Cycles: Crash on AVX2 processors in some corner cases #50050

New Issue

Sergey Sharybin · 2016-11-16T15:40:28+01:00

Sergey Sharybin commented

2016-11-16 15:40:28 +01:00

System Information

Linux Debian jessie, 64bit, GTX 1080 GPU.

Blender Version

Broken on 2a2eb0c with the following patch applied

P424: (An Untitled Masterwork)

diff --git a/intern/cycles/util/util_math.h b/intern/cycles/util/util_math.h
index 3f4d3e0..6cb68b5 100644
--- a/intern/cycles/util/util_math.h
+++ b/intern/cycles/util/util_math.h
@@ -596,8 +596,7 @@ ccl_device_inline float len_squared(const float4& a)
 
 ccl_device_inline float3 normalize(const float3& a)
 {
-	/* TODO(sergey): Disabled for now, causes crashes in certain cases. */
-#if defined(__KERNEL_SSE41__) && defined(__KERNEL_SSE__) && 0
+#if defined(__KERNEL_SSE41__) && defined(__KERNEL_SSE__)
 	__m128 norm = _mm_sqrt_ps(_mm_dp_ps(a.m128, a.m128, 0x7F));
 	return _mm_div_ps(a.m128, norm);
 #else
@@ -798,8 +797,7 @@ ccl_device_inline float4 operator-(const float4& a)
 
 ccl_device_inline float4 operator*(const float4& a, const float4& b)
 {
-	/* TODO(sergey): Disabled for now, causes crashes in certain cases. */
-#if defined(__KERNEL_SSE__) && 0
+#ifdef __KERNEL_SSE__
 	return _mm_mul_ps(a.m128, b.m128);
 #else
 	return make_float4(a.x*b.x, a.y*b.y, a.z*b.z, a.w*b.w);
@@ -847,8 +845,7 @@ ccl_device_inline float4 operator/(const float4& a, const float4& b)
 
 ccl_device_inline float4 operator+(const float4& a, const float4& b)
 {
-	/* TODO(sergey): Disabled for now, causes crashes in certain cases. */
-#if defined(__KERNEL_SSE__) && 0
+#ifdef __KERNEL_SSE__
 	return _mm_add_ps(a.m128, b.m128);
 #else
 	return make_float4(a.x+b.x, a.y+b.y, a.z+b.z, a.w+b.w);
diff --git a/intern/cycles/util/util_transform.h b/intern/cycles/util/util_transform.h
index ea5eb3b..a0695f2 100644
--- a/intern/cycles/util/util_transform.h
+++ b/intern/cycles/util/util_transform.h
@@ -74,7 +74,7 @@ ccl_device_inline float3 transform_perspective(const Transform *t, const float3
 ccl_device_inline float3 transform_point(const Transform *t, const float3 a)
 {
 	/* TODO(sergey): Disabled for now, causes crashes in certain cases. */
-#if defined(__KERNEL_SSE__) && defined(__KERNEL_SSE2__) && 0
+#if defined(__KERNEL_SSE__) && defined(__KERNEL_SSE2__)
 	ssef x, y, z, w, aa;
 	aa = a.m128;
 
@@ -103,8 +103,7 @@ ccl_device_inline float3 transform_point(const Transform *t, const float3 a)
 
 ccl_device_inline float3 transform_direction(const Transform *t, const float3 a)
 {
-	/* TODO(sergey): Disabled for now, causes crashes in certain cases. */
-#if defined(__KERNEL_SSE__) && defined(__KERNEL_SSE2__) && 0
+#if defined(__KERNEL_SSE__) && defined(__KERNEL_SSE2__)
 	ssef x, y, z, w, aa;
 	aa = a.m128;
 	x = _mm_loadu_ps(&t->x.x);

NOTE: It is important to apply this patch because it removes workarounds which we added to

Short description of error

Crash happens in some corner cases when rendering on AVX2 CPU with Blender compiled in Release mode.

