = 1;
+	grid.block[1] = 1;
+	grid.block[2] = 1;
+	grid.grid[0] = 1;
+	grid.grid[1] = 1;
+	grid.grid[2] = 1;
+
+	consts.config = 0;
+	if (index < 0)
+		consts.config |= 4;
+	if (query->b.type == PIPE_QUERY_OCCLUSION_PREDICATE ||
+	    query->b.type == PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE)
+		consts.config |= 8;
+	else if (query->b.type == PIPE_QUERY_SO_OVERFLOW_PREDICATE ||
+		 query->b.type == PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE)
+		consts.config |= 8 | 256;
+	else if (query->b.type == PIPE_QUERY_TIMESTAMP ||
+		 query->b.type == PIPE_QUERY_TIME_ELAPSED)
+		consts.config |= 32;
+
+	switch (result_type) {
+	case PIPE_QUERY_TYPE_U64:
+	case PIPE_QUERY_TYPE_I64:
+		consts.config |= 64;
+		break;
+	case PIPE_QUERY_TYPE_I32:
+		consts.config |= 128;
+		break;
+	case PIPE_QUERY_TYPE_U32:
+		break;
+	}
+
+	sctx->b.flags |= sctx->screen->barrier_flags.cp_to_L2;
+
+	for (qbuf = &query->buffer; qbuf; qbuf = qbuf_prev) {
+		if (query->b.type != PIPE_QUERY_TIMESTAMP) {
+			qbuf_prev = qbuf->previous;
+			consts.result_count = qbuf->results_end / query->result_size;
+			consts.config &= ~3;
+			if (qbuf != &query->buffer)
+				consts.config |= 1;
+			if (qbuf->previous)
+				consts.config |= 2;
+		} else {
+			/* Only read the last timestamp. */
+			qbuf_prev = NULL;
+			consts.result_count = 0;
+			consts.config |= 16;
+			params.start_offset += qbuf->results_end - query->result_size;
+		}
+
+		sctx->b.b.set_constant_buffer(&sctx->b.b, PIPE_SHADER_COMPUTE, 0, &constant_buffer);
+
+		ssbo[0].buffer = &qbuf->buf->b.b;
+		ssbo[0].buffer_offset = params.start_offset;
+		ssbo[0].buffer_size = qbuf->results_end - params.start_offset;
+
+		if (!qbuf->previous) {
+			ssbo[2].buffer = resource;
+			ssbo[2].buffer_offset = offset;
+			ssbo[2].buffer_size = 8;
+
+			((struct r600_resource *)resource)->TC_L2_dirty = true;
+		}
+
+		sctx->b.b.set_shader_buffers(&sctx->b.b, PIPE_SHADER_COMPUTE, 0, 3, ssbo);
+
+		if (wait && qbuf == &query->buffer) {
+			uint64_t va;
+
+			/* Wait for result availability. Wait only for readiness
+			 * of the last entry, since the fence writes should be
+			 * serialized in the CP.
+			 */
+			va = qbuf->buf->gpu_address + qbuf->results_end - query->result_size;
+			va += params.fence_offset;
+
+			si_gfx_wait_fence(sctx, va, 0x80000000, 0x80000000);
+		}
+
+		sctx->b.b.launch_grid(&sctx->b.b, &grid);
+		sctx->b.flags |= SI_CONTEXT_CS_PARTIAL_FLUSH;
+	}
+
+	si_restore_qbo_state(sctx, &saved_state);
+	pipe_resource_reference(&tmp_buffer, NULL);
+}
+
+static void si_render_condition(struct pipe_context *ctx,
+				struct pipe_query *query,
+				boolean condition,
+				enum pipe_render_cond_flag mode)
+{
+	struct si_context *sctx = (struct si_context *)ctx;
+	struct si_query_hw *rquery = (struct si_query_hw *)query;
+	struct r600_atom *atom = &sctx->b.render_cond_atom;
+
+	if (query) {
+		bool needs_workaround = false;
+
+		/* There was a firmware regression in VI which causes successive
+		 * SET_PREDICATION packets to give the wrong answer for
+		 * non-inverted stream overflow predication.
