[Intel-gfx] [PATCH 28/71] drm/i915/chv: Added CHV specific register read and write

Ben Widawsky ben at bwidawsk.net
Fri Apr 18 02:28:33 CEST 2014


On Wed, Apr 09, 2014 at 01:28:26PM +0300, ville.syrjala at linux.intel.com wrote:
> From: Deepak S <deepak.s at intel.com>
> 
> Support to individually control Media/Render well based on the register access.
> Add CHV specific write function to habdle difference between registers
> that are sadowed vs those that need forcewake even for writes.
> 
> v2: Drop write FIFO for CHV and add comman well forcewake (Ville)
> 
> Signed-off-by: Deepak S <deepak.s at intel.com>
> [vsyrjala: Move the register range macros into intel_uncore.c]
> Signed-off-by: Ville Syrjälä <ville.syrjala at linux.intel.com>
> ---
>  drivers/gpu/drm/i915/intel_uncore.c | 139 +++++++++++++++++++++++++++++++++---
>  1 file changed, 131 insertions(+), 8 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/intel_uncore.c b/drivers/gpu/drm/i915/intel_uncore.c
> index 823d699..8e3c686 100644
> --- a/drivers/gpu/drm/i915/intel_uncore.c
> +++ b/drivers/gpu/drm/i915/intel_uncore.c
> @@ -495,6 +495,31 @@ void assert_force_wake_inactive(struct drm_i915_private *dev_priv)
>  	((reg) >= 0x22000 && (reg) < 0x24000) ||\
>  	((reg) >= 0x30000 && (reg) < 0x40000))
>  
> +#define FORCEWAKE_CHV_RENDER_RANGE_OFFSET(reg) \
> +	(((reg) >= 0x2000 && (reg) < 0x4000) ||\
> +	((reg) >= 0x5000 && (reg) < 0x8000) ||\
> +	((reg) >= 0x8300 && (reg) < 0x8500) ||\
> +	((reg) >= 0xB000 && (reg) < 0xC000) ||\
> +	((reg) >= 0xE000 && (reg) < 0xE800))
> +
> +#define FORCEWAKE_CHV_MEDIA_RANGE_OFFSET(reg)\
> +	(((reg) >= 0x8800 && (reg) < 0x8900) ||\
> +	((reg) >= 0xD000 && (reg) < 0xD800) ||\
> +	((reg) >= 0x12000 && (reg) < 0x14000) ||\
> +	((reg) >= 0x1A000 && (reg) < 0x1C000) ||\
> +	((reg) >= 0x1E800 && (reg) < 0x1EA00) ||\
> +	((reg) >= 0x30000 && (reg) < 0x40000))
> +
> +#define FORCEWAKE_CHV_COMMON_RANGE_OFFSET(reg)\
> +	(((reg) >= 0x4000 && (reg) < 0x5000) ||\
> +	((reg) >= 0x8000 && (reg) < 0x8300) ||\
> +	((reg) >= 0x8500 && (reg) < 0x8600) ||\
> +	((reg) >= 0x9000 && (reg) < 0xB000) ||\
> +	((reg) >= 0xC000 && (reg) < 0xc800) ||\
> +	((reg) >= 0xF000 && (reg) < 0x10000) ||\
> +	((reg) >= 0x14000 && (reg) < 0x14400) ||\
> +	((reg) >= 0x22000 && (reg) < 0x24000))
> +

To satisfy both Chris, and Ville, how about:
#define REG_RANGE(reg, start, end) ((reg) >= (start) && (reg) < 0x5000)
	REG_RANGE(reg, 0x4000, 0x5000) || \
	REG_RANGE(reg, 0x8000, 0x8300) || \
	...

By the way, I spent my due diligence trying to find where these ranges
come from, and have been unable. Doc name? I should have all the docs
from Ville.

I can't speak for the code generated, either.