Exact steps for others to reproduce the error

Open one of those files and hit F12. The crash will happen soon after few samples are done.

flythrough_main.blend flythrough_main_rendertest.blend

**System Information** Linux Debian jessie, 64bit, GTX 1080 GPU. **Blender Version** Broken on 2a2eb0c with the following patch applied [P424: (An Untitled Masterwork)](https://archive.blender.org/developer/P424.txt) ``` diff --git a/intern/cycles/util/util_math.h b/intern/cycles/util/util_math.h index 3f4d3e0..6cb68b5 100644 --- a/intern/cycles/util/util_math.h +++ b/intern/cycles/util/util_math.h @@ -596,8 +596,7 @@ ccl_device_inline float len_squared(const float4& a) ccl_device_inline float3 normalize(const float3& a) { - /* TODO(sergey): Disabled for now, causes crashes in certain cases. */ -#if defined(__KERNEL_SSE41__) && defined(__KERNEL_SSE__) && 0 +#if defined(__KERNEL_SSE41__) && defined(__KERNEL_SSE__) __m128 norm = _mm_sqrt_ps(_mm_dp_ps(a.m128, a.m128, 0x7F)); return _mm_div_ps(a.m128, norm); #else @@ -798,8 +797,7 @@ ccl_device_inline float4 operator-(const float4& a) ccl_device_inline float4 operator*(const float4& a, const float4& b) { - /* TODO(sergey): Disabled for now, causes crashes in certain cases. */ -#if defined(__KERNEL_SSE__) && 0 +#ifdef __KERNEL_SSE__ return _mm_mul_ps(a.m128, b.m128); #else return make_float4(a.x*b.x, a.y*b.y, a.z*b.z, a.w*b.w); @@ -847,8 +845,7 @@ ccl_device_inline float4 operator/(const float4& a, const float4& b) ccl_device_inline float4 operator+(const float4& a, const float4& b) { - /* TODO(sergey): Disabled for now, causes crashes in certain cases. */ -#if defined(__KERNEL_SSE__) && 0 +#ifdef __KERNEL_SSE__ return _mm_add_ps(a.m128, b.m128); #else return make_float4(a.x+b.x, a.y+b.y, a.z+b.z, a.w+b.w); diff --git a/intern/cycles/util/util_transform.h b/intern/cycles/util/util_transform.h index ea5eb3b..a0695f2 100644 --- a/intern/cycles/util/util_transform.h +++ b/intern/cycles/util/util_transform.h @@ -74,7 +74,7 @@ ccl_device_inline float3 transform_perspective(const Transform *t, const float3 ccl_device_inline float3 transform_point(const Transform *t, const float3 a) { /* TODO(sergey): Disabled for now, causes crashes in certain cases. */ -#if defined(__KERNEL_SSE__) && defined(__KERNEL_SSE2__) && 0 +#if defined(__KERNEL_SSE__) && defined(__KERNEL_SSE2__) ssef x, y, z, w, aa; aa = a.m128; @@ -103,8 +103,7 @@ ccl_device_inline float3 transform_point(const Transform *t, const float3 a) ccl_device_inline float3 transform_direction(const Transform *t, const float3 a) { - /* TODO(sergey): Disabled for now, causes crashes in certain cases. */ -#if defined(__KERNEL_SSE__) && defined(__KERNEL_SSE2__) && 0 +#if defined(__KERNEL_SSE__) && defined(__KERNEL_SSE2__) ssef x, y, z, w, aa; aa = a.m128; x = _mm_loadu_ps(&t->x.x); ``` NOTE: It is important to apply this patch because it removes workarounds which we added to **Short description of error** Crash happens in some corner cases when rendering on AVX2 CPU with Blender compiled in Release mode. **Exact steps for others to reproduce the error** Open one of those files and hit F12. The crash will happen soon after few samples are done. [flythrough_main.blend](https://archive.blender.org/developer/F401709/flythrough_main.blend) [flythrough_main_rendertest.blend](https://archive.blender.org/developer/F401705/flythrough_main_rendertest.blend)

Sergey Sharybin commented

2016-11-16 15:40:28 +01:00

Changed status to: 'Open'

Sergey Sharybin commented

2016-11-16 15:40:28 +01:00

Added subscriber: @Sergey

Sergey Sharybin self-assigned this 2016-11-16 15:40:39 +01:00

Sergey Sharybin commented

2016-12-02 10:20:34 +01:00

Changed status from 'Open' to: 'Archived'

Sergey Sharybin closed this issue

2016-12-02 10:20:34 +01:00

Sergey Sharybin commented

2016-12-02 10:20:34 +01:00

Can no longer reproduce the crash after light threshold and sampling commits. Could be some compiler issue or so.

Enabled all AVX2 optimizations again to see if we hit this bug again.

Can no longer reproduce the crash after light threshold and sampling commits. Could be some compiler issue or so. Enabled all AVX2 optimizations again to see if we hit this bug again.

Sign in to join this conversation.

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Cycles: Crash on AVX2 processors in some corner cases #50050