+		 */
+		if (((sctx->b.chip_class == VI && sctx->screen->info.pfp_fw_feature < 49) ||
+		     (sctx->b.chip_class == GFX9 && sctx->screen->info.pfp_fw_feature < 38)) &&
+		    !condition &&
+		    (rquery->b.type == PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE ||
+		     (rquery->b.type == PIPE_QUERY_SO_OVERFLOW_PREDICATE &&
+		      (rquery->buffer.previous ||
+		       rquery->buffer.results_end > rquery->result_size)))) {
+			needs_workaround = true;
+		}
+
+		if (needs_workaround && !rquery->workaround_buf) {
+			bool old_force_off = sctx->b.render_cond_force_off;
+			sctx->b.render_cond_force_off = true;
+
+			u_suballocator_alloc(
+				sctx->b.allocator_zeroed_memory, 8, 8,
+				&rquery->workaround_offset,
+				(struct pipe_resource **)&rquery->workaround_buf);
+
+			/* Reset to NULL to avoid a redundant SET_PREDICATION
+			 * from launching the compute grid.
+			 */
+			sctx->b.render_cond = NULL;
+
+			ctx->get_query_result_resource(
+				ctx, query, true, PIPE_QUERY_TYPE_U64, 0,
+				&rquery->workaround_buf->b.b, rquery->workaround_offset);
+
+			/* Settings this in the render cond atom is too late,
+			 * so set it here. */
+			sctx->b.flags |= sctx->screen->barrier_flags.L2_to_cp |
+				       SI_CONTEXT_FLUSH_FOR_RENDER_COND;
+
+			sctx->b.render_cond_force_off = old_force_off;
+		}
+	}
+
+	sctx->b.render_cond = query;
+	sctx->b.render_cond_invert = condition;
+	sctx->b.render_cond_mode = mode;
+
+	si_set_atom_dirty(sctx, atom, query != NULL);
+}
+
+void si_suspend_queries(struct si_context *sctx)
+{
+	struct si_query_hw *query;
+
+	LIST_FOR_EACH_ENTRY(query, &sctx->b.active_queries, list) {
+		si_query_hw_emit_stop(sctx, query);
+	}
+	assert(sctx->b.num_cs_dw_queries_suspend == 0);
+}
+
+void si_resume_queries(struct si_context *sctx)
+{
+	struct si_query_hw *query;
+
+	assert(sctx->b.num_cs_dw_queries_suspend == 0);
+
+	/* Check CS space here. Resuming must not be interrupted by flushes. */
+	si_need_gfx_cs_space(sctx);
+
+	LIST_FOR_EACH_ENTRY(query, &sctx->b.active_queries, list) {
+		si_query_hw_emit_start(sctx, query);
+	}
+}
+
+#define XFULL(name_, query_type_, type_, result_type_, group_id_) \
+	{ \
+		.name = name_, \
+		.query_type = SI_QUERY_##query_type_, \
+		.type = PIPE_DRIVER_QUERY_TYPE_##type_, \
+		.result_type = PIPE_DRIVER_QUERY_RESULT_TYPE_##result_type_, \
+		.group_id = group_id_ \
+	}
+
+#define X(name_, query_type_, type_, result_type_) \
+	XFULL(name_, query_type_, type_, result_type_, ~(unsigned)0)
+
+#define XG(group_, name_, query_type_, type_, result_type_) \
+	XFULL(name_, query_type_, type_, result_type_, SI_QUERY_GROUP_##group_)
+
+static struct pipe_driver_query_info si_driver_query_list[] = {
