<!DOCTYPE html><html><head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
</head>
<body>
It could be that the automated testing tools will run amok on that.<br>
<br>
I need to ask Felix for details but I think the code for the CWSR
trap handlers were converted to C style comments as well because of
that.<br>
<br>
Christian.<br>
<br>
<div class="moz-cite-prefix">Am 04.09.24 um 12:52 schrieb Deucher,
Alexander:<br>
</div>
<blockquote type="cite" cite="mid:BL1PR12MB51440CA6B867F66075BD50ADF79C2@BL1PR12MB5144.namprd12.prod.outlook.com">
<style type="text/css" style="display:none;">P {margin-top:0;margin-bottom:0;}</style>
<p style="font-family:Calibri;font-size:10pt;color:#0000FF;margin:5pt;font-style:normal;font-weight:normal;text-decoration:none;" align="Left">
[AMD Official Use Only - AMD Internal Distribution Only]<br>
</p>
<br>
<div>
<div class="elementToProof" style="font-family: Aptos, Aptos_EmbeddedFont, Aptos_MSFontService, Calibri, Helvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
I think you can drop this patch. This code is not compiled,
it's just for reference, and changing all of the comments will
just make it harder to keep in sync with the internal version.</div>
<div class="elementToProof" style="font-family: Aptos, Aptos_EmbeddedFont, Aptos_MSFontService, Calibri, Helvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
<br>
</div>
<div class="elementToProof" style="font-family: Aptos, Aptos_EmbeddedFont, Aptos_MSFontService, Calibri, Helvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
Alex</div>
<div class="elementToProof" style="font-family: Aptos, Aptos_EmbeddedFont, Aptos_MSFontService, Calibri, Helvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
<br>
</div>
<hr style="display:inline-block;width:98%" tabindex="-1">
<div id="divRplyFwdMsg" dir="ltr"><font style="font-size:11pt" face="Calibri, sans-serif" color="#000000"><b>From:</b>
SHANMUGAM, SRINIVASAN <a class="moz-txt-link-rfc2396E" href="mailto:SRINIVASAN.SHANMUGAM@amd.com"><SRINIVASAN.SHANMUGAM@amd.com></a><br>
<b>Sent:</b> Wednesday, September 4, 2024 5:51 AM<br>
<b>To:</b> Koenig, Christian
<a class="moz-txt-link-rfc2396E" href="mailto:Christian.Koenig@amd.com"><Christian.Koenig@amd.com></a>; Deucher, Alexander
<a class="moz-txt-link-rfc2396E" href="mailto:Alexander.Deucher@amd.com"><Alexander.Deucher@amd.com></a><br>
<b>Cc:</b> <a class="moz-txt-link-abbreviated" href="mailto:amd-gfx@lists.freedesktop.org">amd-gfx@lists.freedesktop.org</a>
<a class="moz-txt-link-rfc2396E" href="mailto:amd-gfx@lists.freedesktop.org"><amd-gfx@lists.freedesktop.org></a>; SHANMUGAM, SRINIVASAN
<a class="moz-txt-link-rfc2396E" href="mailto:SRINIVASAN.SHANMUGAM@amd.com"><SRINIVASAN.SHANMUGAM@amd.com></a><br>
<b>Subject:</b> [PATCH v2] drm/amdgpu/gfx9: Convert `//` to
`/* ... */` in cleaner shader code</font>
<div> </div>
</div>
<div class="BodyFragment"><font size="2"><span style="font-size:11pt;">
<div class="PlainText">This commit updates the comment
style in the cleaner shader code from<br>
`//` to `/* ... */` to adhere to the Linux kernel coding
style.<br>
<br>
The comments describe the operation of the cleaner
shader, which is used<br>
to clean LDS, SGPRs, and VGPRs. The shader uses two
kernels launched<br>
separately to clean VGPRs, LDS, and lower SGPRs, and to
clean remaining<br>
SGPRs.<br>
<br>
Fixes: 3b721dfb2c95 ("drm/amdgpu/gfx9: Add cleaner