>  static void
>  ilk_dummy_write(struct drm_i915_private *dev_priv)
>  {
> @@ -587,7 +612,48 @@ vlv_read##x(struct drm_i915_private *dev_priv, off_t reg, bool trace) { \
>  	REG_READ_FOOTER; \
>  }
>  
> +#define __chv_read(x) \
> +static u##x \
> +chv_read##x(struct drm_i915_private *dev_priv, off_t reg, bool trace) { \
> +	unsigned fwengine = 0; \
> +	REG_READ_HEADER(x); \
> +	if (FORCEWAKE_CHV_RENDER_RANGE_OFFSET(reg)) { \
> +		fwengine = FORCEWAKE_RENDER; \
> +	} \
> +	else if (FORCEWAKE_CHV_MEDIA_RANGE_OFFSET(reg)) { \
> +		fwengine = FORCEWAKE_MEDIA; \
> +	} \
> +	else if (FORCEWAKE_CHV_COMMON_RANGE_OFFSET(reg)) { \
> +		fwengine = FORCEWAKE_ALL; \
> +	} \
> +	if (FORCEWAKE_RENDER & fwengine) { \
> +		if (dev_priv->uncore.fw_rendercount++ == 0) \
> +			(dev_priv)->uncore.funcs.force_wake_get(dev_priv, \
> +								fwengine); \
> +	} \
> +	if (FORCEWAKE_MEDIA & fwengine) { \
> +		if (dev_priv->uncore.fw_mediacount++ == 0) \
> +			(dev_priv)->uncore.funcs.force_wake_get(dev_priv, \
> +								fwengine); \
> +	} \
> +	val = __raw_i915_read##x(dev_priv, reg); \
> +	if (FORCEWAKE_RENDER & fwengine) { \
> +		if (dev_priv->uncore.fw_rendercount++ == 0) \
> +			(dev_priv)->uncore.funcs.force_wake_put(dev_priv, \
> +								fwengine); \
> +	} \
> +	if (FORCEWAKE_MEDIA & fwengine) { \
> +		if (dev_priv->uncore.fw_mediacount++ == 0) \
> +			(dev_priv)->uncore.funcs.force_wake_put(dev_priv, \
> +								fwengine); \
> +	} \
> +	REG_READ_FOOTER; \
> +}
>  
> +__chv_read(8)
> +__chv_read(16)
> +__chv_read(32)
> +__chv_read(64)
>  __vlv_read(8)
>  __vlv_read(16)
>  __vlv_read(32)
> @@ -605,6 +671,7 @@ __gen4_read(16)
>  __gen4_read(32)
>  __gen4_read(64)
>  
> +#undef __chv_read
>  #undef __vlv_read
>  #undef __gen6_read
>  #undef __gen5_read
> @@ -709,6 +776,49 @@ gen8_write##x(struct drm_i915_private *dev_priv, off_t reg, u##x val, bool trace
>  	REG_WRITE_FOOTER; \
>  }
>  
> +#define __chv_write(x) \
> +static void \
> +chv_write##x(struct drm_i915_private *dev_priv, off_t reg, u##x val, bool trace) { \
> +	unsigned fwengine = 0; \
> +	bool __needs_put = !is_gen8_shadowed(dev_priv, reg); \
> +	REG_WRITE_HEADER; \
> +	if (FORCEWAKE_CHV_RENDER_RANGE_OFFSET(reg)) { \
> +		fwengine = FORCEWAKE_RENDER; \
> +	} \
> +	else if (FORCEWAKE_CHV_MEDIA_RANGE_OFFSET(reg)) { \
> +		fwengine = FORCEWAKE_MEDIA; \
> +	} \
> +	else if (FORCEWAKE_CHV_COMMON_RANGE_OFFSET(reg)) { \
> +		fwengine = FORCEWAKE_ALL; \
> +	} \
> +	if (__needs_put && (FORCEWAKE_RENDER & fwengine)) { \
> +			if (dev_priv->uncore.fw_rendercount++ == 0) \
> +				(dev_priv)->uncore.funcs.force_wake_get(dev_priv, \
> +									fwengine); \
> +	} \
> +	if (__needs_put && (FORCEWAKE_MEDIA & fwengine)) { \
> +		if (dev_priv->uncore.fw_mediacount++ == 0) \
> +			(dev_priv)->uncore.funcs.force_wake_get(dev_priv, \
> +								fwengine); \
> +	} \
> +	__raw_i915_write##x(dev_priv, reg, val); \
> +	if (__needs_put && (FORCEWAKE_RENDER & fwengine)) { \
> +			if (dev_priv->uncore.fw_rendercount++ == 0) \
> +				(dev_priv)->uncore.funcs.force_wake_put(dev_priv, \
> +									fwengine); \
> +	} \
> +	if (__needs_put && (FORCEWAKE_MEDIA & fwengine)) { \
> +		if (dev_priv->uncore.fw_mediacount++ == 0) \
> +			(dev_priv)->uncore.funcs.force_wake_put(dev_priv, \
> +								fwengine); \
> +	} \
> +	REG_WRITE_FOOTER; \
> +}