+	X("num-compilations",		NUM_COMPILATIONS,	UINT64, CUMULATIVE),
+	X("num-shaders-created",	NUM_SHADERS_CREATED,	UINT64, CUMULATIVE),
+	X("num-shader-cache-hits",	NUM_SHADER_CACHE_HITS,	UINT64, CUMULATIVE),
+	X("draw-calls",			DRAW_CALLS,		UINT64, AVERAGE),
+	X("decompress-calls",		DECOMPRESS_CALLS,	UINT64, AVERAGE),
+	X("MRT-draw-calls",		MRT_DRAW_CALLS,		UINT64, AVERAGE),
+	X("prim-restart-calls",		PRIM_RESTART_CALLS,	UINT64, AVERAGE),
+	X("spill-draw-calls",		SPILL_DRAW_CALLS,	UINT64, AVERAGE),
+	X("compute-calls",		COMPUTE_CALLS,		UINT64, AVERAGE),
+	X("spill-compute-calls",	SPILL_COMPUTE_CALLS,	UINT64, AVERAGE),
+	X("dma-calls",			DMA_CALLS,		UINT64, AVERAGE),
+	X("cp-dma-calls",		CP_DMA_CALLS,		UINT64, AVERAGE),
+	X("num-vs-flushes",		NUM_VS_FLUSHES,		UINT64, AVERAGE),
+	X("num-ps-flushes",		NUM_PS_FLUSHES,		UINT64, AVERAGE),
+	X("num-cs-flushes",		NUM_CS_FLUSHES,		UINT64, AVERAGE),
+	X("num-CB-cache-flushes",	NUM_CB_CACHE_FLUSHES,	UINT64, AVERAGE),
+	X("num-DB-cache-flushes",	NUM_DB_CACHE_FLUSHES,	UINT64, AVERAGE),
+	X("num-L2-invalidates",		NUM_L2_INVALIDATES,	UINT64, AVERAGE),
+	X("num-L2-writebacks",		NUM_L2_WRITEBACKS,	UINT64, AVERAGE),
+	X("num-resident-handles",	NUM_RESIDENT_HANDLES,	UINT64, AVERAGE),
+	X("tc-offloaded-slots",		TC_OFFLOADED_SLOTS,     UINT64, AVERAGE),
+	X("tc-direct-slots",		TC_DIRECT_SLOTS,	UINT64, AVERAGE),
+	X("tc-num-syncs",		TC_NUM_SYNCS,		UINT64, AVERAGE),
+	X("CS-thread-busy",		CS_THREAD_BUSY,		UINT64, AVERAGE),
+	X("gallium-thread-busy",	GALLIUM_THREAD_BUSY,	UINT64, AVERAGE),
+	X("requested-VRAM",		REQUESTED_VRAM,		BYTES, AVERAGE),
+	X("requested-GTT",		REQUESTED_GTT,		BYTES, AVERAGE),
+	X("mapped-VRAM",		MAPPED_VRAM,		BYTES, AVERAGE),
+	X("mapped-GTT",			MAPPED_GTT,		BYTES, AVERAGE),
+	X("buffer-wait-time",		BUFFER_WAIT_TIME,	MICROSECONDS, CUMULATIVE),
+	X("num-mapped-buffers",		NUM_MAPPED_BUFFERS,	UINT64, AVERAGE),
+	X("num-GFX-IBs",		NUM_GFX_IBS,		UINT64, AVERAGE),
+	X("num-SDMA-IBs",		NUM_SDMA_IBS,		UINT64, AVERAGE),
+	X("GFX-BO-list-size",		GFX_BO_LIST_SIZE,	UINT64, AVERAGE),
+	X("GFX-IB-size",		GFX_IB_SIZE,		UINT64, AVERAGE),
+	X("num-bytes-moved",		NUM_BYTES_MOVED,	BYTES, CUMULATIVE),
+	X("num-evictions",		NUM_EVICTIONS,		UINT64, CUMULATIVE),
+	X("VRAM-CPU-page-faults",	NUM_VRAM_CPU_PAGE_FAULTS, UINT64, CUMULATIVE),
+	X("VRAM-usage",			VRAM_USAGE,		BYTES, AVERAGE),
+	X("VRAM-vis-usage",		VRAM_VIS_USAGE,		BYTES, AVERAGE),
+	X("GTT-usage",			GTT_USAGE,		BYTES, AVERAGE),
+	X("back-buffer-ps-draw-ratio",	BACK_BUFFER_PS_DRAW_RATIO, UINT64, AVERAGE),
+
+	/* GPIN queries are for the benefit of old versions of GPUPerfStudio,
+	 * which use it as a fallback path to detect the GPU type.