shader for GFX9.4.3")<br>
Cc: Christian König <a class="moz-txt-link-rfc2396E" href="mailto:christian.koenig@amd.com"><christian.koenig@amd.com></a><br>
Cc: Alex Deucher <a class="moz-txt-link-rfc2396E" href="mailto:alexander.deucher@amd.com"><alexander.deucher@amd.com></a><br>
Signed-off-by: Srinivasan Shanmugam
<a class="moz-txt-link-rfc2396E" href="mailto:srinivasan.shanmugam@amd.com"><srinivasan.shanmugam@amd.com></a><br>
---<br>
v2:<br>
- Corrected typo for iteraions<br>
- Added fixes tag<br>
<br>
.../amd/amdgpu/gfx_v9_4_3_cleaner_shader.asm | 139
+++++++++---------<br>
1 file changed, 72 insertions(+), 67 deletions(-)<br>
<br>
diff --git
a/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3_cleaner_shader.asm
b/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3_cleaner_shader.asm<br>
index d5325ef80ab0..8951b5a87ae1 100644<br>
---
a/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3_cleaner_shader.asm<br>
+++
b/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3_cleaner_shader.asm<br>
@@ -21,47 +21,52 @@<br>
* OTHER DEALINGS IN THE SOFTWARE.<br>
*/<br>
<br>
-// This shader is to clean LDS, SGPRs and VGPRs. It is
first 64 Dwords or 256 bytes of 192 Dwords cleaner
shader.<br>
-//To turn this shader program on for complitaion change
this to main and lower shader main to main_1<br>
- <br>
-// MI300 : Clear SGPRs, VGPRs and LDS<br>
-// Uses two kernels launched separately:<br>
-// 1. Clean VGPRs, LDS, and lower SGPRs<br>
-// Launches one workgroup per CU, each workgroup
with 4x wave64 per SIMD in the CU<br>
-// Waves are "wave64" and have 128 VGPRs each,
which uses all 512 VGPRs per SIMD<br>
-// Waves in the workgroup share the 64KB of LDS<br>
-// Each wave clears SGPRs 0 - 95. Because there
are 4 waves/SIMD, this is physical SGPRs 0-383<br>
-// Each wave clears 128 VGPRs, so all 512 in the
SIMD<br>
-// The first wave of the workgroup clears its
64KB of LDS<br>
-// The shader starts with "S_BARRIER" to ensure
SPI has launched all waves of the workgroup<br>
-// before any wave in the workgroup could
end. Without this, it is possible not all SGPRs get
cleared.<br>
-// 2. Clean remaining SGPRs<br>
-// Launches a workgroup with 24 waves per
workgroup, yielding 6 waves per SIMD in each CU<br>
-// Waves are allocating 96 SGPRs<br>
-// CP sets up SPI_RESOURCE_RESERVE_* registers
to prevent these waves from allocating SGPRs 0-223.<br>
-// As such, these 6 waves per SIMD are
allocated physical SGPRs 224-799<br>
-// Barriers do not work for >16 waves per
workgroup, so we cannot start with S_BARRIER<br>
-// Instead, the shader starts with an
S_SETHALT 1. Once all waves are launched CP will send
unhalt command<br>
-// The shader then clears all SGPRs allocated to
it, cleaning out physical SGPRs 224-799<br>
+/*<br>
+ * This shader is to clean LDS, SGPRs and VGPRs. It is
first 64 Dwords or 256 bytes of 192 Dwords cleaner
shader.<br>
+ * To turn this shader program on for complitaion
change this to main and lower shader main to main_1<br>
+ *<br>
+ * MI300 : Clear SGPRs, VGPRs and LDS<br>
+ * Uses two kernels launched separately:<br>
+ * 1. Clean VGPRs, LDS, and lower SGPRs<br>
+ * Launches one workgroup per CU, each workgroup
with 4x wave64 per SIMD in the CU<br>
+ * Waves are "wave64" and have 128 VGPRs each,
which uses all 512 VGPRs per SIMD<br>
+ * Waves in the workgroup share the 64KB of LDS<br>
