[Mesa-dev] [PATCH] radv: implement a workaround for VK_EXT_conditional_rendering

[Samuel Pitoiset]

Per the Vulkan spec 1.1.107, the predicate is a 32-bit value. Though
the AMD hardware treats it as a 64-bit value which means it might
fail to discard.

I don't know why this extension has been drafted like that but this
definitely not fit with AMD. The hardware doesn't seem to support
a 32-bit value for the predicate, so we need to implement a workaround.

This fixes an issue when DXVK enables conditional rendering with RADV.

Fixes: e45ba51ea45 ("radv: add support for VK_EXT_conditional_rendering")
Reported-by: Philip Rebohle <philip.rebohle at tu-dortmund.de>
Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas at basnieuwenhuizen.nl>
---
 src/amd/vulkan/radv_cmd_buffer.c | 46 +++++++++++++++++++++++++++++---
 1 file changed, 43 insertions(+), 3 deletions(-)

diff --git a/src/amd/vulkan/radv_cmd_buffer.c b/src/amd/vulkan/radv_cmd_buffer.c
index 7ee5a5ca7dc..9d9f2577f75 100644
--- a/src/amd/vulkan/radv_cmd_buffer.c
+++ b/src/amd/vulkan/radv_cmd_buffer.c
@@ -4932,8 +4932,11 @@ void radv_CmdBeginConditionalRenderingEXT(
 {
 	RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
 	RADV_FROM_HANDLE(radv_buffer, buffer, pConditionalRenderingBegin->buffer);
+	struct radeon_cmdbuf *cs = cmd_buffer->cs;
 	bool draw_visible = true;
-	uint64_t va;
+	uint64_t pred_value = 0;
+	uint64_t va, new_va;
+	unsigned pred_offset;
 
 	va = radv_buffer_get_va(buffer->bo) + pConditionalRenderingBegin->offset;
 
@@ -4949,13 +4952,50 @@ void radv_CmdBeginConditionalRenderingEXT(
 
 	si_emit_cache_flush(cmd_buffer);
 
+	/* From the Vulkan spec 1.1.107:
+	 *
+	 * "If the 32-bit value at offset in buffer memory is zero, then the
+	 *  rendering commands are discarded, otherwise they are executed as
+	 *  normal. If the value of the predicate in buffer memory changes while
+	 *  conditional rendering is active, the rendering commands may be
+	 *  discarded in an implementation-dependent way. Some implementations
+	 *  may latch the value of the predicate upon beginning conditional
+	 *  rendering while others may read it before every rendering command."
+	 *
+	 * But, the AMD hardware treats the predicate as a 64-bit value which
+	 * means we need a workaround in the driver. Luckily, it's not required
+	 * to support if the value changes when predication is active.
+	 *
+	 * The workaround is as follows:
+	 * 1) allocate a 64-value in the upload BO and initialize it to 0
+	 * 2) copy the 32-bit predicate value to the upload BO
+	 * 3) use the new allocated VA address for predication
+	 *
+	 * Based on the conditionalrender demo, it's faster to do the COPY_DATA
+	 * in ME  (+ sync PFP) instead of PFP.
+	 */
+	radv_cmd_buffer_upload_data(cmd_buffer, 8, 16, &pred_value, &pred_offset);
+
+	new_va = radv_buffer_get_va(cmd_buffer->upload.upload_bo) + pred_offset;
+
+	radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
+	radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_SRC_MEM) |
+			COPY_DATA_DST_SEL(COPY_DATA_DST_MEM));
+	radeon_emit(cs, va);
+	radeon_emit(cs, va >> 32);
+	radeon_emit(cs, new_va);
+	radeon_emit(cs, new_va >> 32);
+
+	radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
+	radeon_emit(cs, 0);
+
 	/* Enable predication for this command buffer. */
-	si_emit_set_predication_state(cmd_buffer, draw_visible, va);
+	si_emit_set_predication_state(cmd_buffer, draw_visible, new_va);
 	cmd_buffer->state.predicating = true;
 
 	/* Store conditional rendering user info. */
 	cmd_buffer->state.predication_type = draw_visible;
-	cmd_buffer->state.predication_va = va;
+	cmd_buffer->state.predication_va = new_va;
 }
 
 void radv_CmdEndConditionalRenderingEXT(
-- 
2.21.0