+	 *
+	 * Note: The names of these queries are significant for GPUPerfStudio
+	 * (and possibly their order as well). */
+	XG(GPIN, "GPIN_000",		GPIN_ASIC_ID,		UINT, AVERAGE),
+	XG(GPIN, "GPIN_001",		GPIN_NUM_SIMD,		UINT, AVERAGE),
+	XG(GPIN, "GPIN_002",		GPIN_NUM_RB,		UINT, AVERAGE),
+	XG(GPIN, "GPIN_003",		GPIN_NUM_SPI,		UINT, AVERAGE),
+	XG(GPIN, "GPIN_004",		GPIN_NUM_SE,		UINT, AVERAGE),
+
+	X("temperature",		GPU_TEMPERATURE,	UINT64, AVERAGE),
+	X("shader-clock",		CURRENT_GPU_SCLK,	HZ, AVERAGE),
+	X("memory-clock",		CURRENT_GPU_MCLK,	HZ, AVERAGE),
+
+	/* The following queries must be at the end of the list because their
+	 * availability is adjusted dynamically based on the DRM version. */
+	X("GPU-load",			GPU_LOAD,		UINT64, AVERAGE),
+	X("GPU-shaders-busy",		GPU_SHADERS_BUSY,	UINT64, AVERAGE),
+	X("GPU-ta-busy",		GPU_TA_BUSY,		UINT64, AVERAGE),
+	X("GPU-gds-busy",		GPU_GDS_BUSY,		UINT64, AVERAGE),
+	X("GPU-vgt-busy",		GPU_VGT_BUSY,		UINT64, AVERAGE),
+	X("GPU-ia-busy",		GPU_IA_BUSY,		UINT64, AVERAGE),
+	X("GPU-sx-busy",		GPU_SX_BUSY,		UINT64, AVERAGE),
+	X("GPU-wd-busy",		GPU_WD_BUSY,		UINT64, AVERAGE),
+	X("GPU-bci-busy",		GPU_BCI_BUSY,		UINT64, AVERAGE),
+	X("GPU-sc-busy",		GPU_SC_BUSY,		UINT64, AVERAGE),
+	X("GPU-pa-busy",		GPU_PA_BUSY,		UINT64, AVERAGE),
+	X("GPU-db-busy",		GPU_DB_BUSY,		UINT64, AVERAGE),
+	X("GPU-cp-busy",		GPU_CP_BUSY,		UINT64, AVERAGE),
+	X("GPU-cb-busy",		GPU_CB_BUSY,		UINT64, AVERAGE),
+	X("GPU-sdma-busy",		GPU_SDMA_BUSY,		UINT64, AVERAGE),
+	X("GPU-pfp-busy",		GPU_PFP_BUSY,		UINT64, AVERAGE),
+	X("GPU-meq-busy",		GPU_MEQ_BUSY,		UINT64, AVERAGE),
+	X("GPU-me-busy",		GPU_ME_BUSY,		UINT64, AVERAGE),
+	X("GPU-surf-sync-busy",		GPU_SURF_SYNC_BUSY,	UINT64, AVERAGE),
+	X("GPU-cp-dma-busy",		GPU_CP_DMA_BUSY,	UINT64, AVERAGE),
+	X("GPU-scratch-ram-busy",	GPU_SCRATCH_RAM_BUSY,	UINT64, AVERAGE),
+};
+
+#undef X
+#undef XG
+#undef XFULL
+
+static unsigned si_get_num_queries(struct si_screen *sscreen)
+{
+	if (sscreen->info.drm_major == 2 && sscreen->info.drm_minor >= 42)
+		return ARRAY_SIZE(si_driver_query_list);
+	else if (sscreen->info.drm_major == 3) {
+		if (sscreen->info.chip_class >= VI)
+			return ARRAY_SIZE(si_driver_query_list);
+		else
+			return ARRAY_SIZE(si_driver_query_list) - 7;
+	}
+	else
+		return ARRAY_SIZE(si_driver_query_list) - 25;
+}
+
+static int si_get_driver_query_info(struct pipe_screen *screen,
+				    unsigned index,
+				    struct pipe_driver_query_info *info)
+{
+	struct si_screen *sscreen = (struct si_screen*)screen;
+	unsigned num_queries = si_get_num_queries(sscreen);
+
+	if (!info) {
+		unsigned num_perfcounters =
+			si_get_perfcounter_info(sscreen, 0, NULL);
+
+		return num_queries + num_perfcounters;
+	}
+
+	if (index >= num_queries)