+ * Each wave clears SGPRs 0 - 95. Because there
are 4 waves/SIMD, this is physical SGPRs 0-383<br>
+ * Each wave clears 128 VGPRs, so all 512 in the
SIMD<br>
+ * The first wave of the workgroup clears its
64KB of LDS<br>
+ * The shader starts with "S_BARRIER" to ensure
SPI has launched all waves of the workgroup<br>
+ * before any wave in the workgroup could
end. Without this, it is possible not all SGPRs get
cleared.<br>
+ * 2. Clean remaining SGPRs<br>
+ * Launches a workgroup with 24 waves per
workgroup, yielding 6 waves per SIMD in each CU<br>
+ * Waves are allocating 96 SGPRs<br>
+ * CP sets up SPI_RESOURCE_RESERVE_* registers
to prevent these waves from allocating SGPRs 0-223.<br>
+ * As such, these 6 waves per SIMD are
allocated physical SGPRs 224-799<br>
+ * Barriers do not work for >16 waves per
workgroup, so we cannot start with S_BARRIER<br>
+ * Instead, the shader starts with an S_SETHALT
1. Once all waves are launched CP will send unhalt
command<br>
+ * The shader then clears all SGPRs allocated to
it, cleaning out physical SGPRs 224-799<br>
+ */<br>
<br>
shader main<br>
asic(MI300)<br>
type(CS)<br>
wave_size(64)<br>
-// Note: original source code from SQ team<br>
<br>
-// (theorhetical fastest = ~512clks vgpr + 1536 lds +
~128 sgpr = 2176 clks)<br>
+/*<br>
+ * Note: original source code from SQ team<br>
+ *<br>
+ * (theorhetical fastest = ~512clks vgpr + 1536 lds +
~128 sgpr = 2176 clks)<br>
+ */<br>
<br>
- s_cmp_eq_u32 s0, 1 //
Bit0 is set, sgpr0 is set then clear VGPRS and LDS as FW
set COMPUTE_USER_DATA_3<br>
- s_cbranch_scc0 label_0023 //
Clean VGPRs and LDS if sgpr0 of wave is set, scc = (s3
== 1)<br>
+ s_cmp_eq_u32 s0, 1 /*
Bit0 is set, sgpr0 is set then clear VGPRS and LDS as FW
set COMPUTE_USER_DATA_3 */<br>
+ s_cbranch_scc0 label_0023 /*
Clean VGPRs and LDS if sgpr0 of wave is set, scc = (s3
== 1) */<br>
S_BARRIER<br>
<br>
s_movk_i32 m0, 0x0000<br>
- s_mov_b32 s2, 0x00000078 // Loop 128/8=16 times
(loop unrolled for performance)<br>
- //<br>
- // CLEAR VGPRs<br>
- //<br>
- s_set_gpr_idx_on s2, 0x8 // enable Dest VGPR
indexing<br>
+ s_mov_b32 s2, 0x00000078 /* Loop 128/8=16 times
(loop unrolled for performance) */<br>
+/*<br>
+ * CLEAR VGPRs<br>
+ */<br>
+ s_set_gpr_idx_on s2, 0x8 /* enable Dest VGPR
indexing */<br>
label_0005:<br>
v_mov_b32 v0, 0<br>
v_mov_b32 v1, 0<br>
@@ -75,24 +80,24 @@ label_0005:<br>
s_set_gpr_idx_idx s2<br>
s_cbranch_scc0 label_0005<br>
s_set_gpr_idx_off<br>
- <br>
- //<br>
- //<br>
- <br>
- s_mov_b32 s2, 0x80000000 //
Bit31 is first_wave<br>
- s_and_b32 s2, s2, s1 //
sgpr0 has tg_size (first_wave) term as in ucode only
COMPUTE_PGM_RSRC2.tg_size_en is set<br>
- s_cbranch_scc0 label_clean_sgpr_1 //
Clean LDS if its first wave of ThreadGroup/WorkGroup<br>
- // CLEAR LDS<br>
- //<br>
+<br>
+ s_mov_b32 s2, 0x80000000 /*
Bit31 is first_wave */<br>
+ s_and_b32 s2, s2, s1 /*
sgpr0 has tg_size (first_wave) term as in ucode only
COMPUTE_PGM_RSRC2.tg_size_en is set */<br>
+ s_cbranch_scc0 label_clean_sgpr_1 /*
Clean LDS if its first wave of ThreadGroup/WorkGroup */<br>
+/*<br>
+ * CLEAR LDS<br>
+ */<br>
s_mov_b32 exec_lo, 0xffffffff<br>
s_mov_b32 exec_hi, 0xffffffff<br>
- v_mbcnt_lo_u32_b32 v1, exec_hi, 0 // Set V1
to thread-ID (0..63)<br>
- v_mbcnt_hi_u32_b32 v1, exec_lo, v1 // Set V1
to thread-ID (0..63)<br>
- v_mul_u32_u24 v1, 0x00000008, v1 // * 8,
so each thread is a double-dword address (8byte)<br>