Feels like this would be a lot neater if you let force_wake_put() handle
this complexity. I guess our force_wake_funcs can't handle count. In
that case we could even share the gen8_write family of functions. Was
there some reason this can't work? (Again, forgive laziness)

> +
> +__chv_write(8)
> +__chv_write(16)
> +__chv_write(32)
> +__chv_write(64)

Similar to broadwell (my bad, I know), we probably only actually want to
do this for read/write32. So we potentially could reduce the obj size by
only doing it for that.

>  __gen8_write(8)
>  __gen8_write(16)
>  __gen8_write(32)
> @@ -730,6 +840,7 @@ __gen4_write(16)
>  __gen4_write(32)
>  __gen4_write(64)
>  
> +#undef __chv_write
>  #undef __gen8_write
>  #undef __hsw_write
>  #undef __gen6_write
> @@ -793,14 +904,26 @@ void intel_uncore_init(struct drm_device *dev)
>  
>  	switch (INTEL_INFO(dev)->gen) {
>  	default:
> -		dev_priv->uncore.funcs.mmio_writeb  = gen8_write8;
> -		dev_priv->uncore.funcs.mmio_writew  = gen8_write16;
> -		dev_priv->uncore.funcs.mmio_writel  = gen8_write32;
> -		dev_priv->uncore.funcs.mmio_writeq  = gen8_write64;
> -		dev_priv->uncore.funcs.mmio_readb  = gen6_read8;
> -		dev_priv->uncore.funcs.mmio_readw  = gen6_read16;
> -		dev_priv->uncore.funcs.mmio_readl  = gen6_read32;
> -		dev_priv->uncore.funcs.mmio_readq  = gen6_read64;
> +		if (IS_CHERRYVIEW(dev)) {
> +			dev_priv->uncore.funcs.mmio_writeb  = chv_write8;
> +			dev_priv->uncore.funcs.mmio_writew  = chv_write16;
> +			dev_priv->uncore.funcs.mmio_writel  = chv_write32;
> +			dev_priv->uncore.funcs.mmio_writeq  = chv_write64;
> +			dev_priv->uncore.funcs.mmio_readb  = chv_read8;
> +			dev_priv->uncore.funcs.mmio_readw  = chv_read16;
> +			dev_priv->uncore.funcs.mmio_readl  = chv_read32;
> +			dev_priv->uncore.funcs.mmio_readq  = chv_read64;
> +
> +		} else {
> +			dev_priv->uncore.funcs.mmio_writeb  = gen8_write8;
> +			dev_priv->uncore.funcs.mmio_writew  = gen8_write16;
> +			dev_priv->uncore.funcs.mmio_writel  = gen8_write32;
> +			dev_priv->uncore.funcs.mmio_writeq  = gen8_write64;
> +			dev_priv->uncore.funcs.mmio_readb  = gen6_read8;
> +			dev_priv->uncore.funcs.mmio_readw  = gen6_read16;
> +			dev_priv->uncore.funcs.mmio_readl  = gen6_read32;
> +			dev_priv->uncore.funcs.mmio_readq  = gen6_read64;
> +		}
>  		break;
>  	case 7:
>  	case 6:
> -- 
> 1.8.3.2
> 
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/intel-gfx

-- 
Ben Widawsky, Intel Open Source Technology Center



More information about the Intel-gfx mailing list