+		return si_get_perfcounter_info(sscreen, index - num_queries, info);
+
+	*info = si_driver_query_list[index];
+
+	switch (info->query_type) {
+	case SI_QUERY_REQUESTED_VRAM:
+	case SI_QUERY_VRAM_USAGE:
+	case SI_QUERY_MAPPED_VRAM:
+		info->max_value.u64 = sscreen->info.vram_size;
+		break;
+	case SI_QUERY_REQUESTED_GTT:
+	case SI_QUERY_GTT_USAGE:
+	case SI_QUERY_MAPPED_GTT:
+		info->max_value.u64 = sscreen->info.gart_size;
+		break;
+	case SI_QUERY_GPU_TEMPERATURE:
+		info->max_value.u64 = 125;
+		break;
+	case SI_QUERY_VRAM_VIS_USAGE:
+		info->max_value.u64 = sscreen->info.vram_vis_size;
+		break;
+	}
+
+	if (info->group_id != ~(unsigned)0 && sscreen->perfcounters)
+		info->group_id += sscreen->perfcounters->num_groups;
+
+	return 1;
+}
+
+/* Note: Unfortunately, GPUPerfStudio hardcodes the order of hardware
+ * performance counter groups, so be careful when changing this and related
+ * functions.
+ */
+static int si_get_driver_query_group_info(struct pipe_screen *screen,
+					  unsigned index,
+					  struct pipe_driver_query_group_info *info)
+{
+	struct si_screen *sscreen = (struct si_screen *)screen;
+	unsigned num_pc_groups = 0;
+
+	if (sscreen->perfcounters)
+		num_pc_groups = sscreen->perfcounters->num_groups;
+
+	if (!info)
+		return num_pc_groups + SI_NUM_SW_QUERY_GROUPS;
+
+	if (index < num_pc_groups)
+		return si_get_perfcounter_group_info(sscreen, index, info);
+
+	index -= num_pc_groups;
+	if (index >= SI_NUM_SW_QUERY_GROUPS)
+		return 0;
+
+	info->name = "GPIN";
+	info->max_active_queries = 5;
+	info->num_queries = 5;
+	return 1;
+}
+
+void si_init_query_functions(struct si_context *sctx)
+{
+	sctx->b.b.create_query = si_create_query;
+	sctx->b.b.create_batch_query = si_create_batch_query;
+	sctx->b.b.destroy_query = si_destroy_query;
+	sctx->b.b.begin_query = si_begin_query;
+	sctx->b.b.end_query = si_end_query;
+	sctx->b.b.get_query_result = si_get_query_result;
+	sctx->b.b.get_query_result_resource = si_get_query_result_resource;
+	sctx->b.render_cond_atom.emit = si_emit_query_predication;
+
+	if (((struct si_screen*)sctx->b.b.screen)->info.num_render_backends > 0)
+	    sctx->b.b.render_condition = si_render_condition;
+
+	LIST_INITHEAD(&sctx->b.active_queries);
+}
+
+void si_init_screen_query_functions(struct si_screen *sscreen)
+{
+	sscreen->b.get_driver_query_info = si_get_driver_query_info;
+	sscreen->b.get_driver_query_group_info = si_get_driver_query_group_info;
+}
diff --git a/src/gallium/drivers/radeonsi/si_query.h b/src/gallium/drivers/radeonsi/si_query.h
new file mode 100644
index 00000000000..3f60208e2f8
--- /dev/null
+++ b/src/gallium/drivers/radeonsi/si_query.h
@@ -0,0 +1,318 @@
+/*
+ * Copyright 2015 Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * 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.