- s_mov_b32 s2, 0x00000003f // 64
loop iteraions<br>
+ v_mbcnt_lo_u32_b32 v1, exec_hi, 0 /* Set V1
to thread-ID (0..63) */<br>
+ v_mbcnt_hi_u32_b32 v1, exec_lo, v1 /* Set V1
to thread-ID (0..63) */<br>
+ v_mul_u32_u24 v1, 0x00000008, v1 /* * 8,
so each thread is a double-dword address (8byte) */<br>
+ s_mov_b32 s2, 0x00000003f /* 64
loop iterations */<br>
s_mov_b32 m0, 0xffffffff<br>
- // Clear all of LDS space<br>
- // Each FirstWave of WorkGroup clears 64kbyte block<br>
+/*<br>
+ * Clear all of LDS space<br>
+ * Each FirstWave of WorkGroup clears 64kbyte block<br>
+ */<br>
<br>
label_001F:<br>
ds_write2_b64 v1, v[2:3], v[2:3] offset1:32<br>
@@ -100,11 +105,11 @@ label_001F:<br>
v_add_co_u32 v1, vcc, 0x00000400, v1<br>
s_sub_u32 s2, s2, 1<br>
s_cbranch_scc0 label_001F<br>
- //<br>
- // CLEAR SGPRs<br>
- //<br>
+/*<br>
+ * CLEAR SGPRs<br>
+ */<br>
label_clean_sgpr_1:<br>
- s_mov_b32 m0, 0x0000005c // Loop 96/4=24 times
(loop unrolled for performance)<br>
+ s_mov_b32 m0, 0x0000005c /* Loop 96/4=24 times
(loop unrolled for performance) */<br>
s_nop 0<br>
label_sgpr_loop:<br>
s_movreld_b32 s0, 0<br>
@@ -114,25 +119,25 @@ label_sgpr_loop:<br>
s_sub_u32 m0, m0, 4<br>
s_cbranch_scc0 label_sgpr_loop<br>
<br>
- //clear vcc, flat scratch<br>
- s_mov_b32 flat_scratch_lo, 0 //clear flat scratch
lo SGPR <br>
- s_mov_b32 flat_scratch_hi, 0 //clear flat scratch
hi SGPR <br>
- s_mov_b64 vcc, 0 //clear vcc<br>
- s_mov_b64 ttmp0, 0 //Clear ttmp0 and
ttmp1 <br>
- s_mov_b64 ttmp2, 0 //Clear ttmp2 and
ttmp3 <br>
- s_mov_b64 ttmp4, 0 //Clear ttmp4 and
ttmp5 <br>
- s_mov_b64 ttmp6, 0 //Clear ttmp6 and
ttmp7 <br>
- s_mov_b64 ttmp8, 0 //Clear ttmp8 and
ttmp9 <br>
- s_mov_b64 ttmp10, 0 //Clear ttmp10 and
ttmp11 <br>
- s_mov_b64 ttmp12, 0 //Clear ttmp12 and
ttmp13 <br>
- s_mov_b64 ttmp14, 0 //Clear ttmp14 and
ttmp15 <br>
+ /* clear vcc, flat scratch */<br>
+ s_mov_b32 flat_scratch_lo, 0 /* clear flat scratch
lo SGPR */<br>
+ s_mov_b32 flat_scratch_hi, 0 /* clear flat scratch
hi SGPR */<br>
+ s_mov_b64 vcc, 0 /* clear vcc */<br>
+ s_mov_b64 ttmp0, 0 /* Clear ttmp0 and
ttmp1 */<br>
+ s_mov_b64 ttmp2, 0 /* Clear ttmp2 and
ttmp3 */<br>
+ s_mov_b64 ttmp4, 0 /* Clear ttmp4 and
ttmp5 */<br>
+ s_mov_b64 ttmp6, 0 /* Clear ttmp6 and
ttmp7 */<br>
+ s_mov_b64 ttmp8, 0 /* Clear ttmp8 and
ttmp9 */<br>
+ s_mov_b64 ttmp10, 0 /* Clear ttmp10 and
ttmp11 */<br>
+ s_mov_b64 ttmp12, 0 /* Clear ttmp12 and
ttmp13 */<br>
+ s_mov_b64 ttmp14, 0 /* Clear ttmp14 and
ttmp15 */<br>
s_endpgm<br>
<br>
label_0023:<br>
<br>
s_sethalt 1<br>
<br>
- s_mov_b32 m0, 0x0000005c // Loop 96/4=24 times
(loop unrolled for performance)<br>
+ s_mov_b32 m0, 0x0000005c /* Loop 128/8=16
times (loop unrolled for performance) */<br>
s_nop 0<br>
label_sgpr_loop1:<br>
<br>
@@ -143,10 +148,10 @@ label_sgpr_loop1:<br>
s_sub_u32 m0, m0, 4<br>
s_cbranch_scc0 label_sgpr_loop1<br>
<br>
- //clear vcc, flat scratch<br>
- s_mov_b32 flat_scratch_lo, 0 //clear flat scratch
lo SGPR <br>
- s_mov_b32 flat_scratch_hi, 0 //clear flat scratch
hi SGPR <br>
- s_mov_b64 vcc, 0xee //clear vcc<br>
+ /* clear vcc, flat scratch */<br>
+ s_mov_b32 flat_scratch_lo, 0 /* clear flat scratch
lo SGPR */<br>
+ s_mov_b32 flat_scratch_hi, 0 /* clear flat scratch
hi SGPR */<br>
+ s_mov_b64 vcc, 0xee /* clear vcc */<br>
<br>
s_endpgm<br>
end <br>
-- <br>
2.34.1<br>
<br>
</div>
</span></font></div>
</div>
</blockquote>
<br>
</body>
</html>