+ */
+
+#ifndef SI_QUERY_H
+#define SI_QUERY_H
+
+#include "util/u_threaded_context.h"
+
+struct pipe_context;
+struct pipe_query;
+struct pipe_resource;
+
+struct si_screen;
+struct si_context;
+struct si_query;
+struct si_query_hw;
+struct r600_resource;
+
+enum {
+	SI_QUERY_DRAW_CALLS = PIPE_QUERY_DRIVER_SPECIFIC,
+	SI_QUERY_DECOMPRESS_CALLS,
+	SI_QUERY_MRT_DRAW_CALLS,
+	SI_QUERY_PRIM_RESTART_CALLS,
+	SI_QUERY_SPILL_DRAW_CALLS,
+	SI_QUERY_COMPUTE_CALLS,
+	SI_QUERY_SPILL_COMPUTE_CALLS,
+	SI_QUERY_DMA_CALLS,
+	SI_QUERY_CP_DMA_CALLS,
+	SI_QUERY_NUM_VS_FLUSHES,
+	SI_QUERY_NUM_PS_FLUSHES,
+	SI_QUERY_NUM_CS_FLUSHES,
+	SI_QUERY_NUM_CB_CACHE_FLUSHES,
+	SI_QUERY_NUM_DB_CACHE_FLUSHES,
+	SI_QUERY_NUM_L2_INVALIDATES,
+	SI_QUERY_NUM_L2_WRITEBACKS,
+	SI_QUERY_NUM_RESIDENT_HANDLES,
+	SI_QUERY_TC_OFFLOADED_SLOTS,
+	SI_QUERY_TC_DIRECT_SLOTS,
+	SI_QUERY_TC_NUM_SYNCS,
+	SI_QUERY_CS_THREAD_BUSY,
+	SI_QUERY_GALLIUM_THREAD_BUSY,
+	SI_QUERY_REQUESTED_VRAM,
+	SI_QUERY_REQUESTED_GTT,
+	SI_QUERY_MAPPED_VRAM,
+	SI_QUERY_MAPPED_GTT,
+	SI_QUERY_BUFFER_WAIT_TIME,
+	SI_QUERY_NUM_MAPPED_BUFFERS,
+	SI_QUERY_NUM_GFX_IBS,
+	SI_QUERY_NUM_SDMA_IBS,
+	SI_QUERY_GFX_BO_LIST_SIZE,
+	SI_QUERY_GFX_IB_SIZE,
+	SI_QUERY_NUM_BYTES_MOVED,
+	SI_QUERY_NUM_EVICTIONS,
+	SI_QUERY_NUM_VRAM_CPU_PAGE_FAULTS,
+	SI_QUERY_VRAM_USAGE,
+	SI_QUERY_VRAM_VIS_USAGE,
+	SI_QUERY_GTT_USAGE,
+	SI_QUERY_GPU_TEMPERATURE,
+	SI_QUERY_CURRENT_GPU_SCLK,
+	SI_QUERY_CURRENT_GPU_MCLK,
+	SI_QUERY_GPU_LOAD,
+	SI_QUERY_GPU_SHADERS_BUSY,
+	SI_QUERY_GPU_TA_BUSY,
+	SI_QUERY_GPU_GDS_BUSY,
+	SI_QUERY_GPU_VGT_BUSY,
+	SI_QUERY_GPU_IA_BUSY,
+	SI_QUERY_GPU_SX_BUSY,
+	SI_QUERY_GPU_WD_BUSY,
+	SI_QUERY_GPU_BCI_BUSY,
+	SI_QUERY_GPU_SC_BUSY,
+	SI_QUERY_GPU_PA_BUSY,
+	SI_QUERY_GPU_DB_BUSY,
+	SI_QUERY_GPU_CP_BUSY,
+	SI_QUERY_GPU_CB_BUSY,
+	SI_QUERY_GPU_SDMA_BUSY,
+	SI_QUERY_GPU_PFP_BUSY,
+	SI_QUERY_GPU_MEQ_BUSY,
+	SI_QUERY_GPU_ME_BUSY,
+	SI_QUERY_GPU_SURF_SYNC_BUSY,
+	SI_QUERY_GPU_CP_DMA_BUSY,
+	SI_QUERY_GPU_SCRATCH_RAM_BUSY,
+	SI_QUERY_NUM_COMPILATIONS,
+	SI_QUERY_NUM_SHADERS_CREATED,
+	SI_QUERY_BACK_BUFFER_PS_DRAW_RATIO,
+	SI_QUERY_NUM_SHADER_CACHE_HITS,
+	SI_QUERY_GPIN_ASIC_ID,
+	SI_QUERY_GPIN_NUM_SIMD,
+	SI_QUERY_GPIN_NUM_RB,
+	SI_QUERY_GPIN_NUM_SPI,
+	SI_QUERY_GPIN_NUM_SE,
+
+	SI_QUERY_FIRST_PERFCOUNTER = PIPE_QUERY_DRIVER_SPECIFIC + 100,
+};
+
+enum {
+	SI_QUERY_GROUP_GPIN = 0,
+	SI_NUM_SW_QUERY_GROUPS
+};
+
+struct si_query_ops {
+	void (*destroy)(struct si_screen *, struct si_query *);
+	bool (*begin)(struct si_context *, struct si_query *);
+	bool (*end)(struct si_context *, struct si_query *);
+	bool (*get_result)(struct si_context *,
+			   struct si_query *, bool wait,
+			   union pipe_query_result *result);
+	void (*get_result_resource)(struct si_context *,
+				    struct si_query *, bool wait,
+				    enum pipe_query_value_type result_type,
+				    int index,
+				    struct pipe_resource *resource,
+				    unsigned offset);
+};
+
+struct si_query {
+	struct threaded_query b;
+	struct si_query_ops *ops;
+
+	/* The type of query */
+	unsigned type;
+};
+
+enum {
+	SI_QUERY_HW_FLAG_NO_START = (1 << 0),
+	/* gap */
+	/* whether begin_query doesn't clear the result */
+	SI_QUERY_HW_FLAG_BEGIN_RESUMES = (1 << 2),
+};
+
+struct si_query_hw_ops {
+	bool (*prepare_buffer)(struct si_screen *,
+			       struct si_query_hw *,
+			       struct r600_resource *);
+	void (*emit_start)(struct si_context *,
+			   struct si_query_hw *,
+			   struct r600_resource *buffer, uint64_t va);
+	void (*emit_stop)(struct si_context *,
+			  struct si_query_hw *,
+			  struct r600_resource *buffer, uint64_t va);
+	void (*clear_result)(struct si_query_hw *, union pipe_query_result *);
+	void (*add_result)(struct si_screen *screen,
+			   struct si_query_hw *, void *buffer,
+			   union pipe_query_result *result);
+};
+
+struct si_query_buffer {
+	/* The buffer where query results are stored. */
+	struct r600_resource		*buf;
+	/* Offset of the next free result after current query data */
+	unsigned			results_end;
+	/* If a query buffer is full, a new buffer is created and the old one
+	 * is put in here. When we calculate the result, we sum up the samples
+	 * from all buffers. */
+	struct si_query_buffer	*previous;
+};
+
+struct si_query_hw {
+	struct si_query b;
+	struct si_query_hw_ops *ops;
+	unsigned flags;
+
+	/* The query buffer and how many results are in it. */
+	struct si_query_buffer buffer;
+	/* Size of the result in memory for both begin_query and end_query,
+	 * this can be one or two numbers, or it could even be a size of a structure. */
+	unsigned result_size;
+	/* The number of dwords for end_query. */
+	unsigned num_cs_dw_end;
+	/* Linked list of queries */
+	struct list_head list;
+	/* For transform feedback: which stream the query is for */
+	unsigned stream;
+
+	/* Workaround via compute shader */
+	struct r600_resource *workaround_buf;
+	unsigned workaround_offset;
+};
+
+bool si_query_hw_init(struct si_screen *sscreen,
+		      struct si_query_hw *query);
+void si_query_hw_destroy(struct si_screen *sscreen,
+			 struct si_query *rquery);
+bool si_query_hw_begin(struct si_context *sctx,
+		       struct si_query *rquery);
+bool si_query_hw_end(struct si_context *sctx,
+		     struct si_query *rquery);
+bool si_query_hw_get_result(struct si_context *sctx,
+			    struct si_query *rquery,
+			    bool wait,
+			    union pipe_query_result *result);
+
+/* Performance counters */
+enum {
+	/* This block is part of the shader engine */
+	SI_PC_BLOCK_SE = (1 << 0),
+
+	/* Expose per-instance groups instead of summing all instances (within
+	 * an SE). */
+	SI_PC_BLOCK_INSTANCE_GROUPS = (1 << 1),
+
+	/* Expose per-SE groups instead of summing instances across SEs. */
+	SI_PC_BLOCK_SE_GROUPS = (1 << 2),
+
+	/* Shader block */
+	SI_PC_BLOCK_SHADER = (1 << 3),
+
+	/* Non-shader block with perfcounters windowed by shaders. */
+	SI_PC_BLOCK_SHADER_WINDOWED = (1 << 4),
+};
+
+/* Describes a hardware block with performance counters. Multiple instances of
+ * each block, possibly per-SE, may exist on the chip. Depending on the block
+ * and on the user's configuration, we either
+ *  (a) expose every instance as a performance counter group,
+ *  (b) expose a single performance counter group that reports the sum over all
+ *      instances, or
+ *  (c) expose one performance counter group per instance, but summed over all
+ *      shader engines.
+ */
+struct si_perfcounter_block {
+	const char *basename;
+	unsigned flags;
+	unsigned num_counters;
+	unsigned num_selectors;
+	unsigned num_instances;
+
+	unsigned num_groups;
+	char *group_names;
+	unsigned group_name_stride;
+
+	char *selector_names;
+	unsigned selector_name_stride;
+
+	void *data;
+};
+
+struct si_perfcounters {
+	unsigned num_groups;
+	unsigned num_blocks;
+	struct si_perfcounter_block *blocks;
+
+	unsigned num_stop_cs_dwords;
+	unsigned num_instance_cs_dwords;
+
+	unsigned num_shader_types;
+	const char * const *shader_type_suffixes;
+	const unsigned *shader_type_bits;
+
+	void (*emit_instance)(struct si_context *,
+			      int se, int instance);
+	void (*emit_shaders)(struct si_context *, unsigned shaders);
+	void (*emit_select)(struct si_context *,
+			    struct si_perfcounter_block *,
+			    unsigned count, unsigned *selectors);
+	void (*emit_start)(struct si_context *,
+			  struct r600_resource *buffer, uint64_t va);
+	void (*emit_stop)(struct si_context *,
+			  struct r600_resource *buffer, uint64_t va);
+	void (*emit_read)(struct si_context *,
+			  struct si_perfcounter_block *,
+			  unsigned count, unsigned *selectors,
+			  struct r600_resource *buffer, uint64_t va);
+
+	void (*cleanup)(struct si_screen *);
+
+	bool separate_se;
+	bool separate_instance;
+};
+
+struct pipe_query *si_create_batch_query(struct pipe_context *ctx,
+					 unsigned num_queries,
+					 unsigned *query_types);
+
+int si_get_perfcounter_info(struct si_screen *,
+			    unsigned index,
+			    struct pipe_driver_query_info *info);
+int si_get_perfcounter_group_info(struct si_screen *,
+				  unsigned index,
+				  struct pipe_driver_query_group_info *info);
+
+bool si_perfcounters_init(struct si_perfcounters *, unsigned num_blocks);
+void si_perfcounters_add_block(struct si_screen *,
+			       struct si_perfcounters *,
+			       const char *name, unsigned flags,
+			       unsigned counters, unsigned selectors,
+			       unsigned instances, void *data);
+void si_perfcounters_do_destroy(struct si_perfcounters *);
+void si_query_hw_reset_buffers(struct si_context *sctx,
+			       struct si_query_hw *query);
+
+struct si_qbo_state {
+	void *saved_compute;
+	struct pipe_constant_buffer saved_const0;
+	struct pipe_shader_buffer saved_ssbo[3];
+};
+
+#endif /* SI_QUERY_H */
diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c
index e4de521cac8..60eec54eb3c 100644
--- a/src/gallium/drivers/radeonsi/si_state.c
+++ b/src/gallium/drivers/radeonsi/si_state.c
@@ -24,7 +24,7 @@
 
 #include "si_build_pm4.h"
 #include "gfx9d.h"
-#include "radeon/r600_query.h"
+#include "si_query.h"
 
 #include "util/u_dual_blend.h"
 #include "util/u_format.h"